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

840 lines
30 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"
#ifndef SUPPORT_CLOCK_CACHE
namespace ROCKSDB_NAMESPACE {
std::shared_ptr<Cache> NewClockCache(
size_t /*capacity*/, int /*num_shard_bits*/, bool /*strict_capacity_limit*/,
CacheMetadataChargePolicy /*metadata_charge_policy*/) {
// Clock cache not supported.
return nullptr;
}
} // namespace ROCKSDB_NAMESPACE
#else
#include <assert.h>
#include <atomic>
#include <deque>
// "tbb/concurrent_hash_map.h" requires RTTI if exception is enabled.
// Disable it so users can chooose to disable RTTI.
#ifndef ROCKSDB_USE_RTTI
#define TBB_USE_EXCEPTIONS 0
#endif
#include "cache/sharded_cache.h"
#include "port/lang.h"
#include "port/malloc.h"
#include "port/port.h"
#include "tbb/concurrent_hash_map.h"
#include "util/autovector.h"
Use optimized folly DistributedMutex in LRUCache when available (#10179) Summary: folly DistributedMutex is faster than standard mutexes though imposes some static obligations on usage. See https://github.com/facebook/folly/blob/main/folly/synchronization/DistributedMutex.h for details. Here we use this alternative for our Cache implementations (especially LRUCache) for better locking performance, when RocksDB is compiled with folly. Also added information about which distributed mutex implementation is being used to cache_bench output and to DB LOG. Intended follow-up: * Use DMutex in more places, perhaps improving API to support non-scoped locking * Fix linking with fbcode compiler (needs ROCKSDB_NO_FBCODE=1 currently) Credit: Thanks Siying for reminding me about this line of work that was previously left unfinished. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10179 Test Plan: for correctness, existing tests. CircleCI config updated. Also Meta-internal buck build updated. For performance, ran simultaneous before & after cache_bench. Out of three comparison runs, the middle improvement to ops/sec was +21%: Baseline: USE_CLANG=1 DEBUG_LEVEL=0 make -j24 cache_bench (fbcode compiler) ``` Complete in 20.201 s; Rough parallel ops/sec = 1584062 Thread ops/sec = 107176 Operation latency (ns): Count: 32000000 Average: 9257.9421 StdDev: 122412.04 Min: 134 Median: 3623.0493 Max: 56918500 Percentiles: P50: 3623.05 P75: 10288.02 P99: 30219.35 P99.9: 683522.04 P99.99: 7302791.63 ``` New: (add USE_FOLLY=1) ``` Complete in 16.674 s; Rough parallel ops/sec = 1919135 (+21%) Thread ops/sec = 135487 Operation latency (ns): Count: 32000000 Average: 7304.9294 StdDev: 108530.28 Min: 132 Median: 3777.6012 Max: 91030902 Percentiles: P50: 3777.60 P75: 10169.89 P99: 24504.51 P99.9: 59721.59 P99.99: 1861151.83 ``` Reviewed By: anand1976 Differential Revision: D37182983 Pulled By: pdillinger fbshipit-source-id: a17eb05f25b832b6a2c1356f5c657e831a5af8d1
2 years ago
#include "util/distributed_mutex.h"
namespace ROCKSDB_NAMESPACE {
namespace {
// An implementation of the Cache interface based on CLOCK algorithm, with
// better concurrent performance than LRUCache. The idea of CLOCK algorithm
// is to maintain all cache entries in a circular list, and an iterator
// (the "head") pointing to the last examined entry. Eviction starts from the
// current head. Each entry is given a second chance before eviction, if it
// has been access since last examine. In contrast to LRU, no modification
// to the internal data-structure (except for flipping the usage bit) needs
// to be done upon lookup. This gives us oppertunity to implement a cache
// with better concurrency.
//
// Each cache entry is represented by a cache handle, and all the handles
// are arranged in a circular list, as describe above. Upon erase of an entry,
// we never remove the handle. Instead, the handle is put into a recycle bin
// to be re-use. This is to avoid memory dealocation, which is hard to deal
// with in concurrent environment.
//
// The cache also maintains a concurrent hash map for lookup. Any concurrent
// hash map implementation should do the work. We currently use
// tbb::concurrent_hash_map because it supports concurrent erase.
//
// Each cache handle has the following flags and counters, which are squeeze
// in an atomic interger, to make sure the handle always be in a consistent
// state:
//
// * In-cache bit: whether the entry is reference by the cache itself. If
// an entry is in cache, its key would also be available in the hash map.
// * Usage bit: whether the entry has been access by user since last
// examine for eviction. Can be reset by eviction.
// * Reference count: reference count by user.
//
// An entry can be reference only when it's in cache. An entry can be evicted
// only when it is in cache, has no usage since last examine, and reference
// count is zero.
//
// The follow figure shows a possible layout of the cache. Boxes represents
// cache handles and numbers in each box being in-cache bit, usage bit and
// reference count respectively.
//
// hash map:
// +-------+--------+
// | key | handle |
// +-------+--------+
// | "foo" | 5 |-------------------------------------+
// +-------+--------+ |
// | "bar" | 2 |--+ |
// +-------+--------+ | |
// | |
// head | |
// | | |
// circular list: | | |
// +-------+ +-------+ +-------+ +-------+ +-------+ +-------
// |(0,0,0)|---|(1,1,0)|---|(0,0,0)|---|(0,1,3)|---|(1,0,0)|---| ...
// +-------+ +-------+ +-------+ +-------+ +-------+ +-------
// | |
// +-------+ +-----------+
// | |
// +---+---+
// recycle bin: | 1 | 3 |
// +---+---+
//
// Suppose we try to insert "baz" into the cache at this point and the cache is
// full. The cache will first look for entries to evict, starting from where
// head points to (the second entry). It resets usage bit of the second entry,
// skips the third and fourth entry since they are not in cache, and finally
// evict the fifth entry ("foo"). It looks at recycle bin for available handle,
// grabs handle 3, and insert the key into the handle. The following figure
// shows the resulting layout.
//
// hash map:
// +-------+--------+
// | key | handle |
// +-------+--------+
// | "baz" | 3 |-------------+
// +-------+--------+ |
// | "bar" | 2 |--+ |
// +-------+--------+ | |
// | |
// | | head
// | | |
// circular list: | | |
// +-------+ +-------+ +-------+ +-------+ +-------+ +-------
// |(0,0,0)|---|(1,0,0)|---|(1,0,0)|---|(0,1,3)|---|(0,0,0)|---| ...
// +-------+ +-------+ +-------+ +-------+ +-------+ +-------
// | |
// +-------+ +-----------------------------------+
// | |
// +---+---+
// recycle bin: | 1 | 5 |
// +---+---+
//
// A global mutex guards the circular list, the head, and the recycle bin.
// We additionally require that modifying the hash map needs to hold the mutex.
// As such, Modifying the cache (such as Insert() and Erase()) require to
// hold the mutex. Lookup() only access the hash map and the flags associated
// with each handle, and don't require explicit locking. Release() has to
// acquire the mutex only when it releases the last reference to the entry and
// the entry has been erased from cache explicitly. A future improvement could
// be to remove the mutex completely.
//
// Benchmark:
// We run readrandom db_bench on a test DB of size 13GB, with size of each
// level:
//
// Level Files Size(MB)
// -------------------------
// L0 1 0.01
// L1 18 17.32
// L2 230 182.94
// L3 1186 1833.63
// L4 4602 8140.30
//
// We test with both 32 and 16 read threads, with 2GB cache size (the whole DB
// doesn't fits in) and 64GB cache size (the whole DB can fit in cache), and
// whether to put index and filter blocks in block cache. The benchmark runs
// with
// with RocksDB 4.10. We got the following result:
//
// Threads Cache Cache ClockCache LRUCache
// Size Index/Filter Throughput(MB/s) Hit Throughput(MB/s) Hit
// 32 2GB yes 466.7 85.9% 433.7 86.5%
// 32 2GB no 529.9 72.7% 532.7 73.9%
// 32 64GB yes 649.9 99.9% 507.9 99.9%
// 32 64GB no 740.4 99.9% 662.8 99.9%
// 16 2GB yes 278.4 85.9% 283.4 86.5%
// 16 2GB no 318.6 72.7% 335.8 73.9%
// 16 64GB yes 391.9 99.9% 353.3 99.9%
// 16 64GB no 433.8 99.8% 419.4 99.8%
// Cache entry meta data.
struct CacheHandle {
Slice key;
void* value;
size_t charge;
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
Cache::DeleterFn deleter;
Fix use-after-free threading bug in ClockCache (#8261) Summary: In testing for https://github.com/facebook/rocksdb/issues/8225 I found cache_bench would crash with -use_clock_cache, as well as db_bench -use_clock_cache, but not single-threaded. Smaller cache size hits failure much faster. ASAN reported the failuer as calling malloc_usable_size on the `key` pointer of a ClockCache handle after it was reportedly freed. On detailed inspection I found this bad sequence of operations for a cache entry: state=InCache=1,refs=1 [thread 1] Start ClockCacheShard::Unref (from Release, no mutex) [thread 1] Decrement ref count state=InCache=1,refs=0 [thread 1] Suspend before CalcTotalCharge (no mutex) [thread 2] Start UnsetInCache (from Insert, mutex held) [thread 2] clear InCache bit state=InCache=0,refs=0 [thread 2] Calls RecycleHandle (based on pre-updated state) [thread 2] Returns to Insert which calls Cleanup which deletes `key` [thread 1] Resume ClockCacheShard::Unref [thread 1] Read `key` in CalcTotalCharge To fix this, I've added a field to the handle to store the metadata charge so that we can efficiently remember everything we need from the handle in Unref. We must not read from the handle again if we decrement the count to zero with InCache=1, which means we don't own the entry and someone else could eject/overwrite it immediately. Note before this change, on amd64 sizeof(Handle) == 56 even though there are only 48 bytes of data. Grouping together the uint32_t fields would cut it down to 48, but I've added another uint32_t, which takes it back up to 56. Not a big deal. Also fixed DisownData to cooperate with ASAN as in LRUCache. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8261 Test Plan: Manual + adding use_clock_cache to db_crashtest.py Base performance ./cache_bench -use_clock_cache Complete in 17.060 s; QPS = 2458513 New performance ./cache_bench -use_clock_cache Complete in 17.052 s; QPS = 2459695 Any difference is easily buried in small noise. Crash test shows still more bug(s) in ClockCache, so I'm expecting to disable ClockCache from production code in a follow-up PR (if we can't find and fix the bug(s)) Reviewed By: mrambacher Differential Revision: D28207358 Pulled By: pdillinger fbshipit-source-id: aa7a9322afc6f18f30e462c75dbbe4a1206eb294
4 years ago
uint32_t hash;
// Addition to "charge" to get "total charge" under metadata policy.
