You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
rocksdb/cache/clock_cache.cc

1480 lines
57 KiB

// Copyright (c) 2011-present, Facebook, Inc. All rights reserved.
// This source code is licensed under both the GPLv2 (found in the
// COPYING file in the root directory) and Apache 2.0 License
// (found in the LICENSE.Apache file in the root directory).
//
// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file. See the AUTHORS file for names of contributors.
#include "cache/clock_cache.h"
#include <functional>
Observe and warn about misconfigured HyperClockCache (#10965) Summary: Background. One of the core risks of chosing HyperClockCache is ending up with degraded performance if estimated_entry_charge is very significantly wrong. Too low leads to under-utilized hash table, which wastes a bit of (tracked) memory and likely increases access times due to larger working set size (more TLB misses). Too high leads to fully populated hash table (at some limit with reasonable lookup performance) and not being able to cache as many objects as the memory limit would allow. In either case, performance degradation is graceful/continuous but can be quite significant. For example, cutting block size in half without updating estimated_entry_charge could lead to a large portion of configured block cache memory (up to roughly 1/3) going unused. Fix. This change adds a mechanism through which the DB periodically probes the block cache(s) for "problems" to report, and adds diagnostics to the HyperClockCache for bad estimated_entry_charge. The periodic probing is currently done with DumpStats / stats_dump_period_sec, and diagnostics reported to info_log (normally LOG file). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10965 Test Plan: unit test included. Doesn't cover all the implemented subtleties of reporting, but ensures basics of when to report or not. Also manual testing with db_bench. Create db with ``` ./db_bench --benchmarks=fillrandom,flush --num=3000000 --disable_wal=1 ``` Use and check LOG file for HyperClockCache for various block sizes (used as estimated_entry_charge) ``` ./db_bench --use_existing_db --benchmarks=readrandom --num=3000000 --duration=20 --stats_dump_period_sec=8 --cache_type=hyper_clock_cache -block_size=XXXX ``` Seeing warnings / errors or not as expected. Reviewed By: anand1976 Differential Revision: D41406932 Pulled By: pdillinger fbshipit-source-id: 4ca56162b73017e4b9cec2cad74466f49c27a0a7
2 years ago
#include <numeric>
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
#include "cache/cache_key.h"
HyperClockCache support for SecondaryCache, with refactoring (#11301) Summary: Internally refactors SecondaryCache integration out of LRUCache specifically and into a wrapper/adapter class that works with various Cache implementations. Notably, this relies on separating the notion of async lookup handles from other cache handles, so that HyperClockCache doesn't have to deal with the problem of allocating handles from the hash table for lookups that might fail anyway, and might be on the same key without support for coalescing. (LRUCache's hash table can incorporate previously allocated handles thanks to its pointer indirection.) Specifically, I'm worried about the case in which hundreds of threads try to access the same block and probing in the hash table degrades to linear search on the pile of entries with the same key. This change is a big step in the direction of supporting stacked SecondaryCaches, but there are obstacles to completing that. Especially, there is no SecondaryCache hook for evictions to pass from one to the next. It has been proposed that evictions be transmitted simply as the persisted data (as in SaveToCallback), but given the current structure provided by the CacheItemHelpers, that would require an extra copy of the block data, because there's intentionally no way to ask for a contiguous Slice of the data (to allow for flexibility in storage). `AsyncLookupHandle` and the re-worked `WaitAll()` should be essentially prepared for stacked SecondaryCaches, but several "TODO with stacked secondaries" issues remain in various places. It could be argued that the stacking instead be done as a SecondaryCache adapter that wraps two (or more) SecondaryCaches, but at least with the current API that would require an extra heap allocation on SecondaryCache Lookup for a wrapper SecondaryCacheResultHandle that can transfer a Lookup between secondaries. We could also consider trying to unify the Cache and SecondaryCache APIs, though that might be difficult if `AsyncLookupHandle` is kept a fixed struct. ## cache.h (public API) Moves `secondary_cache` option from LRUCacheOptions to ShardedCacheOptions so that it is applicable to HyperClockCache. ## advanced_cache.h (advanced public API) * Add `Cache::CreateStandalone()` so that the SecondaryCache support wrapper can use it. * Add `SetEvictionCallback()` / `eviction_callback_` so that the SecondaryCache support wrapper can use it. Only a single callback is supported for efficiency. If there is ever a need for more than one, hopefully that can be handled with a broadcast callback wrapper. These are essentially the two "extra" pieces of `Cache` for pulling out specific SecondaryCache support from the `Cache` implementation. I think it's a good trade-off as these are reasonable, limited, and reusable "cut points" into the `Cache` implementations. * Remove async capability from standard `Lookup()` (getting rid of awkward restrictions on pending Handles) and add `AsyncLookupHandle` and `StartAsyncLookup()`. As noted in the comments, the full struct of `AsyncLookupHandle` is exposed so that it can be stack allocated, for efficiency, though more data is being copied around than before, which could impact performance. (Lookup info -> AsyncLookupHandle -> Handle vs. Lookup info -> Handle) I could foresee a future in which a Cache internally saves a pointer to the AsyncLookupHandle, which means it's dangerous to allow it to be copyable or even movable. It also means it's not compatible with std::vector (which I don't like requiring as an API parameter anyway), so `WaitAll()` expects any contiguous array of AsyncLookupHandles. I believe this is best for common case efficiency, while behaving well in other cases also. For example, `WaitAll()` has no effect on default-constructed AsyncLookupHandles, which look like a completed cache miss. ## cacheable_entry.h A couple of functions are obsolete because Cache::Handle can no longer be pending. ## cache.cc Provides default implementations for new or revamped Cache functions, especially appropriate for non-blocking caches. ## secondary_cache_adapter.{h,cc} The full details of the Cache wrapper adding SecondaryCache support. Essentially replicates the SecondaryCache handling that was in LRUCache, but obviously refactored. There is a bit of logic duplication, where Lookup() is essentially a manually optimized version of StartAsyncLookup() and Wait(), but it's roughly a dozen lines of code. ## sharded_cache.h, typed_cache.h, charged_cache.{h,cc}, sim_cache.cc Simply updated for Cache API changes. ## lru_cache.{h,cc} Carefully remove SecondaryCache logic, implement `CreateStandalone` and eviction handler functionality. ## clock_cache.{h,cc} Expose existing `CreateStandalone` functionality, add eviction handler functionality. Light refactoring. ## block_based_table_reader* Mostly re-worked the only usage of async Lookup, which is in BlockBasedTable::MultiGet. Used arrays in place of autovector in some places for efficiency. Simplified some logic by not trying to process some cache results before they're all ready. Created new function `BlockBasedTable::GetCachePriority()` to reduce some pre-existing code duplication (and avoid making it worse). Fixed at least one small bug from the prior confusing mixture of async and sync Lookups. In MaybeReadBlockAndLoadToCache(), called by RetrieveBlock(), called by MultiGet() with wait=false, is_cache_hit for the block_cache_tracer entry would not be set to true if the handle was pending after Lookup and before Wait. ## Intended follow-up work * Figure out if there are any missing stats or block_cache_tracer work in refactored BlockBasedTable::MultiGet * Stacked secondary caches (see above discussion) * See if we can make up for the small MultiGet performance regression. * Study more performance with SecondaryCache * Items evicted from over-full LRUCache in Release were not being demoted to SecondaryCache, and still aren't to minimize unit test churn. Ideally they would be demoted, but it's an exceptional case so not a big deal. * Use CreateStandalone for cache reservations (save unnecessary hash table operations). Not a big deal, but worthy cleanup. * Somehow I got the contract for SecondaryCache::Insert wrong in #10945. (Doesn't take ownership!) That API comment needs to be fixed, but didn't want to mingle that in here. Pull Request resolved: https://github.com/facebook/rocksdb/pull/11301 Test Plan: ## Unit tests Generally updated to include HCC in SecondaryCache tests, though HyperClockCache has some different, less strict behaviors that leads to some tests not really being set up to work with it. Some of the tests remain disabled with it, but I think we have good coverage without them. ## Crash/stress test Updated to use the new combination. ## Performance First, let's check for regression on caches without secondary cache configured. Adding support for the eviction callback is likely to have a tiny effect, but it shouldn't be worrisome. LRUCache could benefit slightly from less logic around SecondaryCache handling. We can test with cache_bench default settings, built with DEBUG_LEVEL=0 and PORTABLE=0. ``` (while :; do base/cache_bench --cache_type=hyper_clock_cache | grep Rough; done) | awk '{ sum += $9; count++; print $0; print "Average: " int(sum / count) }' ``` **Before** this and #11299 (which could also have a small effect), running for about an hour, before & after running concurrently for each cache type: HyperClockCache: 3168662 (average parallel ops/sec) LRUCache: 2940127 **After** this and #11299, running for about an hour: HyperClockCache: 3164862 (average parallel ops/sec) (0.12% slower) LRUCache: 2940928 (0.03% faster) This is an acceptable difference IMHO. Next, let's consider essentially the worst case of new CPU overhead affecting overall performance. MultiGet uses the async lookup interface regardless of whether SecondaryCache or folly are used. We can configure a benchmark where all block cache queries are for data blocks, and all are hits. Create DB and test (before and after tests running simultaneously): ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 TEST_TMPDIR=/dev/shm base/db_bench -benchmarks=multireadrandom[-X30] -readonly -multiread_batched -batch_size=32 -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16 ``` **Before**: multireadrandom [AVG 30 runs] : 3444202 (± 57049) ops/sec; 240.9 (± 4.0) MB/sec multireadrandom [MEDIAN 30 runs] : 3514443 ops/sec; 245.8 MB/sec **After**: multireadrandom [AVG 30 runs] : 3291022 (± 58851) ops/sec; 230.2 (± 4.1) MB/sec multireadrandom [MEDIAN 30 runs] : 3366179 ops/sec; 235.4 MB/sec So that's roughly a 3% regression, on kind of a *worst case* test of MultiGet CPU. Similar story with HyperClockCache: **Before**: multireadrandom [AVG 30 runs] : 3933777 (± 41840) ops/sec; 275.1 (± 2.9) MB/sec multireadrandom [MEDIAN 30 runs] : 3970667 ops/sec; 277.7 MB/sec **After**: multireadrandom [AVG 30 runs] : 3755338 (± 30391) ops/sec; 262.6 (± 2.1) MB/sec multireadrandom [MEDIAN 30 runs] : 3785696 ops/sec; 264.8 MB/sec Roughly a 4-5% regression. Not ideal, but not the whole story, fortunately. Let's also look at Get() in db_bench: ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X30] -readonly -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16 ``` **Before**: readrandom [AVG 30 runs] : 2198685 (± 13412) ops/sec; 153.8 (± 0.9) MB/sec readrandom [MEDIAN 30 runs] : 2209498 ops/sec; 154.5 MB/sec **After**: readrandom [AVG 30 runs] : 2292814 (± 43508) ops/sec; 160.3 (± 3.0) MB/sec readrandom [MEDIAN 30 runs] : 2365181 ops/sec; 165.4 MB/sec That's showing roughly a 4% improvement, perhaps because of the secondary cache code that is no longer part of LRUCache. But weirdly, HyperClockCache is also showing 2-3% improvement: **Before**: readrandom [AVG 30 runs] : 2272333 (± 9992) ops/sec; 158.9 (± 0.7) MB/sec readrandom [MEDIAN 30 runs] : 2273239 ops/sec; 159.0 MB/sec **After**: readrandom [AVG 30 runs] : 2332407 (± 11252) ops/sec; 163.1 (± 0.8) MB/sec readrandom [MEDIAN 30 runs] : 2335329 ops/sec; 163.3 MB/sec Reviewed By: ltamasi Differential Revision: D44177044 Pulled By: pdillinger fbshipit-source-id: e808e48ff3fe2f792a79841ba617be98e48689f5
2 years ago
#include "cache/secondary_cache_adapter.h"
Observe and warn about misconfigured HyperClockCache (#10965) Summary: Background. One of the core risks of chosing HyperClockCache is ending up with degraded performance if estimated_entry_charge is very significantly wrong. Too low leads to under-utilized hash table, which wastes a bit of (tracked) memory and likely increases access times due to larger working set size (more TLB misses). Too high leads to fully populated hash table (at some limit with reasonable lookup performance) and not being able to cache as many objects as the memory limit would allow. In either case, performance degradation is graceful/continuous but can be quite significant. For example, cutting block size in half without updating estimated_entry_charge could lead to a large portion of configured block cache memory (up to roughly 1/3) going unused. Fix. This change adds a mechanism through which the DB periodically probes the block cache(s) for "problems" to report, and adds diagnostics to the HyperClockCache for bad estimated_entry_charge. The periodic probing is currently done with DumpStats / stats_dump_period_sec, and diagnostics reported to info_log (normally LOG file). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10965 Test Plan: unit test included. Doesn't cover all the implemented subtleties of reporting, but ensures basics of when to report or not. Also manual testing with db_bench. Create db with ``` ./db_bench --benchmarks=fillrandom,flush --num=3000000 --disable_wal=1 ``` Use and check LOG file for HyperClockCache for various block sizes (used as estimated_entry_charge) ``` ./db_bench --use_existing_db --benchmarks=readrandom --num=3000000 --duration=20 --stats_dump_period_sec=8 --cache_type=hyper_clock_cache -block_size=XXXX ``` Seeing warnings / errors or not as expected. Reviewed By: anand1976 Differential Revision: D41406932 Pulled By: pdillinger fbshipit-source-id: 4ca56162b73017e4b9cec2cad74466f49c27a0a7
2 years ago
#include "logging/logging.h"
#include "monitoring/perf_context_imp.h"
#include "monitoring/statistics.h"
#include "port/lang.h"
#include "rocksdb/env.h"
#include "util/hash.h"
#include "util/math.h"
#include "util/random.h"
namespace ROCKSDB_NAMESPACE {
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
namespace clock_cache {
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
namespace {
inline uint64_t GetRefcount(uint64_t meta) {
return ((meta >> ClockHandle::kAcquireCounterShift) -
(meta >> ClockHandle::kReleaseCounterShift)) &
ClockHandle::kCounterMask;
}
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
inline uint64_t GetInitialCountdown(Cache::Priority priority) {
// Set initial clock data from priority
// TODO: configuration parameters for priority handling and clock cycle
// count?
switch (priority) {
case Cache::Priority::HIGH:
return ClockHandle::kHighCountdown;
default:
assert(false);
FALLTHROUGH_INTENDED;
case Cache::Priority::LOW:
return ClockHandle::kLowCountdown;
case Cache::Priority::BOTTOM:
return ClockHandle::kBottomCountdown;
}
}
HyperClockCache support for SecondaryCache, with refactoring (#11301) Summary: Internally refactors SecondaryCache integration out of LRUCache specifically and into a wrapper/adapter class that works with various Cache implementations. Notably, this relies on separating the notion of async lookup handles from other cache handles, so that HyperClockCache doesn't have to deal with the problem of allocating handles from the hash table for lookups that might fail anyway, and might be on the same key without support for coalescing. (LRUCache's hash table can incorporate previously allocated handles thanks to its pointer indirection.) Specifically, I'm worried about the case in which hundreds of threads try to access the same block and probing in the hash table degrades to linear search on the pile of entries with the same key. This change is a big step in the direction of supporting stacked SecondaryCaches, but there are obstacles to completing that. Especially, there is no SecondaryCache hook for evictions to pass from one to the next. It has been proposed that evictions be transmitted simply as the persisted data (as in SaveToCallback), but given the current structure provided by the CacheItemHelpers, that would require an extra copy of the block data, because there's intentionally no way to ask for a contiguous Slice of the data (to allow for flexibility in storage). `AsyncLookupHandle` and the re-worked `WaitAll()` should be essentially prepared for stacked SecondaryCaches, but several "TODO with stacked secondaries" issues remain in various places. It could be argued that the stacking instead be done as a SecondaryCache adapter that wraps two (or more) SecondaryCaches, but at least with the current API that would require an extra heap allocation on SecondaryCache Lookup for a wrapper SecondaryCacheResultHandle that can transfer a Lookup between secondaries. We could also consider trying to unify the Cache and SecondaryCache APIs, though that might be difficult if `AsyncLookupHandle` is kept a fixed struct. ## cache.h (public API) Moves `secondary_cache` option from LRUCacheOptions to ShardedCacheOptions so that it is applicable to HyperClockCache. ## advanced_cache.h (advanced public API) * Add `Cache::CreateStandalone()` so that the SecondaryCache support wrapper can use it. * Add `SetEvictionCallback()` / `eviction_callback_` so that the SecondaryCache support wrapper can use it. Only a single callback is supported for efficiency. If there is ever a need for more than one, hopefully that can be handled with a broadcast callback wrapper. These are essentially the two "extra" pieces of `Cache` for pulling out specific SecondaryCache support from the `Cache` implementation. I think it's a good trade-off as these are reasonable, limited, and reusable "cut points" into the `Cache` implementations. * Remove async capability from standard `Lookup()` (getting rid of awkward restrictions on pending Handles) and add `AsyncLookupHandle` and `StartAsyncLookup()`. As noted in the comments, the full struct of `AsyncLookupHandle` is exposed so that it can be stack allocated, for efficiency, though more data is being copied around than before, which could impact performance. (Lookup info -> AsyncLookupHandle -> Handle vs. Lookup info -> Handle) I could foresee a future in which a Cache internally saves a pointer to the AsyncLookupHandle, which means it's dangerous to allow it to be copyable or even movable. It also means it's not compatible with std::vector (which I don't like requiring as an API parameter anyway), so `WaitAll()` expects any contiguous array of AsyncLookupHandles. I believe this is best for common case efficiency, while behaving well in other cases also. For example, `WaitAll()` has no effect on default-constructed AsyncLookupHandles, which look like a completed cache miss. ## cacheable_entry.h A couple of functions are obsolete because Cache::Handle can no longer be pending. ## cache.cc Provides default implementations for new or revamped Cache functions, especially appropriate for non-blocking caches. ## secondary_cache_adapter.{h,cc} The full details of the Cache wrapper adding SecondaryCache support. Essentially replicates the SecondaryCache handling that was in LRUCache, but obviously refactored. There is a bit of logic duplication, where Lookup() is essentially a manually optimized version of StartAsyncLookup() and Wait(), but it's roughly a dozen lines of code. ## sharded_cache.h, typed_cache.h, charged_cache.{h,cc}, sim_cache.cc Simply updated for Cache API changes. ## lru_cache.{h,cc} Carefully remove SecondaryCache logic, implement `CreateStandalone` and eviction handler functionality. ## clock_cache.{h,cc} Expose existing `CreateStandalone` functionality, add eviction handler functionality. Light refactoring. ## block_based_table_reader* Mostly re-worked the only usage of async Lookup, which is in BlockBasedTable::MultiGet. Used arrays in place of autovector in some places for efficiency. Simplified some logic by not trying to process some cache results before they're all ready. Created new function `BlockBasedTable::GetCachePriority()` to reduce some pre-existing code duplication (and avoid making it worse). Fixed at least one small bug from the prior confusing mixture of async and sync Lookups. In MaybeReadBlockAndLoadToCache(), called by RetrieveBlock(), called by MultiGet() with wait=false, is_cache_hit for the block_cache_tracer entry would not be set to true if the handle was pending after Lookup and before Wait. ## Intended follow-up work * Figure out if there are any missing stats or block_cache_tracer work in refactored BlockBasedTable::MultiGet * Stacked secondary caches (see above discussion) * See if we can make up for the small MultiGet performance regression. * Study more performance with SecondaryCache * Items evicted from over-full LRUCache in Release were not being demoted to SecondaryCache, and still aren't to minimize unit test churn. Ideally they would be demoted, but it's an exceptional case so not a big deal. * Use CreateStandalone for cache reservations (save unnecessary hash table operations). Not a big deal, but worthy cleanup. * Somehow I got the contract for SecondaryCache::Insert wrong in #10945. (Doesn't take ownership!) That API comment needs to be fixed, but didn't want to mingle that in here. Pull Request resolved: https://github.com/facebook/rocksdb/pull/11301 Test Plan: ## Unit tests Generally updated to include HCC in SecondaryCache tests, though HyperClockCache has some different, less strict behaviors that leads to some tests not really being set up to work with it. Some of the tests remain disabled with it, but I think we have good coverage without them. ## Crash/stress test Updated to use the new combination. ## Performance First, let's check for regression on caches without secondary cache configured. Adding support for the eviction callback is likely to have a tiny effect, but it shouldn't be worrisome. LRUCache could benefit slightly from less logic around SecondaryCache handling. We can test with cache_bench default settings, built with DEBUG_LEVEL=0 and PORTABLE=0. ``` (while :; do base/cache_bench --cache_type=hyper_clock_cache | grep Rough; done) | awk '{ sum += $9; count++; print $0; print "Average: " int(sum / count) }' ``` **Before** this and #11299 (which could also have a small effect), running for about an hour, before & after running concurrently for each cache type: HyperClockCache: 3168662 (average parallel ops/sec) LRUCache: 2940127 **After** this and #11299, running for about an hour: HyperClockCache: 3164862 (average parallel ops/sec) (0.12% slower) LRUCache: 2940928 (0.03% faster) This is an acceptable difference IMHO. Next, let's consider essentially the worst case of new CPU overhead affecting overall performance. MultiGet uses the async lookup interface regardless of whether SecondaryCache or folly are used. We can configure a benchmark where all block cache queries are for data blocks, and all are hits. Create DB and test (before and after tests running simultaneously): ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 TEST_TMPDIR=/dev/shm base/db_bench -benchmarks=multireadrandom[-X30] -readonly -multiread_batched -batch_size=32 -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16 ``` **Before**: multireadrandom [AVG 30 runs] : 3444202 (± 57049) ops/sec; 240.9 (± 4.0) MB/sec multireadrandom [MEDIAN 30 runs] : 3514443 ops/sec; 245.8 MB/sec **After**: multireadrandom [AVG 30 runs] : 3291022 (± 58851) ops/sec; 230.2 (± 4.1) MB/sec multireadrandom [MEDIAN 30 runs] : 3366179 ops/sec; 235.4 MB/sec So that's roughly a 3% regression, on kind of a *worst case* test of MultiGet CPU. Similar story with HyperClockCache: **Before**: multireadrandom [AVG 30 runs] : 3933777 (± 41840) ops/sec; 275.1 (± 2.9) MB/sec multireadrandom [MEDIAN 30 runs] : 3970667 ops/sec; 277.7 MB/sec **After**: multireadrandom [AVG 30 runs] : 3755338 (± 30391) ops/sec; 262.6 (± 2.1) MB/sec multireadrandom [MEDIAN 30 runs] : 3785696 ops/sec; 264.8 MB/sec Roughly a 4-5% regression. Not ideal, but not the whole story, fortunately. Let's also look at Get() in db_bench: ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X30] -readonly -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16 ``` **Before**: readrandom [AVG 30 runs] : 2198685 (± 13412) ops/sec; 153.8 (± 0.9) MB/sec readrandom [MEDIAN 30 runs] : 2209498 ops/sec; 154.5 MB/sec **After**: readrandom [AVG 30 runs] : 2292814 (± 43508) ops/sec; 160.3 (± 3.0) MB/sec readrandom [MEDIAN 30 runs] : 2365181 ops/sec; 165.4 MB/sec That's showing roughly a 4% improvement, perhaps because of the secondary cache code that is no longer part of LRUCache. But weirdly, HyperClockCache is also showing 2-3% improvement: **Before**: readrandom [AVG 30 runs] : 2272333 (± 9992) ops/sec; 158.9 (± 0.7) MB/sec readrandom [MEDIAN 30 runs] : 2273239 ops/sec; 159.0 MB/sec **After**: readrandom [AVG 30 runs] : 2332407 (± 11252) ops/sec; 163.1 (± 0.8) MB/sec readrandom [MEDIAN 30 runs] : 2335329 ops/sec; 163.3 MB/sec Reviewed By: ltamasi Differential Revision: D44177044 Pulled By: pdillinger fbshipit-source-id: e808e48ff3fe2f792a79841ba617be98e48689f5
2 years ago
inline void MarkEmpty(ClockHandle& h) {
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
#ifndef NDEBUG
// Mark slot as empty, with assertion
uint64_t meta = h.meta.exchange(0, std::memory_order_release);
assert(meta >> ClockHandle::kStateShift == ClockHandle::kStateConstruction);
#else
// Mark slot as empty
h.meta.store(0, std::memory_order_release);
#endif
}
HyperClockCache support for SecondaryCache, with refactoring (#11301) Summary: Internally refactors SecondaryCache integration out of LRUCache specifically and into a wrapper/adapter class that works with various Cache implementations. Notably, this relies on separating the notion of async lookup handles from other cache handles, so that HyperClockCache doesn't have to deal with the problem of allocating handles from the hash table for lookups that might fail anyway, and might be on the same key without support for coalescing. (LRUCache's hash table can incorporate previously allocated handles thanks to its pointer indirection.) Specifically, I'm worried about the case in which hundreds of threads try to access the same block and probing in the hash table degrades to linear search on the pile of entries with the same key. This change is a big step in the direction of supporting stacked SecondaryCaches, but there are obstacles to completing that. Especially, there is no SecondaryCache hook for evictions to pass from one to the next. It has been proposed that evictions be transmitted simply as the persisted data (as in SaveToCallback), but given the current structure provided by the CacheItemHelpers, that would require an extra copy of the block data, because there's intentionally no way to ask for a contiguous Slice of the data (to allow for flexibility in storage). `AsyncLookupHandle` and the re-worked `WaitAll()` should be essentially prepared for stacked SecondaryCaches, but several "TODO with stacked secondaries" issues remain in various places. It could be argued that the stacking instead be done as a SecondaryCache adapter that wraps two (or more) SecondaryCaches, but at least with the current API that would require an extra heap allocation on SecondaryCache Lookup for a wrapper SecondaryCacheResultHandle that can transfer a Lookup between secondaries. We could also consider trying to unify the Cache and SecondaryCache APIs, though that might be difficult if `AsyncLookupHandle` is kept a fixed struct. ## cache.h (public API) Moves `secondary_cache` option from LRUCacheOptions to ShardedCacheOptions so that it is applicable to HyperClockCache. ## advanced_cache.h (advanced public API) * Add `Cache::CreateStandalone()` so that the SecondaryCache support wrapper can use it. * Add `SetEvictionCallback()` / `eviction_callback_` so that the SecondaryCache support wrapper can use it. Only a single callback is supported for efficiency. If there is ever a need for more than one, hopefully that can be handled with a broadcast callback wrapper. These are essentially the two "extra" pieces of `Cache` for pulling out specific SecondaryCache support from the `Cache` implementation. I think it's a good trade-off as these are reasonable, limited, and reusable "cut points" into the `Cache` implementations. * Remove async capability from standard `Lookup()` (getting rid of awkward restrictions on pending Handles) and add `AsyncLookupHandle` and `StartAsyncLookup()`. As noted in the comments, the full struct of `AsyncLookupHandle` is exposed so that it can be stack allocated, for efficiency, though more data is being copied around than before, which could impact performance. (Lookup info -> AsyncLookupHandle -> Handle vs. Lookup info -> Handle) I could foresee a future in which a Cache internally saves a pointer to the AsyncLookupHandle, which means it's dangerous to allow it to be copyable or even movable. It also means it's not compatible with std::vector (which I don't like requiring as an API parameter anyway), so `WaitAll()` expects any contiguous array of AsyncLookupHandles. I believe this is best for common case efficiency, while behaving well in other cases also. For example, `WaitAll()` has no effect on default-constructed AsyncLookupHandles, which look like a completed cache miss. ## cacheable_entry.h A couple of functions are obsolete because Cache::Handle can no longer be pending. ## cache.cc Provides default implementations for new or revamped Cache functions, especially appropriate for non-blocking caches. ## secondary_cache_adapter.{h,cc} The full details of the Cache wrapper adding SecondaryCache support. Essentially replicates the SecondaryCache handling that was in LRUCache, but obviously refactored. There is a bit of logic duplication, where Lookup() is essentially a manually optimized version of StartAsyncLookup() and Wait(), but it's roughly a dozen lines of code. ## sharded_cache.h, typed_cache.h, charged_cache.{h,cc}, sim_cache.cc Simply updated for Cache API changes. ## lru_cache.{h,cc} Carefully remove SecondaryCache logic, implement `CreateStandalone` and eviction handler functionality. ## clock_cache.{h,cc} Expose existing `CreateStandalone` functionality, add eviction handler functionality. Light refactoring. ## block_based_table_reader* Mostly re-worked the only usage of async Lookup, which is in BlockBasedTable::MultiGet. Used arrays in place of autovector in some places for efficiency. Simplified some logic by not trying to process some cache results before they're all ready. Created new function `BlockBasedTable::GetCachePriority()` to reduce some pre-existing code duplication (and avoid making it worse). Fixed at least one small bug from the prior confusing mixture of async and sync Lookups. In MaybeReadBlockAndLoadToCache(), called by RetrieveBlock(), called by MultiGet() with wait=false, is_cache_hit for the block_cache_tracer entry would not be set to true if the handle was pending after Lookup and before Wait. ## Intended follow-up work * Figure out if there are any missing stats or block_cache_tracer work in refactored BlockBasedTable::MultiGet * Stacked secondary caches (see above discussion) * See if we can make up for the small MultiGet performance regression. * Study more performance with SecondaryCache * Items evicted from over-full LRUCache in Release were not being demoted to SecondaryCache, and still aren't to minimize unit test churn. Ideally they would be demoted, but it's an exceptional case so not a big deal. * Use CreateStandalone for cache reservations (save unnecessary hash table operations). Not a big deal, but worthy cleanup. * Somehow I got the contract for SecondaryCache::Insert wrong in #10945. (Doesn't take ownership!) That API comment needs to be fixed, but didn't want to mingle that in here. Pull Request resolved: https://github.com/facebook/rocksdb/pull/11301 Test Plan: ## Unit tests Generally updated to include HCC in SecondaryCache tests, though HyperClockCache has some different, less strict behaviors that leads to some tests not really being set up to work with it. Some of the tests remain disabled with it, but I think we have good coverage without them. ## Crash/stress test Updated to use the new combination. ## Performance First, let's check for regression on caches without secondary cache configured. Adding support for the eviction callback is likely to have a tiny effect, but it shouldn't be worrisome. LRUCache could benefit slightly from less logic around SecondaryCache handling. We can test with cache_bench default settings, built with DEBUG_LEVEL=0 and PORTABLE=0. ``` (while :; do base/cache_bench --cache_type=hyper_clock_cache | grep Rough; done) | awk '{ sum += $9; count++; print $0; print "Average: " int(sum / count) }' ``` **Before** this and #11299 (which could also have a small effect), running for about an hour, before & after running concurrently for each cache type: HyperClockCache: 3168662 (average parallel ops/sec) LRUCache: 2940127 **After** this and #11299, running for about an hour: HyperClockCache: 3164862 (average parallel ops/sec) (0.12% slower) LRUCache: 2940928 (0.03% faster) This is an acceptable difference IMHO. Next, let's consider essentially the worst case of new CPU overhead affecting overall performance. MultiGet uses the async lookup interface regardless of whether SecondaryCache or folly are used. We can configure a benchmark where all block cache queries are for data blocks, and all are hits. Create DB and test (before and after tests running simultaneously): ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 TEST_TMPDIR=/dev/shm base/db_bench -benchmarks=multireadrandom[-X30] -readonly -multiread_batched -batch_size=32 -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16 ``` **Before**: multireadrandom [AVG 30 runs] : 3444202 (± 57049) ops/sec; 240.9 (± 4.0) MB/sec multireadrandom [MEDIAN 30 runs] : 3514443 ops/sec; 245.8 MB/sec **After**: multireadrandom [AVG 30 runs] : 3291022 (± 58851) ops/sec; 230.2 (± 4.1) MB/sec multireadrandom [MEDIAN 30 runs] : 3366179 ops/sec; 235.4 MB/sec So that's roughly a 3% regression, on kind of a *worst case* test of MultiGet CPU. Similar story with HyperClockCache: **Before**: multireadrandom [AVG 30 runs] : 3933777 (± 41840) ops/sec; 275.1 (± 2.9) MB/sec multireadrandom [MEDIAN 30 runs] : 3970667 ops/sec; 277.7 MB/sec **After**: multireadrandom [AVG 30 runs] : 3755338 (± 30391) ops/sec; 262.6 (± 2.1) MB/sec multireadrandom [MEDIAN 30 runs] : 3785696 ops/sec; 264.8 MB/sec Roughly a 4-5% regression. Not ideal, but not the whole story, fortunately. Let's also look at Get() in db_bench: ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X30] -readonly -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16 ``` **Before**: readrandom [AVG 30 runs] : 2198685 (± 13412) ops/sec; 153.8 (± 0.9) MB/sec readrandom [MEDIAN 30 runs] : 2209498 ops/sec; 154.5 MB/sec **After**: readrandom [AVG 30 runs] : 2292814 (± 43508) ops/sec; 160.3 (± 3.0) MB/sec readrandom [MEDIAN 30 runs] : 2365181 ops/sec; 165.4 MB/sec That's showing roughly a 4% improvement, perhaps because of the secondary cache code that is no longer part of LRUCache. But weirdly, HyperClockCache is also showing 2-3% improvement: **Before**: readrandom [AVG 30 runs] : 2272333 (± 9992) ops/sec; 158.9 (± 0.7) MB/sec readrandom [MEDIAN 30 runs] : 2273239 ops/sec; 159.0 MB/sec **After**: readrandom [AVG 30 runs] : 2332407 (± 11252) ops/sec; 163.1 (± 0.8) MB/sec readrandom [MEDIAN 30 runs] : 2335329 ops/sec; 163.3 MB/sec Reviewed By: ltamasi Differential Revision: D44177044 Pulled By: pdillinger fbshipit-source-id: e808e48ff3fe2f792a79841ba617be98e48689f5
2 years ago
inline void FreeDataMarkEmpty(ClockHandle& h, MemoryAllocator* allocator) {
// NOTE: in theory there's more room for parallelism if we copy the handle
// data and delay actions like this until after marking the entry as empty,
// but performance tests only show a regression by copying the few words
// of data.