uint32_t meta_charge;
// Flags and counters associated with the cache handle:
// lowest bit: in-cache bit
// second lowest bit: usage bit
// the rest bits: reference count
// The handle is unused when flags equals to 0. The thread decreases the count
// to 0 is responsible to put the handle back to recycle_ and cleanup memory.
std::atomic<uint32_t> flags;
CacheHandle() = default;
CacheHandle(const CacheHandle& a) { *this = a; }
CacheHandle(const Slice& k, void* v,
void (*del)(const Slice& key, void* value))
: key(k), value(v), deleter(del) {}
CacheHandle& operator=(const CacheHandle& a) {
// Only copy members needed for deletion.
key = a.key;
value = a.value;
deleter = a.deleter;
return *this;
}
Fix use-after-free threading bug in ClockCache (#8261) Summary: In testing for https://github.com/facebook/rocksdb/issues/8225 I found cache_bench would crash with -use_clock_cache, as well as db_bench -use_clock_cache, but not single-threaded. Smaller cache size hits failure much faster. ASAN reported the failuer as calling malloc_usable_size on the `key` pointer of a ClockCache handle after it was reportedly freed. On detailed inspection I found this bad sequence of operations for a cache entry: state=InCache=1,refs=1 [thread 1] Start ClockCacheShard::Unref (from Release, no mutex) [thread 1] Decrement ref count state=InCache=1,refs=0 [thread 1] Suspend before CalcTotalCharge (no mutex) [thread 2] Start UnsetInCache (from Insert, mutex held) [thread 2] clear InCache bit state=InCache=0,refs=0 [thread 2] Calls RecycleHandle (based on pre-updated state) [thread 2] Returns to Insert which calls Cleanup which deletes `key` [thread 1] Resume ClockCacheShard::Unref [thread 1] Read `key` in CalcTotalCharge To fix this, I've added a field to the handle to store the metadata charge so that we can efficiently remember everything we need from the handle in Unref. We must not read from the handle again if we decrement the count to zero with InCache=1, which means we don't own the entry and someone else could eject/overwrite it immediately. Note before this change, on amd64 sizeof(Handle) == 56 even though there are only 48 bytes of data. Grouping together the uint32_t fields would cut it down to 48, but I've added another uint32_t, which takes it back up to 56. Not a big deal. Also fixed DisownData to cooperate with ASAN as in LRUCache. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8261 Test Plan: Manual + adding use_clock_cache to db_crashtest.py Base performance ./cache_bench -use_clock_cache Complete in 17.060 s; QPS = 2458513 New performance ./cache_bench -use_clock_cache Complete in 17.052 s; QPS = 2459695 Any difference is easily buried in small noise. Crash test shows still more bug(s) in ClockCache, so I'm expecting to disable ClockCache from production code in a follow-up PR (if we can't find and fix the bug(s)) Reviewed By: mrambacher Differential Revision: D28207358 Pulled By: pdillinger fbshipit-source-id: aa7a9322afc6f18f30e462c75dbbe4a1206eb294
4 years ago
inline static uint32_t CalcMetadataCharge(
Slice key, CacheMetadataChargePolicy metadata_charge_policy) {
size_t meta_charge = 0;
if (metadata_charge_policy == kFullChargeCacheMetadata) {
meta_charge += sizeof(CacheHandle);
#ifdef ROCKSDB_MALLOC_USABLE_SIZE
meta_charge +=
malloc_usable_size(static_cast<void*>(const_cast<char*>(key.data())));
#else
meta_charge += key.size();
#endif
}
Fix use-after-free threading bug in ClockCache (#8261) Summary: In testing for https://github.com/facebook/rocksdb/issues/8225 I found cache_bench would crash with -use_clock_cache, as well as db_bench -use_clock_cache, but not single-threaded. Smaller cache size hits failure much faster. ASAN reported the failuer as calling malloc_usable_size on the `key` pointer of a ClockCache handle after it was reportedly freed. On detailed inspection I found this bad sequence of operations for a cache entry: state=InCache=1,refs=1 [thread 1] Start ClockCacheShard::Unref (from Release, no mutex) [thread 1] Decrement ref count state=InCache=1,refs=0 [thread 1] Suspend before CalcTotalCharge (no mutex) [thread 2] Start UnsetInCache (from Insert, mutex held) [thread 2] clear InCache bit state=InCache=0,refs=0 [thread 2] Calls RecycleHandle (based on pre-updated state) [thread 2] Returns to Insert which calls Cleanup which deletes `key` [thread 1] Resume ClockCacheShard::Unref [thread 1] Read `key` in CalcTotalCharge To fix this, I've added a field to the handle to store the metadata charge so that we can efficiently remember everything we need from the handle in Unref. We must not read from the handle again if we decrement the count to zero with InCache=1, which means we don't own the entry and someone else could eject/overwrite it immediately. Note before this change, on amd64 sizeof(Handle) == 56 even though there are only 48 bytes of data. Grouping together the uint32_t fields would cut it down to 48, but I've added another uint32_t, which takes it back up to 56. Not a big deal. Also fixed DisownData to cooperate with ASAN as in LRUCache. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8261 Test Plan: Manual + adding use_clock_cache to db_crashtest.py Base performance ./cache_bench -use_clock_cache Complete in 17.060 s; QPS = 2458513 New performance ./cache_bench -use_clock_cache Complete in 17.052 s; QPS = 2459695 Any difference is easily buried in small noise. Crash test shows still more bug(s) in ClockCache, so I'm expecting to disable ClockCache from production code in a follow-up PR (if we can't find and fix the bug(s)) Reviewed By: mrambacher Differential Revision: D28207358 Pulled By: pdillinger fbshipit-source-id: aa7a9322afc6f18f30e462c75dbbe4a1206eb294
4 years ago
assert(meta_charge <= UINT32_MAX);
return static_cast<uint32_t>(meta_charge);
}
Fix use-after-free threading bug in ClockCache (#8261) Summary: In testing for https://github.com/facebook/rocksdb/issues/8225 I found cache_bench would crash with -use_clock_cache, as well as db_bench -use_clock_cache, but not single-threaded. Smaller cache size hits failure much faster. ASAN reported the failuer as calling malloc_usable_size on the `key` pointer of a ClockCache handle after it was reportedly freed. On detailed inspection I found this bad sequence of operations for a cache entry: state=InCache=1,refs=1 [thread 1] Start ClockCacheShard::Unref (from Release, no mutex) [thread 1] Decrement ref count state=InCache=1,refs=0 [thread 1] Suspend before CalcTotalCharge (no mutex) [thread 2] Start UnsetInCache (from Insert, mutex held) [thread 2] clear InCache bit state=InCache=0,refs=0 [thread 2] Calls RecycleHandle (based on pre-updated state) [thread 2] Returns to Insert which calls Cleanup which deletes `key` [thread 1] Resume ClockCacheShard::Unref [thread 1] Read `key` in CalcTotalCharge To fix this, I've added a field to the handle to store the metadata charge so that we can efficiently remember everything we need from the handle in Unref. We must not read from the handle again if we decrement the count to zero with InCache=1, which means we don't own the entry and someone else could eject/overwrite it immediately. Note before this change, on amd64 sizeof(Handle) == 56 even though there are only 48 bytes of data. Grouping together the uint32_t fields would cut it down to 48, but I've added another uint32_t, which takes it back up to 56. Not a big deal. Also fixed DisownData to cooperate with ASAN as in LRUCache. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8261 Test Plan: Manual + adding use_clock_cache to db_crashtest.py Base performance ./cache_bench -use_clock_cache Complete in 17.060 s; QPS = 2458513 New performance ./cache_bench -use_clock_cache Complete in 17.052 s; QPS = 2459695 Any difference is easily buried in small noise. Crash test shows still more bug(s) in ClockCache, so I'm expecting to disable ClockCache from production code in a follow-up PR (if we can't find and fix the bug(s)) Reviewed By: mrambacher Differential Revision: D28207358 Pulled By: pdillinger fbshipit-source-id: aa7a9322afc6f18f30e462c75dbbe4a1206eb294
4 years ago
inline size_t GetTotalCharge() { return charge + meta_charge; }
};
// Key of hash map. We store hash value with the key for convenience.
struct ClockCacheKey {
Slice key;
uint32_t hash_value;
ClockCacheKey() = default;
ClockCacheKey(const Slice& k, uint32_t h) {
key = k;
hash_value = h;
}
static bool equal(const ClockCacheKey& a, const ClockCacheKey& b) {
return a.hash_value == b.hash_value && a.key == b.key;
}
static size_t hash(const ClockCacheKey& a) {
return static_cast<size_t>(a.hash_value);
}
};
struct CleanupContext {
// List of values to be deleted, along with the key and deleter.
autovector<CacheHandle> to_delete_value;
// List of keys to be deleted.
autovector<const char*> to_delete_key;
};
// A cache shard which maintains its own CLOCK cache.
class ClockCacheShard final : public CacheShard {
public:
// Hash map type.
using HashTable =
tbb::concurrent_hash_map<ClockCacheKey, CacheHandle*, ClockCacheKey>;
ClockCacheShard();
~ClockCacheShard() override;
// Interfaces
void SetCapacity(size_t capacity) override;
void SetStrictCapacityLimit(bool strict_capacity_limit) override;
Status Insert(const Slice& key, uint32_t hash, void* value, size_t charge,
void (*deleter)(const Slice& key, void* value),
Cache::Handle** handle, Cache::Priority priority) override;
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
Status Insert(const Slice& key, uint32_t hash, void* value,
const Cache::CacheItemHelper* helper, size_t charge,
Cache::Handle** handle, Cache::Priority priority) override {
return Insert(key, hash, value, charge, helper->del_cb, handle, priority);
}
Cache::Handle* Lookup(const Slice& key, uint32_t hash) override;
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
Cache::Handle* Lookup(const Slice& key, uint32_t hash,
const Cache::CacheItemHelper* /*helper*/,
const Cache::CreateCallback& /*create_cb*/,
Cache::Priority /*priority*/, bool /*wait*/,
Statistics* /*stats*/) override {
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
return Lookup(key, hash);
}
bool Release(Cache::Handle* handle, bool /*useful*/,
bool erase_if_last_ref) override {
return Release(handle, erase_if_last_ref);
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
}
bool IsReady(Cache::Handle* /*handle*/) override { return true; }
void Wait(Cache::Handle* /*handle*/) override {}
// If the entry in in cache, increase reference count and return true.