h.FreeData(allocator);
MarkEmpty(h);
}
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
inline bool ClockUpdate(ClockHandle& h) {
uint64_t meta = h.meta.load(std::memory_order_relaxed);
uint64_t acquire_count =
(meta >> ClockHandle::kAcquireCounterShift) & ClockHandle::kCounterMask;
uint64_t release_count =
(meta >> ClockHandle::kReleaseCounterShift) & ClockHandle::kCounterMask;
// fprintf(stderr, "ClockUpdate @ %p: %lu %lu %u\n", &h, acquire_count,
// release_count, (unsigned)(meta >> ClockHandle::kStateShift));
if (acquire_count != release_count) {
// Only clock update entries with no outstanding refs
return false;
}
if (!((meta >> ClockHandle::kStateShift) & ClockHandle::kStateShareableBit)) {
// Only clock update Shareable entries
return false;
}
if ((meta >> ClockHandle::kStateShift == ClockHandle::kStateVisible) &&
acquire_count > 0) {
// Decrement clock
uint64_t new_count =
std::min(acquire_count - 1, uint64_t{ClockHandle::kMaxCountdown} - 1);
// Compare-exchange in the decremented clock info, but
// not aggressively
uint64_t new_meta =
(uint64_t{ClockHandle::kStateVisible} << ClockHandle::kStateShift) |
(new_count << ClockHandle::kReleaseCounterShift) |
(new_count << ClockHandle::kAcquireCounterShift);
h.meta.compare_exchange_strong(meta, new_meta, std::memory_order_relaxed);
return false;
}
// Otherwise, remove entry (either unreferenced invisible or
// unreferenced and expired visible).
if (h.meta.compare_exchange_strong(
meta,
uint64_t{ClockHandle::kStateConstruction} << ClockHandle::kStateShift,
std::memory_order_acquire)) {
// Took ownership.
return true;
} else {
// Compare-exchange failing probably
// indicates the entry was used, so skip it in that case.
return false;
}
}
} // namespace
Major Cache refactoring, CPU efficiency improvement (#10975) Summary: This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache). The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below. * static_cast lines of code +29 -35 (net removed 6) * reinterpret_cast lines of code +6 -32 (net removed 26) ## cache.h and secondary_cache.h * Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications: * Simpler for implementations to deal with just one Insert and one Lookup. * Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters * Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428. * Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks). * It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below). * I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc. * Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation. * Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.) * Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.) * Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774) * Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object. * Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change. ## typed_cache.h Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae). The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used. * PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value. * BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter. * FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue. * For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`. These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.) Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it. ## block_cache.h This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table. ## block_based_table_reader.cc Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation. The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions. ## block_based_table_builder.cc, cache_dump_load_impl.cc Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.) ## Everything else Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975 Test Plan: tests updated Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache): 34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844 34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594 34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297 34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523 34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602 34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293 34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926 34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488 233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984 233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922 233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559 233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93 233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418 233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273 233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691 233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82 1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55 1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02 1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45 1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24 1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92 1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78 1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36 1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83 Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn. Reviewed By: anand1976 Differential Revision: D42417818 Pulled By: pdillinger fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
void ClockHandleBasicData::FreeData(MemoryAllocator* allocator) const {
if (helper->del_cb) {
helper->del_cb(value, allocator);
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
}
}
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
HyperClockTable::HyperClockTable(
size_t capacity, bool /*strict_capacity_limit*/,
Major Cache refactoring, CPU efficiency improvement (#10975) Summary: This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache). The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below. * static_cast lines of code +29 -35 (net removed 6) * reinterpret_cast lines of code +6 -32 (net removed 26) ## cache.h and secondary_cache.h * Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications: * Simpler for implementations to deal with just one Insert and one Lookup. * Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters * Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428. * Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks). * It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below). * I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc. * Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation. * Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.) * Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.) * Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774) * Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object. * Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change. ## typed_cache.h Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae). The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used. * PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value. * BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter. * FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue. * For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`. These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.) Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it. ## block_cache.h This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table. ## block_based_table_reader.cc Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation. The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions. ## block_based_table_builder.cc, cache_dump_load_impl.cc Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.) ## Everything else Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975 Test Plan: tests updated Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache): 34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844 34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594 34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297 34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523 34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602 34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293 34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926 34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488 233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984 233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922 233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559 233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93 233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418 233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273 233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691 233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82 1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55 1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02 1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45 1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24 1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92 1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78 1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36 1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83 Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn. Reviewed By: anand1976 Differential Revision: D42417818 Pulled By: pdillinger fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
CacheMetadataChargePolicy metadata_charge_policy,
HyperClockCache support for SecondaryCache, with refactoring (#11301) Summary: Internally refactors SecondaryCache integration out of LRUCache specifically and into a wrapper/adapter class that works with various Cache implementations. Notably, this relies on separating the notion of async lookup handles from other cache handles, so that HyperClockCache doesn't have to deal with the problem of allocating handles from the hash table for lookups that might fail anyway, and might be on the same key without support for coalescing. (LRUCache's hash table can incorporate previously allocated handles thanks to its pointer indirection.) Specifically, I'm worried about the case in which hundreds of threads try to access the same block and probing in the hash table degrades to linear search on the pile of entries with the same key. This change is a big step in the direction of supporting stacked SecondaryCaches, but there are obstacles to completing that. Especially, there is no SecondaryCache hook for evictions to pass from one to the next. It has been proposed that evictions be transmitted simply as the persisted data (as in SaveToCallback), but given the current structure provided by the CacheItemHelpers, that would require an extra copy of the block data, because there's intentionally no way to ask for a contiguous Slice of the data (to allow for flexibility in storage). `AsyncLookupHandle` and the re-worked `WaitAll()` should be essentially prepared for stacked SecondaryCaches, but several "TODO with stacked secondaries" issues remain in various places. It could be argued that the stacking instead be done as a SecondaryCache adapter that wraps two (or more) SecondaryCaches, but at least with the current API that would require an extra heap allocation on SecondaryCache Lookup for a wrapper SecondaryCacheResultHandle that can transfer a Lookup between secondaries. We could also consider trying to unify the Cache and SecondaryCache APIs, though that might be difficult if `AsyncLookupHandle` is kept a fixed struct. ## cache.h (public API) Moves `secondary_cache` option from LRUCacheOptions to ShardedCacheOptions so that it is applicable to HyperClockCache. ## advanced_cache.h (advanced public API) * Add `Cache::CreateStandalone()` so that the SecondaryCache support wrapper can use it. * Add `SetEvictionCallback()` / `eviction_callback_` so that the SecondaryCache support wrapper can use it. Only a single callback is supported for efficiency. If there is ever a need for more than one, hopefully that can be handled with a broadcast callback wrapper. These are essentially the two "extra" pieces of `Cache` for pulling out specific SecondaryCache support from the `Cache` implementation. I think it's a good trade-off as these are reasonable, limited, and reusable "cut points" into the `Cache` implementations. * Remove async capability from standard `Lookup()` (getting rid of awkward restrictions on pending Handles) and add `AsyncLookupHandle` and `StartAsyncLookup()`. As noted in the comments, the full struct of `AsyncLookupHandle` is exposed so that it can be stack allocated, for efficiency, though more data is being copied around than before, which could impact performance. (Lookup info -> AsyncLookupHandle -> Handle vs. Lookup info -> Handle) I could foresee a future in which a Cache internally saves a pointer to the AsyncLookupHandle, which means it's dangerous to allow it to be copyable or even movable. It also means it's not compatible with std::vector (which I don't like requiring as an API parameter anyway), so `WaitAll()` expects any contiguous array of AsyncLookupHandles. I believe this is best for common case efficiency, while behaving well in other cases also. For example, `WaitAll()` has no effect on default-constructed AsyncLookupHandles, which look like a completed cache miss. ## cacheable_entry.h A couple of functions are obsolete because Cache::Handle can no longer be pending. ## cache.cc Provides default implementations for new or revamped Cache functions, especially appropriate for non-blocking caches. ## secondary_cache_adapter.{h,cc} The full details of the Cache wrapper adding SecondaryCache support. Essentially replicates the SecondaryCache handling that was in LRUCache, but obviously refactored. There is a bit of logic duplication, where Lookup() is essentially a manually optimized version of StartAsyncLookup() and Wait(), but it's roughly a dozen lines of code. ## sharded_cache.h, typed_cache.h, charged_cache.{h,cc}, sim_cache.cc Simply updated for Cache API changes. ## lru_cache.{h,cc} Carefully remove SecondaryCache logic, implement `CreateStandalone` and eviction handler functionality. ## clock_cache.{h,cc} Expose existing `CreateStandalone` functionality, add eviction handler functionality. Light refactoring. ## block_based_table_reader* Mostly re-worked the only usage of async Lookup, which is in BlockBasedTable::MultiGet. Used arrays in place of autovector in some places for efficiency. Simplified some logic by not trying to process some cache results before they're all ready. Created new function `BlockBasedTable::GetCachePriority()` to reduce some pre-existing code duplication (and avoid making it worse). Fixed at least one small bug from the prior confusing mixture of async and sync Lookups. In MaybeReadBlockAndLoadToCache(), called by RetrieveBlock(), called by MultiGet() with wait=false, is_cache_hit for the block_cache_tracer entry would not be set to true if the handle was pending after Lookup and before Wait. ## Intended follow-up work * Figure out if there are any missing stats or block_cache_tracer work in refactored BlockBasedTable::MultiGet * Stacked secondary caches (see above discussion) * See if we can make up for the small MultiGet performance regression. * Study more performance with SecondaryCache * Items evicted from over-full LRUCache in Release were not being demoted to SecondaryCache, and still aren't to minimize unit test churn. Ideally they would be demoted, but it's an exceptional case so not a big deal. * Use CreateStandalone for cache reservations (save unnecessary hash table operations). Not a big deal, but worthy cleanup. * Somehow I got the contract for SecondaryCache::Insert wrong in #10945. (Doesn't take ownership!) That API comment needs to be fixed, but didn't want to mingle that in here. Pull Request resolved: https://github.com/facebook/rocksdb/pull/11301 Test Plan: ## Unit tests Generally updated to include HCC in SecondaryCache tests, though HyperClockCache has some different, less strict behaviors that leads to some tests not really being set up to work with it. Some of the tests remain disabled with it, but I think we have good coverage without them. ## Crash/stress test Updated to use the new combination. ## Performance First, let's check for regression on caches without secondary cache configured. Adding support for the eviction callback is likely to have a tiny effect, but it shouldn't be worrisome. LRUCache could benefit slightly from less logic around SecondaryCache handling. We can test with cache_bench default settings, built with DEBUG_LEVEL=0 and PORTABLE=0. ``` (while :; do base/cache_bench --cache_type=hyper_clock_cache | grep Rough; done) | awk '{ sum += $9; count++; print $0; print "Average: " int(sum / count) }' ``` **Before** this and #11299 (which could also have a small effect), running for about an hour, before & after running concurrently for each cache type: HyperClockCache: 3168662 (average parallel ops/sec) LRUCache: 2940127 **After** this and #11299, running for about an hour: HyperClockCache: 3164862 (average parallel ops/sec) (0.12% slower) LRUCache: 2940928 (0.03% faster) This is an acceptable difference IMHO. Next, let's consider essentially the worst case of new CPU overhead affecting overall performance. MultiGet uses the async lookup interface regardless of whether SecondaryCache or folly are used. We can configure a benchmark where all block cache queries are for data blocks, and all are hits. Create DB and test (before and after tests running simultaneously): ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 TEST_TMPDIR=/dev/shm base/db_bench -benchmarks=multireadrandom[-X30] -readonly -multiread_batched -batch_size=32 -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16 ``` **Before**: multireadrandom [AVG 30 runs] : 3444202 (± 57049) ops/sec; 240.9 (± 4.0) MB/sec multireadrandom [MEDIAN 30 runs] : 3514443 ops/sec; 245.8 MB/sec **After**: multireadrandom [AVG 30 runs] : 3291022 (± 58851) ops/sec; 230.2 (± 4.1) MB/sec multireadrandom [MEDIAN 30 runs] : 3366179 ops/sec; 235.4 MB/sec So that's roughly a 3% regression, on kind of a *worst case* test of MultiGet CPU. Similar story with HyperClockCache: **Before**: multireadrandom [AVG 30 runs] : 3933777 (± 41840) ops/sec; 275.1 (± 2.9) MB/sec multireadrandom [MEDIAN 30 runs] : 3970667 ops/sec; 277.7 MB/sec **After**: multireadrandom [AVG 30 runs] : 3755338 (± 30391) ops/sec; 262.6 (± 2.1) MB/sec multireadrandom [MEDIAN 30 runs] : 3785696 ops/sec; 264.8 MB/sec Roughly a 4-5% regression. Not ideal, but not the whole story, fortunately. Let's also look at Get() in db_bench: ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X30] -readonly -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16 ``` **Before**: readrandom [AVG 30 runs] : 2198685 (± 13412) ops/sec; 153.8 (± 0.9) MB/sec readrandom [MEDIAN 30 runs] : 2209498 ops/sec; 154.5 MB/sec **After**: readrandom [AVG 30 runs] : 2292814 (± 43508) ops/sec; 160.3 (± 3.0) MB/sec readrandom [MEDIAN 30 runs] : 2365181 ops/sec; 165.4 MB/sec That's showing roughly a 4% improvement, perhaps because of the secondary cache code that is no longer part of LRUCache. But weirdly, HyperClockCache is also showing 2-3% improvement: **Before**: readrandom [AVG 30 runs] : 2272333 (± 9992) ops/sec; 158.9 (± 0.7) MB/sec readrandom [MEDIAN 30 runs] : 2273239 ops/sec; 159.0 MB/sec **After**: readrandom [AVG 30 runs] : 2332407 (± 11252) ops/sec; 163.1 (± 0.8) MB/sec readrandom [MEDIAN 30 runs] : 2335329 ops/sec; 163.3 MB/sec Reviewed By: ltamasi Differential Revision: D44177044 Pulled By: pdillinger fbshipit-source-id: e808e48ff3fe2f792a79841ba617be98e48689f5
2 years ago
MemoryAllocator* allocator,
const Cache::EvictionCallback* eviction_callback, const Opts& opts)
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
: length_bits_(CalcHashBits(capacity, opts.estimated_value_size,
metadata_charge_policy)),
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
length_bits_mask_((size_t{1} << length_bits_) - 1),
occupancy_limit_(static_cast<size_t>((uint64_t{1} << length_bits_) *
kStrictLoadFactor)),
Major Cache refactoring, CPU efficiency improvement (#10975) Summary: This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache). The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below. * static_cast lines of code +29 -35 (net removed 6) * reinterpret_cast lines of code +6 -32 (net removed 26) ## cache.h and secondary_cache.h * Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications: * Simpler for implementations to deal with just one Insert and one Lookup. * Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters * Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428. * Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks). * It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below). * I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc. * Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation. * Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.) * Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.) * Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774) * Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object. * Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change. ## typed_cache.h Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae). The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used. * PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value. * BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter. * FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue. * For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`. These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.) Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it. ## block_cache.h This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table. ## block_based_table_reader.cc Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation. The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions. ## block_based_table_builder.cc, cache_dump_load_impl.cc Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.) ## Everything else Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975 Test Plan: tests updated Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache): 34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844 34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594 34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297 34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523 34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602 34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293 34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926 34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488 233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984 233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922 233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559 233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93 233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418 233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273 233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691 233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82 1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55 1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02 1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45 1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24 1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92 1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78 1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36 1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83 Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn. Reviewed By: anand1976 Differential Revision: D42417818 Pulled By: pdillinger fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
array_(new HandleImpl[size_t{1} << length_bits_]),
HyperClockCache support for SecondaryCache, with refactoring (#11301) Summary: Internally refactors SecondaryCache integration out of LRUCache specifically and into a wrapper/adapter class that works with various Cache implementations. Notably, this relies on separating the notion of async lookup handles from other cache handles, so that HyperClockCache doesn't have to deal with the problem of allocating handles from the hash table for lookups that might fail anyway, and might be on the same key without support for coalescing. (LRUCache's hash table can incorporate previously allocated handles thanks to its pointer indirection.) Specifically, I'm worried about the case in which hundreds of threads try to access the same block and probing in the hash table degrades to linear search on the pile of entries with the same key. This change is a big step in the direction of supporting stacked SecondaryCaches, but there are obstacles to completing that. Especially, there is no SecondaryCache hook for evictions to pass from one to the next. It has been proposed that evictions be transmitted simply as the persisted data (as in SaveToCallback), but given the current structure provided by the CacheItemHelpers, that would require an extra copy of the block data, because there's intentionally no way to ask for a contiguous Slice of the data (to allow for flexibility in storage). `AsyncLookupHandle` and the re-worked `WaitAll()` should be essentially prepared for stacked SecondaryCaches, but several "TODO with stacked secondaries" issues remain in various places. It could be argued that the stacking instead be done as a SecondaryCache adapter that wraps two (or more) SecondaryCaches, but at least with the current API that would require an extra heap allocation on SecondaryCache Lookup for a wrapper SecondaryCacheResultHandle that can transfer a Lookup between secondaries. We could also consider trying to unify the Cache and SecondaryCache APIs, though that might be difficult if `AsyncLookupHandle` is kept a fixed struct. ## cache.h (public API) Moves `secondary_cache` option from LRUCacheOptions to ShardedCacheOptions so that it is applicable to HyperClockCache. ## advanced_cache.h (advanced public API) * Add `Cache::CreateStandalone()` so that the SecondaryCache support wrapper can use it. * Add `SetEvictionCallback()` / `eviction_callback_` so that the SecondaryCache support wrapper can use it. Only a single callback is supported for efficiency. If there is ever a need for more than one, hopefully that can be handled with a broadcast callback wrapper. These are essentially the two "extra" pieces of `Cache` for pulling out specific SecondaryCache support from the `Cache` implementation. I think it's a good trade-off as these are reasonable, limited, and reusable "cut points" into the `Cache` implementations. * Remove async capability from standard `Lookup()` (getting rid of awkward restrictions on pending Handles) and add `AsyncLookupHandle` and `StartAsyncLookup()`. As noted in the comments, the full struct of `AsyncLookupHandle` is exposed so that it can be stack allocated, for efficiency, though more data is being copied around than before, which could impact performance. (Lookup info -> AsyncLookupHandle -> Handle vs. Lookup info -> Handle) I could foresee a future in which a Cache internally saves a pointer to the AsyncLookupHandle, which means it's dangerous to allow it to be copyable or even movable. It also means it's not compatible with std::vector (which I don't like requiring as an API parameter anyway), so `WaitAll()` expects any contiguous array of AsyncLookupHandles. I believe this is best for common case efficiency, while behaving well in other cases also. For example, `WaitAll()` has no effect on default-constructed AsyncLookupHandles, which look like a completed cache miss. ## cacheable_entry.h A couple of functions are obsolete because Cache::Handle can no longer be pending. ## cache.cc Provides default implementations for new or revamped Cache functions, especially appropriate for non-blocking caches. ## secondary_cache_adapter.{h,cc} The full details of the Cache wrapper adding SecondaryCache support. Essentially replicates the SecondaryCache handling that was in LRUCache, but obviously refactored. There is a bit of logic duplication, where Lookup() is essentially a manually optimized version of StartAsyncLookup() and Wait(), but it's roughly a dozen lines of code. ## sharded_cache.h, typed_cache.h, charged_cache.{h,cc}, sim_cache.cc Simply updated for Cache API changes. ## lru_cache.{h,cc} Carefully remove SecondaryCache logic, implement `CreateStandalone` and eviction handler functionality. ## clock_cache.{h,cc} Expose existing `CreateStandalone` functionality, add eviction handler functionality. Light refactoring. ## block_based_table_reader* Mostly re-worked the only usage of async Lookup, which is in BlockBasedTable::MultiGet. Used arrays in place of autovector in some places for efficiency. Simplified some logic by not trying to process some cache results before they're all ready. Created new function `BlockBasedTable::GetCachePriority()` to reduce some pre-existing code duplication (and avoid making it worse). Fixed at least one small bug from the prior confusing mixture of async and sync Lookups. In MaybeReadBlockAndLoadToCache(), called by RetrieveBlock(), called by MultiGet() with wait=false, is_cache_hit for the block_cache_tracer entry would not be set to true if the handle was pending after Lookup and before Wait. ## Intended follow-up work * Figure out if there are any missing stats or block_cache_tracer work in refactored BlockBasedTable::MultiGet * Stacked secondary caches (see above discussion) * See if we can make up for the small MultiGet performance regression. * Study more performance with SecondaryCache * Items evicted from over-full LRUCache in Release were not being demoted to SecondaryCache, and still aren't to minimize unit test churn. Ideally they would be demoted, but it's an exceptional case so not a big deal. * Use CreateStandalone for cache reservations (save unnecessary hash table operations). Not a big deal, but worthy cleanup. * Somehow I got the contract for SecondaryCache::Insert wrong in #10945. (Doesn't take ownership!) That API comment needs to be fixed, but didn't want to mingle that in here. Pull Request resolved: https://github.com/facebook/rocksdb/pull/11301 Test Plan: ## Unit tests Generally updated to include HCC in SecondaryCache tests, though HyperClockCache has some different, less strict behaviors that leads to some tests not really being set up to work with it. Some of the tests remain disabled with it, but I think we have good coverage without them. ## Crash/stress test Updated to use the new combination. ## Performance First, let's check for regression on caches without secondary cache configured. Adding support for the eviction callback is likely to have a tiny effect, but it shouldn't be worrisome. LRUCache could benefit slightly from less logic around SecondaryCache handling. We can test with cache_bench default settings, built with DEBUG_LEVEL=0 and PORTABLE=0. ``` (while :; do base/cache_bench --cache_type=hyper_clock_cache | grep Rough; done) | awk '{ sum += $9; count++; print $0; print "Average: " int(sum / count) }' ``` **Before** this and #11299 (which could also have a small effect), running for about an hour, before & after running concurrently for each cache type: HyperClockCache: 3168662 (average parallel ops/sec) LRUCache: 2940127 **After** this and #11299, running for about an hour: HyperClockCache: 3164862 (average parallel ops/sec) (0.12% slower) LRUCache: 2940928 (0.03% faster) This is an acceptable difference IMHO. Next, let's consider essentially the worst case of new CPU overhead affecting overall performance. MultiGet uses the async lookup interface regardless of whether SecondaryCache or folly are used. We can configure a benchmark where all block cache queries are for data blocks, and all are hits. Create DB and test (before and after tests running simultaneously): ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 TEST_TMPDIR=/dev/shm base/db_bench -benchmarks=multireadrandom[-X30] -readonly -multiread_batched -batch_size=32 -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16 ``` **Before**: multireadrandom [AVG 30 runs] : 3444202 (± 57049) ops/sec; 240.9 (± 4.0) MB/sec multireadrandom [MEDIAN 30 runs] : 3514443 ops/sec; 245.8 MB/sec **After**: multireadrandom [AVG 30 runs] : 3291022 (± 58851) ops/sec; 230.2 (± 4.1) MB/sec multireadrandom [MEDIAN 30 runs] : 3366179 ops/sec; 235.4 MB/sec So that's roughly a 3% regression, on kind of a *worst case* test of MultiGet CPU. Similar story with HyperClockCache: **Before**: multireadrandom [AVG 30 runs] : 3933777 (± 41840) ops/sec; 275.1 (± 2.9) MB/sec multireadrandom [MEDIAN 30 runs] : 3970667 ops/sec; 277.7 MB/sec **After**: multireadrandom [AVG 30 runs] : 3755338 (± 30391) ops/sec; 262.6 (± 2.1) MB/sec multireadrandom [MEDIAN 30 runs] : 3785696 ops/sec; 264.8 MB/sec Roughly a 4-5% regression. Not ideal, but not the whole story, fortunately. Let's also look at Get() in db_bench: ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X30] -readonly -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16 ``` **Before**: readrandom [AVG 30 runs] : 2198685 (± 13412) ops/sec; 153.8 (± 0.9) MB/sec readrandom [MEDIAN 30 runs] : 2209498 ops/sec; 154.5 MB/sec **After**: readrandom [AVG 30 runs] : 2292814 (± 43508) ops/sec; 160.3 (± 3.0) MB/sec readrandom [MEDIAN 30 runs] : 2365181 ops/sec; 165.4 MB/sec That's showing roughly a 4% improvement, perhaps because of the secondary cache code that is no longer part of LRUCache. But weirdly, HyperClockCache is also showing 2-3% improvement: **Before**: readrandom [AVG 30 runs] : 2272333 (± 9992) ops/sec; 158.9 (± 0.7) MB/sec readrandom [MEDIAN 30 runs] : 2273239 ops/sec; 159.0 MB/sec **After**: readrandom [AVG 30 runs] : 2332407 (± 11252) ops/sec; 163.1 (± 0.8) MB/sec readrandom [MEDIAN 30 runs] : 2335329 ops/sec; 163.3 MB/sec Reviewed By: ltamasi Differential Revision: D44177044 Pulled By: pdillinger fbshipit-source-id: e808e48ff3fe2f792a79841ba617be98e48689f5
2 years ago
allocator_(allocator),
eviction_callback_(*eviction_callback) {
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
if (metadata_charge_policy ==
CacheMetadataChargePolicy::kFullChargeCacheMetadata) {
usage_ += size_t{GetTableSize()} * sizeof(HandleImpl);
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
}
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
static_assert(sizeof(HandleImpl) == 64U,
"Expecting size / alignment with common cache line size");
}
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
HyperClockTable::~HyperClockTable() {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
// Assumes there are no references or active operations on any slot/element
// in the table.
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
for (size_t i = 0; i < GetTableSize(); i++) {
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
HandleImpl& h = array_[i];
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
switch (h.meta >> ClockHandle::kStateShift) {
case ClockHandle::kStateEmpty:
// noop
break;
case ClockHandle::kStateInvisible: // rare but possible
case ClockHandle::kStateVisible:
assert(GetRefcount(h.meta) == 0);
Major Cache refactoring, CPU efficiency improvement (#10975) Summary: This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache). The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below. * static_cast lines of code +29 -35 (net removed 6) * reinterpret_cast lines of code +6 -32 (net removed 26) ## cache.h and secondary_cache.h * Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications: * Simpler for implementations to deal with just one Insert and one Lookup. * Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters * Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428. * Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks). * It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below). * I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc. * Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation. * Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.) * Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.) * Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774) * Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object. * Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change. ## typed_cache.h Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae). The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used. * PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value. * BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter. * FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue. * For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`. These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.) Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it. ## block_cache.h This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table. ## block_based_table_reader.cc Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation. The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions. ## block_based_table_builder.cc, cache_dump_load_impl.cc Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.) ## Everything else Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975 Test Plan: tests updated Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache): 34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844 34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594 34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297 34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523 34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602 34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293 34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926 34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488 233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984 233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922 233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559 233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93 233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418 233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273 233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691 233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82 1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55 1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02 1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45 1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24 1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92 1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78 1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36 1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83 Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn. Reviewed By: anand1976 Differential Revision: D42417818 Pulled By: pdillinger fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
h.FreeData(allocator_);
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
#ifndef NDEBUG
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
Rollback(h.hashed_key, &h);
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
ReclaimEntryUsage(h.GetTotalCharge());
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
#endif
break;
// otherwise
default:
assert(false);
break;
}
}
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
#ifndef NDEBUG
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
for (size_t i = 0; i < GetTableSize(); i++) {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
assert(array_[i].displacements.load() == 0);
}
#endif
assert(usage_.load() == 0 ||
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
usage_.load() == size_t{GetTableSize()} * sizeof(HandleImpl));
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
assert(occupancy_ == 0);
}
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
// If an entry doesn't receive clock updates but is repeatedly referenced &
// released, the acquire and release counters could overflow without some
// intervention. This is that intervention, which should be inexpensive
// because it only incurs a simple, very predictable check. (Applying a bit
// mask in addition to an increment to every Release likely would be
// relatively expensive, because it's an extra atomic update.)
//
// We do have to assume that we never have many millions of simultaneous
// references to a cache handle, because we cannot represent so many
// references with the difference in counters, masked to the number of
// counter bits. Similarly, we assume there aren't millions of threads
// holding transient references (which might be "undone" rather than
// released by the way).
//
// Consider these possible states for each counter:
// low: less than kMaxCountdown
// medium: kMaxCountdown to half way to overflow + kMaxCountdown
// high: half way to overflow + kMaxCountdown, or greater
//
// And these possible states for the combination of counters:
// acquire / release
// ------- -------
// low low - Normal / common, with caveats (see below)
// medium low - Can happen while holding some refs
// high low - Violates assumptions (too many refs)
// low medium - Violates assumptions (refs underflow, etc.)
// medium medium - Normal (very read heavy cache)
// high medium - Can happen while holding some refs
// low high - This function is supposed to prevent
// medium high - Violates assumptions (refs underflow, etc.)
// high high - Needs CorrectNearOverflow
//
// Basically, this function detects (high, high) state (inferred from
// release alone being high) and bumps it back down to (medium, medium)
// state with the same refcount and the same logical countdown counter
// (everything > kMaxCountdown is logically the same). Note that bumping
// down to (low, low) would modify the countdown counter, so is "reserved"
// in a sense.
//
// If near-overflow correction is triggered here, there's no guarantee
// that another thread hasn't freed the entry and replaced it with another.
// Therefore, it must be the case that the correction does not affect
// entries unless they are very old (many millions of acquire-release cycles).
// (Our bit manipulation is indeed idempotent and only affects entries in
// exceptional cases.) We assume a pre-empted thread will not stall that long.
// If it did, the state could be corrupted in the (unlikely) case that the top
// bit of the acquire counter is set but not the release counter, and thus
// we only clear the top bit of the acquire counter on resumption. It would
// then appear that there are too many refs and the entry would be permanently
// pinned (which is not terrible for an exceptionally rare occurrence), unless
// it is referenced enough (at least kMaxCountdown more times) for the release
// counter to reach "high" state again and bumped back to "medium." (This
// motivates only checking for release counter in high state, not both in high
// state.)
inline void CorrectNearOverflow(uint64_t old_meta,
std::atomic<uint64_t>& meta) {
// We clear both top-most counter bits at the same time.
constexpr uint64_t kCounterTopBit = uint64_t{1}
<< (ClockHandle::kCounterNumBits - 1);
constexpr uint64_t kClearBits =
(kCounterTopBit << ClockHandle::kAcquireCounterShift) |
(kCounterTopBit << ClockHandle::kReleaseCounterShift);
// A simple check that allows us to initiate clearing the top bits for
// a large portion of the "high" state space on release counter.
constexpr uint64_t kCheckBits =
(kCounterTopBit | (ClockHandle::kMaxCountdown + 1))
<< ClockHandle::kReleaseCounterShift;
if (UNLIKELY(old_meta & kCheckBits)) {
meta.fetch_and(~kClearBits, std::memory_order_relaxed);
}
}
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
inline Status HyperClockTable::ChargeUsageMaybeEvictStrict(
size_t total_charge, size_t capacity, bool need_evict_for_occupancy) {
if (total_charge > capacity) {
return Status::MemoryLimit(
"Cache entry too large for a single cache shard: " +
std::to_string(total_charge) + " > " + std::to_string(capacity));
}
// Grab any available capacity, and free up any more required.
size_t old_usage = usage_.load(std::memory_order_relaxed);
size_t new_usage;
if (LIKELY(old_usage != capacity)) {
do {
new_usage = std::min(capacity, old_usage + total_charge);
} while (!usage_.compare_exchange_weak(old_usage, new_usage,
std::memory_order_relaxed));
} else {
new_usage = old_usage;
}
// How much do we need to evict then?
size_t need_evict_charge = old_usage + total_charge - new_usage;
size_t request_evict_charge = need_evict_charge;
if (UNLIKELY(need_evict_for_occupancy) && request_evict_charge == 0) {
// Require at least 1 eviction.
request_evict_charge = 1;
}
if (request_evict_charge > 0) {
size_t evicted_charge = 0;
size_t evicted_count = 0;
Evict(request_evict_charge, &evicted_charge, &evicted_count);
occupancy_.fetch_sub(evicted_count, std::memory_order_release);
if (LIKELY(evicted_charge > need_evict_charge)) {
assert(evicted_count > 0);
// Evicted more than enough
usage_.fetch_sub(evicted_charge - need_evict_charge,
std::memory_order_relaxed);
} else if (evicted_charge < need_evict_charge ||
(UNLIKELY(need_evict_for_occupancy) && evicted_count == 0)) {
// Roll back to old usage minus evicted
usage_.fetch_sub(evicted_charge + (new_usage - old_usage),
std::memory_order_relaxed);
if (evicted_charge < need_evict_charge) {
return Status::MemoryLimit(
"Insert failed because unable to evict entries to stay within "
"capacity limit.");
} else {
return Status::MemoryLimit(
"Insert failed because unable to evict entries to stay within "
"table occupancy limit.");
}
}
// If we needed to evict something and we are proceeding, we must have
// evicted something.
assert(evicted_count > 0);
}
return Status::OK();
}
inline bool HyperClockTable::ChargeUsageMaybeEvictNonStrict(
size_t total_charge, size_t capacity, bool need_evict_for_occupancy) {
// For simplicity, we consider that either the cache can accept the insert
// with no evictions, or we must evict enough to make (at least) enough
// space. It could lead to unnecessary failures or excessive evictions in
// some extreme cases, but allows a fast, simple protocol. If we allow a
// race to get us over capacity, then we might never get back to capacity
// limit if the sizes of entries allow each insertion to evict the minimum
// charge. Thus, we should evict some extra if it's not a signifcant
// portion of the shard capacity. This can have the side benefit of
// involving fewer threads in eviction.
size_t old_usage = usage_.load(std::memory_order_relaxed);
size_t need_evict_charge;
// NOTE: if total_charge > old_usage, there isn't yet enough to evict
// `total_charge` amount. Even if we only try to evict `old_usage` amount,
// there's likely something referenced and we would eat CPU looking for
// enough to evict.
if (old_usage + total_charge <= capacity || total_charge > old_usage) {
// Good enough for me (might run over with a race)
need_evict_charge = 0;
} else {
// Try to evict enough space, and maybe some extra
need_evict_charge = total_charge;
if (old_usage > capacity) {
// Not too much to avoid thundering herd while avoiding strict
// synchronization, such as the compare_exchange used with strict
// capacity limit.
need_evict_charge += std::min(capacity / 1024, total_charge) + 1;
}
}
if (UNLIKELY(need_evict_for_occupancy) && need_evict_charge == 0) {
// Special case: require at least 1 eviction if we only have to
// deal with occupancy
need_evict_charge = 1;
}
size_t evicted_charge = 0;
size_t evicted_count = 0;
if (need_evict_charge > 0) {
Evict(need_evict_charge, &evicted_charge, &evicted_count);
// Deal with potential occupancy deficit
if (UNLIKELY(need_evict_for_occupancy) && evicted_count == 0) {
assert(evicted_charge == 0);
// Can't meet occupancy requirement
return false;
} else {
// Update occupancy for evictions
occupancy_.fetch_sub(evicted_count, std::memory_order_release);
}
}
// Track new usage even if we weren't able to evict enough
usage_.fetch_add(total_charge - evicted_charge, std::memory_order_relaxed);
// No underflow
assert(usage_.load(std::memory_order_relaxed) < SIZE_MAX / 2);
// Success
return true;
}
inline HyperClockTable::HandleImpl* HyperClockTable::StandaloneInsert(
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
const ClockHandleBasicData& proto) {
// Heap allocated separate from table
HandleImpl* h = new HandleImpl();
ClockHandleBasicData* h_alias = h;
*h_alias = proto;
h->SetStandalone();
// Single reference (standalone entries only created if returning a refed
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
// Handle back to user)
uint64_t meta = uint64_t{ClockHandle::kStateInvisible}
<< ClockHandle::kStateShift;
meta |= uint64_t{1} << ClockHandle::kAcquireCounterShift;
h->meta.store(meta, std::memory_order_release);
// Keep track of how much of usage is standalone
standalone_usage_.fetch_add(proto.GetTotalCharge(),
std::memory_order_relaxed);
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
return h;
}
Status HyperClockTable::Insert(const ClockHandleBasicData& proto,
HandleImpl** handle, Cache::Priority priority,
size_t capacity, bool strict_capacity_limit) {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
// Do we have the available occupancy? Optimistically assume we do
// and deal with it if we don't.
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
size_t old_occupancy = occupancy_.fetch_add(1, std::memory_order_acquire);
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
auto revert_occupancy_fn = [&]() {
occupancy_.fetch_sub(1, std::memory_order_relaxed);
};
// Whether we over-committed and need an eviction to make up for it
bool need_evict_for_occupancy = old_occupancy >= occupancy_limit_;
// Usage/capacity handling is somewhat different depending on
// strict_capacity_limit, but mostly pessimistic.
bool use_standalone_insert = false;
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
const size_t total_charge = proto.GetTotalCharge();
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
if (strict_capacity_limit) {
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
Status s = ChargeUsageMaybeEvictStrict(total_charge, capacity,
need_evict_for_occupancy);
if (!s.ok()) {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
revert_occupancy_fn();
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
return s;
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
}
} else {
// Case strict_capacity_limit == false
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
bool success = ChargeUsageMaybeEvictNonStrict(total_charge, capacity,
need_evict_for_occupancy);
if (!success) {
revert_occupancy_fn();
if (handle == nullptr) {
// Don't insert the entry but still return ok, as if the entry
// inserted into cache and evicted immediately.
Major Cache refactoring, CPU efficiency improvement (#10975) Summary: This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache). The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below. * static_cast lines of code +29 -35 (net removed 6) * reinterpret_cast lines of code +6 -32 (net removed 26) ## cache.h and secondary_cache.h * Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications: * Simpler for implementations to deal with just one Insert and one Lookup. * Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters * Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428. * Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks). * It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below). * I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc. * Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation. * Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.) * Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.) * Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774) * Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object. * Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change. ## typed_cache.h Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae). The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used. * PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value. * BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter. * FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue. * For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`. These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.) Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it. ## block_cache.h This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table. ## block_based_table_reader.cc Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation. The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions. ## block_based_table_builder.cc, cache_dump_load_impl.cc Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.) ## Everything else Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975 Test Plan: tests updated Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache): 34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844 34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594 34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297 34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523 34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602 34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293 34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926 34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488 233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984 233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922 233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559 233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93 233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418 233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273 233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691 233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82 1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55 1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02 1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45 1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24 1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92 1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78 1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36 1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83 Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn. Reviewed By: anand1976 Differential Revision: D42417818 Pulled By: pdillinger fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
proto.FreeData(allocator_);
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
return Status::OK();
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
} else {
// Need to track usage of fallback standalone insert
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
usage_.fetch_add(total_charge, std::memory_order_relaxed);
use_standalone_insert = true;
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
}
}
}
auto revert_usage_fn = [&]() {
usage_.fetch_sub(total_charge, std::memory_order_relaxed);
// No underflow
assert(usage_.load(std::memory_order_relaxed) < SIZE_MAX / 2);
};
if (!use_standalone_insert) {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
// Attempt a table insert, but abort if we find an existing entry for the
// key. If we were to overwrite old entries, we would either
// * Have to gain ownership over an existing entry to overwrite it, which
// would only work if there are no outstanding (read) references and would
// create a small gap in availability of the entry (old or new) to lookups.
// * Have to insert into a suboptimal location (more probes) so that the
// old entry can be kept around as well.
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
uint64_t initial_countdown = GetInitialCountdown(priority);
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
assert(initial_countdown > 0);
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
size_t probe = 0;
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
HandleImpl* e = FindSlot(
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
proto.hashed_key,
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
[&](HandleImpl* h) {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
// Optimistically transition the slot from "empty" to
// "under construction" (no effect on other states)
uint64_t old_meta =
h->meta.fetch_or(uint64_t{ClockHandle::kStateOccupiedBit}
<< ClockHandle::kStateShift,
std::memory_order_acq_rel);
uint64_t old_state = old_meta >> ClockHandle::kStateShift;
if (old_state == ClockHandle::kStateEmpty) {
// We've started inserting into an available slot, and taken
// ownership Save data fields
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
ClockHandleBasicData* h_alias = h;
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
*h_alias = proto;
// Transition from "under construction" state to "visible" state
uint64_t new_meta = uint64_t{ClockHandle::kStateVisible}
<< ClockHandle::kStateShift;
// Maybe with an outstanding reference
new_meta |= initial_countdown << ClockHandle::kAcquireCounterShift;
new_meta |= (initial_countdown - (handle != nullptr))
<< ClockHandle::kReleaseCounterShift;
#ifndef NDEBUG
// Save the state transition, with assertion
old_meta = h->meta.exchange(new_meta, std::memory_order_release);
assert(old_meta >> ClockHandle::kStateShift ==
ClockHandle::kStateConstruction);
#else
// Save the state transition
h->meta.store(new_meta, std::memory_order_release);
#endif
return true;
} else if (old_state != ClockHandle::kStateVisible) {
// Slot not usable / touchable now
return false;
}
// Existing, visible entry, which might be a match.
// But first, we need to acquire a ref to read it. In fact, number of
// refs for initial countdown, so that we boost the clock state if
// this is a match.
old_meta = h->meta.fetch_add(
ClockHandle::kAcquireIncrement * initial_countdown,
std::memory_order_acq_rel);
// Like Lookup
if ((old_meta >> ClockHandle::kStateShift) ==
ClockHandle::kStateVisible) {
// Acquired a read reference
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
if (h->hashed_key == proto.hashed_key) {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
// Match. Release in a way that boosts the clock state
old_meta = h->meta.fetch_add(
ClockHandle::kReleaseIncrement * initial_countdown,
std::memory_order_acq_rel);
// Correct for possible (but rare) overflow
CorrectNearOverflow(old_meta, h->meta);
// Insert standalone instead (only if return handle needed)
use_standalone_insert = true;
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
return true;
} else {
// Mismatch. Pretend we never took the reference
old_meta = h->meta.fetch_sub(
ClockHandle::kAcquireIncrement * initial_countdown,
std::memory_order_acq_rel);
}
} else if (UNLIKELY((old_meta >> ClockHandle::kStateShift) ==
ClockHandle::kStateInvisible)) {
// Pretend we never took the reference
// WART: there's a tiny chance we release last ref to invisible
// entry here. If that happens, we let eviction take care of it.
old_meta = h->meta.fetch_sub(
ClockHandle::kAcquireIncrement * initial_countdown,
std::memory_order_acq_rel);
} else {
// For other states, incrementing the acquire counter has no effect
// so we don't need to undo it.
// Slot not usable / touchable now.
}
(void)old_meta;
return false;
},
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
[&](HandleImpl* /*h*/) { return false; },
[&](HandleImpl* h) {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
h->displacements.fetch_add(1, std::memory_order_relaxed);
},
probe);
if (e == nullptr) {
// Occupancy check and never abort FindSlot above should generally
// prevent this, except it's theoretically possible for other threads
// to evict and replace entries in the right order to hit every slot
// when it is populated. Assuming random hashing, the chance of that
// should be no higher than pow(kStrictLoadFactor, n) for n slots.
// That should be infeasible for roughly n >= 256, so if this assertion
// fails, that suggests something is going wrong.
assert(GetTableSize() < 256);
use_standalone_insert = true;
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
}
if (!use_standalone_insert) {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
// Successfully inserted
if (handle) {
*handle = e;
}
return Status::OK();
}
// Roll back table insertion
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
Rollback(proto.hashed_key, e);
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
revert_occupancy_fn();
// Maybe fall back on standalone insert
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
if (handle == nullptr) {
revert_usage_fn();
// As if unrefed entry immdiately evicted
Major Cache refactoring, CPU efficiency improvement (#10975) Summary: This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache). The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below. * static_cast lines of code +29 -35 (net removed 6) * reinterpret_cast lines of code +6 -32 (net removed 26) ## cache.h and secondary_cache.h * Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications: * Simpler for implementations to deal with just one Insert and one Lookup. * Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters * Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428. * Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks). * It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below). * I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc. * Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation. * Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.) * Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.) * Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774) * Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object. * Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change. ## typed_cache.h Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae). The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used. * PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value. * BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter. * FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue. * For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`. These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.) Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it. ## block_cache.h This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table. ## block_based_table_reader.cc Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation. The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions. ## block_based_table_builder.cc, cache_dump_load_impl.cc Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.) ## Everything else Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975 Test Plan: tests updated Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache): 34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844 34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594 34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297 34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523 34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602 34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293 34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926 34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488 233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984 233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922 233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559 233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93 233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418 233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273 233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691 233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82 1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55 1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02 1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45 1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24 1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92 1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78 1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36 1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83 Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn. Reviewed By: anand1976 Differential Revision: D42417818 Pulled By: pdillinger fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
proto.FreeData(allocator_);
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
return Status::OK();
}
}
// Run standalone insert
assert(use_standalone_insert);
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
*handle = StandaloneInsert(proto);
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
// The OkOverwritten status is used to count "redundant" insertions into
// block cache. This implementation doesn't strictly check for redundant
// insertions, but we instead are probably interested in how many insertions
// didn't go into the table (instead "standalone"), which could be redundant
// Insert or some other reason (use_standalone_insert reasons above).
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
return Status::OkOverwritten();
}
HyperClockCache support for SecondaryCache, with refactoring (#11301) Summary: Internally refactors SecondaryCache integration out of LRUCache specifically and into a wrapper/adapter class that works with various Cache implementations. Notably, this relies on separating the notion of async lookup handles from other cache handles, so that HyperClockCache doesn't have to deal with the problem of allocating handles from the hash table for lookups that might fail anyway, and might be on the same key without support for coalescing. (LRUCache's hash table can incorporate previously allocated handles thanks to its pointer indirection.) Specifically, I'm worried about the case in which hundreds of threads try to access the same block and probing in the hash table degrades to linear search on the pile of entries with the same key. This change is a big step in the direction of supporting stacked SecondaryCaches, but there are obstacles to completing that. Especially, there is no SecondaryCache hook for evictions to pass from one to the next. It has been proposed that evictions be transmitted simply as the persisted data (as in SaveToCallback), but given the current structure provided by the CacheItemHelpers, that would require an extra copy of the block data, because there's intentionally no way to ask for a contiguous Slice of the data (to allow for flexibility in storage). `AsyncLookupHandle` and the re-worked `WaitAll()` should be essentially prepared for stacked SecondaryCaches, but several "TODO with stacked secondaries" issues remain in various places. It could be argued that the stacking instead be done as a SecondaryCache adapter that wraps two (or more) SecondaryCaches, but at least with the current API that would require an extra heap allocation on SecondaryCache Lookup for a wrapper SecondaryCacheResultHandle that can transfer a Lookup between secondaries. We could also consider trying to unify the Cache and SecondaryCache APIs, though that might be difficult if `AsyncLookupHandle` is kept a fixed struct. ## cache.h (public API) Moves `secondary_cache` option from LRUCacheOptions to ShardedCacheOptions so that it is applicable to HyperClockCache. ## advanced_cache.h (advanced public API) * Add `Cache::CreateStandalone()` so that the SecondaryCache support wrapper can use it. * Add `SetEvictionCallback()` / `eviction_callback_` so that the SecondaryCache support wrapper can use it. Only a single callback is supported for efficiency. If there is ever a need for more than one, hopefully that can be handled with a broadcast callback wrapper. These are essentially the two "extra" pieces of `Cache` for pulling out specific SecondaryCache support from the `Cache` implementation. I think it's a good trade-off as these are reasonable, limited, and reusable "cut points" into the `Cache` implementations. * Remove async capability from standard `Lookup()` (getting rid of awkward restrictions on pending Handles) and add `AsyncLookupHandle` and `StartAsyncLookup()`. As noted in the comments, the full struct of `AsyncLookupHandle` is exposed so that it can be stack allocated, for efficiency, though more data is being copied around than before, which could impact performance. (Lookup info -> AsyncLookupHandle -> Handle vs. Lookup info -> Handle) I could foresee a future in which a Cache internally saves a pointer to the AsyncLookupHandle, which means it's dangerous to allow it to be copyable or even movable. It also means it's not compatible with std::vector (which I don't like requiring as an API parameter anyway), so `WaitAll()` expects any contiguous array of AsyncLookupHandles. I believe this is best for common case efficiency, while behaving well in other cases also. For example, `WaitAll()` has no effect on default-constructed AsyncLookupHandles, which look like a completed cache miss. ## cacheable_entry.h A couple of functions are obsolete because Cache::Handle can no longer be pending. ## cache.cc Provides default implementations for new or revamped Cache functions, especially appropriate for non-blocking caches. ## secondary_cache_adapter.{h,cc} The full details of the Cache wrapper adding SecondaryCache support. Essentially replicates the SecondaryCache handling that was in LRUCache, but obviously refactored. There is a bit of logic duplication, where Lookup() is essentially a manually optimized version of StartAsyncLookup() and Wait(), but it's roughly a dozen lines of code. ## sharded_cache.h, typed_cache.h, charged_cache.{h,cc}, sim_cache.cc Simply updated for Cache API changes. ## lru_cache.{h,cc} Carefully remove SecondaryCache logic, implement `CreateStandalone` and eviction handler functionality. ## clock_cache.{h,cc} Expose existing `CreateStandalone` functionality, add eviction handler functionality. Light refactoring. ## block_based_table_reader* Mostly re-worked the only usage of async Lookup, which is in BlockBasedTable::MultiGet. Used arrays in place of autovector in some places for efficiency. Simplified some logic by not trying to process some cache results before they're all ready. Created new function `BlockBasedTable::GetCachePriority()` to reduce some pre-existing code duplication (and avoid making it worse). Fixed at least one small bug from the prior confusing mixture of async and sync Lookups. In MaybeReadBlockAndLoadToCache(), called by RetrieveBlock(), called by MultiGet() with wait=false, is_cache_hit for the block_cache_tracer entry would not be set to true if the handle was pending after Lookup and before Wait. ## Intended follow-up work * Figure out if there are any missing stats or block_cache_tracer work in refactored BlockBasedTable::MultiGet * Stacked secondary caches (see above discussion) * See if we can make up for the small MultiGet performance regression. * Study more performance with SecondaryCache * Items evicted from over-full LRUCache in Release were not being demoted to SecondaryCache, and still aren't to minimize unit test churn. Ideally they would be demoted, but it's an exceptional case so not a big deal. * Use CreateStandalone for cache reservations (save unnecessary hash table operations). Not a big deal, but worthy cleanup. * Somehow I got the contract for SecondaryCache::Insert wrong in #10945. (Doesn't take ownership!) That API comment needs to be fixed, but didn't want to mingle that in here. Pull Request resolved: https://github.com/facebook/rocksdb/pull/11301 Test Plan: ## Unit tests Generally updated to include HCC in SecondaryCache tests, though HyperClockCache has some different, less strict behaviors that leads to some tests not really being set up to work with it. Some of the tests remain disabled with it, but I think we have good coverage without them. ## Crash/stress test Updated to use the new combination. ## Performance First, let's check for regression on caches without secondary cache configured. Adding support for the eviction callback is likely to have a tiny effect, but it shouldn't be worrisome. LRUCache could benefit slightly from less logic around SecondaryCache handling. We can test with cache_bench default settings, built with DEBUG_LEVEL=0 and PORTABLE=0. ``` (while :; do base/cache_bench --cache_type=hyper_clock_cache | grep Rough; done) | awk '{ sum += $9; count++; print $0; print "Average: " int(sum / count) }' ``` **Before** this and #11299 (which could also have a small effect), running for about an hour, before & after running concurrently for each cache type: HyperClockCache: 3168662 (average parallel ops/sec) LRUCache: 2940127 **After** this and #11299, running for about an hour: HyperClockCache: 3164862 (average parallel ops/sec) (0.12% slower) LRUCache: 2940928 (0.03% faster) This is an acceptable difference IMHO. Next, let's consider essentially the worst case of new CPU overhead affecting overall performance. MultiGet uses the async lookup interface regardless of whether SecondaryCache or folly are used. We can configure a benchmark where all block cache queries are for data blocks, and all are hits. Create DB and test (before and after tests running simultaneously): ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 TEST_TMPDIR=/dev/shm base/db_bench -benchmarks=multireadrandom[-X30] -readonly -multiread_batched -batch_size=32 -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16 ``` **Before**: multireadrandom [AVG 30 runs] : 3444202 (± 57049) ops/sec; 240.9 (± 4.0) MB/sec multireadrandom [MEDIAN 30 runs] : 3514443 ops/sec; 245.8 MB/sec **After**: multireadrandom [AVG 30 runs] : 3291022 (± 58851) ops/sec; 230.2 (± 4.1) MB/sec multireadrandom [MEDIAN 30 runs] : 3366179 ops/sec; 235.4 MB/sec So that's roughly a 3% regression, on kind of a *worst case* test of MultiGet CPU. Similar story with HyperClockCache: **Before**: multireadrandom [AVG 30 runs] : 3933777 (± 41840) ops/sec; 275.1 (± 2.9) MB/sec multireadrandom [MEDIAN 30 runs] : 3970667 ops/sec; 277.7 MB/sec **After**: multireadrandom [AVG 30 runs] : 3755338 (± 30391) ops/sec; 262.6 (± 2.1) MB/sec multireadrandom [MEDIAN 30 runs] : 3785696 ops/sec; 264.8 MB/sec Roughly a 4-5% regression. Not ideal, but not the whole story, fortunately. Let's also look at Get() in db_bench: ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X30] -readonly -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16 ``` **Before**: readrandom [AVG 30 runs] : 2198685 (± 13412) ops/sec; 153.8 (± 0.9) MB/sec readrandom [MEDIAN 30 runs] : 2209498 ops/sec; 154.5 MB/sec **After**: readrandom [AVG 30 runs] : 2292814 (± 43508) ops/sec; 160.3 (± 3.0) MB/sec readrandom [MEDIAN 30 runs] : 2365181 ops/sec; 165.4 MB/sec That's showing roughly a 4% improvement, perhaps because of the secondary cache code that is no longer part of LRUCache. But weirdly, HyperClockCache is also showing 2-3% improvement: **Before**: readrandom [AVG 30 runs] : 2272333 (± 9992) ops/sec; 158.9 (± 0.7) MB/sec readrandom [MEDIAN 30 runs] : 2273239 ops/sec; 159.0 MB/sec **After**: readrandom [AVG 30 runs] : 2332407 (± 11252) ops/sec; 163.1 (± 0.8) MB/sec readrandom [MEDIAN 30 runs] : 2335329 ops/sec; 163.3 MB/sec Reviewed By: ltamasi Differential Revision: D44177044 Pulled By: pdillinger fbshipit-source-id: e808e48ff3fe2f792a79841ba617be98e48689f5
2 years ago
HyperClockTable::HandleImpl* HyperClockTable::CreateStandalone(
ClockHandleBasicData& proto, size_t capacity, bool strict_capacity_limit,
bool allow_uncharged) {
const size_t total_charge = proto.GetTotalCharge();
if (strict_capacity_limit) {
Status s = ChargeUsageMaybeEvictStrict(total_charge, capacity,
/*need_evict_for_occupancy=*/false);
if (!s.ok()) {
if (allow_uncharged) {
proto.total_charge = 0;
} else {
return nullptr;
}
}
} else {
// Case strict_capacity_limit == false
bool success =
ChargeUsageMaybeEvictNonStrict(total_charge, capacity,
/*need_evict_for_occupancy=*/false);
if (!success) {
// Force the issue
usage_.fetch_add(total_charge, std::memory_order_relaxed);
}
}
return StandaloneInsert(proto);
}
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
HyperClockTable::HandleImpl* HyperClockTable::Lookup(
const UniqueId64x2& hashed_key) {
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
size_t probe = 0;
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
HandleImpl* e = FindSlot(
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
hashed_key,
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
[&](HandleImpl* h) {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
// Mostly branch-free version (similar performance)
/*
uint64_t old_meta = h->meta.fetch_add(ClockHandle::kAcquireIncrement,
std::memory_order_acquire);
bool Shareable = (old_meta >> (ClockHandle::kStateShift + 1)) & 1U;
bool visible = (old_meta >> ClockHandle::kStateShift) & 1U;
bool match = (h->key == key) & visible;
h->meta.fetch_sub(static_cast<uint64_t>(Shareable & !match) <<
ClockHandle::kAcquireCounterShift, std::memory_order_release); return
match;
*/
// Optimistic lookup should pay off when the table is relatively
// sparse.
constexpr bool kOptimisticLookup = true;
uint64_t old_meta;
if (!kOptimisticLookup) {
old_meta = h->meta.load(std::memory_order_acquire);
if ((old_meta >> ClockHandle::kStateShift) !=
ClockHandle::kStateVisible) {
return false;
}
}
// (Optimistically) increment acquire counter
old_meta = h->meta.fetch_add(ClockHandle::kAcquireIncrement,
std::memory_order_acquire);
// Check if it's an entry visible to lookups
if ((old_meta >> ClockHandle::kStateShift) ==
ClockHandle::kStateVisible) {
// Acquired a read reference
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
if (h->hashed_key == hashed_key) {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
// Match
return true;
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
} else {
// Mismatch. Pretend we never took the reference
old_meta = h->meta.fetch_sub(ClockHandle::kAcquireIncrement,
std::memory_order_release);
}
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
} else if (UNLIKELY((old_meta >> ClockHandle::kStateShift) ==
ClockHandle::kStateInvisible)) {
// Pretend we never took the reference
// WART: there's a tiny chance we release last ref to invisible
// entry here. If that happens, we let eviction take care of it.
old_meta = h->meta.fetch_sub(ClockHandle::kAcquireIncrement,
std::memory_order_release);
} else {
// For other states, incrementing the acquire counter has no effect
// so we don't need to undo it. Furthermore, we cannot safely undo
// it because we did not acquire a read reference to lock the
// entry in a Shareable state.
}
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
(void)old_meta;
return false;
},
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
[&](HandleImpl* h) {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
return h->displacements.load(std::memory_order_relaxed) == 0;
},
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
[&](HandleImpl* /*h*/) {}, probe);
return e;
}
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
bool HyperClockTable::Release(HandleImpl* h, bool useful,
bool erase_if_last_ref) {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
// In contrast with LRUCache's Release, this function won't delete the handle
// when the cache is above capacity and the reference is the last one. Space
// is only freed up by EvictFromClock (called by Insert when space is needed)
// and Erase. We do this to avoid an extra atomic read of the variable usage_.