// Return false otherwise.
//
// Not necessary to hold mutex_ before being called.
bool Ref(Cache::Handle* handle) override;
bool Release(Cache::Handle* handle, bool erase_if_last_ref = false) override;
void Erase(const Slice& key, uint32_t hash) override;
bool EraseAndConfirm(const Slice& key, uint32_t hash,
CleanupContext* context);
size_t GetUsage() const override;
size_t GetPinnedUsage() const override;
void EraseUnRefEntries() override;
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
void ApplyToSomeEntries(
const std::function<void(const Slice& key, void* value, size_t charge,
DeleterFn deleter)>& callback,
uint32_t average_entries_per_lock, uint32_t* state) override;
private:
static const uint32_t kInCacheBit = 1;
static const uint32_t kUsageBit = 2;
static const uint32_t kRefsOffset = 2;
static const uint32_t kOneRef = 1 << kRefsOffset;
// Helper functions to extract cache handle flags and counters.
static bool InCache(uint32_t flags) { return flags & kInCacheBit; }
static bool HasUsage(uint32_t flags) { return flags & kUsageBit; }
static uint32_t CountRefs(uint32_t flags) { return flags >> kRefsOffset; }
// Decrease reference count of the entry. If this decreases the count to 0,
// recycle the entry. If set_usage is true, also set the usage bit.
//
// returns true if a value is erased.
//
// Not necessary to hold mutex_ before being called.
bool Unref(CacheHandle* handle, bool set_usage, CleanupContext* context);
// Unset in-cache bit of the entry. Recycle the handle if necessary.
//
// returns true if a value is erased.
//
// Has to hold mutex_ before being called.
bool UnsetInCache(CacheHandle* handle, CleanupContext* context);
// Put the handle back to recycle_ list, and put the value associated with
// it into to-be-deleted list. It doesn't cleanup the key as it might be
// reused by another handle.
//
// Has to hold mutex_ before being called.
void RecycleHandle(CacheHandle* handle, CleanupContext* context);
// Delete keys and values in to-be-deleted list. Call the method without
// holding mutex, as destructors can be expensive.
void Cleanup(const CleanupContext& context);
// Examine the handle for eviction. If the handle is in cache, usage bit is
// not set, and referece count is 0, evict it from cache. Otherwise unset
// the usage bit.
//
// Has to hold mutex_ before being called.
bool TryEvict(CacheHandle* value, CleanupContext* context);
// Scan through the circular list, evict entries until we get enough capacity
// for new cache entry of specific size. Return true if success, false
// otherwise.
//
// Has to hold mutex_ before being called.
bool EvictFromCache(size_t charge, CleanupContext* context);
CacheHandle* Insert(const Slice& key, uint32_t hash, void* value,
size_t change,
void (*deleter)(const Slice& key, void* value),
Stats for redundant insertions into block cache (#6681) Summary: Since read threads do not coordinate on loading data into block cache, two threads between Lookup and Insert can end up loading and inserting the same data. This is particularly concerning with cache_index_and_filter_blocks since those are hot and more likely to be race targets if ejected from (or not pre-populated in) the cache. Particularly with moves toward disaggregated / network storage, the cost of redundant retrieval might be high, and we should at least have some hard statistics from which we can estimate impact. Example with full filter thrashing "cliff": $ ./db_bench --benchmarks=fillrandom --num=15000000 --cache_index_and_filter_blocks -bloom_bits=10 ... $ ./db_bench --db=/tmp/rocksdbtest-172704/dbbench --use_existing_db --benchmarks=readrandom,stats --num=200000 --cache_index_and_filter_blocks --cache_size=$((130 * 1024 * 1024)) --bloom_bits=10 --threads=16 -statistics 2>&1 | egrep '^rocksdb.block.cache.(.*add|.*redundant)' | grep -v compress | sort rocksdb.block.cache.add COUNT : 14181 rocksdb.block.cache.add.failures COUNT : 0 rocksdb.block.cache.add.redundant COUNT : 476 rocksdb.block.cache.data.add COUNT : 12749 rocksdb.block.cache.data.add.redundant COUNT : 18 rocksdb.block.cache.filter.add COUNT : 1003 rocksdb.block.cache.filter.add.redundant COUNT : 217 rocksdb.block.cache.index.add COUNT : 429 rocksdb.block.cache.index.add.redundant COUNT : 241 $ ./db_bench --db=/tmp/rocksdbtest-172704/dbbench --use_existing_db --benchmarks=readrandom,stats --num=200000 --cache_index_and_filter_blocks --cache_size=$((120 * 1024 * 1024)) --bloom_bits=10 --threads=16 -statistics 2>&1 | egrep '^rocksdb.block.cache.(.*add|.*redundant)' | grep -v compress | sort rocksdb.block.cache.add COUNT : 1182223 rocksdb.block.cache.add.failures COUNT : 0 rocksdb.block.cache.add.redundant COUNT : 302728 rocksdb.block.cache.data.add COUNT : 31425 rocksdb.block.cache.data.add.redundant COUNT : 12 rocksdb.block.cache.filter.add COUNT : 795455 rocksdb.block.cache.filter.add.redundant COUNT : 130238 rocksdb.block.cache.index.add COUNT : 355343 rocksdb.block.cache.index.add.redundant COUNT : 172478 Pull Request resolved: https://github.com/facebook/rocksdb/pull/6681 Test Plan: Some manual testing (above) and unit test covering key metrics is included Reviewed By: ltamasi Differential Revision: D21134113 Pulled By: pdillinger fbshipit-source-id: c11497b5f00f4ffdfe919823904e52d0a1a91d87
5 years ago
bool hold_reference, CleanupContext* context,
bool* overwritten);
// Guards list_, head_, and recycle_. In addition, updating table_ also has
// to hold the mutex, to avoid the cache being in inconsistent state.
Use optimized folly DistributedMutex in LRUCache when available (#10179) Summary: folly DistributedMutex is faster than standard mutexes though imposes some static obligations on usage. See https://github.com/facebook/folly/blob/main/folly/synchronization/DistributedMutex.h for details. Here we use this alternative for our Cache implementations (especially LRUCache) for better locking performance, when RocksDB is compiled with folly. Also added information about which distributed mutex implementation is being used to cache_bench output and to DB LOG. Intended follow-up: * Use DMutex in more places, perhaps improving API to support non-scoped locking * Fix linking with fbcode compiler (needs ROCKSDB_NO_FBCODE=1 currently) Credit: Thanks Siying for reminding me about this line of work that was previously left unfinished. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10179 Test Plan: for correctness, existing tests. CircleCI config updated. Also Meta-internal buck build updated. For performance, ran simultaneous before & after cache_bench. Out of three comparison runs, the middle improvement to ops/sec was +21%: Baseline: USE_CLANG=1 DEBUG_LEVEL=0 make -j24 cache_bench (fbcode compiler) ``` Complete in 20.201 s; Rough parallel ops/sec = 1584062 Thread ops/sec = 107176 Operation latency (ns): Count: 32000000 Average: 9257.9421 StdDev: 122412.04 Min: 134 Median: 3623.0493 Max: 56918500 Percentiles: P50: 3623.05 P75: 10288.02 P99: 30219.35 P99.9: 683522.04 P99.99: 7302791.63 ``` New: (add USE_FOLLY=1) ``` Complete in 16.674 s; Rough parallel ops/sec = 1919135 (+21%) Thread ops/sec = 135487 Operation latency (ns): Count: 32000000 Average: 7304.9294 StdDev: 108530.28 Min: 132 Median: 3777.6012 Max: 91030902 Percentiles: P50: 3777.60 P75: 10169.89 P99: 24504.51 P99.9: 59721.59 P99.99: 1861151.83 ``` Reviewed By: anand1976 Differential Revision: D37182983 Pulled By: pdillinger fbshipit-source-id: a17eb05f25b832b6a2c1356f5c657e831a5af8d1
2 years ago
mutable DMutex mutex_;
// The circular list of cache handles. Initially the list is empty. Once a
// handle is needed by insertion, and no more handles are available in
// recycle bin, one more handle is appended to the end.
//
// We use std::deque for the circular list because we want to make sure
// pointers to handles are valid through out the life-cycle of the cache
// (in contrast to std::vector), and be able to grow the list (in contrast
// to statically allocated arrays).
std::deque<CacheHandle> list_;
// Pointer to the next handle in the circular list to be examine for
// eviction.
size_t head_;
// Recycle bin of cache handles.
autovector<CacheHandle*> recycle_;
// Maximum cache size.
std::atomic<size_t> capacity_;
// Current total size of the cache.
std::atomic<size_t> usage_;
// Total un-released cache size.
std::atomic<size_t> pinned_usage_;
// Whether allow insert into cache if cache is full.
std::atomic<bool> strict_capacity_limit_;
// Hash table (tbb::concurrent_hash_map) for lookup.