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
uint64_t old_meta;
if (useful) {
// Increment release counter to indicate was used
old_meta = h->meta.fetch_add(ClockHandle::kReleaseIncrement,
std::memory_order_release);
} else {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
// Decrement acquire counter to pretend it never happened
old_meta = h->meta.fetch_sub(ClockHandle::kAcquireIncrement,
std::memory_order_release);
}
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
assert((old_meta >> ClockHandle::kStateShift) &
ClockHandle::kStateShareableBit);
// No underflow
assert(((old_meta >> ClockHandle::kAcquireCounterShift) &
ClockHandle::kCounterMask) !=
((old_meta >> ClockHandle::kReleaseCounterShift) &
ClockHandle::kCounterMask));
if (erase_if_last_ref || UNLIKELY(old_meta >> ClockHandle::kStateShift ==
ClockHandle::kStateInvisible)) {
// Update for last fetch_add op
if (useful) {
old_meta += ClockHandle::kReleaseIncrement;
} else {
old_meta -= ClockHandle::kAcquireIncrement;
}
// Take ownership if no refs
do {
if (GetRefcount(old_meta) != 0) {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
// Not last ref at some point in time during this Release call
// Correct for possible (but rare) overflow
CorrectNearOverflow(old_meta, h->meta);
return false;
}
if ((old_meta & (uint64_t{ClockHandle::kStateShareableBit}
<< ClockHandle::kStateShift)) == 0) {
// Someone else took ownership
return false;
}
// Note that there's a small chance that we release, another thread
// replaces this entry with another, reaches zero refs, and then we end
// up erasing that other entry. That's an acceptable risk / imprecision.
} while (!h->meta.compare_exchange_weak(
old_meta,
uint64_t{ClockHandle::kStateConstruction} << ClockHandle::kStateShift,
std::memory_order_acquire));
// Took ownership
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
size_t total_charge = h->GetTotalCharge();
if (UNLIKELY(h->IsStandalone())) {
Major Cache refactoring, CPU efficiency improvement (#10975) Summary: This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache). The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below. * static_cast lines of code +29 -35 (net removed 6) * reinterpret_cast lines of code +6 -32 (net removed 26) ## cache.h and secondary_cache.h * Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications: * Simpler for implementations to deal with just one Insert and one Lookup. * Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters * Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428. * Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks). * It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below). * I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc. * Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation. * Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.) * Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.) * Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774) * Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object. * Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change. ## typed_cache.h Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae). The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used. * PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value. * BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter. * FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue. * For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`. These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.) Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it. ## block_cache.h This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table. ## block_based_table_reader.cc Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation. The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions. ## block_based_table_builder.cc, cache_dump_load_impl.cc Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.) ## Everything else Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975 Test Plan: tests updated Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache): 34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844 34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594 34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297 34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523 34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602 34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293 34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926 34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488 233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984 233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922 233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559 233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93 233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418 233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273 233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691 233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82 1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55 1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02 1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45 1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24 1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92 1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78 1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36 1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83 Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn. Reviewed By: anand1976 Differential Revision: D42417818 Pulled By: pdillinger fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
h->FreeData(allocator_);
// Delete standalone handle
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
delete h;
standalone_usage_.fetch_sub(total_charge, std::memory_order_relaxed);
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
usage_.fetch_sub(total_charge, std::memory_order_relaxed);
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
} else {
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
Rollback(h->hashed_key, h);
Major Cache refactoring, CPU efficiency improvement (#10975) Summary: This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache). The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below. * static_cast lines of code +29 -35 (net removed 6) * reinterpret_cast lines of code +6 -32 (net removed 26) ## cache.h and secondary_cache.h * Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications: * Simpler for implementations to deal with just one Insert and one Lookup. * Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters * Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428. * Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks). * It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below). * I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc. * Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation. * Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.) * Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.) * Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774) * Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object. * Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change. ## typed_cache.h Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae). The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used. * PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value. * BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter. * FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue. * For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`. These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.) Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it. ## block_cache.h This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table. ## block_based_table_reader.cc Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation. The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions. ## block_based_table_builder.cc, cache_dump_load_impl.cc Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.) ## Everything else Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975 Test Plan: tests updated Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache): 34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844 34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594 34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297 34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523 34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602 34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293 34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926 34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488 233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984 233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922 233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559 233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93 233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418 233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273 233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691 233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82 1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55 1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02 1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45 1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24 1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92 1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78 1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36 1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83 Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn. Reviewed By: anand1976 Differential Revision: D42417818 Pulled By: pdillinger fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
FreeDataMarkEmpty(*h, allocator_);
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
ReclaimEntryUsage(total_charge);
}
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
return true;
} else {
// Correct for possible (but rare) overflow
CorrectNearOverflow(old_meta, h->meta);
return false;
}
}
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
void HyperClockTable::Ref(HandleImpl& h) {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
// Increment acquire counter
uint64_t old_meta = h.meta.fetch_add(ClockHandle::kAcquireIncrement,
std::memory_order_acquire);
assert((old_meta >> ClockHandle::kStateShift) &
ClockHandle::kStateShareableBit);
// Must have already had a reference
assert(GetRefcount(old_meta) > 0);
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
(void)old_meta;
}
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
void HyperClockTable::TEST_RefN(HandleImpl& h, size_t n) {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
// Increment acquire counter
uint64_t old_meta = h.meta.fetch_add(n * ClockHandle::kAcquireIncrement,
std::memory_order_acquire);
assert((old_meta >> ClockHandle::kStateShift) &
ClockHandle::kStateShareableBit);
(void)old_meta;
}
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
void HyperClockTable::TEST_ReleaseN(HandleImpl* h, size_t n) {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
if (n > 0) {
// Split into n - 1 and 1 steps.
uint64_t old_meta = h->meta.fetch_add(
(n - 1) * ClockHandle::kReleaseIncrement, std::memory_order_acquire);
assert((old_meta >> ClockHandle::kStateShift) &
ClockHandle::kStateShareableBit);
(void)old_meta;
Release(h, /*useful*/ true, /*erase_if_last_ref*/ false);
}
}
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
void HyperClockTable::Erase(const UniqueId64x2& hashed_key) {
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
size_t probe = 0;
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
(void)FindSlot(
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
hashed_key,
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
[&](HandleImpl* h) {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
// Could be multiple entries in rare cases. Erase them all.
// Optimistically increment acquire counter
uint64_t old_meta = h->meta.fetch_add(ClockHandle::kAcquireIncrement,
std::memory_order_acquire);
// Check if it's an entry visible to lookups
if ((old_meta >> ClockHandle::kStateShift) ==
ClockHandle::kStateVisible) {
// Acquired a read reference
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
if (h->hashed_key == hashed_key) {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
// Match. Set invisible.
old_meta =
h->meta.fetch_and(~(uint64_t{ClockHandle::kStateVisibleBit}
<< ClockHandle::kStateShift),
std::memory_order_acq_rel);
// Apply update to local copy
old_meta &= ~(uint64_t{ClockHandle::kStateVisibleBit}
<< ClockHandle::kStateShift);
for (;;) {
uint64_t refcount = GetRefcount(old_meta);
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
assert(refcount > 0);
if (refcount > 1) {
// Not last ref at some point in time during this Erase call
// Pretend we never took the reference
h->meta.fetch_sub(ClockHandle::kAcquireIncrement,
std::memory_order_release);
break;
} else if (h->meta.compare_exchange_weak(
old_meta,
uint64_t{ClockHandle::kStateConstruction}
<< ClockHandle::kStateShift,
std::memory_order_acq_rel)) {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
// Took ownership
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
assert(hashed_key == h->hashed_key);
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
size_t total_charge = h->GetTotalCharge();
Major Cache refactoring, CPU efficiency improvement (#10975) Summary: This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache). The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below. * static_cast lines of code +29 -35 (net removed 6) * reinterpret_cast lines of code +6 -32 (net removed 26) ## cache.h and secondary_cache.h * Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications: * Simpler for implementations to deal with just one Insert and one Lookup. * Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters * Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428. * Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks). * It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below). * I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc. * Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation. * Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.) * Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.) * Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774) * Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object. * Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change. ## typed_cache.h Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae). The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used. * PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value. * BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter. * FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue. * For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`. These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.) Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it. ## block_cache.h This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table. ## block_based_table_reader.cc Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation. The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions. ## block_based_table_builder.cc, cache_dump_load_impl.cc Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.) ## Everything else Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975 Test Plan: tests updated Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache): 34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844 34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594 34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297 34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523 34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602 34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293 34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926 34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488 233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984 233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922 233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559 233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93 233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418 233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273 233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691 233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82 1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55 1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02 1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45 1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24 1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92 1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78 1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36 1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83 Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn. Reviewed By: anand1976 Differential Revision: D42417818 Pulled By: pdillinger fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
FreeDataMarkEmpty(*h, allocator_);
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
ReclaimEntryUsage(total_charge);
// We already have a copy of hashed_key in this case, so OK to
// delay Rollback until after releasing the entry
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
Rollback(hashed_key, h);
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
break;
}
}
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
} else {
// Mismatch. Pretend we never took the reference
h->meta.fetch_sub(ClockHandle::kAcquireIncrement,
std::memory_order_release);
}
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
} else if (UNLIKELY((old_meta >> ClockHandle::kStateShift) ==
ClockHandle::kStateInvisible)) {
// Pretend we never took the reference
// WART: there's a tiny chance we release last ref to invisible
// entry here. If that happens, we let eviction take care of it.
h->meta.fetch_sub(ClockHandle::kAcquireIncrement,
std::memory_order_release);
} else {
// For other states, incrementing the acquire counter has no effect
// so we don't need to undo it.
}
return false;
},
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
[&](HandleImpl* h) {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
return h->displacements.load(std::memory_order_relaxed) == 0;
},
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
[&](HandleImpl* /*h*/) {}, probe);
}
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
void HyperClockTable::ConstApplyToEntriesRange(
std::function<void(const HandleImpl&)> func, size_t index_begin,
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
size_t index_end, bool apply_if_will_be_deleted) const {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
uint64_t check_state_mask = ClockHandle::kStateShareableBit;
if (!apply_if_will_be_deleted) {
check_state_mask |= ClockHandle::kStateVisibleBit;
}
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
for (size_t i = index_begin; i < index_end; i++) {
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
HandleImpl& h = array_[i];
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
// Note: to avoid using compare_exchange, we have to be extra careful.
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
uint64_t old_meta = h.meta.load(std::memory_order_relaxed);
// Check if it's an entry visible to lookups
if ((old_meta >> ClockHandle::kStateShift) & check_state_mask) {
// Increment acquire counter. Note: it's possible that the entry has
// completely changed since we loaded old_meta, but incrementing acquire
// count is always safe. (Similar to optimistic Lookup here.)
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
old_meta = h.meta.fetch_add(ClockHandle::kAcquireIncrement,
std::memory_order_acquire);
// Check whether we actually acquired a reference.
if ((old_meta >> ClockHandle::kStateShift) &
ClockHandle::kStateShareableBit) {
// Apply func if appropriate
if ((old_meta >> ClockHandle::kStateShift) & check_state_mask) {
func(h);
}
// Pretend we never took the reference
h.meta.fetch_sub(ClockHandle::kAcquireIncrement,
std::memory_order_release);
// No net change, so don't need to check for overflow
} else {
// For other states, incrementing the acquire counter has no effect
// so we don't need to undo it. Furthermore, we cannot safely undo
// it because we did not acquire a read reference to lock the
// entry in a Shareable state.
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
}
}
}
}
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
void HyperClockTable::EraseUnRefEntries() {
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
for (size_t i = 0; i <= this->length_bits_mask_; i++) {
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
HandleImpl& h = array_[i];
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
uint64_t old_meta = h.meta.load(std::memory_order_relaxed);
if (old_meta & (uint64_t{ClockHandle::kStateShareableBit}
<< ClockHandle::kStateShift) &&
GetRefcount(old_meta) == 0 &&
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
h.meta.compare_exchange_strong(old_meta,
uint64_t{ClockHandle::kStateConstruction}
<< ClockHandle::kStateShift,
std::memory_order_acquire)) {
// Took ownership
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
size_t total_charge = h.GetTotalCharge();
Rollback(h.hashed_key, &h);
Major Cache refactoring, CPU efficiency improvement (#10975) Summary: This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache). The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below. * static_cast lines of code +29 -35 (net removed 6) * reinterpret_cast lines of code +6 -32 (net removed 26) ## cache.h and secondary_cache.h * Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications: * Simpler for implementations to deal with just one Insert and one Lookup. * Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters * Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428. * Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks). * It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below). * I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc. * Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation. * Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.) * Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.) * Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774) * Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object. * Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change. ## typed_cache.h Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae). The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used. * PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value. * BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter. * FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue. * For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`. These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.) Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it. ## block_cache.h This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table. ## block_based_table_reader.cc Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation. The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions. ## block_based_table_builder.cc, cache_dump_load_impl.cc Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.) ## Everything else Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975 Test Plan: tests updated Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache): 34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844 34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594 34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297 34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523 34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602 34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293 34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926 34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488 233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984 233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922 233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559 233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93 233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418 233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273 233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691 233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82 1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55 1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02 1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45 1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24 1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92 1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78 1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36 1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83 Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn. Reviewed By: anand1976 Differential Revision: D42417818 Pulled By: pdillinger fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
FreeDataMarkEmpty(h, allocator_);
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
ReclaimEntryUsage(total_charge);
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
}
}
}
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
inline HyperClockTable::HandleImpl* HyperClockTable::FindSlot(
const UniqueId64x2& hashed_key, std::function<bool(HandleImpl*)> match_fn,
std::function<bool(HandleImpl*)> abort_fn,
std::function<void(HandleImpl*)> update_fn, size_t& probe) {
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
// NOTE: upper 32 bits of hashed_key[0] is used for sharding
//
// We use double-hashing probing. Every probe in the sequence is a
// pseudorandom integer, computed as a linear function of two random hashes,
// which we call base and increment. Specifically, the i-th probe is base + i
// * increment modulo the table size.
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
size_t base = static_cast<size_t>(hashed_key[1]);
// We use an odd increment, which is relatively prime with the power-of-two
// table size. This implies that we cycle back to the first probe only
// after probing every slot exactly once.
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
// TODO: we could also reconsider linear probing, though locality benefits
// are limited because each slot is a full cache line
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
size_t increment = static_cast<size_t>(hashed_key[0]) | 1U;
size_t current = ModTableSize(base + probe * increment);
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
while (probe <= length_bits_mask_) {
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
HandleImpl* h = &array_[current];
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
if (match_fn(h)) {
probe++;
return h;
}
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
if (abort_fn(h)) {
return nullptr;
}
probe++;
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
update_fn(h);
current = ModTableSize(current + increment);
}
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
// We looped back.
return nullptr;
}
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
inline void HyperClockTable::Rollback(const UniqueId64x2& hashed_key,
const HandleImpl* h) {
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
size_t current = ModTableSize(hashed_key[1]);
size_t increment = static_cast<size_t>(hashed_key[0]) | 1U;
while (&array_[current] != h) {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
array_[current].displacements.fetch_sub(1, std::memory_order_relaxed);
current = ModTableSize(current + increment);
}
}
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
inline void HyperClockTable::ReclaimEntryUsage(size_t total_charge) {
auto old_occupancy = occupancy_.fetch_sub(1U, std::memory_order_release);
(void)old_occupancy;
// No underflow
assert(old_occupancy > 0);
auto old_usage = usage_.fetch_sub(total_charge, std::memory_order_relaxed);
(void)old_usage;
// No underflow
assert(old_usage >= total_charge);
}
inline void HyperClockTable::Evict(size_t requested_charge,
size_t* freed_charge, size_t* freed_count) {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
// precondition
assert(requested_charge > 0);
// TODO: make a tuning parameter?
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
constexpr size_t step_size = 4;
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
// First (concurrent) increment clock pointer
uint64_t old_clock_pointer =
clock_pointer_.fetch_add(step_size, std::memory_order_relaxed);
// Cap the eviction effort at this thread (along with those operating in
// parallel) circling through the whole structure kMaxCountdown times.
// In other words, this eviction run must find something/anything that is
// unreferenced at start of and during the eviction run that isn't reclaimed
// by a concurrent eviction run.
uint64_t max_clock_pointer =
old_clock_pointer + (ClockHandle::kMaxCountdown << length_bits_);
HyperClockCache support for SecondaryCache, with refactoring (#11301) Summary: Internally refactors SecondaryCache integration out of LRUCache specifically and into a wrapper/adapter class that works with various Cache implementations. Notably, this relies on separating the notion of async lookup handles from other cache handles, so that HyperClockCache doesn't have to deal with the problem of allocating handles from the hash table for lookups that might fail anyway, and might be on the same key without support for coalescing. (LRUCache's hash table can incorporate previously allocated handles thanks to its pointer indirection.) Specifically, I'm worried about the case in which hundreds of threads try to access the same block and probing in the hash table degrades to linear search on the pile of entries with the same key. This change is a big step in the direction of supporting stacked SecondaryCaches, but there are obstacles to completing that. Especially, there is no SecondaryCache hook for evictions to pass from one to the next. It has been proposed that evictions be transmitted simply as the persisted data (as in SaveToCallback), but given the current structure provided by the CacheItemHelpers, that would require an extra copy of the block data, because there's intentionally no way to ask for a contiguous Slice of the data (to allow for flexibility in storage). `AsyncLookupHandle` and the re-worked `WaitAll()` should be essentially prepared for stacked SecondaryCaches, but several "TODO with stacked secondaries" issues remain in various places. It could be argued that the stacking instead be done as a SecondaryCache adapter that wraps two (or more) SecondaryCaches, but at least with the current API that would require an extra heap allocation on SecondaryCache Lookup for a wrapper SecondaryCacheResultHandle that can transfer a Lookup between secondaries. We could also consider trying to unify the Cache and SecondaryCache APIs, though that might be difficult if `AsyncLookupHandle` is kept a fixed struct. ## cache.h (public API) Moves `secondary_cache` option from LRUCacheOptions to ShardedCacheOptions so that it is applicable to HyperClockCache. ## advanced_cache.h (advanced public API) * Add `Cache::CreateStandalone()` so that the SecondaryCache support wrapper can use it. * Add `SetEvictionCallback()` / `eviction_callback_` so that the SecondaryCache support wrapper can use it. Only a single callback is supported for efficiency. If there is ever a need for more than one, hopefully that can be handled with a broadcast callback wrapper. These are essentially the two "extra" pieces of `Cache` for pulling out specific SecondaryCache support from the `Cache` implementation. I think it's a good trade-off as these are reasonable, limited, and reusable "cut points" into the `Cache` implementations. * Remove async capability from standard `Lookup()` (getting rid of awkward restrictions on pending Handles) and add `AsyncLookupHandle` and `StartAsyncLookup()`. As noted in the comments, the full struct of `AsyncLookupHandle` is exposed so that it can be stack allocated, for efficiency, though more data is being copied around than before, which could impact performance. (Lookup info -> AsyncLookupHandle -> Handle vs. Lookup info -> Handle) I could foresee a future in which a Cache internally saves a pointer to the AsyncLookupHandle, which means it's dangerous to allow it to be copyable or even movable. It also means it's not compatible with std::vector (which I don't like requiring as an API parameter anyway), so `WaitAll()` expects any contiguous array of AsyncLookupHandles. I believe this is best for common case efficiency, while behaving well in other cases also. For example, `WaitAll()` has no effect on default-constructed AsyncLookupHandles, which look like a completed cache miss. ## cacheable_entry.h A couple of functions are obsolete because Cache::Handle can no longer be pending. ## cache.cc Provides default implementations for new or revamped Cache functions, especially appropriate for non-blocking caches. ## secondary_cache_adapter.{h,cc} The full details of the Cache wrapper adding SecondaryCache support. Essentially replicates the SecondaryCache handling that was in LRUCache, but obviously refactored. There is a bit of logic duplication, where Lookup() is essentially a manually optimized version of StartAsyncLookup() and Wait(), but it's roughly a dozen lines of code. ## sharded_cache.h, typed_cache.h, charged_cache.{h,cc}, sim_cache.cc Simply updated for Cache API changes. ## lru_cache.{h,cc} Carefully remove SecondaryCache logic, implement `CreateStandalone` and eviction handler functionality. ## clock_cache.{h,cc} Expose existing `CreateStandalone` functionality, add eviction handler functionality. Light refactoring. ## block_based_table_reader* Mostly re-worked the only usage of async Lookup, which is in BlockBasedTable::MultiGet. Used arrays in place of autovector in some places for efficiency. Simplified some logic by not trying to process some cache results before they're all ready. Created new function `BlockBasedTable::GetCachePriority()` to reduce some pre-existing code duplication (and avoid making it worse). Fixed at least one small bug from the prior confusing mixture of async and sync Lookups. In MaybeReadBlockAndLoadToCache(), called by RetrieveBlock(), called by MultiGet() with wait=false, is_cache_hit for the block_cache_tracer entry would not be set to true if the handle was pending after Lookup and before Wait. ## Intended follow-up work * Figure out if there are any missing stats or block_cache_tracer work in refactored BlockBasedTable::MultiGet * Stacked secondary caches (see above discussion) * See if we can make up for the small MultiGet performance regression. * Study more performance with SecondaryCache * Items evicted from over-full LRUCache in Release were not being demoted to SecondaryCache, and still aren't to minimize unit test churn. Ideally they would be demoted, but it's an exceptional case so not a big deal. * Use CreateStandalone for cache reservations (save unnecessary hash table operations). Not a big deal, but worthy cleanup. * Somehow I got the contract for SecondaryCache::Insert wrong in #10945. (Doesn't take ownership!) That API comment needs to be fixed, but didn't want to mingle that in here. Pull Request resolved: https://github.com/facebook/rocksdb/pull/11301 Test Plan: ## Unit tests Generally updated to include HCC in SecondaryCache tests, though HyperClockCache has some different, less strict behaviors that leads to some tests not really being set up to work with it. Some of the tests remain disabled with it, but I think we have good coverage without them. ## Crash/stress test Updated to use the new combination. ## Performance First, let's check for regression on caches without secondary cache configured. Adding support for the eviction callback is likely to have a tiny effect, but it shouldn't be worrisome. LRUCache could benefit slightly from less logic around SecondaryCache handling. We can test with cache_bench default settings, built with DEBUG_LEVEL=0 and PORTABLE=0. ``` (while :; do base/cache_bench --cache_type=hyper_clock_cache | grep Rough; done) | awk '{ sum += $9; count++; print $0; print "Average: " int(sum / count) }' ``` **Before** this and #11299 (which could also have a small effect), running for about an hour, before & after running concurrently for each cache type: HyperClockCache: 3168662 (average parallel ops/sec) LRUCache: 2940127 **After** this and #11299, running for about an hour: HyperClockCache: 3164862 (average parallel ops/sec) (0.12% slower) LRUCache: 2940928 (0.03% faster) This is an acceptable difference IMHO. Next, let's consider essentially the worst case of new CPU overhead affecting overall performance. MultiGet uses the async lookup interface regardless of whether SecondaryCache or folly are used. We can configure a benchmark where all block cache queries are for data blocks, and all are hits. Create DB and test (before and after tests running simultaneously): ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 TEST_TMPDIR=/dev/shm base/db_bench -benchmarks=multireadrandom[-X30] -readonly -multiread_batched -batch_size=32 -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16 ``` **Before**: multireadrandom [AVG 30 runs] : 3444202 (± 57049) ops/sec; 240.9 (± 4.0) MB/sec multireadrandom [MEDIAN 30 runs] : 3514443 ops/sec; 245.8 MB/sec **After**: multireadrandom [AVG 30 runs] : 3291022 (± 58851) ops/sec; 230.2 (± 4.1) MB/sec multireadrandom [MEDIAN 30 runs] : 3366179 ops/sec; 235.4 MB/sec So that's roughly a 3% regression, on kind of a *worst case* test of MultiGet CPU. Similar story with HyperClockCache: **Before**: multireadrandom [AVG 30 runs] : 3933777 (± 41840) ops/sec; 275.1 (± 2.9) MB/sec multireadrandom [MEDIAN 30 runs] : 3970667 ops/sec; 277.7 MB/sec **After**: multireadrandom [AVG 30 runs] : 3755338 (± 30391) ops/sec; 262.6 (± 2.1) MB/sec multireadrandom [MEDIAN 30 runs] : 3785696 ops/sec; 264.8 MB/sec Roughly a 4-5% regression. Not ideal, but not the whole story, fortunately. Let's also look at Get() in db_bench: ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X30] -readonly -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16 ``` **Before**: readrandom [AVG 30 runs] : 2198685 (± 13412) ops/sec; 153.8 (± 0.9) MB/sec readrandom [MEDIAN 30 runs] : 2209498 ops/sec; 154.5 MB/sec **After**: readrandom [AVG 30 runs] : 2292814 (± 43508) ops/sec; 160.3 (± 3.0) MB/sec readrandom [MEDIAN 30 runs] : 2365181 ops/sec; 165.4 MB/sec That's showing roughly a 4% improvement, perhaps because of the secondary cache code that is no longer part of LRUCache. But weirdly, HyperClockCache is also showing 2-3% improvement: **Before**: readrandom [AVG 30 runs] : 2272333 (± 9992) ops/sec; 158.9 (± 0.7) MB/sec readrandom [MEDIAN 30 runs] : 2273239 ops/sec; 159.0 MB/sec **After**: readrandom [AVG 30 runs] : 2332407 (± 11252) ops/sec; 163.1 (± 0.8) MB/sec readrandom [MEDIAN 30 runs] : 2335329 ops/sec; 163.3 MB/sec Reviewed By: ltamasi Differential Revision: D44177044 Pulled By: pdillinger fbshipit-source-id: e808e48ff3fe2f792a79841ba617be98e48689f5
2 years ago
// For key reconstructed from hash
UniqueId64x2 unhashed;
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
for (;;) {
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
for (size_t i = 0; i < step_size; i++) {
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
HandleImpl& h = array_[ModTableSize(Lower32of64(old_clock_pointer + i))];
bool evicting = ClockUpdate(h);
if (evicting) {
Rollback(h.hashed_key, &h);
*freed_charge += h.GetTotalCharge();
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
*freed_count += 1;
HyperClockCache support for SecondaryCache, with refactoring (#11301) Summary: Internally refactors SecondaryCache integration out of LRUCache specifically and into a wrapper/adapter class that works with various Cache implementations. Notably, this relies on separating the notion of async lookup handles from other cache handles, so that HyperClockCache doesn't have to deal with the problem of allocating handles from the hash table for lookups that might fail anyway, and might be on the same key without support for coalescing. (LRUCache's hash table can incorporate previously allocated handles thanks to its pointer indirection.) Specifically, I'm worried about the case in which hundreds of threads try to access the same block and probing in the hash table degrades to linear search on the pile of entries with the same key. This change is a big step in the direction of supporting stacked SecondaryCaches, but there are obstacles to completing that. Especially, there is no SecondaryCache hook for evictions to pass from one to the next. It has been proposed that evictions be transmitted simply as the persisted data (as in SaveToCallback), but given the current structure provided by the CacheItemHelpers, that would require an extra copy of the block data, because there's intentionally no way to ask for a contiguous Slice of the data (to allow for flexibility in storage). `AsyncLookupHandle` and the re-worked `WaitAll()` should be essentially prepared for stacked SecondaryCaches, but several "TODO with stacked secondaries" issues remain in various places. It could be argued that the stacking instead be done as a SecondaryCache adapter that wraps two (or more) SecondaryCaches, but at least with the current API that would require an extra heap allocation on SecondaryCache Lookup for a wrapper SecondaryCacheResultHandle that can transfer a Lookup between secondaries. We could also consider trying to unify the Cache and SecondaryCache APIs, though that might be difficult if `AsyncLookupHandle` is kept a fixed struct. ## cache.h (public API) Moves `secondary_cache` option from LRUCacheOptions to ShardedCacheOptions so that it is applicable to HyperClockCache. ## advanced_cache.h (advanced public API) * Add `Cache::CreateStandalone()` so that the SecondaryCache support wrapper can use it. * Add `SetEvictionCallback()` / `eviction_callback_` so that the SecondaryCache support wrapper can use it. Only a single callback is supported for efficiency. If there is ever a need for more than one, hopefully that can be handled with a broadcast callback wrapper. These are essentially the two "extra" pieces of `Cache` for pulling out specific SecondaryCache support from the `Cache` implementation. I think it's a good trade-off as these are reasonable, limited, and reusable "cut points" into the `Cache` implementations. * Remove async capability from standard `Lookup()` (getting rid of awkward restrictions on pending Handles) and add `AsyncLookupHandle` and `StartAsyncLookup()`. As noted in the comments, the full struct of `AsyncLookupHandle` is exposed so that it can be stack allocated, for efficiency, though more data is being copied around than before, which could impact performance. (Lookup info -> AsyncLookupHandle -> Handle vs. Lookup info -> Handle) I could foresee a future in which a Cache internally saves a pointer to the AsyncLookupHandle, which means it's dangerous to allow it to be copyable or even movable. It also means it's not compatible with std::vector (which I don't like requiring as an API parameter anyway), so `WaitAll()` expects any contiguous array of AsyncLookupHandles. I believe this is best for common case efficiency, while behaving well in other cases also. For example, `WaitAll()` has no effect on default-constructed AsyncLookupHandles, which look like a completed cache miss. ## cacheable_entry.h A couple of functions are obsolete because Cache::Handle can no longer be pending. ## cache.cc Provides default implementations for new or revamped Cache functions, especially appropriate for non-blocking caches. ## secondary_cache_adapter.{h,cc} The full details of the Cache wrapper adding SecondaryCache support. Essentially replicates the SecondaryCache handling that was in LRUCache, but obviously refactored. There is a bit of logic duplication, where Lookup() is essentially a manually optimized version of StartAsyncLookup() and Wait(), but it's roughly a dozen lines of code. ## sharded_cache.h, typed_cache.h, charged_cache.{h,cc}, sim_cache.cc Simply updated for Cache API changes. ## lru_cache.{h,cc} Carefully remove SecondaryCache logic, implement `CreateStandalone` and eviction handler functionality. ## clock_cache.{h,cc} Expose existing `CreateStandalone` functionality, add eviction handler functionality. Light refactoring. ## block_based_table_reader* Mostly re-worked the only usage of async Lookup, which is in BlockBasedTable::MultiGet. Used arrays in place of autovector in some places for efficiency. Simplified some logic by not trying to process some cache results before they're all ready. Created new function `BlockBasedTable::GetCachePriority()` to reduce some pre-existing code duplication (and avoid making it worse). Fixed at least one small bug from the prior confusing mixture of async and sync Lookups. In MaybeReadBlockAndLoadToCache(), called by RetrieveBlock(), called by MultiGet() with wait=false, is_cache_hit for the block_cache_tracer entry would not be set to true if the handle was pending after Lookup and before Wait. ## Intended follow-up work * Figure out if there are any missing stats or block_cache_tracer work in refactored BlockBasedTable::MultiGet * Stacked secondary caches (see above discussion) * See if we can make up for the small MultiGet performance regression. * Study more performance with SecondaryCache * Items evicted from over-full LRUCache in Release were not being demoted to SecondaryCache, and still aren't to minimize unit test churn. Ideally they would be demoted, but it's an exceptional case so not a big deal. * Use CreateStandalone for cache reservations (save unnecessary hash table operations). Not a big deal, but worthy cleanup. * Somehow I got the contract for SecondaryCache::Insert wrong in #10945. (Doesn't take ownership!) That API comment needs to be fixed, but didn't want to mingle that in here. Pull Request resolved: https://github.com/facebook/rocksdb/pull/11301 Test Plan: ## Unit tests Generally updated to include HCC in SecondaryCache tests, though HyperClockCache has some different, less strict behaviors that leads to some tests not really being set up to work with it. Some of the tests remain disabled with it, but I think we have good coverage without them. ## Crash/stress test Updated to use the new combination. ## Performance First, let's check for regression on caches without secondary cache configured. Adding support for the eviction callback is likely to have a tiny effect, but it shouldn't be worrisome. LRUCache could benefit slightly from less logic around SecondaryCache handling. We can test with cache_bench default settings, built with DEBUG_LEVEL=0 and PORTABLE=0. ``` (while :; do base/cache_bench --cache_type=hyper_clock_cache | grep Rough; done) | awk '{ sum += $9; count++; print $0; print "Average: " int(sum / count) }' ``` **Before** this and #11299 (which could also have a small effect), running for about an hour, before & after running concurrently for each cache type: HyperClockCache: 3168662 (average parallel ops/sec) LRUCache: 2940127 **After** this and #11299, running for about an hour: HyperClockCache: 3164862 (average parallel ops/sec) (0.12% slower) LRUCache: 2940928 (0.03% faster) This is an acceptable difference IMHO. Next, let's consider essentially the worst case of new CPU overhead affecting overall performance. MultiGet uses the async lookup interface regardless of whether SecondaryCache or folly are used. We can configure a benchmark where all block cache queries are for data blocks, and all are hits. Create DB and test (before and after tests running simultaneously): ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 TEST_TMPDIR=/dev/shm base/db_bench -benchmarks=multireadrandom[-X30] -readonly -multiread_batched -batch_size=32 -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16 ``` **Before**: multireadrandom [AVG 30 runs] : 3444202 (± 57049) ops/sec; 240.9 (± 4.0) MB/sec multireadrandom [MEDIAN 30 runs] : 3514443 ops/sec; 245.8 MB/sec **After**: multireadrandom [AVG 30 runs] : 3291022 (± 58851) ops/sec; 230.2 (± 4.1) MB/sec multireadrandom [MEDIAN 30 runs] : 3366179 ops/sec; 235.4 MB/sec So that's roughly a 3% regression, on kind of a *worst case* test of MultiGet CPU. Similar story with HyperClockCache: **Before**: multireadrandom [AVG 30 runs] : 3933777 (± 41840) ops/sec; 275.1 (± 2.9) MB/sec multireadrandom [MEDIAN 30 runs] : 3970667 ops/sec; 277.7 MB/sec **After**: multireadrandom [AVG 30 runs] : 3755338 (± 30391) ops/sec; 262.6 (± 2.1) MB/sec multireadrandom [MEDIAN 30 runs] : 3785696 ops/sec; 264.8 MB/sec Roughly a 4-5% regression. Not ideal, but not the whole story, fortunately. Let's also look at Get() in db_bench: ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X30] -readonly -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16 ``` **Before**: readrandom [AVG 30 runs] : 2198685 (± 13412) ops/sec; 153.8 (± 0.9) MB/sec readrandom [MEDIAN 30 runs] : 2209498 ops/sec; 154.5 MB/sec **After**: readrandom [AVG 30 runs] : 2292814 (± 43508) ops/sec; 160.3 (± 3.0) MB/sec readrandom [MEDIAN 30 runs] : 2365181 ops/sec; 165.4 MB/sec That's showing roughly a 4% improvement, perhaps because of the secondary cache code that is no longer part of LRUCache. But weirdly, HyperClockCache is also showing 2-3% improvement: **Before**: readrandom [AVG 30 runs] : 2272333 (± 9992) ops/sec; 158.9 (± 0.7) MB/sec readrandom [MEDIAN 30 runs] : 2273239 ops/sec; 159.0 MB/sec **After**: readrandom [AVG 30 runs] : 2332407 (± 11252) ops/sec; 163.1 (± 0.8) MB/sec readrandom [MEDIAN 30 runs] : 2335329 ops/sec; 163.3 MB/sec Reviewed By: ltamasi Differential Revision: D44177044 Pulled By: pdillinger fbshipit-source-id: e808e48ff3fe2f792a79841ba617be98e48689f5
2 years ago
bool took_ownership = false;
if (eviction_callback_) {
took_ownership =
eviction_callback_(ClockCacheShard<HyperClockTable>::ReverseHash(
h.GetHash(), &unhashed),
reinterpret_cast<Cache::Handle*>(&h));
}
if (!took_ownership) {
h.FreeData(allocator_);
}
MarkEmpty(h);
}
}
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
// Loop exit condition
if (*freed_charge >= requested_charge) {
return;
}
if (old_clock_pointer >= max_clock_pointer) {
return;
}
// Advance clock pointer (concurrently)
old_clock_pointer =
clock_pointer_.fetch_add(step_size, std::memory_order_relaxed);
}
}
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
template <class Table>
ClockCacheShard<Table>::ClockCacheShard(
size_t capacity, bool strict_capacity_limit,
CacheMetadataChargePolicy metadata_charge_policy,
HyperClockCache support for SecondaryCache, with refactoring (#11301) Summary: Internally refactors SecondaryCache integration out of LRUCache specifically and into a wrapper/adapter class that works with various Cache implementations. Notably, this relies on separating the notion of async lookup handles from other cache handles, so that HyperClockCache doesn't have to deal with the problem of allocating handles from the hash table for lookups that might fail anyway, and might be on the same key without support for coalescing. (LRUCache's hash table can incorporate previously allocated handles thanks to its pointer indirection.) Specifically, I'm worried about the case in which hundreds of threads try to access the same block and probing in the hash table degrades to linear search on the pile of entries with the same key. This change is a big step in the direction of supporting stacked SecondaryCaches, but there are obstacles to completing that. Especially, there is no SecondaryCache hook for evictions to pass from one to the next. It has been proposed that evictions be transmitted simply as the persisted data (as in SaveToCallback), but given the current structure provided by the CacheItemHelpers, that would require an extra copy of the block data, because there's intentionally no way to ask for a contiguous Slice of the data (to allow for flexibility in storage). `AsyncLookupHandle` and the re-worked `WaitAll()` should be essentially prepared for stacked SecondaryCaches, but several "TODO with stacked secondaries" issues remain in various places. It could be argued that the stacking instead be done as a SecondaryCache adapter that wraps two (or more) SecondaryCaches, but at least with the current API that would require an extra heap allocation on SecondaryCache Lookup for a wrapper SecondaryCacheResultHandle that can transfer a Lookup between secondaries. We could also consider trying to unify the Cache and SecondaryCache APIs, though that might be difficult if `AsyncLookupHandle` is kept a fixed struct. ## cache.h (public API) Moves `secondary_cache` option from LRUCacheOptions to ShardedCacheOptions so that it is applicable to HyperClockCache. ## advanced_cache.h (advanced public API) * Add `Cache::CreateStandalone()` so that the SecondaryCache support wrapper can use it. * Add `SetEvictionCallback()` / `eviction_callback_` so that the SecondaryCache support wrapper can use it. Only a single callback is supported for efficiency. If there is ever a need for more than one, hopefully that can be handled with a broadcast callback wrapper. These are essentially the two "extra" pieces of `Cache` for pulling out specific SecondaryCache support from the `Cache` implementation. I think it's a good trade-off as these are reasonable, limited, and reusable "cut points" into the `Cache` implementations. * Remove async capability from standard `Lookup()` (getting rid of awkward restrictions on pending Handles) and add `AsyncLookupHandle` and `StartAsyncLookup()`. As noted in the comments, the full struct of `AsyncLookupHandle` is exposed so that it can be stack allocated, for efficiency, though more data is being copied around than before, which could impact performance. (Lookup info -> AsyncLookupHandle -> Handle vs. Lookup info -> Handle) I could foresee a future in which a Cache internally saves a pointer to the AsyncLookupHandle, which means it's dangerous to allow it to be copyable or even movable. It also means it's not compatible with std::vector (which I don't like requiring as an API parameter anyway), so `WaitAll()` expects any contiguous array of AsyncLookupHandles. I believe this is best for common case efficiency, while behaving well in other cases also. For example, `WaitAll()` has no effect on default-constructed AsyncLookupHandles, which look like a completed cache miss. ## cacheable_entry.h A couple of functions are obsolete because Cache::Handle can no longer be pending. ## cache.cc Provides default implementations for new or revamped Cache functions, especially appropriate for non-blocking caches. ## secondary_cache_adapter.{h,cc} The full details of the Cache wrapper adding SecondaryCache support. Essentially replicates the SecondaryCache handling that was in LRUCache, but obviously refactored. There is a bit of logic duplication, where Lookup() is essentially a manually optimized version of StartAsyncLookup() and Wait(), but it's roughly a dozen lines of code. ## sharded_cache.h, typed_cache.h, charged_cache.{h,cc}, sim_cache.cc Simply updated for Cache API changes. ## lru_cache.{h,cc} Carefully remove SecondaryCache logic, implement `CreateStandalone` and eviction handler functionality. ## clock_cache.{h,cc} Expose existing `CreateStandalone` functionality, add eviction handler functionality. Light refactoring. ## block_based_table_reader* Mostly re-worked the only usage of async Lookup, which is in BlockBasedTable::MultiGet. Used arrays in place of autovector in some places for efficiency. Simplified some logic by not trying to process some cache results before they're all ready. Created new function `BlockBasedTable::GetCachePriority()` to reduce some pre-existing code duplication (and avoid making it worse). Fixed at least one small bug from the prior confusing mixture of async and sync Lookups. In MaybeReadBlockAndLoadToCache(), called by RetrieveBlock(), called by MultiGet() with wait=false, is_cache_hit for the block_cache_tracer entry would not be set to true if the handle was pending after Lookup and before Wait. ## Intended follow-up work * Figure out if there are any missing stats or block_cache_tracer work in refactored BlockBasedTable::MultiGet * Stacked secondary caches (see above discussion) * See if we can make up for the small MultiGet performance regression. * Study more performance with SecondaryCache * Items evicted from over-full LRUCache in Release were not being demoted to SecondaryCache, and still aren't to minimize unit test churn. Ideally they would be demoted, but it's an exceptional case so not a big deal. * Use CreateStandalone for cache reservations (save unnecessary hash table operations). Not a big deal, but worthy cleanup. * Somehow I got the contract for SecondaryCache::Insert wrong in #10945. (Doesn't take ownership!) That API comment needs to be fixed, but didn't want to mingle that in here. Pull Request resolved: https://github.com/facebook/rocksdb/pull/11301 Test Plan: ## Unit tests Generally updated to include HCC in SecondaryCache tests, though HyperClockCache has some different, less strict behaviors that leads to some tests not really being set up to work with it. Some of the tests remain disabled with it, but I think we have good coverage without them. ## Crash/stress test Updated to use the new combination. ## Performance First, let's check for regression on caches without secondary cache configured. Adding support for the eviction callback is likely to have a tiny effect, but it shouldn't be worrisome. LRUCache could benefit slightly from less logic around SecondaryCache handling. We can test with cache_bench default settings, built with DEBUG_LEVEL=0 and PORTABLE=0. ``` (while :; do base/cache_bench --cache_type=hyper_clock_cache | grep Rough; done) | awk '{ sum += $9; count++; print $0; print "Average: " int(sum / count) }' ``` **Before** this and #11299 (which could also have a small effect), running for about an hour, before & after running concurrently for each cache type: HyperClockCache: 3168662 (average parallel ops/sec) LRUCache: 2940127 **After** this and #11299, running for about an hour: HyperClockCache: 3164862 (average parallel ops/sec) (0.12% slower) LRUCache: 2940928 (0.03% faster) This is an acceptable difference IMHO. Next, let's consider essentially the worst case of new CPU overhead affecting overall performance. MultiGet uses the async lookup interface regardless of whether SecondaryCache or folly are used. We can configure a benchmark where all block cache queries are for data blocks, and all are hits. Create DB and test (before and after tests running simultaneously): ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 TEST_TMPDIR=/dev/shm base/db_bench -benchmarks=multireadrandom[-X30] -readonly -multiread_batched -batch_size=32 -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16 ``` **Before**: multireadrandom [AVG 30 runs] : 3444202 (± 57049) ops/sec; 240.9 (± 4.0) MB/sec multireadrandom [MEDIAN 30 runs] : 3514443 ops/sec; 245.8 MB/sec **After**: multireadrandom [AVG 30 runs] : 3291022 (± 58851) ops/sec; 230.2 (± 4.1) MB/sec multireadrandom [MEDIAN 30 runs] : 3366179 ops/sec; 235.4 MB/sec So that's roughly a 3% regression, on kind of a *worst case* test of MultiGet CPU. Similar story with HyperClockCache: **Before**: multireadrandom [AVG 30 runs] : 3933777 (± 41840) ops/sec; 275.1 (± 2.9) MB/sec multireadrandom [MEDIAN 30 runs] : 3970667 ops/sec; 277.7 MB/sec **After**: multireadrandom [AVG 30 runs] : 3755338 (± 30391) ops/sec; 262.6 (± 2.1) MB/sec multireadrandom [MEDIAN 30 runs] : 3785696 ops/sec; 264.8 MB/sec Roughly a 4-5% regression. Not ideal, but not the whole story, fortunately. Let's also look at Get() in db_bench: ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X30] -readonly -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16 ``` **Before**: readrandom [AVG 30 runs] : 2198685 (± 13412) ops/sec; 153.8 (± 0.9) MB/sec readrandom [MEDIAN 30 runs] : 2209498 ops/sec; 154.5 MB/sec **After**: readrandom [AVG 30 runs] : 2292814 (± 43508) ops/sec; 160.3 (± 3.0) MB/sec readrandom [MEDIAN 30 runs] : 2365181 ops/sec; 165.4 MB/sec That's showing roughly a 4% improvement, perhaps because of the secondary cache code that is no longer part of LRUCache. But weirdly, HyperClockCache is also showing 2-3% improvement: **Before**: readrandom [AVG 30 runs] : 2272333 (± 9992) ops/sec; 158.9 (± 0.7) MB/sec readrandom [MEDIAN 30 runs] : 2273239 ops/sec; 159.0 MB/sec **After**: readrandom [AVG 30 runs] : 2332407 (± 11252) ops/sec; 163.1 (± 0.8) MB/sec readrandom [MEDIAN 30 runs] : 2335329 ops/sec; 163.3 MB/sec Reviewed By: ltamasi Differential Revision: D44177044 Pulled By: pdillinger fbshipit-source-id: e808e48ff3fe2f792a79841ba617be98e48689f5
2 years ago
MemoryAllocator* allocator,
const Cache::EvictionCallback* eviction_callback,
const typename Table::Opts& opts)
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
: CacheShardBase(metadata_charge_policy),
Major Cache refactoring, CPU efficiency improvement (#10975) Summary: This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache). The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below. * static_cast lines of code +29 -35 (net removed 6) * reinterpret_cast lines of code +6 -32 (net removed 26) ## cache.h and secondary_cache.h * Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications: * Simpler for implementations to deal with just one Insert and one Lookup. * Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters * Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428. * Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks). * It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below). * I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc. * Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation. * Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.) * Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.) * Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774) * Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object. * Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change. ## typed_cache.h Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae). The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used. * PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value. * BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter. * FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue. * For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`. These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.) Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it. ## block_cache.h This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table. ## block_based_table_reader.cc Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation. The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions. ## block_based_table_builder.cc, cache_dump_load_impl.cc Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.) ## Everything else Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975 Test Plan: tests updated Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache): 34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844 34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594 34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297 34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523 34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602 34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293 34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926 34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488 233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984 233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922 233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559 233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93 233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418 233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273 233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691 233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82 1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55 1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02 1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45 1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24 1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92 1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78 1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36 1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83 Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn. Reviewed By: anand1976 Differential Revision: D42417818 Pulled By: pdillinger fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
table_(capacity, strict_capacity_limit, metadata_charge_policy, allocator,
HyperClockCache support for SecondaryCache, with refactoring (#11301) Summary: Internally refactors SecondaryCache integration out of LRUCache specifically and into a wrapper/adapter class that works with various Cache implementations. Notably, this relies on separating the notion of async lookup handles from other cache handles, so that HyperClockCache doesn't have to deal with the problem of allocating handles from the hash table for lookups that might fail anyway, and might be on the same key without support for coalescing. (LRUCache's hash table can incorporate previously allocated handles thanks to its pointer indirection.) Specifically, I'm worried about the case in which hundreds of threads try to access the same block and probing in the hash table degrades to linear search on the pile of entries with the same key. This change is a big step in the direction of supporting stacked SecondaryCaches, but there are obstacles to completing that. Especially, there is no SecondaryCache hook for evictions to pass from one to the next. It has been proposed that evictions be transmitted simply as the persisted data (as in SaveToCallback), but given the current structure provided by the CacheItemHelpers, that would require an extra copy of the block data, because there's intentionally no way to ask for a contiguous Slice of the data (to allow for flexibility in storage). `AsyncLookupHandle` and the re-worked `WaitAll()` should be essentially prepared for stacked SecondaryCaches, but several "TODO with stacked secondaries" issues remain in various places. It could be argued that the stacking instead be done as a SecondaryCache adapter that wraps two (or more) SecondaryCaches, but at least with the current API that would require an extra heap allocation on SecondaryCache Lookup for a wrapper SecondaryCacheResultHandle that can transfer a Lookup between secondaries. We could also consider trying to unify the Cache and SecondaryCache APIs, though that might be difficult if `AsyncLookupHandle` is kept a fixed struct. ## cache.h (public API) Moves `secondary_cache` option from LRUCacheOptions to ShardedCacheOptions so that it is applicable to HyperClockCache. ## advanced_cache.h (advanced public API) * Add `Cache::CreateStandalone()` so that the SecondaryCache support wrapper can use it. * Add `SetEvictionCallback()` / `eviction_callback_` so that the SecondaryCache support wrapper can use it. Only a single callback is supported for efficiency. If there is ever a need for more than one, hopefully that can be handled with a broadcast callback wrapper. These are essentially the two "extra" pieces of `Cache` for pulling out specific SecondaryCache support from the `Cache` implementation. I think it's a good trade-off as these are reasonable, limited, and reusable "cut points" into the `Cache` implementations. * Remove async capability from standard `Lookup()` (getting rid of awkward restrictions on pending Handles) and add `AsyncLookupHandle` and `StartAsyncLookup()`. As noted in the comments, the full struct of `AsyncLookupHandle` is exposed so that it can be stack allocated, for efficiency, though more data is being copied around than before, which could impact performance. (Lookup info -> AsyncLookupHandle -> Handle vs. Lookup info -> Handle) I could foresee a future in which a Cache internally saves a pointer to the AsyncLookupHandle, which means it's dangerous to allow it to be copyable or even movable. It also means it's not compatible with std::vector (which I don't like requiring as an API parameter anyway), so `WaitAll()` expects any contiguous array of AsyncLookupHandles. I believe this is best for common case efficiency, while behaving well in other cases also. For example, `WaitAll()` has no effect on default-constructed AsyncLookupHandles, which look like a completed cache miss. ## cacheable_entry.h A couple of functions are obsolete because Cache::Handle can no longer be pending. ## cache.cc Provides default implementations for new or revamped Cache functions, especially appropriate for non-blocking caches. ## secondary_cache_adapter.{h,cc} The full details of the Cache wrapper adding SecondaryCache support. Essentially replicates the SecondaryCache handling that was in LRUCache, but obviously refactored. There is a bit of logic duplication, where Lookup() is essentially a manually optimized version of StartAsyncLookup() and Wait(), but it's roughly a dozen lines of code. ## sharded_cache.h, typed_cache.h, charged_cache.{h,cc}, sim_cache.cc Simply updated for Cache API changes. ## lru_cache.{h,cc} Carefully remove SecondaryCache logic, implement `CreateStandalone` and eviction handler functionality. ## clock_cache.{h,cc} Expose existing `CreateStandalone` functionality, add eviction handler functionality. Light refactoring. ## block_based_table_reader* Mostly re-worked the only usage of async Lookup, which is in BlockBasedTable::MultiGet. Used arrays in place of autovector in some places for efficiency. Simplified some logic by not trying to process some cache results before they're all ready. Created new function `BlockBasedTable::GetCachePriority()` to reduce some pre-existing code duplication (and avoid making it worse). Fixed at least one small bug from the prior confusing mixture of async and sync Lookups. In MaybeReadBlockAndLoadToCache(), called by RetrieveBlock(), called by MultiGet() with wait=false, is_cache_hit for the block_cache_tracer entry would not be set to true if the handle was pending after Lookup and before Wait. ## Intended follow-up work * Figure out if there are any missing stats or block_cache_tracer work in refactored BlockBasedTable::MultiGet * Stacked secondary caches (see above discussion) * See if we can make up for the small MultiGet performance regression. * Study more performance with SecondaryCache * Items evicted from over-full LRUCache in Release were not being demoted to SecondaryCache, and still aren't to minimize unit test churn. Ideally they would be demoted, but it's an exceptional case so not a big deal. * Use CreateStandalone for cache reservations (save unnecessary hash table operations). Not a big deal, but worthy cleanup. * Somehow I got the contract for SecondaryCache::Insert wrong in #10945. (Doesn't take ownership!) That API comment needs to be fixed, but didn't want to mingle that in here. Pull Request resolved: https://github.com/facebook/rocksdb/pull/11301 Test Plan: ## Unit tests Generally updated to include HCC in SecondaryCache tests, though HyperClockCache has some different, less strict behaviors that leads to some tests not really being set up to work with it. Some of the tests remain disabled with it, but I think we have good coverage without them. ## Crash/stress test Updated to use the new combination. ## Performance First, let's check for regression on caches without secondary cache configured. Adding support for the eviction callback is likely to have a tiny effect, but it shouldn't be worrisome. LRUCache could benefit slightly from less logic around SecondaryCache handling. We can test with cache_bench default settings, built with DEBUG_LEVEL=0 and PORTABLE=0. ``` (while :; do base/cache_bench --cache_type=hyper_clock_cache | grep Rough; done) | awk '{ sum += $9; count++; print $0; print "Average: " int(sum / count) }' ``` **Before** this and #11299 (which could also have a small effect), running for about an hour, before & after running concurrently for each cache type: HyperClockCache: 3168662 (average parallel ops/sec) LRUCache: 2940127 **After** this and #11299, running for about an hour: HyperClockCache: 3164862 (average parallel ops/sec) (0.12% slower) LRUCache: 2940928 (0.03% faster) This is an acceptable difference IMHO. Next, let's consider essentially the worst case of new CPU overhead affecting overall performance. MultiGet uses the async lookup interface regardless of whether SecondaryCache or folly are used. We can configure a benchmark where all block cache queries are for data blocks, and all are hits. Create DB and test (before and after tests running simultaneously): ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 TEST_TMPDIR=/dev/shm base/db_bench -benchmarks=multireadrandom[-X30] -readonly -multiread_batched -batch_size=32 -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16 ``` **Before**: multireadrandom [AVG 30 runs] : 3444202 (± 57049) ops/sec; 240.9 (± 4.0) MB/sec multireadrandom [MEDIAN 30 runs] : 3514443 ops/sec; 245.8 MB/sec **After**: multireadrandom [AVG 30 runs] : 3291022 (± 58851) ops/sec; 230.2 (± 4.1) MB/sec multireadrandom [MEDIAN 30 runs] : 3366179 ops/sec; 235.4 MB/sec So that's roughly a 3% regression, on kind of a *worst case* test of MultiGet CPU. Similar story with HyperClockCache: **Before**: multireadrandom [AVG 30 runs] : 3933777 (± 41840) ops/sec; 275.1 (± 2.9) MB/sec multireadrandom [MEDIAN 30 runs] : 3970667 ops/sec; 277.7 MB/sec **After**: multireadrandom [AVG 30 runs] : 3755338 (± 30391) ops/sec; 262.6 (± 2.1) MB/sec multireadrandom [MEDIAN 30 runs] : 3785696 ops/sec; 264.8 MB/sec Roughly a 4-5% regression. Not ideal, but not the whole story, fortunately. Let's also look at Get() in db_bench: ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X30] -readonly -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16 ``` **Before**: readrandom [AVG 30 runs] : 2198685 (± 13412) ops/sec; 153.8 (± 0.9) MB/sec readrandom [MEDIAN 30 runs] : 2209498 ops/sec; 154.5 MB/sec **After**: readrandom [AVG 30 runs] : 2292814 (± 43508) ops/sec; 160.3 (± 3.0) MB/sec readrandom [MEDIAN 30 runs] : 2365181 ops/sec; 165.4 MB/sec That's showing roughly a 4% improvement, perhaps because of the secondary cache code that is no longer part of LRUCache. But weirdly, HyperClockCache is also showing 2-3% improvement: **Before**: readrandom [AVG 30 runs] : 2272333 (± 9992) ops/sec; 158.9 (± 0.7) MB/sec readrandom [MEDIAN 30 runs] : 2273239 ops/sec; 159.0 MB/sec **After**: readrandom [AVG 30 runs] : 2332407 (± 11252) ops/sec; 163.1 (± 0.8) MB/sec readrandom [MEDIAN 30 runs] : 2335329 ops/sec; 163.3 MB/sec Reviewed By: ltamasi Differential Revision: D44177044 Pulled By: pdillinger fbshipit-source-id: e808e48ff3fe2f792a79841ba617be98e48689f5
2 years ago
eviction_callback, opts),
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
capacity_(capacity),
strict_capacity_limit_(strict_capacity_limit) {
// Initial charge metadata should not exceed capacity
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
assert(table_.GetUsage() <= capacity_ || capacity_ < sizeof(HandleImpl));
}
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
template <class Table>
void ClockCacheShard<Table>::EraseUnRefEntries() {
table_.EraseUnRefEntries();
}
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
template <class Table>
void ClockCacheShard<Table>::ApplyToSomeEntries(
Major Cache refactoring, CPU efficiency improvement (#10975) Summary: This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache). The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below. * static_cast lines of code +29 -35 (net removed 6) * reinterpret_cast lines of code +6 -32 (net removed 26) ## cache.h and secondary_cache.h * Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications: * Simpler for implementations to deal with just one Insert and one Lookup. * Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters * Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428. * Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks). * It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below). * I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc. * Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation. * Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.) * Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.) * Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774) * Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object. * Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change. ## typed_cache.h Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae). The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used. * PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value. * BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter. * FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue. * For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`. These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.) Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it. ## block_cache.h This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table. ## block_based_table_reader.cc Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation. The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions. ## block_based_table_builder.cc, cache_dump_load_impl.cc Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.) ## Everything else Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975 Test Plan: tests updated Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache): 34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844 34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594 34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297 34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523 34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602 34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293 34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926 34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488 233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984 233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922 233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559 233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93 233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418 233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273 233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691 233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82 1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55 1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02 1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45 1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24 1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92 1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78 1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36 1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83 Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn. Reviewed By: anand1976 Differential Revision: D42417818 Pulled By: pdillinger fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
const std::function<void(const Slice& key, Cache::ObjectPtr value,
size_t charge,
const Cache::CacheItemHelper* helper)>& callback,
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
size_t average_entries_per_lock, size_t* state) {
// The state is essentially going to be the starting hash, which works
// nicely even if we resize between calls because we use upper-most
// hash bits for table indexes.