HashTable table_;
};
ClockCacheShard::ClockCacheShard()
: head_(0), usage_(0), pinned_usage_(0), strict_capacity_limit_(false) {}
ClockCacheShard::~ClockCacheShard() {
for (auto& handle : list_) {
uint32_t flags = handle.flags.load(std::memory_order_relaxed);
if (InCache(flags) || CountRefs(flags) > 0) {
if (handle.deleter != nullptr) {
(*handle.deleter)(handle.key, handle.value);
}
delete[] handle.key.data();
}
}
}
size_t ClockCacheShard::GetUsage() const {
return usage_.load(std::memory_order_relaxed);
}
size_t ClockCacheShard::GetPinnedUsage() const {
return pinned_usage_.load(std::memory_order_relaxed);
}
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
void ClockCacheShard::ApplyToSomeEntries(
const std::function<void(const Slice& key, void* value, size_t charge,
DeleterFn deleter)>& callback,
uint32_t average_entries_per_lock, uint32_t* state) {
assert(average_entries_per_lock > 0);
Use optimized folly DistributedMutex in LRUCache when available (#10179) Summary: folly DistributedMutex is faster than standard mutexes though imposes some static obligations on usage. See https://github.com/facebook/folly/blob/main/folly/synchronization/DistributedMutex.h for details. Here we use this alternative for our Cache implementations (especially LRUCache) for better locking performance, when RocksDB is compiled with folly. Also added information about which distributed mutex implementation is being used to cache_bench output and to DB LOG. Intended follow-up: * Use DMutex in more places, perhaps improving API to support non-scoped locking * Fix linking with fbcode compiler (needs ROCKSDB_NO_FBCODE=1 currently) Credit: Thanks Siying for reminding me about this line of work that was previously left unfinished. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10179 Test Plan: for correctness, existing tests. CircleCI config updated. Also Meta-internal buck build updated. For performance, ran simultaneous before & after cache_bench. Out of three comparison runs, the middle improvement to ops/sec was +21%: Baseline: USE_CLANG=1 DEBUG_LEVEL=0 make -j24 cache_bench (fbcode compiler) ``` Complete in 20.201 s; Rough parallel ops/sec = 1584062 Thread ops/sec = 107176 Operation latency (ns): Count: 32000000 Average: 9257.9421 StdDev: 122412.04 Min: 134 Median: 3623.0493 Max: 56918500 Percentiles: P50: 3623.05 P75: 10288.02 P99: 30219.35 P99.9: 683522.04 P99.99: 7302791.63 ``` New: (add USE_FOLLY=1) ``` Complete in 16.674 s; Rough parallel ops/sec = 1919135 (+21%) Thread ops/sec = 135487 Operation latency (ns): Count: 32000000 Average: 7304.9294 StdDev: 108530.28 Min: 132 Median: 3777.6012 Max: 91030902 Percentiles: P50: 3777.60 P75: 10169.89 P99: 24504.51 P99.9: 59721.59 P99.99: 1861151.83 ``` Reviewed By: anand1976 Differential Revision: D37182983 Pulled By: pdillinger fbshipit-source-id: a17eb05f25b832b6a2c1356f5c657e831a5af8d1
2 years ago
DMutexLock l(mutex_);
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
// Figure out the range to iterate, update `state`
size_t list_size = list_.size();
size_t start_idx = *state;
size_t end_idx = start_idx + average_entries_per_lock;
if (start_idx > list_size) {
// Shouldn't reach here, but recoverable
assert(false);
// Mark finished with all
*state = UINT32_MAX;
return;
}
if (end_idx >= list_size || end_idx >= UINT32_MAX) {
// This also includes the hypothetical case of >4 billion
// cache handles.
end_idx = list_size;
// Mark finished with all
*state = UINT32_MAX;
} else {
*state = static_cast<uint32_t>(end_idx);
}
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
// Do the iteration
auto cur = list_.begin() + start_idx;
auto end = list_.begin() + end_idx;
for (; cur != end; ++cur) {
const CacheHandle& handle = *cur;
// Use relaxed semantics instead of acquire semantics since we are
// holding mutex
uint32_t flags = handle.flags.load(std::memory_order_relaxed);
if (InCache(flags)) {
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
callback(handle.key, handle.value, handle.charge, handle.deleter);
}
}
}
void ClockCacheShard::RecycleHandle(CacheHandle* handle,
CleanupContext* context) {
mutex_.AssertHeld();
assert(!InCache(handle->flags) && CountRefs(handle->flags) == 0);
context->to_delete_key.push_back(handle->key.data());
context->to_delete_value.emplace_back(*handle);
Fix use-after-free threading bug in ClockCache (#8261) Summary: In testing for https://github.com/facebook/rocksdb/issues/8225 I found cache_bench would crash with -use_clock_cache, as well as db_bench -use_clock_cache, but not single-threaded. Smaller cache size hits failure much faster. ASAN reported the failuer as calling malloc_usable_size on the `key` pointer of a ClockCache handle after it was reportedly freed. On detailed inspection I found this bad sequence of operations for a cache entry: state=InCache=1,refs=1 [thread 1] Start ClockCacheShard::Unref (from Release, no mutex) [thread 1] Decrement ref count state=InCache=1,refs=0 [thread 1] Suspend before CalcTotalCharge (no mutex) [thread 2] Start UnsetInCache (from Insert, mutex held) [thread 2] clear InCache bit state=InCache=0,refs=0 [thread 2] Calls RecycleHandle (based on pre-updated state) [thread 2] Returns to Insert which calls Cleanup which deletes `key` [thread 1] Resume ClockCacheShard::Unref [thread 1] Read `key` in CalcTotalCharge To fix this, I've added a field to the handle to store the metadata charge so that we can efficiently remember everything we need from the handle in Unref. We must not read from the handle again if we decrement the count to zero with InCache=1, which means we don't own the entry and someone else could eject/overwrite it immediately. Note before this change, on amd64 sizeof(Handle) == 56 even though there are only 48 bytes of data. Grouping together the uint32_t fields would cut it down to 48, but I've added another uint32_t, which takes it back up to 56. Not a big deal. Also fixed DisownData to cooperate with ASAN as in LRUCache. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8261 Test Plan: Manual + adding use_clock_cache to db_crashtest.py Base performance ./cache_bench -use_clock_cache Complete in 17.060 s; QPS = 2458513 New performance ./cache_bench -use_clock_cache Complete in 17.052 s; QPS = 2459695 Any difference is easily buried in small noise. Crash test shows still more bug(s) in ClockCache, so I'm expecting to disable ClockCache from production code in a follow-up PR (if we can't find and fix the bug(s)) Reviewed By: mrambacher Differential Revision: D28207358 Pulled By: pdillinger fbshipit-source-id: aa7a9322afc6f18f30e462c75dbbe4a1206eb294
4 years ago
size_t total_charge = handle->GetTotalCharge();
// clearing `handle` fields would go here but not strictly required
recycle_.push_back(handle);
usage_.fetch_sub(total_charge, std::memory_order_relaxed);
}
void ClockCacheShard::Cleanup(const CleanupContext& context) {
for (const CacheHandle& handle : context.to_delete_value) {
if (handle.deleter) {
(*handle.deleter)(handle.key, handle.value);
}
}
for (const char* key : context.to_delete_key) {
delete[] key;
}
}
bool ClockCacheShard::Ref(Cache::Handle* h) {
auto handle = reinterpret_cast<CacheHandle*>(h);
// CAS loop to increase reference count.
uint32_t flags = handle->flags.load(std::memory_order_relaxed);
while (InCache(flags)) {
// Use acquire semantics on success, as further operations on the cache
// entry has to be order after reference count is increased.
if (handle->flags.compare_exchange_weak(flags, flags + kOneRef,
std::memory_order_acquire,
std::memory_order_relaxed)) {
if (CountRefs(flags) == 0) {
// No reference count before the operation.
Fix use-after-free threading bug in ClockCache (#8261) Summary: In testing for https://github.com/facebook/rocksdb/issues/8225 I found cache_bench would crash with -use_clock_cache, as well as db_bench -use_clock_cache, but not single-threaded. Smaller cache size hits failure much faster. ASAN reported the failuer as calling malloc_usable_size on the `key` pointer of a ClockCache handle after it was reportedly freed. On detailed inspection I found this bad sequence of operations for a cache entry: state=InCache=1,refs=1 [thread 1] Start ClockCacheShard::Unref (from Release, no mutex) [thread 1] Decrement ref count state=InCache=1,refs=0 [thread 1] Suspend before CalcTotalCharge (no mutex) [thread 2] Start UnsetInCache (from Insert, mutex held) [thread 2] clear InCache bit state=InCache=0,refs=0 [thread 2] Calls RecycleHandle (based on pre-updated state) [thread 2] Returns to Insert which calls Cleanup which deletes `key` [thread 1] Resume ClockCacheShard::Unref [thread 1] Read `key` in CalcTotalCharge To fix this, I've added a field to the handle to store the metadata charge so that we can efficiently remember everything we need from the handle in Unref. We must not read from the handle again if we decrement the count to zero with InCache=1, which means we don't own the entry and someone else could eject/overwrite it immediately. Note before this change, on amd64 sizeof(Handle) == 56 even though there are only 48 bytes of data. Grouping together the uint32_t fields would cut it down to 48, but I've added another uint32_t, which takes it back up to 56. Not a big deal. Also fixed DisownData to cooperate with ASAN as in LRUCache. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8261 Test Plan: Manual + adding use_clock_cache to db_crashtest.py Base performance ./cache_bench -use_clock_cache Complete in 17.060 s; QPS = 2458513 New performance ./cache_bench -use_clock_cache Complete in 17.052 s; QPS = 2459695 Any difference is easily buried in small noise. Crash test shows still more bug(s) in ClockCache, so I'm expecting to disable ClockCache from production code in a follow-up PR (if we can't find and fix the bug(s)) Reviewed By: mrambacher Differential Revision: D28207358 Pulled By: pdillinger fbshipit-source-id: aa7a9322afc6f18f30e462c75dbbe4a1206eb294
4 years ago
size_t total_charge = handle->GetTotalCharge();
pinned_usage_.fetch_add(total_charge, std::memory_order_relaxed);
}
return true;
}
}
return false;
}
bool ClockCacheShard::Unref(CacheHandle* handle, bool set_usage,
CleanupContext* context) {
if (set_usage) {
handle->flags.fetch_or(kUsageBit, std::memory_order_relaxed);
}
Fix use-after-free threading bug in ClockCache (#8261) Summary: In testing for https://github.com/facebook/rocksdb/issues/8225 I found cache_bench would crash with -use_clock_cache, as well as db_bench -use_clock_cache, but not single-threaded. Smaller cache size hits failure much faster. ASAN reported the failuer as calling malloc_usable_size on the `key` pointer of a ClockCache handle after it was reportedly freed. On detailed inspection I found this bad sequence of operations for a cache entry: state=InCache=1,refs=1 [thread 1] Start ClockCacheShard::Unref (from Release, no mutex) [thread 1] Decrement ref count state=InCache=1,refs=0 [thread 1] Suspend before CalcTotalCharge (no mutex) [thread 2] Start UnsetInCache (from Insert, mutex held) [thread 2] clear InCache bit state=InCache=0,refs=0 [thread 2] Calls RecycleHandle (based on pre-updated state) [thread 2] Returns to Insert which calls Cleanup which deletes `key` [thread 1] Resume ClockCacheShard::Unref [thread 1] Read `key` in CalcTotalCharge To fix this, I've added a field to the handle to store the metadata charge so that we can efficiently remember everything we need from the handle in Unref. We must not read from the handle again if we decrement the count to zero with InCache=1, which means we don't own the entry and someone else could eject/overwrite it immediately. Note before this change, on amd64 sizeof(Handle) == 56 even though there are only 48 bytes of data. Grouping together the uint32_t fields would cut it down to 48, but I've added another uint32_t, which takes it back up to 56. Not a big deal. Also fixed DisownData to cooperate with ASAN as in LRUCache. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8261 Test Plan: Manual + adding use_clock_cache to db_crashtest.py Base performance ./cache_bench -use_clock_cache Complete in 17.060 s; QPS = 2458513 New performance ./cache_bench -use_clock_cache Complete in 17.052 s; QPS = 2459695 Any difference is easily buried in small noise. Crash test shows still more bug(s) in ClockCache, so I'm expecting to disable ClockCache from production code in a follow-up PR (if we can't find and fix the bug(s)) Reviewed By: mrambacher Differential Revision: D28207358 Pulled By: pdillinger fbshipit-source-id: aa7a9322afc6f18f30e462c75dbbe4a1206eb294
4 years ago
// If the handle reaches state refs=0 and InCache=true after this
// atomic operation then we cannot access `handle` afterward, because
// it could be evicted before we access the `handle`.