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
size_t length_bits = table_.GetLengthBits();
size_t length = table_.GetTableSize();
New Cache API for gathering statistics (#8225) Summary: Adds a new Cache::ApplyToAllEntries API that we expect to use (in follow-up PRs) for efficiently gathering block cache statistics. Notable features vs. old ApplyToAllCacheEntries: * Includes key and deleter (in addition to value and charge). We could have passed in a Handle but then more virtual function calls would be needed to get the "fields" of each entry. We expect to use the 'deleter' to identify the origin of entries, perhaps even more. * Heavily tuned to minimize latency impact on operating cache. It does this by iterating over small sections of each cache shard while cycling through the shards. * Supports tuning roughly how many entries to operate on for each lock acquire and release, to control the impact on the latency of other operations without excessive lock acquire & release. The right balance can depend on the cost of the callback. Good default seems to be around 256. * There should be no need to disable thread safety. (I would expect uncontended locks to be sufficiently fast.) I have enhanced cache_bench to validate this approach: * Reports a histogram of ns per operation, so we can look at the ditribution of times, not just throughput (average). * Can add a thread for simulated "gather stats" which calls ApplyToAllEntries at a specified interval. We also generate a histogram of time to run ApplyToAllEntries. To make the iteration over some entries of each shard work as cleanly as possible, even with resize between next set of entries, I have re-arranged which hash bits are used for sharding and which for indexing within a shard. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8225 Test Plan: A couple of unit tests are added, but primary validation is manual, as the primary risk is to performance. The primary validation is using cache_bench to ensure that neither the minor hashing changes nor the simulated stats gathering significantly impact QPS or latency distribution. Note that adding op latency histogram seriously impacts the benchmark QPS, so for a fair baseline, we need the cache_bench changes (except remove simulated stat gathering to make it compile). In short, we don't see any reproducible difference in ops/sec or op latency unless we are gathering stats nearly continuously. Test uses 10GB block cache with 8KB values to be somewhat realistic in the number of items to iterate over. Baseline typical output: ``` Complete in 92.017 s; Rough parallel ops/sec = 869401 Thread ops/sec = 54662 Operation latency (ns): Count: 80000000 Average: 11223.9494 StdDev: 29.61 Min: 0 Median: 7759.3973 Max: 9620500 Percentiles: P50: 7759.40 P75: 14190.73 P99: 46922.75 P99.9: 77509.84 P99.99: 217030.58 ------------------------------------------------------ [ 0, 1 ] 68 0.000% 0.000% ( 2900, 4400 ] 89 0.000% 0.000% ( 4400, 6600 ] 33630240 42.038% 42.038% ######## ( 6600, 9900 ] 18129842 22.662% 64.700% ##### ( 9900, 14000 ] 7877533 9.847% 74.547% ## ( 14000, 22000 ] 15193238 18.992% 93.539% #### ( 22000, 33000 ] 3037061 3.796% 97.335% # ( 33000, 50000 ] 1626316 2.033% 99.368% ( 50000, 75000 ] 421532 0.527% 99.895% ( 75000, 110000 ] 56910 0.071% 99.966% ( 110000, 170000 ] 16134 0.020% 99.986% ( 170000, 250000 ] 5166 0.006% 99.993% ( 250000, 380000 ] 3017 0.004% 99.996% ( 380000, 570000 ] 1337 0.002% 99.998% ( 570000, 860000 ] 805 0.001% 99.999% ( 860000, 1200000 ] 319 0.000% 100.000% ( 1200000, 1900000 ] 231 0.000% 100.000% ( 1900000, 2900000 ] 100 0.000% 100.000% ( 2900000, 4300000 ] 39 0.000% 100.000% ( 4300000, 6500000 ] 16 0.000% 100.000% ( 6500000, 9800000 ] 7 0.000% 100.000% ``` New, gather_stats=false. Median thread ops/sec of 5 runs: ``` Complete in 92.030 s; Rough parallel ops/sec = 869285 Thread ops/sec = 54458 Operation latency (ns): Count: 80000000 Average: 11298.1027 StdDev: 42.18 Min: 0 Median: 7722.0822 Max: 6398720 Percentiles: P50: 7722.08 P75: 14294.68 P99: 47522.95 P99.9: 85292.16 P99.99: 228077.78 ------------------------------------------------------ [ 0, 1 ] 109 0.000% 0.000% ( 2900, 4400 ] 793 0.001% 0.001% ( 4400, 6600 ] 34054563 42.568% 42.569% ######### ( 6600, 9900 ] 17482646 21.853% 64.423% #### ( 9900, 14000 ] 7908180 9.885% 74.308% ## ( 14000, 22000 ] 15032072 18.790% 93.098% #### ( 22000, 33000 ] 3237834 4.047% 97.145% # ( 33000, 50000 ] 1736882 2.171% 99.316% ( 50000, 75000 ] 446851 0.559% 99.875% ( 75000, 110000 ] 68251 0.085% 99.960% ( 110000, 170000 ] 18592 0.023% 99.983% ( 170000, 250000 ] 7200 0.009% 99.992% ( 250000, 380000 ] 3334 0.004% 99.997% ( 380000, 570000 ] 1393 0.002% 99.998% ( 570000, 860000 ] 700 0.001% 99.999% ( 860000, 1200000 ] 293 0.000% 100.000% ( 1200000, 1900000 ] 196 0.000% 100.000% ( 1900000, 2900000 ] 69 0.000% 100.000% ( 2900000, 4300000 ] 32 0.000% 100.000% ( 4300000, 6500000 ] 10 0.000% 100.000% ``` New, gather_stats=true, 1 second delay between scans. Scans take about 1 second here so it's spending about 50% time scanning. Still the effect on ops/sec and latency seems to be in the noise. Median thread ops/sec of 5 runs: ``` Complete in 91.890 s; Rough parallel ops/sec = 870608 Thread ops/sec = 54551 Operation latency (ns): Count: 80000000 Average: 11311.2629 StdDev: 45.28 Min: 0 Median: 7686.5458 Max: 10018340 Percentiles: P50: 7686.55 P75: 14481.95 P99: 47232.60 P99.9: 79230.18 P99.99: 232998.86 ------------------------------------------------------ [ 0, 1 ] 71 0.000% 0.000% ( 2900, 4400 ] 291 0.000% 0.000% ( 4400, 6600 ] 34492060 43.115% 43.116% ######### ( 6600, 9900 ] 16727328 20.909% 64.025% #### ( 9900, 14000 ] 7845828 9.807% 73.832% ## ( 14000, 22000 ] 15510654 19.388% 93.220% #### ( 22000, 33000 ] 3216533 4.021% 97.241% # ( 33000, 50000 ] 1680859 2.101% 99.342% ( 50000, 75000 ] 439059 0.549% 99.891% ( 75000, 110000 ] 60540 0.076% 99.967% ( 110000, 170000 ] 14649 0.018% 99.985% ( 170000, 250000 ] 5242 0.007% 99.991% ( 250000, 380000 ] 3260 0.004% 99.995% ( 380000, 570000 ] 1599 0.002% 99.997% ( 570000, 860000 ] 1043 0.001% 99.999% ( 860000, 1200000 ] 471 0.001% 99.999% ( 1200000, 1900000 ] 275 0.000% 100.000% ( 1900000, 2900000 ] 143 0.000% 100.000% ( 2900000, 4300000 ] 60 0.000% 100.000% ( 4300000, 6500000 ] 27 0.000% 100.000% ( 6500000, 9800000 ] 7 0.000% 100.000% ( 9800000, 14000000 ] 1 0.000% 100.000% Gather stats latency (us): Count: 46 Average: 980387.5870 StdDev: 60911.18 Min: 879155 Median: 1033777.7778 Max: 1261431 Percentiles: P50: 1033777.78 P75: 1120666.67 P99: 1261431.00 P99.9: 1261431.00 P99.99: 1261431.00 ------------------------------------------------------ ( 860000, 1200000 ] 45 97.826% 97.826% #################### ( 1200000, 1900000 ] 1 2.174% 100.000% Most recent cache entry stats: Number of entries: 1295133 Total charge: 9.88 GB Average key size: 23.4982 Average charge: 8.00 KB Unique deleters: 3 ``` Reviewed By: mrambacher Differential Revision: D28295742 Pulled By: pdillinger fbshipit-source-id: bbc4a552f91ba0fe10e5cc025c42cef5a81f2b95
4 years ago
assert(average_entries_per_lock > 0);
// Assuming we are called with same average_entries_per_lock repeatedly,
// this simplifies some logic (index_end will not overflow).
assert(average_entries_per_lock < length || *state == 0);
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
size_t index_begin = *state >> (sizeof(size_t) * 8u - length_bits);
size_t index_end = index_begin + average_entries_per_lock;
if (index_end >= length) {
// Going to end.
index_end = length;
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
*state = SIZE_MAX;
New Cache API for gathering statistics (#8225) Summary: Adds a new Cache::ApplyToAllEntries API that we expect to use (in follow-up PRs) for efficiently gathering block cache statistics. Notable features vs. old ApplyToAllCacheEntries: * Includes key and deleter (in addition to value and charge). We could have passed in a Handle but then more virtual function calls would be needed to get the "fields" of each entry. We expect to use the 'deleter' to identify the origin of entries, perhaps even more. * Heavily tuned to minimize latency impact on operating cache. It does this by iterating over small sections of each cache shard while cycling through the shards. * Supports tuning roughly how many entries to operate on for each lock acquire and release, to control the impact on the latency of other operations without excessive lock acquire & release. The right balance can depend on the cost of the callback. Good default seems to be around 256. * There should be no need to disable thread safety. (I would expect uncontended locks to be sufficiently fast.) I have enhanced cache_bench to validate this approach: * Reports a histogram of ns per operation, so we can look at the ditribution of times, not just throughput (average). * Can add a thread for simulated "gather stats" which calls ApplyToAllEntries at a specified interval. We also generate a histogram of time to run ApplyToAllEntries. To make the iteration over some entries of each shard work as cleanly as possible, even with resize between next set of entries, I have re-arranged which hash bits are used for sharding and which for indexing within a shard. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8225 Test Plan: A couple of unit tests are added, but primary validation is manual, as the primary risk is to performance. The primary validation is using cache_bench to ensure that neither the minor hashing changes nor the simulated stats gathering significantly impact QPS or latency distribution. Note that adding op latency histogram seriously impacts the benchmark QPS, so for a fair baseline, we need the cache_bench changes (except remove simulated stat gathering to make it compile). In short, we don't see any reproducible difference in ops/sec or op latency unless we are gathering stats nearly continuously. Test uses 10GB block cache with 8KB values to be somewhat realistic in the number of items to iterate over. Baseline typical output: ``` Complete in 92.017 s; Rough parallel ops/sec = 869401 Thread ops/sec = 54662 Operation latency (ns): Count: 80000000 Average: 11223.9494 StdDev: 29.61 Min: 0 Median: 7759.3973 Max: 9620500 Percentiles: P50: 7759.40 P75: 14190.73 P99: 46922.75 P99.9: 77509.84 P99.99: 217030.58 ------------------------------------------------------ [ 0, 1 ] 68 0.000% 0.000% ( 2900, 4400 ] 89 0.000% 0.000% ( 4400, 6600 ] 33630240 42.038% 42.038% ######## ( 6600, 9900 ] 18129842 22.662% 64.700% ##### ( 9900, 14000 ] 7877533 9.847% 74.547% ## ( 14000, 22000 ] 15193238 18.992% 93.539% #### ( 22000, 33000 ] 3037061 3.796% 97.335% # ( 33000, 50000 ] 1626316 2.033% 99.368% ( 50000, 75000 ] 421532 0.527% 99.895% ( 75000, 110000 ] 56910 0.071% 99.966% ( 110000, 170000 ] 16134 0.020% 99.986% ( 170000, 250000 ] 5166 0.006% 99.993% ( 250000, 380000 ] 3017 0.004% 99.996% ( 380000, 570000 ] 1337 0.002% 99.998% ( 570000, 860000 ] 805 0.001% 99.999% ( 860000, 1200000 ] 319 0.000% 100.000% ( 1200000, 1900000 ] 231 0.000% 100.000% ( 1900000, 2900000 ] 100 0.000% 100.000% ( 2900000, 4300000 ] 39 0.000% 100.000% ( 4300000, 6500000 ] 16 0.000% 100.000% ( 6500000, 9800000 ] 7 0.000% 100.000% ``` New, gather_stats=false. Median thread ops/sec of 5 runs: ``` Complete in 92.030 s; Rough parallel ops/sec = 869285 Thread ops/sec = 54458 Operation latency (ns): Count: 80000000 Average: 11298.1027 StdDev: 42.18 Min: 0 Median: 7722.0822 Max: 6398720 Percentiles: P50: 7722.08 P75: 14294.68 P99: 47522.95 P99.9: 85292.16 P99.99: 228077.78 ------------------------------------------------------ [ 0, 1 ] 109 0.000% 0.000% ( 2900, 4400 ] 793 0.001% 0.001% ( 4400, 6600 ] 34054563 42.568% 42.569% ######### ( 6600, 9900 ] 17482646 21.853% 64.423% #### ( 9900, 14000 ] 7908180 9.885% 74.308% ## ( 14000, 22000 ] 15032072 18.790% 93.098% #### ( 22000, 33000 ] 3237834 4.047% 97.145% # ( 33000, 50000 ] 1736882 2.171% 99.316% ( 50000, 75000 ] 446851 0.559% 99.875% ( 75000, 110000 ] 68251 0.085% 99.960% ( 110000, 170000 ] 18592 0.023% 99.983% ( 170000, 250000 ] 7200 0.009% 99.992% ( 250000, 380000 ] 3334 0.004% 99.997% ( 380000, 570000 ] 1393 0.002% 99.998% ( 570000, 860000 ] 700 0.001% 99.999% ( 860000, 1200000 ] 293 0.000% 100.000% ( 1200000, 1900000 ] 196 0.000% 100.000% ( 1900000, 2900000 ] 69 0.000% 100.000% ( 2900000, 4300000 ] 32 0.000% 100.000% ( 4300000, 6500000 ] 10 0.000% 100.000% ``` New, gather_stats=true, 1 second delay between scans. Scans take about 1 second here so it's spending about 50% time scanning. Still the effect on ops/sec and latency seems to be in the noise. Median thread ops/sec of 5 runs: ``` Complete in 91.890 s; Rough parallel ops/sec = 870608 Thread ops/sec = 54551 Operation latency (ns): Count: 80000000 Average: 11311.2629 StdDev: 45.28 Min: 0 Median: 7686.5458 Max: 10018340 Percentiles: P50: 7686.55 P75: 14481.95 P99: 47232.60 P99.9: 79230.18 P99.99: 232998.86 ------------------------------------------------------ [ 0, 1 ] 71 0.000% 0.000% ( 2900, 4400 ] 291 0.000% 0.000% ( 4400, 6600 ] 34492060 43.115% 43.116% ######### ( 6600, 9900 ] 16727328 20.909% 64.025% #### ( 9900, 14000 ] 7845828 9.807% 73.832% ## ( 14000, 22000 ] 15510654 19.388% 93.220% #### ( 22000, 33000 ] 3216533 4.021% 97.241% # ( 33000, 50000 ] 1680859 2.101% 99.342% ( 50000, 75000 ] 439059 0.549% 99.891% ( 75000, 110000 ] 60540 0.076% 99.967% ( 110000, 170000 ] 14649 0.018% 99.985% ( 170000, 250000 ] 5242 0.007% 99.991% ( 250000, 380000 ] 3260 0.004% 99.995% ( 380000, 570000 ] 1599 0.002% 99.997% ( 570000, 860000 ] 1043 0.001% 99.999% ( 860000, 1200000 ] 471 0.001% 99.999% ( 1200000, 1900000 ] 275 0.000% 100.000% ( 1900000, 2900000 ] 143 0.000% 100.000% ( 2900000, 4300000 ] 60 0.000% 100.000% ( 4300000, 6500000 ] 27 0.000% 100.000% ( 6500000, 9800000 ] 7 0.000% 100.000% ( 9800000, 14000000 ] 1 0.000% 100.000% Gather stats latency (us): Count: 46 Average: 980387.5870 StdDev: 60911.18 Min: 879155 Median: 1033777.7778 Max: 1261431 Percentiles: P50: 1033777.78 P75: 1120666.67 P99: 1261431.00 P99.9: 1261431.00 P99.99: 1261431.00 ------------------------------------------------------ ( 860000, 1200000 ] 45 97.826% 97.826% #################### ( 1200000, 1900000 ] 1 2.174% 100.000% Most recent cache entry stats: Number of entries: 1295133 Total charge: 9.88 GB Average key size: 23.4982 Average charge: 8.00 KB Unique deleters: 3 ``` Reviewed By: mrambacher Differential Revision: D28295742 Pulled By: pdillinger fbshipit-source-id: bbc4a552f91ba0fe10e5cc025c42cef5a81f2b95
4 years ago
} else {
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
*state = index_end << (sizeof(size_t) * 8u - length_bits);
}
New Cache API for gathering statistics (#8225) Summary: Adds a new Cache::ApplyToAllEntries API that we expect to use (in follow-up PRs) for efficiently gathering block cache statistics. Notable features vs. old ApplyToAllCacheEntries: * Includes key and deleter (in addition to value and charge). We could have passed in a Handle but then more virtual function calls would be needed to get the "fields" of each entry. We expect to use the 'deleter' to identify the origin of entries, perhaps even more. * Heavily tuned to minimize latency impact on operating cache. It does this by iterating over small sections of each cache shard while cycling through the shards. * Supports tuning roughly how many entries to operate on for each lock acquire and release, to control the impact on the latency of other operations without excessive lock acquire & release. The right balance can depend on the cost of the callback. Good default seems to be around 256. * There should be no need to disable thread safety. (I would expect uncontended locks to be sufficiently fast.) I have enhanced cache_bench to validate this approach: * Reports a histogram of ns per operation, so we can look at the ditribution of times, not just throughput (average). * Can add a thread for simulated "gather stats" which calls ApplyToAllEntries at a specified interval. We also generate a histogram of time to run ApplyToAllEntries. To make the iteration over some entries of each shard work as cleanly as possible, even with resize between next set of entries, I have re-arranged which hash bits are used for sharding and which for indexing within a shard. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8225 Test Plan: A couple of unit tests are added, but primary validation is manual, as the primary risk is to performance. The primary validation is using cache_bench to ensure that neither the minor hashing changes nor the simulated stats gathering significantly impact QPS or latency distribution. Note that adding op latency histogram seriously impacts the benchmark QPS, so for a fair baseline, we need the cache_bench changes (except remove simulated stat gathering to make it compile). In short, we don't see any reproducible difference in ops/sec or op latency unless we are gathering stats nearly continuously. Test uses 10GB block cache with 8KB values to be somewhat realistic in the number of items to iterate over. Baseline typical output: ``` Complete in 92.017 s; Rough parallel ops/sec = 869401 Thread ops/sec = 54662 Operation latency (ns): Count: 80000000 Average: 11223.9494 StdDev: 29.61 Min: 0 Median: 7759.3973 Max: 9620500 Percentiles: P50: 7759.40 P75: 14190.73 P99: 46922.75 P99.9: 77509.84 P99.99: 217030.58 ------------------------------------------------------ [ 0, 1 ] 68 0.000% 0.000% ( 2900, 4400 ] 89 0.000% 0.000% ( 4400, 6600 ] 33630240 42.038% 42.038% ######## ( 6600, 9900 ] 18129842 22.662% 64.700% ##### ( 9900, 14000 ] 7877533 9.847% 74.547% ## ( 14000, 22000 ] 15193238 18.992% 93.539% #### ( 22000, 33000 ] 3037061 3.796% 97.335% # ( 33000, 50000 ] 1626316 2.033% 99.368% ( 50000, 75000 ] 421532 0.527% 99.895% ( 75000, 110000 ] 56910 0.071% 99.966% ( 110000, 170000 ] 16134 0.020% 99.986% ( 170000, 250000 ] 5166 0.006% 99.993% ( 250000, 380000 ] 3017 0.004% 99.996% ( 380000, 570000 ] 1337 0.002% 99.998% ( 570000, 860000 ] 805 0.001% 99.999% ( 860000, 1200000 ] 319 0.000% 100.000% ( 1200000, 1900000 ] 231 0.000% 100.000% ( 1900000, 2900000 ] 100 0.000% 100.000% ( 2900000, 4300000 ] 39 0.000% 100.000% ( 4300000, 6500000 ] 16 0.000% 100.000% ( 6500000, 9800000 ] 7 0.000% 100.000% ``` New, gather_stats=false. Median thread ops/sec of 5 runs: ``` Complete in 92.030 s; Rough parallel ops/sec = 869285 Thread ops/sec = 54458 Operation latency (ns): Count: 80000000 Average: 11298.1027 StdDev: 42.18 Min: 0 Median: 7722.0822 Max: 6398720 Percentiles: P50: 7722.08 P75: 14294.68 P99: 47522.95 P99.9: 85292.16 P99.99: 228077.78 ------------------------------------------------------ [ 0, 1 ] 109 0.000% 0.000% ( 2900, 4400 ] 793 0.001% 0.001% ( 4400, 6600 ] 34054563 42.568% 42.569% ######### ( 6600, 9900 ] 17482646 21.853% 64.423% #### ( 9900, 14000 ] 7908180 9.885% 74.308% ## ( 14000, 22000 ] 15032072 18.790% 93.098% #### ( 22000, 33000 ] 3237834 4.047% 97.145% # ( 33000, 50000 ] 1736882 2.171% 99.316% ( 50000, 75000 ] 446851 0.559% 99.875% ( 75000, 110000 ] 68251 0.085% 99.960% ( 110000, 170000 ] 18592 0.023% 99.983% ( 170000, 250000 ] 7200 0.009% 99.992% ( 250000, 380000 ] 3334 0.004% 99.997% ( 380000, 570000 ] 1393 0.002% 99.998% ( 570000, 860000 ] 700 0.001% 99.999% ( 860000, 1200000 ] 293 0.000% 100.000% ( 1200000, 1900000 ] 196 0.000% 100.000% ( 1900000, 2900000 ] 69 0.000% 100.000% ( 2900000, 4300000 ] 32 0.000% 100.000% ( 4300000, 6500000 ] 10 0.000% 100.000% ``` New, gather_stats=true, 1 second delay between scans. Scans take about 1 second here so it's spending about 50% time scanning. Still the effect on ops/sec and latency seems to be in the noise. Median thread ops/sec of 5 runs: ``` Complete in 91.890 s; Rough parallel ops/sec = 870608 Thread ops/sec = 54551 Operation latency (ns): Count: 80000000 Average: 11311.2629 StdDev: 45.28 Min: 0 Median: 7686.5458 Max: 10018340 Percentiles: P50: 7686.55 P75: 14481.95 P99: 47232.60 P99.9: 79230.18 P99.99: 232998.86 ------------------------------------------------------ [ 0, 1 ] 71 0.000% 0.000% ( 2900, 4400 ] 291 0.000% 0.000% ( 4400, 6600 ] 34492060 43.115% 43.116% ######### ( 6600, 9900 ] 16727328 20.909% 64.025% #### ( 9900, 14000 ] 7845828 9.807% 73.832% ## ( 14000, 22000 ] 15510654 19.388% 93.220% #### ( 22000, 33000 ] 3216533 4.021% 97.241% # ( 33000, 50000 ] 1680859 2.101% 99.342% ( 50000, 75000 ] 439059 0.549% 99.891% ( 75000, 110000 ] 60540 0.076% 99.967% ( 110000, 170000 ] 14649 0.018% 99.985% ( 170000, 250000 ] 5242 0.007% 99.991% ( 250000, 380000 ] 3260 0.004% 99.995% ( 380000, 570000 ] 1599 0.002% 99.997% ( 570000, 860000 ] 1043 0.001% 99.999% ( 860000, 1200000 ] 471 0.001% 99.999% ( 1200000, 1900000 ] 275 0.000% 100.000% ( 1900000, 2900000 ] 143 0.000% 100.000% ( 2900000, 4300000 ] 60 0.000% 100.000% ( 4300000, 6500000 ] 27 0.000% 100.000% ( 6500000, 9800000 ] 7 0.000% 100.000% ( 9800000, 14000000 ] 1 0.000% 100.000% Gather stats latency (us): Count: 46 Average: 980387.5870 StdDev: 60911.18 Min: 879155 Median: 1033777.7778 Max: 1261431 Percentiles: P50: 1033777.78 P75: 1120666.67 P99: 1261431.00 P99.9: 1261431.00 P99.99: 1261431.00 ------------------------------------------------------ ( 860000, 1200000 ] 45 97.826% 97.826% #################### ( 1200000, 1900000 ] 1 2.174% 100.000% Most recent cache entry stats: Number of entries: 1295133 Total charge: 9.88 GB Average key size: 23.4982 Average charge: 8.00 KB Unique deleters: 3 ``` Reviewed By: mrambacher Differential Revision: D28295742 Pulled By: pdillinger fbshipit-source-id: bbc4a552f91ba0fe10e5cc025c42cef5a81f2b95
4 years ago
table_.ConstApplyToEntriesRange(
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
[callback](const HandleImpl& h) {
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
UniqueId64x2 unhashed;
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
callback(ReverseHash(h.hashed_key, &unhashed), h.value,
Major Cache refactoring, CPU efficiency improvement (#10975) Summary: This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache). The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below. * static_cast lines of code +29 -35 (net removed 6) * reinterpret_cast lines of code +6 -32 (net removed 26) ## cache.h and secondary_cache.h * Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications: * Simpler for implementations to deal with just one Insert and one Lookup. * Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters * Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428. * Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks). * It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below). * I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc. * Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation. * Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.) * Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.) * Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774) * Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object. * Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change. ## typed_cache.h Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae). The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used. * PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value. * BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter. * FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue. * For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`. These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.) Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it. ## block_cache.h This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table. ## block_based_table_reader.cc Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation. The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions. ## block_based_table_builder.cc, cache_dump_load_impl.cc Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.) ## Everything else Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975 Test Plan: tests updated Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache): 34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844 34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594 34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297 34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523 34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602 34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293 34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926 34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488 233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984 233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922 233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559 233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93 233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418 233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273 233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691 233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82 1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55 1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02 1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45 1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24 1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92 1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78 1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36 1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83 Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn. Reviewed By: anand1976 Differential Revision: D42417818 Pulled By: pdillinger fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
h.GetTotalCharge(), h.helper);
},
index_begin, index_end, false);
}
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
int HyperClockTable::CalcHashBits(
size_t capacity, size_t estimated_value_size,
CacheMetadataChargePolicy metadata_charge_policy) {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
double average_slot_charge = estimated_value_size * kLoadFactor;
if (metadata_charge_policy == kFullChargeCacheMetadata) {
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
average_slot_charge += sizeof(HandleImpl);
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
}
assert(average_slot_charge > 0.0);
uint64_t num_slots =
static_cast<uint64_t>(capacity / average_slot_charge + 0.999999);
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
int hash_bits = FloorLog2((num_slots << 1) - 1);
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
if (metadata_charge_policy == kFullChargeCacheMetadata) {
// For very small estimated value sizes, it's possible to overshoot
while (hash_bits > 0 &&
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
uint64_t{sizeof(HandleImpl)} << hash_bits > capacity) {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
hash_bits--;
}
}
return hash_bits;
}
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
template <class Table>
void ClockCacheShard<Table>::SetCapacity(size_t capacity) {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
capacity_.store(capacity, std::memory_order_relaxed);
// next Insert will take care of any necessary evictions
}
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
template <class Table>
void ClockCacheShard<Table>::SetStrictCapacityLimit(
bool strict_capacity_limit) {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
strict_capacity_limit_.store(strict_capacity_limit,
std::memory_order_relaxed);
// next Insert will take care of any necessary evictions
}
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
template <class Table>
Status ClockCacheShard<Table>::Insert(const Slice& key,
const UniqueId64x2& hashed_key,
Major Cache refactoring, CPU efficiency improvement (#10975) Summary: This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache). The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below. * static_cast lines of code +29 -35 (net removed 6) * reinterpret_cast lines of code +6 -32 (net removed 26) ## cache.h and secondary_cache.h * Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications: * Simpler for implementations to deal with just one Insert and one Lookup. * Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters * Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428. * Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks). * It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below). * I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc. * Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation. * Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.) * Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.) * Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774) * Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object. * Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change. ## typed_cache.h Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae). The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used. * PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value. * BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter. * FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue. * For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`. These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.) Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it. ## block_cache.h This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table. ## block_based_table_reader.cc Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation. The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions. ## block_based_table_builder.cc, cache_dump_load_impl.cc Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.) ## Everything else Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975 Test Plan: tests updated Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache): 34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844 34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594 34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297 34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523 34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602 34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293 34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926 34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488 233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984 233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922 233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559 233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93 233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418 233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273 233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691 233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82 1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55 1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02 1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45 1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24 1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92 1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78 1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36 1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83 Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn. Reviewed By: anand1976 Differential Revision: D42417818 Pulled By: pdillinger fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
Cache::ObjectPtr value,
const Cache::CacheItemHelper* helper,
size_t charge, HandleImpl** handle,
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
Cache::Priority priority) {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
if (UNLIKELY(key.size() != kCacheKeySize)) {
return Status::NotSupported("ClockCache only supports key size " +
std::to_string(kCacheKeySize) + "B");
}
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
ClockHandleBasicData proto;
proto.hashed_key = hashed_key;
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
proto.value = value;
Major Cache refactoring, CPU efficiency improvement (#10975) Summary: This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache). The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below. * static_cast lines of code +29 -35 (net removed 6) * reinterpret_cast lines of code +6 -32 (net removed 26) ## cache.h and secondary_cache.h * Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications: * Simpler for implementations to deal with just one Insert and one Lookup. * Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters * Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428. * Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks). * It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below). * I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc. * Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation. * Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.) * Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.) * Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774) * Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object. * Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change. ## typed_cache.h Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae). The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used. * PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value. * BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter. * FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue. * For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`. These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.) Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it. ## block_cache.h This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table. ## block_based_table_reader.cc Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation. The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions. ## block_based_table_builder.cc, cache_dump_load_impl.cc Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.) ## Everything else Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975 Test Plan: tests updated Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache): 34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844 34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594 34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297 34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523 34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602 34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293 34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926 34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488 233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984 233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922 233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559 233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93 233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418 233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273 233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691 233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82 1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55 1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02 1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45 1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24 1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92 1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78 1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36 1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83 Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn. Reviewed By: anand1976 Differential Revision: D42417818 Pulled By: pdillinger fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
proto.helper = helper;
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
proto.total_charge = charge;
HyperClockCache support for SecondaryCache, with refactoring (#11301) Summary: Internally refactors SecondaryCache integration out of LRUCache specifically and into a wrapper/adapter class that works with various Cache implementations. Notably, this relies on separating the notion of async lookup handles from other cache handles, so that HyperClockCache doesn't have to deal with the problem of allocating handles from the hash table for lookups that might fail anyway, and might be on the same key without support for coalescing. (LRUCache's hash table can incorporate previously allocated handles thanks to its pointer indirection.) Specifically, I'm worried about the case in which hundreds of threads try to access the same block and probing in the hash table degrades to linear search on the pile of entries with the same key. This change is a big step in the direction of supporting stacked SecondaryCaches, but there are obstacles to completing that. Especially, there is no SecondaryCache hook for evictions to pass from one to the next. It has been proposed that evictions be transmitted simply as the persisted data (as in SaveToCallback), but given the current structure provided by the CacheItemHelpers, that would require an extra copy of the block data, because there's intentionally no way to ask for a contiguous Slice of the data (to allow for flexibility in storage). `AsyncLookupHandle` and the re-worked `WaitAll()` should be essentially prepared for stacked SecondaryCaches, but several "TODO with stacked secondaries" issues remain in various places. It could be argued that the stacking instead be done as a SecondaryCache adapter that wraps two (or more) SecondaryCaches, but at least with the current API that would require an extra heap allocation on SecondaryCache Lookup for a wrapper SecondaryCacheResultHandle that can transfer a Lookup between secondaries. We could also consider trying to unify the Cache and SecondaryCache APIs, though that might be difficult if `AsyncLookupHandle` is kept a fixed struct. ## cache.h (public API) Moves `secondary_cache` option from LRUCacheOptions to ShardedCacheOptions so that it is applicable to HyperClockCache. ## advanced_cache.h (advanced public API) * Add `Cache::CreateStandalone()` so that the SecondaryCache support wrapper can use it. * Add `SetEvictionCallback()` / `eviction_callback_` so that the SecondaryCache support wrapper can use it. Only a single callback is supported for efficiency. If there is ever a need for more than one, hopefully that can be handled with a broadcast callback wrapper. These are essentially the two "extra" pieces of `Cache` for pulling out specific SecondaryCache support from the `Cache` implementation. I think it's a good trade-off as these are reasonable, limited, and reusable "cut points" into the `Cache` implementations. * Remove async capability from standard `Lookup()` (getting rid of awkward restrictions on pending Handles) and add `AsyncLookupHandle` and `StartAsyncLookup()`. As noted in the comments, the full struct of `AsyncLookupHandle` is exposed so that it can be stack allocated, for efficiency, though more data is being copied around than before, which could impact performance. (Lookup info -> AsyncLookupHandle -> Handle vs. Lookup info -> Handle) I could foresee a future in which a Cache internally saves a pointer to the AsyncLookupHandle, which means it's dangerous to allow it to be copyable or even movable. It also means it's not compatible with std::vector (which I don't like requiring as an API parameter anyway), so `WaitAll()` expects any contiguous array of AsyncLookupHandles. I believe this is best for common case efficiency, while behaving well in other cases also. For example, `WaitAll()` has no effect on default-constructed AsyncLookupHandles, which look like a completed cache miss. ## cacheable_entry.h A couple of functions are obsolete because Cache::Handle can no longer be pending. ## cache.cc Provides default implementations for new or revamped Cache functions, especially appropriate for non-blocking caches. ## secondary_cache_adapter.{h,cc} The full details of the Cache wrapper adding SecondaryCache support. Essentially replicates the SecondaryCache handling that was in LRUCache, but obviously refactored. There is a bit of logic duplication, where Lookup() is essentially a manually optimized version of StartAsyncLookup() and Wait(), but it's roughly a dozen lines of code. ## sharded_cache.h, typed_cache.h, charged_cache.{h,cc}, sim_cache.cc Simply updated for Cache API changes. ## lru_cache.{h,cc} Carefully remove SecondaryCache logic, implement `CreateStandalone` and eviction handler functionality. ## clock_cache.{h,cc} Expose existing `CreateStandalone` functionality, add eviction handler functionality. Light refactoring. ## block_based_table_reader* Mostly re-worked the only usage of async Lookup, which is in BlockBasedTable::MultiGet. Used arrays in place of autovector in some places for efficiency. Simplified some logic by not trying to process some cache results before they're all ready. Created new function `BlockBasedTable::GetCachePriority()` to reduce some pre-existing code duplication (and avoid making it worse). Fixed at least one small bug from the prior confusing mixture of async and sync Lookups. In MaybeReadBlockAndLoadToCache(), called by RetrieveBlock(), called by MultiGet() with wait=false, is_cache_hit for the block_cache_tracer entry would not be set to true if the handle was pending after Lookup and before Wait. ## Intended follow-up work * Figure out if there are any missing stats or block_cache_tracer work in refactored BlockBasedTable::MultiGet * Stacked secondary caches (see above discussion) * See if we can make up for the small MultiGet performance regression. * Study more performance with SecondaryCache * Items evicted from over-full LRUCache in Release were not being demoted to SecondaryCache, and still aren't to minimize unit test churn. Ideally they would be demoted, but it's an exceptional case so not a big deal. * Use CreateStandalone for cache reservations (save unnecessary hash table operations). Not a big deal, but worthy cleanup. * Somehow I got the contract for SecondaryCache::Insert wrong in #10945. (Doesn't take ownership!) That API comment needs to be fixed, but didn't want to mingle that in here. Pull Request resolved: https://github.com/facebook/rocksdb/pull/11301 Test Plan: ## Unit tests Generally updated to include HCC in SecondaryCache tests, though HyperClockCache has some different, less strict behaviors that leads to some tests not really being set up to work with it. Some of the tests remain disabled with it, but I think we have good coverage without them. ## Crash/stress test Updated to use the new combination. ## Performance First, let's check for regression on caches without secondary cache configured. Adding support for the eviction callback is likely to have a tiny effect, but it shouldn't be worrisome. LRUCache could benefit slightly from less logic around SecondaryCache handling. We can test with cache_bench default settings, built with DEBUG_LEVEL=0 and PORTABLE=0. ``` (while :; do base/cache_bench --cache_type=hyper_clock_cache | grep Rough; done) | awk '{ sum += $9; count++; print $0; print "Average: " int(sum / count) }' ``` **Before** this and #11299 (which could also have a small effect), running for about an hour, before & after running concurrently for each cache type: HyperClockCache: 3168662 (average parallel ops/sec) LRUCache: 2940127 **After** this and #11299, running for about an hour: HyperClockCache: 3164862 (average parallel ops/sec) (0.12% slower) LRUCache: 2940928 (0.03% faster) This is an acceptable difference IMHO. Next, let's consider essentially the worst case of new CPU overhead affecting overall performance. MultiGet uses the async lookup interface regardless of whether SecondaryCache or folly are used. We can configure a benchmark where all block cache queries are for data blocks, and all are hits. Create DB and test (before and after tests running simultaneously): ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 TEST_TMPDIR=/dev/shm base/db_bench -benchmarks=multireadrandom[-X30] -readonly -multiread_batched -batch_size=32 -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16 ``` **Before**: multireadrandom [AVG 30 runs] : 3444202 (± 57049) ops/sec; 240.9 (± 4.0) MB/sec multireadrandom [MEDIAN 30 runs] : 3514443 ops/sec; 245.8 MB/sec **After**: multireadrandom [AVG 30 runs] : 3291022 (± 58851) ops/sec; 230.2 (± 4.1) MB/sec multireadrandom [MEDIAN 30 runs] : 3366179 ops/sec; 235.4 MB/sec So that's roughly a 3% regression, on kind of a *worst case* test of MultiGet CPU. Similar story with HyperClockCache: **Before**: multireadrandom [AVG 30 runs] : 3933777 (± 41840) ops/sec; 275.1 (± 2.9) MB/sec multireadrandom [MEDIAN 30 runs] : 3970667 ops/sec; 277.7 MB/sec **After**: multireadrandom [AVG 30 runs] : 3755338 (± 30391) ops/sec; 262.6 (± 2.1) MB/sec multireadrandom [MEDIAN 30 runs] : 3785696 ops/sec; 264.8 MB/sec Roughly a 4-5% regression. Not ideal, but not the whole story, fortunately. Let's also look at Get() in db_bench: ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X30] -readonly -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16 ``` **Before**: readrandom [AVG 30 runs] : 2198685 (± 13412) ops/sec; 153.8 (± 0.9) MB/sec readrandom [MEDIAN 30 runs] : 2209498 ops/sec; 154.5 MB/sec **After**: readrandom [AVG 30 runs] : 2292814 (± 43508) ops/sec; 160.3 (± 3.0) MB/sec readrandom [MEDIAN 30 runs] : 2365181 ops/sec; 165.4 MB/sec That's showing roughly a 4% improvement, perhaps because of the secondary cache code that is no longer part of LRUCache. But weirdly, HyperClockCache is also showing 2-3% improvement: **Before**: readrandom [AVG 30 runs] : 2272333 (± 9992) ops/sec; 158.9 (± 0.7) MB/sec readrandom [MEDIAN 30 runs] : 2273239 ops/sec; 159.0 MB/sec **After**: readrandom [AVG 30 runs] : 2332407 (± 11252) ops/sec; 163.1 (± 0.8) MB/sec readrandom [MEDIAN 30 runs] : 2335329 ops/sec; 163.3 MB/sec Reviewed By: ltamasi Differential Revision: D44177044 Pulled By: pdillinger fbshipit-source-id: e808e48ff3fe2f792a79841ba617be98e48689f5
2 years ago
return table_.Insert(proto, handle, priority,
capacity_.load(std::memory_order_relaxed),
strict_capacity_limit_.load(std::memory_order_relaxed));
}
template <class Table>
typename ClockCacheShard<Table>::HandleImpl*
ClockCacheShard<Table>::CreateStandalone(const Slice& key,
const UniqueId64x2& hashed_key,
Cache::ObjectPtr obj,
const Cache::CacheItemHelper* helper,
size_t charge, bool allow_uncharged) {
if (UNLIKELY(key.size() != kCacheKeySize)) {
return nullptr;
}
ClockHandleBasicData proto;
proto.hashed_key = hashed_key;
proto.value = obj;
proto.helper = helper;
proto.total_charge = charge;
return table_.CreateStandalone(
proto, capacity_.load(std::memory_order_relaxed),
strict_capacity_limit_.load(std::memory_order_relaxed), allow_uncharged);
}
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
template <class Table>
typename ClockCacheShard<Table>::HandleImpl* ClockCacheShard<Table>::Lookup(
const Slice& key, const UniqueId64x2& hashed_key) {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
if (UNLIKELY(key.size() != kCacheKeySize)) {
return nullptr;
}
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
return table_.Lookup(hashed_key);
}
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
template <class Table>
bool ClockCacheShard<Table>::Ref(HandleImpl* h) {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
if (h == nullptr) {
return false;
}
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
table_.Ref(*h);
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
return true;
}
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
template <class Table>
bool ClockCacheShard<Table>::Release(HandleImpl* handle, bool useful,
bool erase_if_last_ref) {
if (handle == nullptr) {
return false;
}
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
return table_.Release(handle, useful, erase_if_last_ref);
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
}
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
template <class Table>
void ClockCacheShard<Table>::TEST_RefN(HandleImpl* h, size_t n) {
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
table_.TEST_RefN(*h, n);
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
}
Towards a production-quality ClockCache (#10418) Summary: In this PR we bring ClockCache closer to production quality. We implement the following changes: 1. Fixed a few bugs in ClockCache. 2. ClockCache now fully supports ``strict_capacity_limit == false``: When an insertion over capacity is commanded, we allocate a handle separately from the hash table. 3. ClockCache now runs on almost every test in cache_test. The only exceptions are a test where either the LRU policy is required, and a test that dynamically increases the table capacity. 4. ClockCache now supports dynamically decreasing capacity via SetCapacity. (This is easy: we shrink the capacity upper bound and run the clock algorithm.) 5. Old FastLRUCache tests in lru_cache_test.cc are now also used on ClockCache. As a byproduct of 1. and 2. we are able to turn on ClockCache in the stress tests. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10418 Test Plan: - ``make -j24 USE_CLANG=1 COMPILE_WITH_ASAN=1 COMPILE_WITH_UBSAN=1 check`` - ``make -j24 USE_CLANG=1 COMPILE_WITH_TSAN=1 check`` - ``make -j24 USE_CLANG=1 COMPILE_WITH_ASAN=1 COMPILE_WITH_UBSAN=1 CRASH_TEST_EXT_ARGS="--duration=960 --cache_type=clock_cache" blackbox_crash_test_with_atomic_flush`` - ``make -j24 USE_CLANG=1 COMPILE_WITH_TSAN=1 CRASH_TEST_EXT_ARGS="--duration=960 --cache_type=clock_cache" blackbox_crash_test_with_atomic_flush`` Reviewed By: pdillinger Differential Revision: D38170673 Pulled By: guidotag fbshipit-source-id: 508987b9dc9d9d68f1a03eefac769820b680340a
2 years ago
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
template <class Table>
void ClockCacheShard<Table>::TEST_ReleaseN(HandleImpl* h, size_t n) {
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
table_.TEST_ReleaseN(h, n);
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
}
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
template <class Table>
bool ClockCacheShard<Table>::Release(HandleImpl* handle,
bool erase_if_last_ref) {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
return Release(handle, /*useful=*/true, erase_if_last_ref);
}
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
template <class Table>
void ClockCacheShard<Table>::Erase(const Slice& key,
const UniqueId64x2& hashed_key) {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
if (UNLIKELY(key.size() != kCacheKeySize)) {
return;
}
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
table_.Erase(hashed_key);
}
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
template <class Table>
size_t ClockCacheShard<Table>::GetUsage() const {
return table_.GetUsage();
}
Observe and warn about misconfigured HyperClockCache (#10965) Summary: Background. One of the core risks of chosing HyperClockCache is ending up with degraded performance if estimated_entry_charge is very significantly wrong. Too low leads to under-utilized hash table, which wastes a bit of (tracked) memory and likely increases access times due to larger working set size (more TLB misses). Too high leads to fully populated hash table (at some limit with reasonable lookup performance) and not being able to cache as many objects as the memory limit would allow. In either case, performance degradation is graceful/continuous but can be quite significant. For example, cutting block size in half without updating estimated_entry_charge could lead to a large portion of configured block cache memory (up to roughly 1/3) going unused. Fix. This change adds a mechanism through which the DB periodically probes the block cache(s) for "problems" to report, and adds diagnostics to the HyperClockCache for bad estimated_entry_charge. The periodic probing is currently done with DumpStats / stats_dump_period_sec, and diagnostics reported to info_log (normally LOG file). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10965 Test Plan: unit test included. Doesn't cover all the implemented subtleties of reporting, but ensures basics of when to report or not. Also manual testing with db_bench. Create db with ``` ./db_bench --benchmarks=fillrandom,flush --num=3000000 --disable_wal=1 ``` Use and check LOG file for HyperClockCache for various block sizes (used as estimated_entry_charge) ``` ./db_bench --use_existing_db --benchmarks=readrandom --num=3000000 --duration=20 --stats_dump_period_sec=8 --cache_type=hyper_clock_cache -block_size=XXXX ``` Seeing warnings / errors or not as expected. Reviewed By: anand1976 Differential Revision: D41406932 Pulled By: pdillinger fbshipit-source-id: 4ca56162b73017e4b9cec2cad74466f49c27a0a7
2 years ago
template <class Table>
size_t ClockCacheShard<Table>::GetStandaloneUsage() const {
return table_.GetStandaloneUsage();
Observe and warn about misconfigured HyperClockCache (#10965) Summary: Background. One of the core risks of chosing HyperClockCache is ending up with degraded performance if estimated_entry_charge is very significantly wrong. Too low leads to under-utilized hash table, which wastes a bit of (tracked) memory and likely increases access times due to larger working set size (more TLB misses). Too high leads to fully populated hash table (at some limit with reasonable lookup performance) and not being able to cache as many objects as the memory limit would allow. In either case, performance degradation is graceful/continuous but can be quite significant. For example, cutting block size in half without updating estimated_entry_charge could lead to a large portion of configured block cache memory (up to roughly 1/3) going unused. Fix. This change adds a mechanism through which the DB periodically probes the block cache(s) for "problems" to report, and adds diagnostics to the HyperClockCache for bad estimated_entry_charge. The periodic probing is currently done with DumpStats / stats_dump_period_sec, and diagnostics reported to info_log (normally LOG file). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10965 Test Plan: unit test included. Doesn't cover all the implemented subtleties of reporting, but ensures basics of when to report or not. Also manual testing with db_bench. Create db with ``` ./db_bench --benchmarks=fillrandom,flush --num=3000000 --disable_wal=1 ``` Use and check LOG file for HyperClockCache for various block sizes (used as estimated_entry_charge) ``` ./db_bench --use_existing_db --benchmarks=readrandom --num=3000000 --duration=20 --stats_dump_period_sec=8 --cache_type=hyper_clock_cache -block_size=XXXX ``` Seeing warnings / errors or not as expected. Reviewed By: anand1976 Differential Revision: D41406932 Pulled By: pdillinger fbshipit-source-id: 4ca56162b73017e4b9cec2cad74466f49c27a0a7
2 years ago
}
template <class Table>
size_t ClockCacheShard<Table>::GetCapacity() const {
return capacity_;
}
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
template <class Table>
size_t ClockCacheShard<Table>::GetPinnedUsage() const {
// Computes the pinned usage by scanning the whole hash table. This
// is slow, but avoids keeping an exact counter on the clock usage,
// i.e., the number of not externally referenced elements.
// Why avoid this counter? Because Lookup removes elements from the clock
// list, so it would need to update the pinned usage every time,
// which creates additional synchronization costs.
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
size_t table_pinned_usage = 0;
const bool charge_metadata =
metadata_charge_policy_ == kFullChargeCacheMetadata;
table_.ConstApplyToEntriesRange(
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
[&table_pinned_usage, charge_metadata](const HandleImpl& h) {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
uint64_t meta = h.meta.load(std::memory_order_relaxed);
uint64_t refcount = GetRefcount(meta);
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
// Holding one ref for ConstApplyToEntriesRange
assert(refcount > 0);
if (refcount > 1) {
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
table_pinned_usage += h.GetTotalCharge();
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
if (charge_metadata) {
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
table_pinned_usage += sizeof(HandleImpl);
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
}
}
},
0, table_.GetTableSize(), true);
return table_pinned_usage + table_.GetStandaloneUsage();
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
}
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
template <class Table>
size_t ClockCacheShard<Table>::GetOccupancyCount() const {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
return table_.GetOccupancy();
}
Observe and warn about misconfigured HyperClockCache (#10965) Summary: Background. One of the core risks of chosing HyperClockCache is ending up with degraded performance if estimated_entry_charge is very significantly wrong. Too low leads to under-utilized hash table, which wastes a bit of (tracked) memory and likely increases access times due to larger working set size (more TLB misses). Too high leads to fully populated hash table (at some limit with reasonable lookup performance) and not being able to cache as many objects as the memory limit would allow. In either case, performance degradation is graceful/continuous but can be quite significant. For example, cutting block size in half without updating estimated_entry_charge could lead to a large portion of configured block cache memory (up to roughly 1/3) going unused. Fix. This change adds a mechanism through which the DB periodically probes the block cache(s) for "problems" to report, and adds diagnostics to the HyperClockCache for bad estimated_entry_charge. The periodic probing is currently done with DumpStats / stats_dump_period_sec, and diagnostics reported to info_log (normally LOG file). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10965 Test Plan: unit test included. Doesn't cover all the implemented subtleties of reporting, but ensures basics of when to report or not. Also manual testing with db_bench. Create db with ``` ./db_bench --benchmarks=fillrandom,flush --num=3000000 --disable_wal=1 ``` Use and check LOG file for HyperClockCache for various block sizes (used as estimated_entry_charge) ``` ./db_bench --use_existing_db --benchmarks=readrandom --num=3000000 --duration=20 --stats_dump_period_sec=8 --cache_type=hyper_clock_cache -block_size=XXXX ``` Seeing warnings / errors or not as expected. Reviewed By: anand1976 Differential Revision: D41406932 Pulled By: pdillinger fbshipit-source-id: 4ca56162b73017e4b9cec2cad74466f49c27a0a7
2 years ago
template <class Table>
size_t ClockCacheShard<Table>::GetOccupancyLimit() const {
return table_.GetOccupancyLimit();
}
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
template <class Table>
size_t ClockCacheShard<Table>::GetTableAddressCount() const {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
return table_.GetTableSize();
}
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
// Explicit instantiation
template class ClockCacheShard<HyperClockTable>;
HyperClockCache::HyperClockCache(const HyperClockCacheOptions& opts)
: ShardedCache(opts) {
assert(opts.estimated_entry_charge > 0 ||
opts.metadata_charge_policy != kDontChargeCacheMetadata);
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
// TODO: should not need to go through two levels of pointer indirection to
// get to table entries
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
size_t per_shard = GetPerShardCapacity();
Major Cache refactoring, CPU efficiency improvement (#10975) Summary: This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache). The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below. * static_cast lines of code +29 -35 (net removed 6) * reinterpret_cast lines of code +6 -32 (net removed 26) ## cache.h and secondary_cache.h * Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications: * Simpler for implementations to deal with just one Insert and one Lookup. * Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters * Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428. * Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks). * It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below). * I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc. * Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation. * Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.) * Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.) * Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774) * Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object. * Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change. ## typed_cache.h Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae). The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used. * PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value. * BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter. * FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue. * For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`. These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.) Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it. ## block_cache.h This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table. ## block_based_table_reader.cc Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation. The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions. ## block_based_table_builder.cc, cache_dump_load_impl.cc Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.) ## Everything else Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975 Test Plan: tests updated Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache): 34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844 34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594 34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297 34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523 34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602 34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293 34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926 34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488 233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984 233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922 233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559 233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93 233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418 233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273 233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691 233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82 1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55 1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02 1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45 1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24 1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92 1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78 1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36 1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83 Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn. Reviewed By: anand1976 Differential Revision: D42417818 Pulled By: pdillinger fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
MemoryAllocator* alloc = this->memory_allocator();
InitShards([&](Shard* cs) {
HyperClockTable::Opts table_opts;
table_opts.estimated_value_size = opts.estimated_entry_charge;
new (cs) Shard(per_shard, opts.strict_capacity_limit,
opts.metadata_charge_policy, alloc, &eviction_callback_,
table_opts);
Refactor ShardedCache for more sharing, static polymorphism (#10801) Summary: The motivations for this change include * Free up space in ClockHandle so that we can add data for secondary cache handling while still keeping within single cache line (64 byte) size. * This change frees up space by eliminating the need for the `hash` field by making the fixed-size key itself a hash, using a 128-bit bijective (lossless) hash. * Generally more customizability of ShardedCache (such as hashing) without worrying about virtual call overheads * ShardedCache now uses static polymorphism (template) instead of dynamic polymorphism (virtual overrides) for the CacheShard. No obvious performance benefit is seen from the change (as mostly expected; most calls to virtual functions in CacheShard could already be optimized to static calls), but offers more flexibility without incurring the runtime cost of adhering to a common interface (without type parameters or static callbacks). * You'll also notice less `reinterpret_cast`ing and other boilerplate in the Cache implementations, as this can go in ShardedCache. More detail: * Don't have LRUCacheShard maintain `std::shared_ptr<SecondaryCache>` copies (extra refcount) when LRUCache can be in charge of keeping a `shared_ptr`. * Renamed `capacity_mutex_` to `config_mutex_` to better represent the scope of what it guards. * Some preparation for 64-bit hash and indexing in LRUCache, but didn't include the full change because of slight performance regression. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10801 Test Plan: Unit test updates were non-trivial because of major changes to the ClockCacheShard interface in handling of key vs. hash. Performance: Create with `TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X1000] -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=16 ``` Before: `readrandom [AVG 150 runs] : 321147 (± 253) ops/sec` After: `readrandom [AVG 150 runs] : 321530 (± 326) ops/sec` So possibly ~0.1% improvement. And with `-cache_type=hyper_clock_cache`: Before: `readrandom [AVG 30 runs] : 614126 (± 7978) ops/sec` After: `readrandom [AVG 30 runs] : 645349 (± 8087) ops/sec` So roughly 5% improvement! Reviewed By: anand1976 Differential Revision: D40252236 Pulled By: pdillinger fbshipit-source-id: ff8fc70ef569585edc95bcbaaa0386f61355ae5b
2 years ago
});
}
Major Cache refactoring, CPU efficiency improvement (#10975) Summary: This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache). The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below. * static_cast lines of code +29 -35 (net removed 6) * reinterpret_cast lines of code +6 -32 (net removed 26) ## cache.h and secondary_cache.h * Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications: * Simpler for implementations to deal with just one Insert and one Lookup. * Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters * Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428. * Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks). * It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below). * I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc. * Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation. * Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.) * Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.) * Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774) * Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object. * Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change. ## typed_cache.h Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae). The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used. * PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value. * BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter. * FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue. * For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`. These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.) Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it. ## block_cache.h This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table. ## block_based_table_reader.cc Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation. The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions. ## block_based_table_builder.cc, cache_dump_load_impl.cc Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.) ## Everything else Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975 Test Plan: tests updated Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache): 34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844 34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594 34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297 34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523 34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602 34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293 34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926 34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488 233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984 233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922 233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559 233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93 233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418 233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273 233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691 233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82 1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55 1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02 1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45 1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24 1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92 1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78 1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36 1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83 Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn. Reviewed By: anand1976 Differential Revision: D42417818 Pulled By: pdillinger fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
Cache::ObjectPtr HyperClockCache::Value(Handle* handle) {
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
return reinterpret_cast<const HandleImpl*>(handle)->value;
}
Use deleters to label cache entries and collect stats (#8297) Summary: This change gathers and publishes statistics about the kinds of items in block cache. This is especially important for profiling relative usage of cache by index vs. filter vs. data blocks. It works by iterating over the cache during periodic stats dump (InternalStats, stats_dump_period_sec) or on demand when DB::Get(Map)Property(kBlockCacheEntryStats), except that for efficiency and sharing among column families, saved data from the last scan is used when the data is not considered too old. The new information can be seen in info LOG, for example: Block cache LRUCache@0x7fca62229330 capacity: 95.37 MB collections: 8 last_copies: 0 last_secs: 0.00178 secs_since: 0 Block cache entry stats(count,size,portion): DataBlock(7092,28.24 MB,29.6136%) FilterBlock(215,867.90 KB,0.888728%) FilterMetaBlock(2,5.31 KB,0.00544%) IndexBlock(217,180.11 KB,0.184432%) WriteBuffer(1,256.00 KB,0.262144%) Misc(1,0.00 KB,0%) And also through DB::GetProperty and GetMapProperty (here using ldb just for demonstration): $ ./ldb --db=/dev/shm/dbbench/ get_property rocksdb.block-cache-entry-stats rocksdb.block-cache-entry-stats.bytes.data-block: 0 rocksdb.block-cache-entry-stats.bytes.deprecated-filter-block: 0 rocksdb.block-cache-entry-stats.bytes.filter-block: 0 rocksdb.block-cache-entry-stats.bytes.filter-meta-block: 0 rocksdb.block-cache-entry-stats.bytes.index-block: 178992 rocksdb.block-cache-entry-stats.bytes.misc: 0 rocksdb.block-cache-entry-stats.bytes.other-block: 0 rocksdb.block-cache-entry-stats.bytes.write-buffer: 0 rocksdb.block-cache-entry-stats.capacity: 8388608 rocksdb.block-cache-entry-stats.count.data-block: 0 rocksdb.block-cache-entry-stats.count.deprecated-filter-block: 0 rocksdb.block-cache-entry-stats.count.filter-block: 0 rocksdb.block-cache-entry-stats.count.filter-meta-block: 0 rocksdb.block-cache-entry-stats.count.index-block: 215 rocksdb.block-cache-entry-stats.count.misc: 1 rocksdb.block-cache-entry-stats.count.other-block: 0 rocksdb.block-cache-entry-stats.count.write-buffer: 0 rocksdb.block-cache-entry-stats.id: LRUCache@0x7f3636661290 rocksdb.block-cache-entry-stats.percent.data-block: 0.000000 rocksdb.block-cache-entry-stats.percent.deprecated-filter-block: 0.000000 rocksdb.block-cache-entry-stats.percent.filter-block: 0.000000 rocksdb.block-cache-entry-stats.percent.filter-meta-block: 0.000000 rocksdb.block-cache-entry-stats.percent.index-block: 2.133751 rocksdb.block-cache-entry-stats.percent.misc: 0.000000 rocksdb.block-cache-entry-stats.percent.other-block: 0.000000 rocksdb.block-cache-entry-stats.percent.write-buffer: 0.000000 rocksdb.block-cache-entry-stats.secs_for_last_collection: 0.000052 rocksdb.block-cache-entry-stats.secs_since_last_collection: 0 Solution detail - We need some way to flag what kind of blocks each entry belongs to, preferably without changing the Cache API. One of the complications is that Cache is a general interface that could have other users that don't adhere to whichever convention we decide on for keys and values. Or we would pay for an extra field in the Handle that would only be used for this purpose. This change uses a back-door approach, the deleter, to indicate the "role" of a Cache entry (in addition to the value type, implicitly). This has the added benefit of ensuring proper code origin whenever we recognize a particular role for a cache entry; if the entry came from some other part of the code, it will use an unrecognized deleter, which we simply attribute to the "Misc" role. An internal API makes for simple instantiation and automatic registration of Cache deleters for a given value type and "role". Another internal API, CacheEntryStatsCollector, solves the problem of caching the results of a scan and sharing them, to ensure scans are neither excessive nor redundant so as not to harm Cache performance. Because code is added to BlocklikeTraits, it is pulled out of block_based_table_reader.cc into its own file. This is a reformulation of https://github.com/facebook/rocksdb/issues/8276, without the type checking option (could still be added), and with actual stat gathering. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8297 Test Plan: manual testing with db_bench, and a couple of basic unit tests Reviewed By: ltamasi Differential Revision: D28488721 Pulled By: pdillinger fbshipit-source-id: 472f524a9691b5afb107934be2d41d84f2b129fb
4 years ago
Call experimental new clock cache HyperClockCache (#10684) Summary: This change establishes a distinctive name for the experimental new lock-free clock cache (originally developed by guidotag and revamped in PR https://github.com/facebook/rocksdb/issues/10626). A few reasons: * We want to make it clear that this is a fundamentally different implementation vs. the old clock cache, to avoid people saying "I already tried clock cache." * We want to highlight the key feature: it's fast (especially under parallel load) * Because it requires an estimated charge per entry, it is not drop-in API compatible with old clock cache. This estimate might always be required for highest performance, and giving it a distinct name should reduce confusion about the distinct API requirements. * We might develop a variant requiring the same estimate parameter but with LRU eviction. In that case, using the name HyperLRUCache should make things more clear. (FastLRUCache is just a prototype that might soon be removed.) Some API detail: * To reduce copy-pasting parameter lists, etc. as in LRUCache construction, I have a `MakeSharedCache()` function on `HyperClockCacheOptions` instead of `NewHyperClockCache()`. * Changes -cache_type=clock_cache to -cache_type=hyper_clock_cache for applicable tools. I think this is more consistent / sustainable for reasons already stated. For performance tests see https://github.com/facebook/rocksdb/pull/10626 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10684 Test Plan: no interesting functional changes; tests updated Reviewed By: anand1976 Differential Revision: D39547800 Pulled By: pdillinger fbshipit-source-id: 5c0fe1b5cf3cb680ab369b928c8569682b9795bf
2 years ago
size_t HyperClockCache::GetCharge(Handle* handle) const {
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
return reinterpret_cast<const HandleImpl*>(handle)->GetTotalCharge();
}
Major Cache refactoring, CPU efficiency improvement (#10975) Summary: This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache). The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below. * static_cast lines of code +29 -35 (net removed 6) * reinterpret_cast lines of code +6 -32 (net removed 26) ## cache.h and secondary_cache.h * Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications: * Simpler for implementations to deal with just one Insert and one Lookup. * Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters * Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428. * Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks). * It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below). * I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc. * Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation. * Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.) * Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.) * Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774) * Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object. * Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change. ## typed_cache.h Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae). The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used. * PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value. * BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter. * FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue. * For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`. These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.) Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it. ## block_cache.h This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table. ## block_based_table_reader.cc Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation. The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions. ## block_based_table_builder.cc, cache_dump_load_impl.cc Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.) ## Everything else Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975 Test Plan: tests updated Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache): 34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844 34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594 34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297 34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523 34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602 34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293 34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926 34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488 233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984 233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922 233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559 233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93 233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418 233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273 233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691 233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82 1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55 1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02 1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45 1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24 1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92 1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78 1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36 1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83 Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn. Reviewed By: anand1976 Differential Revision: D42417818 Pulled By: pdillinger fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
const Cache::CacheItemHelper* HyperClockCache::GetCacheItemHelper(
Handle* handle) const {
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
auto h = reinterpret_cast<const HandleImpl*>(handle);
Major Cache refactoring, CPU efficiency improvement (#10975) Summary: This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache). The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below. * static_cast lines of code +29 -35 (net removed 6) * reinterpret_cast lines of code +6 -32 (net removed 26) ## cache.h and secondary_cache.h * Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications: * Simpler for implementations to deal with just one Insert and one Lookup. * Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters * Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428. * Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks). * It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below). * I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc. * Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation. * Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.) * Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.) * Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774) * Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object. * Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change. ## typed_cache.h Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae). The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used. * PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value. * BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter. * FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue. * For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`. These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.) Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it. ## block_cache.h This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table. ## block_based_table_reader.cc Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation. The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions. ## block_based_table_builder.cc, cache_dump_load_impl.cc Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.) ## Everything else Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975 Test Plan: tests updated Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache): 34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844 34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594 34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297 34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523 34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602 34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293 34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926 34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488 233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984 233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922 233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559 233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93 233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418 233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273 233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691 233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82 1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55 1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02 1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45 1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24 1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92 1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78 1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36 1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83 Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn. Reviewed By: anand1976 Differential Revision: D42417818 Pulled By: pdillinger fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
return h->helper;
}
Initial support for secondary cache in LRUCache (#8271) Summary: Defined the abstract interface for a secondary cache in include/rocksdb/secondary_cache.h, and updated LRUCacheOptions to take a std::shared_ptr<SecondaryCache>. An item is initially inserted into the LRU (primary) cache. When it ages out and evicted from memory, its inserted into the secondary cache. On a LRU cache miss and successful lookup in the secondary cache, the item is promoted to the LRU cache. Only support synchronous lookup currently. The secondary cache would be used to implement a persistent (flash cache) or compressed cache. Tests: Results from cache_bench and db_bench don't show any regression due to these changes. cache_bench results before and after this change - Command ```./cache_bench -ops_per_thread=10000000 -threads=1``` Before ```Complete in 40.688 s; QPS = 245774``` ```Complete in 40.486 s; QPS = 246996``` ```Complete in 42.019 s; QPS = 237989``` After ```Complete in 40.672 s; QPS = 245869``` ```Complete in 44.622 s; QPS = 224107``` ```Complete in 42.445 s; QPS = 235599``` db_bench results before this change, and with this change + https://github.com/facebook/rocksdb/issues/8213 and https://github.com/facebook/rocksdb/issues/8191 - Commands ```./db_bench --benchmarks="fillseq,compact" -num=30000000 -key_size=32 -value_size=256 -use_direct_io_for_flush_and_compaction=true -db=/home/anand76/nvm_cache/db -partition_index_and_filters=true``` ```./db_bench -db=/home/anand76/nvm_cache/db -use_existing_db=true -benchmarks=readrandom -num=30000000 -key_size=32 -value_size=256 -use_direct_reads=true -cache_size=1073741824 -cache_numshardbits=6 -cache_index_and_filter_blocks=true -read_random_exp_range=17 -statistics -partition_index_and_filters=true -threads=16 -duration=300``` Before ``` DB path: [/home/anand76/nvm_cache/db] readrandom : 80.702 micros/op 198104 ops/sec; 54.4 MB/s (3708999 of 3708999 found) ``` ``` DB path: [/home/anand76/nvm_cache/db] readrandom : 87.124 micros/op 183625 ops/sec; 50.4 MB/s (3439999 of 3439999 found) ``` After ``` DB path: [/home/anand76/nvm_cache/db] readrandom : 77.653 micros/op 206025 ops/sec; 56.6 MB/s (3866999 of 3866999 found) ``` ``` DB path: [/home/anand76/nvm_cache/db] readrandom : 84.962 micros/op 188299 ops/sec; 51.7 MB/s (3535999 of 3535999 found) ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/8271 Reviewed By: zhichao-cao Differential Revision: D28357511 Pulled By: anand1976 fbshipit-source-id: d1cfa236f00e649a18c53328be10a8062a4b6da2
4 years ago
Observe and warn about misconfigured HyperClockCache (#10965) Summary: Background. One of the core risks of chosing HyperClockCache is ending up with degraded performance if estimated_entry_charge is very significantly wrong. Too low leads to under-utilized hash table, which wastes a bit of (tracked) memory and likely increases access times due to larger working set size (more TLB misses). Too high leads to fully populated hash table (at some limit with reasonable lookup performance) and not being able to cache as many objects as the memory limit would allow. In either case, performance degradation is graceful/continuous but can be quite significant. For example, cutting block size in half without updating estimated_entry_charge could lead to a large portion of configured block cache memory (up to roughly 1/3) going unused. Fix. This change adds a mechanism through which the DB periodically probes the block cache(s) for "problems" to report, and adds diagnostics to the HyperClockCache for bad estimated_entry_charge. The periodic probing is currently done with DumpStats / stats_dump_period_sec, and diagnostics reported to info_log (normally LOG file). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10965 Test Plan: unit test included. Doesn't cover all the implemented subtleties of reporting, but ensures basics of when to report or not. Also manual testing with db_bench. Create db with ``` ./db_bench --benchmarks=fillrandom,flush --num=3000000 --disable_wal=1 ``` Use and check LOG file for HyperClockCache for various block sizes (used as estimated_entry_charge) ``` ./db_bench --use_existing_db --benchmarks=readrandom --num=3000000 --duration=20 --stats_dump_period_sec=8 --cache_type=hyper_clock_cache -block_size=XXXX ``` Seeing warnings / errors or not as expected. Reviewed By: anand1976 Differential Revision: D41406932 Pulled By: pdillinger fbshipit-source-id: 4ca56162b73017e4b9cec2cad74466f49c27a0a7
2 years ago
namespace {
// For each cache shard, estimate what the table load factor would be if
// cache filled to capacity with average entries. This is considered
// indicative of a potential problem if the shard is essentially operating
// "at limit", which we define as high actual usage (>80% of capacity)
// or actual occupancy very close to limit (>95% of limit).