size_t total_charge = handle->GetTotalCharge();
// Use acquire-release semantics as previous operations on the cache entry
// has to be order before reference count is decreased, and potential cleanup
// of the entry has to be order after.
uint32_t flags = handle->flags.fetch_sub(kOneRef, std::memory_order_acq_rel);
assert(CountRefs(flags) > 0);
if (CountRefs(flags) == 1) {
// this is the last reference.
pinned_usage_.fetch_sub(total_charge, std::memory_order_relaxed);
// Cleanup if it is the last reference.
if (!InCache(flags)) {
Use optimized folly DistributedMutex in LRUCache when available (#10179) Summary: folly DistributedMutex is faster than standard mutexes though imposes some static obligations on usage. See https://github.com/facebook/folly/blob/main/folly/synchronization/DistributedMutex.h for details. Here we use this alternative for our Cache implementations (especially LRUCache) for better locking performance, when RocksDB is compiled with folly. Also added information about which distributed mutex implementation is being used to cache_bench output and to DB LOG. Intended follow-up: * Use DMutex in more places, perhaps improving API to support non-scoped locking * Fix linking with fbcode compiler (needs ROCKSDB_NO_FBCODE=1 currently) Credit: Thanks Siying for reminding me about this line of work that was previously left unfinished. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10179 Test Plan: for correctness, existing tests. CircleCI config updated. Also Meta-internal buck build updated. For performance, ran simultaneous before & after cache_bench. Out of three comparison runs, the middle improvement to ops/sec was +21%: Baseline: USE_CLANG=1 DEBUG_LEVEL=0 make -j24 cache_bench (fbcode compiler) ``` Complete in 20.201 s; Rough parallel ops/sec = 1584062 Thread ops/sec = 107176 Operation latency (ns): Count: 32000000 Average: 9257.9421 StdDev: 122412.04 Min: 134 Median: 3623.0493 Max: 56918500 Percentiles: P50: 3623.05 P75: 10288.02 P99: 30219.35 P99.9: 683522.04 P99.99: 7302791.63 ``` New: (add USE_FOLLY=1) ``` Complete in 16.674 s; Rough parallel ops/sec = 1919135 (+21%) Thread ops/sec = 135487 Operation latency (ns): Count: 32000000 Average: 7304.9294 StdDev: 108530.28 Min: 132 Median: 3777.6012 Max: 91030902 Percentiles: P50: 3777.60 P75: 10169.89 P99: 24504.51 P99.9: 59721.59 P99.99: 1861151.83 ``` Reviewed By: anand1976 Differential Revision: D37182983 Pulled By: pdillinger fbshipit-source-id: a17eb05f25b832b6a2c1356f5c657e831a5af8d1
2 years ago
DMutexLock l(mutex_);
RecycleHandle(handle, context);
}
}
return context->to_delete_value.size();
}
bool ClockCacheShard::UnsetInCache(CacheHandle* handle,
CleanupContext* context) {
mutex_.AssertHeld();
// Use acquire-release semantics as previous operations on the cache entry
// has to be order before reference count is decreased, and potential cleanup
// of the entry has to be order after.
uint32_t flags =
handle->flags.fetch_and(~kInCacheBit, std::memory_order_acq_rel);
// Cleanup if it is the last reference.
if (InCache(flags) && CountRefs(flags) == 0) {
RecycleHandle(handle, context);
}
return context->to_delete_value.size();
}
bool ClockCacheShard::TryEvict(CacheHandle* handle, CleanupContext* context) {
mutex_.AssertHeld();
uint32_t flags = kInCacheBit;
if (handle->flags.compare_exchange_strong(flags, 0, std::memory_order_acquire,
std::memory_order_relaxed)) {
bool erased __attribute__((__unused__)) =
table_.erase(ClockCacheKey(handle->key, handle->hash));
assert(erased);
RecycleHandle(handle, context);
return true;
}
handle->flags.fetch_and(~kUsageBit, std::memory_order_relaxed);
return false;
}
bool ClockCacheShard::EvictFromCache(size_t charge, CleanupContext* context) {
size_t usage = usage_.load(std::memory_order_relaxed);
size_t capacity = capacity_.load(std::memory_order_relaxed);
if (usage == 0) {
return charge <= capacity;
}
size_t new_head = head_;
bool second_iteration = false;
while (usage + charge > capacity) {
assert(new_head < list_.size());
if (TryEvict(&list_[new_head], context)) {
usage = usage_.load(std::memory_order_relaxed);
}
new_head = (new_head + 1 >= list_.size()) ? 0 : new_head + 1;
if (new_head == head_) {
if (second_iteration) {
return false;
} else {
second_iteration = true;
}
}
}
head_ = new_head;
return true;
}
void ClockCacheShard::SetCapacity(size_t capacity) {
CleanupContext context;
{
Use optimized folly DistributedMutex in LRUCache when available (#10179) Summary: folly DistributedMutex is faster than standard mutexes though imposes some static obligations on usage. See https://github.com/facebook/folly/blob/main/folly/synchronization/DistributedMutex.h for details. Here we use this alternative for our Cache implementations (especially LRUCache) for better locking performance, when RocksDB is compiled with folly. Also added information about which distributed mutex implementation is being used to cache_bench output and to DB LOG. Intended follow-up: * Use DMutex in more places, perhaps improving API to support non-scoped locking * Fix linking with fbcode compiler (needs ROCKSDB_NO_FBCODE=1 currently) Credit: Thanks Siying for reminding me about this line of work that was previously left unfinished. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10179 Test Plan: for correctness, existing tests. CircleCI config updated. Also Meta-internal buck build updated. For performance, ran simultaneous before & after cache_bench. Out of three comparison runs, the middle improvement to ops/sec was +21%: Baseline: USE_CLANG=1 DEBUG_LEVEL=0 make -j24 cache_bench (fbcode compiler) ``` Complete in 20.201 s; Rough parallel ops/sec = 1584062 Thread ops/sec = 107176 Operation latency (ns): Count: 32000000 Average: 9257.9421 StdDev: 122412.04 Min: 134 Median: 3623.0493 Max: 56918500 Percentiles: P50: 3623.05 P75: 10288.02 P99: 30219.35 P99.9: 683522.04 P99.99: 7302791.63 ``` New: (add USE_FOLLY=1) ``` Complete in 16.674 s; Rough parallel ops/sec = 1919135 (+21%) Thread ops/sec = 135487 Operation latency (ns): Count: 32000000 Average: 7304.9294 StdDev: 108530.28 Min: 132 Median: 3777.6012 Max: 91030902 Percentiles: P50: 3777.60 P75: 10169.89 P99: 24504.51 P99.9: 59721.59 P99.99: 1861151.83 ``` Reviewed By: anand1976 Differential Revision: D37182983 Pulled By: pdillinger fbshipit-source-id: a17eb05f25b832b6a2c1356f5c657e831a5af8d1
2 years ago
DMutexLock l(mutex_);
capacity_.store(capacity, std::memory_order_relaxed);
EvictFromCache(0, &context);
}
Cleanup(context);
}
void ClockCacheShard::SetStrictCapacityLimit(bool strict_capacity_limit) {
strict_capacity_limit_.store(strict_capacity_limit,
std::memory_order_relaxed);
}
CacheHandle* ClockCacheShard::Insert(
const Slice& key, uint32_t hash, void* value, size_t charge,
void (*deleter)(const Slice& key, void* value), bool hold_reference,
Stats for redundant insertions into block cache (#6681) Summary: Since read threads do not coordinate on loading data into block cache, two threads between Lookup and Insert can end up loading and inserting the same data. This is particularly concerning with cache_index_and_filter_blocks since those are hot and more likely to be race targets if ejected from (or not pre-populated in) the cache. Particularly with moves toward disaggregated / network storage, the cost of redundant retrieval might be high, and we should at least have some hard statistics from which we can estimate impact. Example with full filter thrashing "cliff": $ ./db_bench --benchmarks=fillrandom --num=15000000 --cache_index_and_filter_blocks -bloom_bits=10 ... $ ./db_bench --db=/tmp/rocksdbtest-172704/dbbench --use_existing_db --benchmarks=readrandom,stats --num=200000 --cache_index_and_filter_blocks --cache_size=$((130 * 1024 * 1024)) --bloom_bits=10 --threads=16 -statistics 2>&1 | egrep '^rocksdb.block.cache.(.*add|.*redundant)' | grep -v compress | sort rocksdb.block.cache.add COUNT : 14181 rocksdb.block.cache.add.failures COUNT : 0 rocksdb.block.cache.add.redundant COUNT : 476 rocksdb.block.cache.data.add COUNT : 12749 rocksdb.block.cache.data.add.redundant COUNT : 18 rocksdb.block.cache.filter.add COUNT : 1003 rocksdb.block.cache.filter.add.redundant COUNT : 217 rocksdb.block.cache.index.add COUNT : 429 rocksdb.block.cache.index.add.redundant COUNT : 241 $ ./db_bench --db=/tmp/rocksdbtest-172704/dbbench --use_existing_db --benchmarks=readrandom,stats --num=200000 --cache_index_and_filter_blocks --cache_size=$((120 * 1024 * 1024)) --bloom_bits=10 --threads=16 -statistics 2>&1 | egrep '^rocksdb.block.cache.(.*add|.*redundant)' | grep -v compress | sort rocksdb.block.cache.add COUNT : 1182223 rocksdb.block.cache.add.failures COUNT : 0 rocksdb.block.cache.add.redundant COUNT : 302728 rocksdb.block.cache.data.add COUNT : 31425 rocksdb.block.cache.data.add.redundant COUNT : 12 rocksdb.block.cache.filter.add COUNT : 795455 rocksdb.block.cache.filter.add.redundant COUNT : 130238 rocksdb.block.cache.index.add COUNT : 355343 rocksdb.block.cache.index.add.redundant COUNT : 172478 Pull Request resolved: https://github.com/facebook/rocksdb/pull/6681 Test Plan: Some manual testing (above) and unit test covering key metrics is included Reviewed By: ltamasi Differential Revision: D21134113 Pulled By: pdillinger fbshipit-source-id: c11497b5f00f4ffdfe919823904e52d0a1a91d87
5 years ago
CleanupContext* context, bool* overwritten) {
assert(overwritten != nullptr && *overwritten == false);
Fix use-after-free threading bug in ClockCache (#8261) Summary: In testing for https://github.com/facebook/rocksdb/issues/8225 I found cache_bench would crash with -use_clock_cache, as well as db_bench -use_clock_cache, but not single-threaded. Smaller cache size hits failure much faster. ASAN reported the failuer as calling malloc_usable_size on the `key` pointer of a ClockCache handle after it was reportedly freed. On detailed inspection I found this bad sequence of operations for a cache entry: state=InCache=1,refs=1 [thread 1] Start ClockCacheShard::Unref (from Release, no mutex) [thread 1] Decrement ref count state=InCache=1,refs=0 [thread 1] Suspend before CalcTotalCharge (no mutex) [thread 2] Start UnsetInCache (from Insert, mutex held) [thread 2] clear InCache bit state=InCache=0,refs=0 [thread 2] Calls RecycleHandle (based on pre-updated state) [thread 2] Returns to Insert which calls Cleanup which deletes `key` [thread 1] Resume ClockCacheShard::Unref [thread 1] Read `key` in CalcTotalCharge To fix this, I've added a field to the handle to store the metadata charge so that we can efficiently remember everything we need from the handle in Unref. We must not read from the handle again if we decrement the count to zero with InCache=1, which means we don't own the entry and someone else could eject/overwrite it immediately. Note before this change, on amd64 sizeof(Handle) == 56 even though there are only 48 bytes of data. Grouping together the uint32_t fields would cut it down to 48, but I've added another uint32_t, which takes it back up to 56. Not a big deal. Also fixed DisownData to cooperate with ASAN as in LRUCache. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8261 Test Plan: Manual + adding use_clock_cache to db_crashtest.py Base performance ./cache_bench -use_clock_cache Complete in 17.060 s; QPS = 2458513 New performance ./cache_bench -use_clock_cache Complete in 17.052 s; QPS = 2459695 Any difference is easily buried in small noise. Crash test shows still more bug(s) in ClockCache, so I'm expecting to disable ClockCache from production code in a follow-up PR (if we can't find and fix the bug(s)) Reviewed By: mrambacher Differential Revision: D28207358 Pulled By: pdillinger fbshipit-source-id: aa7a9322afc6f18f30e462c75dbbe4a1206eb294
4 years ago
uint32_t meta_charge =
CacheHandle::CalcMetadataCharge(key, metadata_charge_policy_);
size_t total_charge = charge + meta_charge;
Use optimized folly DistributedMutex in LRUCache when available (#10179) Summary: folly DistributedMutex is faster than standard mutexes though imposes some static obligations on usage. See https://github.com/facebook/folly/blob/main/folly/synchronization/DistributedMutex.h for details. Here we use this alternative for our Cache implementations (especially LRUCache) for better locking performance, when RocksDB is compiled with folly. Also added information about which distributed mutex implementation is being used to cache_bench output and to DB LOG. Intended follow-up: * Use DMutex in more places, perhaps improving API to support non-scoped locking * Fix linking with fbcode compiler (needs ROCKSDB_NO_FBCODE=1 currently) Credit: Thanks Siying for reminding me about this line of work that was previously left unfinished. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10179 Test Plan: for correctness, existing tests. CircleCI config updated. Also Meta-internal buck build updated. For performance, ran simultaneous before & after cache_bench. Out of three comparison runs, the middle improvement to ops/sec was +21%: Baseline: USE_CLANG=1 DEBUG_LEVEL=0 make -j24 cache_bench (fbcode compiler) ``` Complete in 20.201 s; Rough parallel ops/sec = 1584062 Thread ops/sec = 107176 Operation latency (ns): Count: 32000000 Average: 9257.9421 StdDev: 122412.04 Min: 134 Median: 3623.0493 Max: 56918500 Percentiles: P50: 3623.05 P75: 10288.02 P99: 30219.35 P99.9: 683522.04 P99.99: 7302791.63 ``` New: (add USE_FOLLY=1) ``` Complete in 16.674 s; Rough parallel ops/sec = 1919135 (+21%) Thread ops/sec = 135487 Operation latency (ns): Count: 32000000 Average: 7304.9294 StdDev: 108530.28 Min: 132 Median: 3777.6012 Max: 91030902 Percentiles: P50: 3777.60 P75: 10169.89 P99: 24504.51 P99.9: 59721.59 P99.99: 1861151.83 ``` Reviewed By: anand1976 Differential Revision: D37182983 Pulled By: pdillinger fbshipit-source-id: a17eb05f25b832b6a2c1356f5c657e831a5af8d1
2 years ago
DMutexLock l(mutex_);
bool success = EvictFromCache(total_charge, context);
bool strict = strict_capacity_limit_.load(std::memory_order_relaxed);
if (!success && (strict || !hold_reference)) {
context->to_delete_key.push_back(key.data());
if (!hold_reference) {
context->to_delete_value.emplace_back(key, value, deleter);
}
return nullptr;
}
// Grab available handle from recycle bin. If recycle bin is empty, create
// and append new handle to end of circular list.
CacheHandle* handle = nullptr;
if (!recycle_.empty()) {
handle = recycle_.back();
recycle_.pop_back();
} else {
list_.emplace_back();
handle = &list_.back();
}
// Fill handle.
handle->key = key;
handle->hash = hash;
handle->value = value;
handle->charge = charge;
Fix use-after-free threading bug in ClockCache (#8261) Summary: In testing for https://github.com/facebook/rocksdb/issues/8225 I found cache_bench would crash with -use_clock_cache, as well as db_bench -use_clock_cache, but not single-threaded. Smaller cache size hits failure much faster. ASAN reported the failuer as calling malloc_usable_size on the `key` pointer of a ClockCache handle after it was reportedly freed. On detailed inspection I found this bad sequence of operations for a cache entry: state=InCache=1,refs=1 [thread 1] Start ClockCacheShard::Unref (from Release, no mutex) [thread 1] Decrement ref count state=InCache=1,refs=0 [thread 1] Suspend before CalcTotalCharge (no mutex) [thread 2] Start UnsetInCache (from Insert, mutex held) [thread 2] clear InCache bit state=InCache=0,refs=0 [thread 2] Calls RecycleHandle (based on pre-updated state) [thread 2] Returns to Insert which calls Cleanup which deletes `key` [thread 1] Resume ClockCacheShard::Unref [thread 1] Read `key` in CalcTotalCharge To fix this, I've added a field to the handle to store the metadata charge so that we can efficiently remember everything we need from the handle in Unref. We must not read from the handle again if we decrement the count to zero with InCache=1, which means we don't own the entry and someone else could eject/overwrite it immediately. Note before this change, on amd64 sizeof(Handle) == 56 even though there are only 48 bytes of data. Grouping together the uint32_t fields would cut it down to 48, but I've added another uint32_t, which takes it back up to 56. Not a big deal. Also fixed DisownData to cooperate with ASAN as in LRUCache. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8261 Test Plan: Manual + adding use_clock_cache to db_crashtest.py Base performance ./cache_bench -use_clock_cache Complete in 17.060 s; QPS = 2458513 New performance ./cache_bench -use_clock_cache Complete in 17.052 s; QPS = 2459695 Any difference is easily buried in small noise. Crash test shows still more bug(s) in ClockCache, so I'm expecting to disable ClockCache from production code in a follow-up PR (if we can't find and fix the bug(s)) Reviewed By: mrambacher Differential Revision: D28207358 Pulled By: pdillinger fbshipit-source-id: aa7a9322afc6f18f30e462c75dbbe4a1206eb294
4 years ago
handle->meta_charge = meta_charge;
handle->deleter = deleter;
uint32_t flags = hold_reference ? kInCacheBit + kOneRef : kInCacheBit;
Fix use-after-free threading bug in ClockCache (#8261) Summary: In testing for https://github.com/facebook/rocksdb/issues/8225 I found cache_bench would crash with -use_clock_cache, as well as db_bench -use_clock_cache, but not single-threaded. Smaller cache size hits failure much faster. ASAN reported the failuer as calling malloc_usable_size on the `key` pointer of a ClockCache handle after it was reportedly freed. On detailed inspection I found this bad sequence of operations for a cache entry: state=InCache=1,refs=1 [thread 1] Start ClockCacheShard::Unref (from Release, no mutex) [thread 1] Decrement ref count state=InCache=1,refs=0 [thread 1] Suspend before CalcTotalCharge (no mutex) [thread 2] Start UnsetInCache (from Insert, mutex held) [thread 2] clear InCache bit state=InCache=0,refs=0 [thread 2] Calls RecycleHandle (based on pre-updated state) [thread 2] Returns to Insert which calls Cleanup which deletes `key` [thread 1] Resume ClockCacheShard::Unref [thread 1] Read `key` in CalcTotalCharge To fix this, I've added a field to the handle to store the metadata charge so that we can efficiently remember everything we need from the handle in Unref. We must not read from the handle again if we decrement the count to zero with InCache=1, which means we don't own the entry and someone else could eject/overwrite it immediately. Note before this change, on amd64 sizeof(Handle) == 56 even though there are only 48 bytes of data. Grouping together the uint32_t fields would cut it down to 48, but I've added another uint32_t, which takes it back up to 56. Not a big deal. Also fixed DisownData to cooperate with ASAN as in LRUCache. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8261 Test Plan: Manual + adding use_clock_cache to db_crashtest.py Base performance ./cache_bench -use_clock_cache Complete in 17.060 s; QPS = 2458513 New performance ./cache_bench -use_clock_cache Complete in 17.052 s; QPS = 2459695 Any difference is easily buried in small noise. Crash test shows still more bug(s) in ClockCache, so I'm expecting to disable ClockCache from production code in a follow-up PR (if we can't find and fix the bug(s)) Reviewed By: mrambacher Differential Revision: D28207358 Pulled By: pdillinger fbshipit-source-id: aa7a9322afc6f18f30e462c75dbbe4a1206eb294
4 years ago
// TODO investigate+fix suspected race condition:
// [thread 1] Lookup starts, up to Ref()
// [thread 2] Erase/evict the entry just looked up
// [thread 1] Ref() the handle, even though it's in the recycle bin
// [thread 2] Insert with recycling that handle
// Here we obliterate the other thread's Ref
// Possible fix: never blindly overwrite the flags, but only make
// relative updates (fetch_add, etc).