// Also, for each shard compute the recommended estimated_entry_charge,
// and keep the minimum one for use as overall recommendation.
void AddShardEvaluation(const HyperClockCache::Shard& shard,
std::vector<double>& predicted_load_factors,
size_t& min_recommendation) {
size_t usage = shard.GetUsage() - shard.GetStandaloneUsage();
Observe and warn about misconfigured HyperClockCache (#10965) Summary: Background. One of the core risks of chosing HyperClockCache is ending up with degraded performance if estimated_entry_charge is very significantly wrong. Too low leads to under-utilized hash table, which wastes a bit of (tracked) memory and likely increases access times due to larger working set size (more TLB misses). Too high leads to fully populated hash table (at some limit with reasonable lookup performance) and not being able to cache as many objects as the memory limit would allow. In either case, performance degradation is graceful/continuous but can be quite significant. For example, cutting block size in half without updating estimated_entry_charge could lead to a large portion of configured block cache memory (up to roughly 1/3) going unused. Fix. This change adds a mechanism through which the DB periodically probes the block cache(s) for "problems" to report, and adds diagnostics to the HyperClockCache for bad estimated_entry_charge. The periodic probing is currently done with DumpStats / stats_dump_period_sec, and diagnostics reported to info_log (normally LOG file). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10965 Test Plan: unit test included. Doesn't cover all the implemented subtleties of reporting, but ensures basics of when to report or not. Also manual testing with db_bench. Create db with ``` ./db_bench --benchmarks=fillrandom,flush --num=3000000 --disable_wal=1 ``` Use and check LOG file for HyperClockCache for various block sizes (used as estimated_entry_charge) ``` ./db_bench --use_existing_db --benchmarks=readrandom --num=3000000 --duration=20 --stats_dump_period_sec=8 --cache_type=hyper_clock_cache -block_size=XXXX ``` Seeing warnings / errors or not as expected. Reviewed By: anand1976 Differential Revision: D41406932 Pulled By: pdillinger fbshipit-source-id: 4ca56162b73017e4b9cec2cad74466f49c27a0a7
2 years ago
size_t capacity = shard.GetCapacity();
double usage_ratio = 1.0 * usage / capacity;
size_t occupancy = shard.GetOccupancyCount();
size_t occ_limit = shard.GetOccupancyLimit();
double occ_ratio = 1.0 * occupancy / occ_limit;
if (usage == 0 || occupancy == 0 || (usage_ratio < 0.8 && occ_ratio < 0.95)) {
// Skip as described above
return;
}
// If filled to capacity, what would the occupancy ratio be?
double ratio = occ_ratio / usage_ratio;
// Given max load factor, what that load factor be?
double lf = ratio * kStrictLoadFactor;
predicted_load_factors.push_back(lf);
// Update min_recommendation also
size_t recommendation = usage / occupancy;
min_recommendation = std::min(min_recommendation, recommendation);
}
} // namespace
void HyperClockCache::ReportProblems(
const std::shared_ptr<Logger>& info_log) const {
uint32_t shard_count = GetNumShards();
std::vector<double> predicted_load_factors;
size_t min_recommendation = SIZE_MAX;
const_cast<HyperClockCache*>(this)->ForEachShard(
[&](HyperClockCache::Shard* shard) {
AddShardEvaluation(*shard, predicted_load_factors, min_recommendation);
});
if (predicted_load_factors.empty()) {
// None operating "at limit" -> nothing to report
return;
}
std::sort(predicted_load_factors.begin(), predicted_load_factors.end());
// First, if the average load factor is within spec, we aren't going to
// complain about a few shards being out of spec.
// NOTE: this is only the average among cache shards operating "at limit,"
// which should be representative of what we care about. It it normal, even
// desirable, for a cache to operate "at limit" so this should not create
// selection bias. See AddShardEvaluation().
// TODO: Consider detecting cases where decreasing the number of shards
// would be good, e.g. serious imbalance among shards.
double average_load_factor =
std::accumulate(predicted_load_factors.begin(),
predicted_load_factors.end(), 0.0) /
shard_count;
constexpr double kLowSpecLoadFactor = kLoadFactor / 2;
constexpr double kMidSpecLoadFactor = kLoadFactor / 1.414;
if (average_load_factor > kLoadFactor) {
// Out of spec => Consider reporting load factor too high
// Estimate effective overall capacity loss due to enforcing occupancy limit
double lost_portion = 0.0;
int over_count = 0;
for (double lf : predicted_load_factors) {
if (lf > kStrictLoadFactor) {
++over_count;
lost_portion += (lf - kStrictLoadFactor) / lf / shard_count;
}
}
// >= 20% loss -> error
// >= 10% loss -> consistent warning
// >= 1% loss -> intermittent warning
InfoLogLevel level = InfoLogLevel::INFO_LEVEL;
bool report = true;
if (lost_portion > 0.2) {
level = InfoLogLevel::ERROR_LEVEL;
} else if (lost_portion > 0.1) {
level = InfoLogLevel::WARN_LEVEL;
} else if (lost_portion > 0.01) {
int report_percent = static_cast<int>(lost_portion * 100.0);
if (Random::GetTLSInstance()->PercentTrue(report_percent)) {
level = InfoLogLevel::WARN_LEVEL;
}
} else {
// don't report
report = false;
}
if (report) {
ROCKS_LOG_AT_LEVEL(
info_log, level,
"HyperClockCache@%p unable to use estimated %.1f%% capacity because "
"of "
"full occupancy in %d/%u cache shards (estimated_entry_charge too "
"high). Recommend estimated_entry_charge=%zu",
this, lost_portion * 100.0, over_count, (unsigned)shard_count,
min_recommendation);
}
} else if (average_load_factor < kLowSpecLoadFactor) {
// Out of spec => Consider reporting load factor too low
// But cautiously because low is not as big of a problem.
// Only report if highest occupancy shard is also below
// spec and only if average is substantially out of spec
if (predicted_load_factors.back() < kLowSpecLoadFactor &&
average_load_factor < kLowSpecLoadFactor / 1.414) {
InfoLogLevel level = InfoLogLevel::INFO_LEVEL;
if (average_load_factor < kLowSpecLoadFactor / 2) {
level = InfoLogLevel::WARN_LEVEL;
}
ROCKS_LOG_AT_LEVEL(
info_log, level,
"HyperClockCache@%p table has low occupancy at full capacity. Higher "
"estimated_entry_charge (about %.1fx) would likely improve "
"performance. Recommend estimated_entry_charge=%zu",
this, kMidSpecLoadFactor / average_load_factor, min_recommendation);
}
}
}
Refactor (Hyper)ClockCache code (#10887) Summary: For clean-up and in preparation for some other anticipated changes, including * A new dynamically-scaling variant of HyperClockCache * SecondaryCache support for HyperClockCache This change does some refactoring for current and future code sharing and reusability. (Including follow-up on https://github.com/facebook/rocksdb/issues/10843) ## clock_cache.h * TBD whether new variant will be a HyperClockCache or use some other name, so namespace is just clock_cache for the family of structures. * A number of helper functions introduced and used. * Pre-emptively split ClockHandle (shared among lock-free clock cache variants) and HandleImpl (specific to a kind of Table), and introduce template to plug new Table implementation into ClockCacheShard. ## clock_cache.cc * Mostly using helper functions. Some things like `Rollback()` and `FreeDataMarkEmpty()` were not combined because `Rollback()` is Table-specific while `FreeDataMarkEmpty()` can be used with different table implementations. * Performance testing indicated that despite more opportunities for parallelism, making a local copy of handle data for processing after marking an entry empty was slower than doing that processing before marking the entry empty (but after marking it "under construction"), thus avoiding a few words of copying data. At least for now, this answers the "TODO? Delay freeing?" questions (no). Pull Request resolved: https://github.com/facebook/rocksdb/pull/10887 Test Plan: fixed a unit testing gap; other minor test updates for refactoring No functionality change ## Performance Same setup as https://github.com/facebook/rocksdb/issues/10801: Before: `readrandom [AVG 81 runs] : 627992 (± 5124) ops/sec` After: `readrandom [AVG 81 runs] : 637512 (± 4866) ops/sec` I've been getting some inconsistent results on restarts like the system is not being fair to the two processes, so I'm not sure there's such a real difference. Reviewed By: anand1976 Differential Revision: D40959240 Pulled By: pdillinger fbshipit-source-id: 0a8f3646b3bdb5bc7aaad60b26790b0779189949
2 years ago
} // namespace clock_cache
Call experimental new clock cache HyperClockCache (#10684) Summary: This change establishes a distinctive name for the experimental new lock-free clock cache (originally developed by guidotag and revamped in PR https://github.com/facebook/rocksdb/issues/10626). A few reasons: * We want to make it clear that this is a fundamentally different implementation vs. the old clock cache, to avoid people saying "I already tried clock cache." * We want to highlight the key feature: it's fast (especially under parallel load) * Because it requires an estimated charge per entry, it is not drop-in API compatible with old clock cache. This estimate might always be required for highest performance, and giving it a distinct name should reduce confusion about the distinct API requirements. * We might develop a variant requiring the same estimate parameter but with LRU eviction. In that case, using the name HyperLRUCache should make things more clear. (FastLRUCache is just a prototype that might soon be removed.) Some API detail: * To reduce copy-pasting parameter lists, etc. as in LRUCache construction, I have a `MakeSharedCache()` function on `HyperClockCacheOptions` instead of `NewHyperClockCache()`. * Changes -cache_type=clock_cache to -cache_type=hyper_clock_cache for applicable tools. I think this is more consistent / sustainable for reasons already stated. For performance tests see https://github.com/facebook/rocksdb/pull/10626 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10684 Test Plan: no interesting functional changes; tests updated Reviewed By: anand1976 Differential Revision: D39547800 Pulled By: pdillinger fbshipit-source-id: 5c0fe1b5cf3cb680ab369b928c8569682b9795bf
2 years ago
// DEPRECATED (see public API)
std::shared_ptr<Cache> NewClockCache(
size_t capacity, int num_shard_bits, bool strict_capacity_limit,
CacheMetadataChargePolicy metadata_charge_policy) {
return NewLRUCache(capacity, num_shard_bits, strict_capacity_limit,
/* high_pri_pool_ratio */ 0.5, nullptr,
kDefaultToAdaptiveMutex, metadata_charge_policy,
/* low_pri_pool_ratio */ 0.0);
}
Call experimental new clock cache HyperClockCache (#10684) Summary: This change establishes a distinctive name for the experimental new lock-free clock cache (originally developed by guidotag and revamped in PR https://github.com/facebook/rocksdb/issues/10626). A few reasons: * We want to make it clear that this is a fundamentally different implementation vs. the old clock cache, to avoid people saying "I already tried clock cache." * We want to highlight the key feature: it's fast (especially under parallel load) * Because it requires an estimated charge per entry, it is not drop-in API compatible with old clock cache. This estimate might always be required for highest performance, and giving it a distinct name should reduce confusion about the distinct API requirements. * We might develop a variant requiring the same estimate parameter but with LRU eviction. In that case, using the name HyperLRUCache should make things more clear. (FastLRUCache is just a prototype that might soon be removed.) Some API detail: * To reduce copy-pasting parameter lists, etc. as in LRUCache construction, I have a `MakeSharedCache()` function on `HyperClockCacheOptions` instead of `NewHyperClockCache()`. * Changes -cache_type=clock_cache to -cache_type=hyper_clock_cache for applicable tools. I think this is more consistent / sustainable for reasons already stated. For performance tests see https://github.com/facebook/rocksdb/pull/10626 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10684 Test Plan: no interesting functional changes; tests updated Reviewed By: anand1976 Differential Revision: D39547800 Pulled By: pdillinger fbshipit-source-id: 5c0fe1b5cf3cb680ab369b928c8569682b9795bf
2 years ago
std::shared_ptr<Cache> HyperClockCacheOptions::MakeSharedCache() const {
// For sanitized options
HyperClockCacheOptions opts = *this;
if (opts.num_shard_bits >= 20) {
return nullptr; // The cache cannot be sharded into too many fine pieces.
}
if (opts.num_shard_bits < 0) {
Revamp, optimize new experimental clock cache (#10626) Summary: * Consolidates most metadata into a single word per slot so that more can be accomplished with a single atomic update. In the common case, Lookup was previously about 4 atomic updates, now just 1 atomic update. Common case Release was previously 1 atomic read + 1 atomic update, now just 1 atomic update. * Eliminate spins / waits / yields, which likely threaten some "lock free" benefits. Compare-exchange loops are only used in explicit Erase, and strict_capacity_limit=true Insert. Eviction uses opportunistic compare- exchange. * Relaxes some aggressiveness and guarantees. For example, * Duplicate Inserts will sometimes go undetected and the shadow duplicate will age out with eviction. * In many cases, the older Inserted value for a given cache key will be kept (i.e. Insert does not support overwrite). * Entries explicitly erased (rather than evicted) might not be freed immediately in some rare cases. * With strict_capacity_limit=false, capacity limit is not tracked/enforced as precisely as LRUCache, but is self-correcting and should only deviate by a very small number of extra or fewer entries. * Use smaller "computed default" number of cache shards in many cases, because benefits to larger usage tracking / eviction pools outweigh the small cost of more lock-free atomic contention. The improvement in CPU and I/O is dramatic in some limit-memory cases. * Even without the sharding change, the eviction algorithm is likely more effective than LRU overall because it's more stateful, even though the "hot path" state tracking for it is essentially free with ref counting. It is like a generalized CLOCK with aging (see code comments). I don't have performance numbers showing a specific improvement, but in theory, for a Poisson access pattern to each block, keeping some state allows better estimation of time to next access (Poisson interval) than strict LRU. The bounded randomness in CLOCK can also reduce "cliff" effect for repeated range scans approaching and exceeding cache size. ## Hot path algorithm comparison Rough descriptions, focusing on number and kind of atomic operations: * Old `Lookup()` (2-5 atomic updates per probe): ``` Loop: Increment internal ref count at slot If possible hit: Check flags atomic (and non-atomic fields) If cache hit: Three distinct updates to 'flags' atomic Increment refs for internal-to-external Return Decrement internal ref count while atomic read 'displacements' > 0 ``` * New `Lookup()` (1-2 atomic updates per probe): ``` Loop: Increment acquire counter in meta word (optimistic) If visible entry (already read meta word): If match (read non-atomic fields): Return Else: Decrement acquire counter in meta word Else if invisible entry (rare, already read meta word): Decrement acquire counter in meta word while atomic read 'displacements' > 0 ``` * Old `Release()` (1 atomic update, conditional on atomic read, rarely more): ``` Read atomic ref count If last reference and invisible (rare): Use CAS etc. to remove Return Else: Decrement ref count ``` * New `Release()` (1 unconditional atomic update, rarely more): ``` Increment release counter in meta word If last reference and invisible (rare): Use CAS etc. to remove Return ``` ## Performance test setup Build DB with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` Test with ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics ``` Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations: base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6) folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry) gt_clock: experimental ClockCache before this change new_clock: experimental ClockCache with this change ## Performance test results First test "hot path" read performance, with block cache large enough for whole DB: 4181MB 1thread base -> kops/s: 47.761 4181MB 1thread folly -> kops/s: 45.877 4181MB 1thread gt_clock -> kops/s: 51.092 4181MB 1thread new_clock -> kops/s: 53.944 4181MB 16thread base -> kops/s: 284.567 4181MB 16thread folly -> kops/s: 249.015 4181MB 16thread gt_clock -> kops/s: 743.762 4181MB 16thread new_clock -> kops/s: 861.821 4181MB 24thread base -> kops/s: 303.415 4181MB 24thread folly -> kops/s: 266.548 4181MB 24thread gt_clock -> kops/s: 975.706 4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944) 4181MB 32thread base -> kops/s: 311.251 4181MB 32thread folly -> kops/s: 274.952 4181MB 32thread gt_clock -> kops/s: 1045.98 4181MB 32thread new_clock -> kops/s: 1370.38 4181MB 48thread base -> kops/s: 310.504 4181MB 48thread folly -> kops/s: 268.322 4181MB 48thread gt_clock -> kops/s: 1195.65 4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944) 4181MB 64thread base -> kops/s: 307.839 4181MB 64thread folly -> kops/s: 272.172 4181MB 64thread gt_clock -> kops/s: 1204.47 4181MB 64thread new_clock -> kops/s: 1615.37 4181MB 128thread base -> kops/s: 310.934 4181MB 128thread folly -> kops/s: 267.468 4181MB 128thread gt_clock -> kops/s: 1188.75 4181MB 128thread new_clock -> kops/s: 1595.46 Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x. Now test a large block cache with low miss ratio, but some eviction is required: 1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23 1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43 1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4 1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56 1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59 1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8 1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89 1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45 1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98 1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91 1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26 1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63 610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137 610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996 610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934 610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5 610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402 610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742 610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062 610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453 610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457 610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426 610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273 610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812 The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.) Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc. 233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371 233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293 233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844 233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461 233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227 233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738 233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688 233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402 233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84 233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785 233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94 233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016 89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086 89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984 89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441 89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754 89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812 89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418 89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422 89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293 89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43 89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824 89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32 89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223 ^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.) Even smaller cache size: 34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914 34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281 34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523 34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125 34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48 34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531 34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465 34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793 34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484 34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457 34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41 34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52 As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn: 13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328 13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633 13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684 13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383 13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492 13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863 13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121 13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758 13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539 13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098 13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77 13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27 gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention: 13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852 13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516 13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688 13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707 13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57 13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219 13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871 13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626 Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN Reviewed By: anand1976 Differential Revision: D39368406 Pulled By: pdillinger fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
2 years ago
// Use larger shard size to reduce risk of large entries clustering
// or skewing individual shards.
constexpr size_t min_shard_size = 32U * 1024U * 1024U;
opts.num_shard_bits =
GetDefaultCacheShardBits(opts.capacity, min_shard_size);
}
std::shared_ptr<Cache> cache =
std::make_shared<clock_cache::HyperClockCache>(opts);
if (opts.secondary_cache) {
cache = std::make_shared<CacheWithSecondaryAdapter>(cache,
opts.secondary_cache);
HyperClockCache support for SecondaryCache, with refactoring (#11301) Summary: Internally refactors SecondaryCache integration out of LRUCache specifically and into a wrapper/adapter class that works with various Cache implementations. Notably, this relies on separating the notion of async lookup handles from other cache handles, so that HyperClockCache doesn't have to deal with the problem of allocating handles from the hash table for lookups that might fail anyway, and might be on the same key without support for coalescing. (LRUCache's hash table can incorporate previously allocated handles thanks to its pointer indirection.) Specifically, I'm worried about the case in which hundreds of threads try to access the same block and probing in the hash table degrades to linear search on the pile of entries with the same key. This change is a big step in the direction of supporting stacked SecondaryCaches, but there are obstacles to completing that. Especially, there is no SecondaryCache hook for evictions to pass from one to the next. It has been proposed that evictions be transmitted simply as the persisted data (as in SaveToCallback), but given the current structure provided by the CacheItemHelpers, that would require an extra copy of the block data, because there's intentionally no way to ask for a contiguous Slice of the data (to allow for flexibility in storage). `AsyncLookupHandle` and the re-worked `WaitAll()` should be essentially prepared for stacked SecondaryCaches, but several "TODO with stacked secondaries" issues remain in various places. It could be argued that the stacking instead be done as a SecondaryCache adapter that wraps two (or more) SecondaryCaches, but at least with the current API that would require an extra heap allocation on SecondaryCache Lookup for a wrapper SecondaryCacheResultHandle that can transfer a Lookup between secondaries. We could also consider trying to unify the Cache and SecondaryCache APIs, though that might be difficult if `AsyncLookupHandle` is kept a fixed struct. ## cache.h (public API) Moves `secondary_cache` option from LRUCacheOptions to ShardedCacheOptions so that it is applicable to HyperClockCache. ## advanced_cache.h (advanced public API) * Add `Cache::CreateStandalone()` so that the SecondaryCache support wrapper can use it. * Add `SetEvictionCallback()` / `eviction_callback_` so that the SecondaryCache support wrapper can use it. Only a single callback is supported for efficiency. If there is ever a need for more than one, hopefully that can be handled with a broadcast callback wrapper. These are essentially the two "extra" pieces of `Cache` for pulling out specific SecondaryCache support from the `Cache` implementation. I think it's a good trade-off as these are reasonable, limited, and reusable "cut points" into the `Cache` implementations. * Remove async capability from standard `Lookup()` (getting rid of awkward restrictions on pending Handles) and add `AsyncLookupHandle` and `StartAsyncLookup()`. As noted in the comments, the full struct of `AsyncLookupHandle` is exposed so that it can be stack allocated, for efficiency, though more data is being copied around than before, which could impact performance. (Lookup info -> AsyncLookupHandle -> Handle vs. Lookup info -> Handle) I could foresee a future in which a Cache internally saves a pointer to the AsyncLookupHandle, which means it's dangerous to allow it to be copyable or even movable. It also means it's not compatible with std::vector (which I don't like requiring as an API parameter anyway), so `WaitAll()` expects any contiguous array of AsyncLookupHandles. I believe this is best for common case efficiency, while behaving well in other cases also. For example, `WaitAll()` has no effect on default-constructed AsyncLookupHandles, which look like a completed cache miss. ## cacheable_entry.h A couple of functions are obsolete because Cache::Handle can no longer be pending. ## cache.cc Provides default implementations for new or revamped Cache functions, especially appropriate for non-blocking caches. ## secondary_cache_adapter.{h,cc} The full details of the Cache wrapper adding SecondaryCache support. Essentially replicates the SecondaryCache handling that was in LRUCache, but obviously refactored. There is a bit of logic duplication, where Lookup() is essentially a manually optimized version of StartAsyncLookup() and Wait(), but it's roughly a dozen lines of code. ## sharded_cache.h, typed_cache.h, charged_cache.{h,cc}, sim_cache.cc Simply updated for Cache API changes. ## lru_cache.{h,cc} Carefully remove SecondaryCache logic, implement `CreateStandalone` and eviction handler functionality. ## clock_cache.{h,cc} Expose existing `CreateStandalone` functionality, add eviction handler functionality. Light refactoring. ## block_based_table_reader* Mostly re-worked the only usage of async Lookup, which is in BlockBasedTable::MultiGet. Used arrays in place of autovector in some places for efficiency. Simplified some logic by not trying to process some cache results before they're all ready. Created new function `BlockBasedTable::GetCachePriority()` to reduce some pre-existing code duplication (and avoid making it worse). Fixed at least one small bug from the prior confusing mixture of async and sync Lookups. In MaybeReadBlockAndLoadToCache(), called by RetrieveBlock(), called by MultiGet() with wait=false, is_cache_hit for the block_cache_tracer entry would not be set to true if the handle was pending after Lookup and before Wait. ## Intended follow-up work * Figure out if there are any missing stats or block_cache_tracer work in refactored BlockBasedTable::MultiGet * Stacked secondary caches (see above discussion) * See if we can make up for the small MultiGet performance regression. * Study more performance with SecondaryCache * Items evicted from over-full LRUCache in Release were not being demoted to SecondaryCache, and still aren't to minimize unit test churn. Ideally they would be demoted, but it's an exceptional case so not a big deal. * Use CreateStandalone for cache reservations (save unnecessary hash table operations). Not a big deal, but worthy cleanup. * Somehow I got the contract for SecondaryCache::Insert wrong in #10945. (Doesn't take ownership!) That API comment needs to be fixed, but didn't want to mingle that in here. Pull Request resolved: https://github.com/facebook/rocksdb/pull/11301 Test Plan: ## Unit tests Generally updated to include HCC in SecondaryCache tests, though HyperClockCache has some different, less strict behaviors that leads to some tests not really being set up to work with it. Some of the tests remain disabled with it, but I think we have good coverage without them. ## Crash/stress test Updated to use the new combination. ## Performance First, let's check for regression on caches without secondary cache configured. Adding support for the eviction callback is likely to have a tiny effect, but it shouldn't be worrisome. LRUCache could benefit slightly from less logic around SecondaryCache handling. We can test with cache_bench default settings, built with DEBUG_LEVEL=0 and PORTABLE=0. ``` (while :; do base/cache_bench --cache_type=hyper_clock_cache | grep Rough; done) | awk '{ sum += $9; count++; print $0; print "Average: " int(sum / count) }' ``` **Before** this and #11299 (which could also have a small effect), running for about an hour, before & after running concurrently for each cache type: HyperClockCache: 3168662 (average parallel ops/sec) LRUCache: 2940127 **After** this and #11299, running for about an hour: HyperClockCache: 3164862 (average parallel ops/sec) (0.12% slower) LRUCache: 2940928 (0.03% faster) This is an acceptable difference IMHO. Next, let's consider essentially the worst case of new CPU overhead affecting overall performance. MultiGet uses the async lookup interface regardless of whether SecondaryCache or folly are used. We can configure a benchmark where all block cache queries are for data blocks, and all are hits. Create DB and test (before and after tests running simultaneously): ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 TEST_TMPDIR=/dev/shm base/db_bench -benchmarks=multireadrandom[-X30] -readonly -multiread_batched -batch_size=32 -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16 ``` **Before**: multireadrandom [AVG 30 runs] : 3444202 (± 57049) ops/sec; 240.9 (± 4.0) MB/sec multireadrandom [MEDIAN 30 runs] : 3514443 ops/sec; 245.8 MB/sec **After**: multireadrandom [AVG 30 runs] : 3291022 (± 58851) ops/sec; 230.2 (± 4.1) MB/sec multireadrandom [MEDIAN 30 runs] : 3366179 ops/sec; 235.4 MB/sec So that's roughly a 3% regression, on kind of a *worst case* test of MultiGet CPU. Similar story with HyperClockCache: **Before**: multireadrandom [AVG 30 runs] : 3933777 (± 41840) ops/sec; 275.1 (± 2.9) MB/sec multireadrandom [MEDIAN 30 runs] : 3970667 ops/sec; 277.7 MB/sec **After**: multireadrandom [AVG 30 runs] : 3755338 (± 30391) ops/sec; 262.6 (± 2.1) MB/sec multireadrandom [MEDIAN 30 runs] : 3785696 ops/sec; 264.8 MB/sec Roughly a 4-5% regression. Not ideal, but not the whole story, fortunately. Let's also look at Get() in db_bench: ``` TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X30] -readonly -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16 ``` **Before**: readrandom [AVG 30 runs] : 2198685 (± 13412) ops/sec; 153.8 (± 0.9) MB/sec readrandom [MEDIAN 30 runs] : 2209498 ops/sec; 154.5 MB/sec **After**: readrandom [AVG 30 runs] : 2292814 (± 43508) ops/sec; 160.3 (± 3.0) MB/sec readrandom [MEDIAN 30 runs] : 2365181 ops/sec; 165.4 MB/sec That's showing roughly a 4% improvement, perhaps because of the secondary cache code that is no longer part of LRUCache. But weirdly, HyperClockCache is also showing 2-3% improvement: **Before**: readrandom [AVG 30 runs] : 2272333 (± 9992) ops/sec; 158.9 (± 0.7) MB/sec readrandom [MEDIAN 30 runs] : 2273239 ops/sec; 159.0 MB/sec **After**: readrandom [AVG 30 runs] : 2332407 (± 11252) ops/sec; 163.1 (± 0.8) MB/sec readrandom [MEDIAN 30 runs] : 2335329 ops/sec; 163.3 MB/sec Reviewed By: ltamasi Differential Revision: D44177044 Pulled By: pdillinger fbshipit-source-id: e808e48ff3fe2f792a79841ba617be98e48689f5
2 years ago
}
return cache;
}
} // namespace ROCKSDB_NAMESPACE