handle->flags.store(flags, std::memory_order_relaxed);
HashTable::accessor accessor;
if (table_.find(accessor, ClockCacheKey(key, hash))) {
Stats for redundant insertions into block cache (#6681) Summary: Since read threads do not coordinate on loading data into block cache, two threads between Lookup and Insert can end up loading and inserting the same data. This is particularly concerning with cache_index_and_filter_blocks since those are hot and more likely to be race targets if ejected from (or not pre-populated in) the cache. Particularly with moves toward disaggregated / network storage, the cost of redundant retrieval might be high, and we should at least have some hard statistics from which we can estimate impact. Example with full filter thrashing "cliff": $ ./db_bench --benchmarks=fillrandom --num=15000000 --cache_index_and_filter_blocks -bloom_bits=10 ... $ ./db_bench --db=/tmp/rocksdbtest-172704/dbbench --use_existing_db --benchmarks=readrandom,stats --num=200000 --cache_index_and_filter_blocks --cache_size=$((130 * 1024 * 1024)) --bloom_bits=10 --threads=16 -statistics 2>&1 | egrep '^rocksdb.block.cache.(.*add|.*redundant)' | grep -v compress | sort rocksdb.block.cache.add COUNT : 14181 rocksdb.block.cache.add.failures COUNT : 0 rocksdb.block.cache.add.redundant COUNT : 476 rocksdb.block.cache.data.add COUNT : 12749 rocksdb.block.cache.data.add.redundant COUNT : 18 rocksdb.block.cache.filter.add COUNT : 1003 rocksdb.block.cache.filter.add.redundant COUNT : 217 rocksdb.block.cache.index.add COUNT : 429 rocksdb.block.cache.index.add.redundant COUNT : 241 $ ./db_bench --db=/tmp/rocksdbtest-172704/dbbench --use_existing_db --benchmarks=readrandom,stats --num=200000 --cache_index_and_filter_blocks --cache_size=$((120 * 1024 * 1024)) --bloom_bits=10 --threads=16 -statistics 2>&1 | egrep '^rocksdb.block.cache.(.*add|.*redundant)' | grep -v compress | sort rocksdb.block.cache.add COUNT : 1182223 rocksdb.block.cache.add.failures COUNT : 0 rocksdb.block.cache.add.redundant COUNT : 302728 rocksdb.block.cache.data.add COUNT : 31425 rocksdb.block.cache.data.add.redundant COUNT : 12 rocksdb.block.cache.filter.add COUNT : 795455 rocksdb.block.cache.filter.add.redundant COUNT : 130238 rocksdb.block.cache.index.add COUNT : 355343 rocksdb.block.cache.index.add.redundant COUNT : 172478 Pull Request resolved: https://github.com/facebook/rocksdb/pull/6681 Test Plan: Some manual testing (above) and unit test covering key metrics is included Reviewed By: ltamasi Differential Revision: D21134113 Pulled By: pdillinger fbshipit-source-id: c11497b5f00f4ffdfe919823904e52d0a1a91d87
5 years ago
*overwritten = true;
CacheHandle* existing_handle = accessor->second;
table_.erase(accessor);
UnsetInCache(existing_handle, context);
}
table_.insert(HashTable::value_type(ClockCacheKey(key, hash), handle));
if (hold_reference) {
pinned_usage_.fetch_add(total_charge, std::memory_order_relaxed);
}
usage_.fetch_add(total_charge, std::memory_order_relaxed);
return handle;
}
Status ClockCacheShard::Insert(const Slice& key, uint32_t hash, void* value,
size_t charge,
void (*deleter)(const Slice& key, void* value),
Cache::Handle** out_handle,
Cache::Priority /*priority*/) {
CleanupContext context;
HashTable::accessor accessor;
char* key_data = new char[key.size()];
memcpy(key_data, key.data(), key.size());
Slice key_copy(key_data, key.size());
Stats for redundant insertions into block cache (#6681) Summary: Since read threads do not coordinate on loading data into block cache, two threads between Lookup and Insert can end up loading and inserting the same data. This is particularly concerning with cache_index_and_filter_blocks since those are hot and more likely to be race targets if ejected from (or not pre-populated in) the cache. Particularly with moves toward disaggregated / network storage, the cost of redundant retrieval might be high, and we should at least have some hard statistics from which we can estimate impact. Example with full filter thrashing "cliff": $ ./db_bench --benchmarks=fillrandom --num=15000000 --cache_index_and_filter_blocks -bloom_bits=10 ... $ ./db_bench --db=/tmp/rocksdbtest-172704/dbbench --use_existing_db --benchmarks=readrandom,stats --num=200000 --cache_index_and_filter_blocks --cache_size=$((130 * 1024 * 1024)) --bloom_bits=10 --threads=16 -statistics 2>&1 | egrep '^rocksdb.block.cache.(.*add|.*redundant)' | grep -v compress | sort rocksdb.block.cache.add COUNT : 14181 rocksdb.block.cache.add.failures COUNT : 0 rocksdb.block.cache.add.redundant COUNT : 476 rocksdb.block.cache.data.add COUNT : 12749 rocksdb.block.cache.data.add.redundant COUNT : 18 rocksdb.block.cache.filter.add COUNT : 1003 rocksdb.block.cache.filter.add.redundant COUNT : 217 rocksdb.block.cache.index.add COUNT : 429 rocksdb.block.cache.index.add.redundant COUNT : 241 $ ./db_bench --db=/tmp/rocksdbtest-172704/dbbench --use_existing_db --benchmarks=readrandom,stats --num=200000 --cache_index_and_filter_blocks --cache_size=$((120 * 1024 * 1024)) --bloom_bits=10 --threads=16 -statistics 2>&1 | egrep '^rocksdb.block.cache.(.*add|.*redundant)' | grep -v compress | sort rocksdb.block.cache.add COUNT : 1182223 rocksdb.block.cache.add.failures COUNT : 0 rocksdb.block.cache.add.redundant COUNT : 302728 rocksdb.block.cache.data.add COUNT : 31425 rocksdb.block.cache.data.add.redundant COUNT : 12 rocksdb.block.cache.filter.add COUNT : 795455 rocksdb.block.cache.filter.add.redundant COUNT : 130238 rocksdb.block.cache.index.add COUNT : 355343 rocksdb.block.cache.index.add.redundant COUNT : 172478 Pull Request resolved: https://github.com/facebook/rocksdb/pull/6681 Test Plan: Some manual testing (above) and unit test covering key metrics is included Reviewed By: ltamasi Differential Revision: D21134113 Pulled By: pdillinger fbshipit-source-id: c11497b5f00f4ffdfe919823904e52d0a1a91d87
5 years ago
bool overwritten = false;
CacheHandle* handle = Insert(key_copy, hash, value, charge, deleter,
Stats for redundant insertions into block cache (#6681) Summary: Since read threads do not coordinate on loading data into block cache, two threads between Lookup and Insert can end up loading and inserting the same data. This is particularly concerning with cache_index_and_filter_blocks since those are hot and more likely to be race targets if ejected from (or not pre-populated in) the cache. Particularly with moves toward disaggregated / network storage, the cost of redundant retrieval might be high, and we should at least have some hard statistics from which we can estimate impact. Example with full filter thrashing "cliff": $ ./db_bench --benchmarks=fillrandom --num=15000000 --cache_index_and_filter_blocks -bloom_bits=10 ... $ ./db_bench --db=/tmp/rocksdbtest-172704/dbbench --use_existing_db --benchmarks=readrandom,stats --num=200000 --cache_index_and_filter_blocks --cache_size=$((130 * 1024 * 1024)) --bloom_bits=10 --threads=16 -statistics 2>&1 | egrep '^rocksdb.block.cache.(.*add|.*redundant)' | grep -v compress | sort rocksdb.block.cache.add COUNT : 14181 rocksdb.block.cache.add.failures COUNT : 0 rocksdb.block.cache.add.redundant COUNT : 476 rocksdb.block.cache.data.add COUNT : 12749 rocksdb.block.cache.data.add.redundant COUNT : 18 rocksdb.block.cache.filter.add COUNT : 1003 rocksdb.block.cache.filter.add.redundant COUNT : 217 rocksdb.block.cache.index.add COUNT : 429 rocksdb.block.cache.index.add.redundant COUNT : 241 $ ./db_bench --db=/tmp/rocksdbtest-172704/dbbench --use_existing_db --benchmarks=readrandom,stats --num=200000 --cache_index_and_filter_blocks --cache_size=$((120 * 1024 * 1024)) --bloom_bits=10 --threads=16 -statistics 2>&1 | egrep '^rocksdb.block.cache.(.*add|.*redundant)' | grep -v compress | sort rocksdb.block.cache.add COUNT : 1182223 rocksdb.block.cache.add.failures COUNT : 0 rocksdb.block.cache.add.redundant COUNT : 302728 rocksdb.block.cache.data.add COUNT : 31425 rocksdb.block.cache.data.add.redundant COUNT : 12 rocksdb.block.cache.filter.add COUNT : 795455 rocksdb.block.cache.filter.add.redundant COUNT : 130238 rocksdb.block.cache.index.add COUNT : 355343 rocksdb.block.cache.index.add.redundant COUNT : 172478 Pull Request resolved: https://github.com/facebook/rocksdb/pull/6681 Test Plan: Some manual testing (above) and unit test covering key metrics is included Reviewed By: ltamasi Differential Revision: D21134113 Pulled By: pdillinger fbshipit-source-id: c11497b5f00f4ffdfe919823904e52d0a1a91d87
5 years ago
out_handle != nullptr, &context, &overwritten);
Status s;
if (out_handle != nullptr) {
if (handle == nullptr) {
s = Status::Incomplete("Insert failed due to CLOCK cache being full.");
} else {
*out_handle = reinterpret_cast<Cache::Handle*>(handle);
}
}
Stats for redundant insertions into block cache (#6681) Summary: Since read threads do not coordinate on loading data into block cache, two threads between Lookup and Insert can end up loading and inserting the same data. This is particularly concerning with cache_index_and_filter_blocks since those are hot and more likely to be race targets if ejected from (or not pre-populated in) the cache. Particularly with moves toward disaggregated / network storage, the cost of redundant retrieval might be high, and we should at least have some hard statistics from which we can estimate impact. Example with full filter thrashing "cliff": $ ./db_bench --benchmarks=fillrandom --num=15000000 --cache_index_and_filter_blocks -bloom_bits=10 ... $ ./db_bench --db=/tmp/rocksdbtest-172704/dbbench --use_existing_db --benchmarks=readrandom,stats --num=200000 --cache_index_and_filter_blocks --cache_size=$((130 * 1024 * 1024)) --bloom_bits=10 --threads=16 -statistics 2>&1 | egrep '^rocksdb.block.cache.(.*add|.*redundant)' | grep -v compress | sort rocksdb.block.cache.add COUNT : 14181 rocksdb.block.cache.add.failures COUNT : 0 rocksdb.block.cache.add.redundant COUNT : 476 rocksdb.block.cache.data.add COUNT : 12749 rocksdb.block.cache.data.add.redundant COUNT : 18 rocksdb.block.cache.filter.add COUNT : 1003 rocksdb.block.cache.filter.add.redundant COUNT : 217 rocksdb.block.cache.index.add COUNT : 429 rocksdb.block.cache.index.add.redundant COUNT : 241 $ ./db_bench --db=/tmp/rocksdbtest-172704/dbbench --use_existing_db --benchmarks=readrandom,stats --num=200000 --cache_index_and_filter_blocks --cache_size=$((120 * 1024 * 1024)) --bloom_bits=10 --threads=16 -statistics 2>&1 | egrep '^rocksdb.block.cache.(.*add|.*redundant)' | grep -v compress | sort rocksdb.block.cache.add COUNT : 1182223 rocksdb.block.cache.add.failures COUNT : 0 rocksdb.block.cache.add.redundant COUNT : 302728 rocksdb.block.cache.data.add COUNT : 31425 rocksdb.block.cache.data.add.redundant COUNT : 12 rocksdb.block.cache.filter.add COUNT : 795455 rocksdb.block.cache.filter.add.redundant COUNT : 130238 rocksdb.block.cache.index.add COUNT : 355343 rocksdb.block.cache.index.add.redundant COUNT : 172478 Pull Request resolved: https://github.com/facebook/rocksdb/pull/6681 Test Plan: Some manual testing (above) and unit test covering key metrics is included Reviewed By: ltamasi Differential Revision: D21134113 Pulled By: pdillinger fbshipit-source-id: c11497b5f00f4ffdfe919823904e52d0a1a91d87
5 years ago
if (overwritten) {
assert(s.ok());
s = Status::OkOverwritten();
}
Cleanup(context);
return s;
}
Cache::Handle* ClockCacheShard::Lookup(const Slice& key, uint32_t hash) {
HashTable::const_accessor accessor;
if (!table_.find(accessor, ClockCacheKey(key, hash))) {
return nullptr;
}
CacheHandle* handle = accessor->second;
accessor.release();
// Ref() could fail if another thread sneak in and evict/erase the cache
// entry before we are able to hold reference.
if (!Ref(reinterpret_cast<Cache::Handle*>(handle))) {
return nullptr;
}
// Double check the key since the handle may now representing another key
// if other threads sneak in, evict/erase the entry and re-used the handle
// for another cache entry.
if (hash != handle->hash || key != handle->key) {
CleanupContext context;
Unref(handle, false, &context);
// It is possible Unref() delete the entry, so we need to cleanup.
Cleanup(context);
return nullptr;
}
return reinterpret_cast<Cache::Handle*>(handle);
}
bool ClockCacheShard::Release(Cache::Handle* h, bool erase_if_last_ref) {
CleanupContext context;
CacheHandle* handle = reinterpret_cast<CacheHandle*>(h);
bool erased = Unref(handle, true, &context);
if (erase_if_last_ref && !erased) {
erased = EraseAndConfirm(handle->key, handle->hash, &context);
}
Cleanup(context);
return erased;
}
void ClockCacheShard::Erase(const Slice& key, uint32_t hash) {
CleanupContext context;
EraseAndConfirm(key, hash, &context);
Cleanup(context);
}
bool ClockCacheShard::EraseAndConfirm(const Slice& key, uint32_t hash,
CleanupContext* context) {
Use optimized folly DistributedMutex in LRUCache when available (#10179) Summary: folly DistributedMutex is faster than standard mutexes though imposes some static obligations on usage. See https://github.com/facebook/folly/blob/main/folly/synchronization/DistributedMutex.h for details. Here we use this alternative for our Cache implementations (especially LRUCache) for better locking performance, when RocksDB is compiled with folly. Also added information about which distributed mutex implementation is being used to cache_bench output and to DB LOG. Intended follow-up: * Use DMutex in more places, perhaps improving API to support non-scoped locking * Fix linking with fbcode compiler (needs ROCKSDB_NO_FBCODE=1 currently) Credit: Thanks Siying for reminding me about this line of work that was previously left unfinished. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10179 Test Plan: for correctness, existing tests. CircleCI config updated. Also Meta-internal buck build updated. For performance, ran simultaneous before & after cache_bench. Out of three comparison runs, the middle improvement to ops/sec was +21%: Baseline: USE_CLANG=1 DEBUG_LEVEL=0 make -j24 cache_bench (fbcode compiler) ``` Complete in 20.201 s; Rough parallel ops/sec = 1584062 Thread ops/sec = 107176 Operation latency (ns): Count: 32000000 Average: 9257.9421 StdDev: 122412.04 Min: 134 Median: 3623.0493 Max: 56918500 Percentiles: P50: 3623.05 P75: 10288.02 P99: 30219.35 P99.9: 683522.04 P99.99: 7302791.63 ``` New: (add USE_FOLLY=1) ``` Complete in 16.674 s; Rough parallel ops/sec = 1919135 (+21%) Thread ops/sec = 135487 Operation latency (ns): Count: 32000000 Average: 7304.9294 StdDev: 108530.28 Min: 132 Median: 3777.6012 Max: 91030902 Percentiles: P50: 3777.60 P75: 10169.89 P99: 24504.51 P99.9: 59721.59 P99.99: 1861151.83 ``` Reviewed By: anand1976 Differential Revision: D37182983 Pulled By: pdillinger fbshipit-source-id: a17eb05f25b832b6a2c1356f5c657e831a5af8d1
2 years ago
DMutexLock l(mutex_);
HashTable::accessor accessor;
bool erased = false;
if (table_.find(accessor, ClockCacheKey(key, hash))) {
CacheHandle* handle = accessor->second;
table_.erase(accessor);
erased = UnsetInCache(handle, context);
}
return erased;
}
void ClockCacheShard::EraseUnRefEntries() {
CleanupContext context;
{
Use optimized folly DistributedMutex in LRUCache when available (#10179) Summary: folly DistributedMutex is faster than standard mutexes though imposes some static obligations on usage. See https://github.com/facebook/folly/blob/main/folly/synchronization/DistributedMutex.h for details. Here we use this alternative for our Cache implementations (especially LRUCache) for better locking performance, when RocksDB is compiled with folly. Also added information about which distributed mutex implementation is being used to cache_bench output and to DB LOG. Intended follow-up: * Use DMutex in more places, perhaps improving API to support non-scoped locking * Fix linking with fbcode compiler (needs ROCKSDB_NO_FBCODE=1 currently) Credit: Thanks Siying for reminding me about this line of work that was previously left unfinished. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10179 Test Plan: for correctness, existing tests. CircleCI config updated. Also Meta-internal buck build updated. For performance, ran simultaneous before & after cache_bench. Out of three comparison runs, the middle improvement to ops/sec was +21%: Baseline: USE_CLANG=1 DEBUG_LEVEL=0 make -j24 cache_bench (fbcode compiler) ``` Complete in 20.201 s; Rough parallel ops/sec = 1584062 Thread ops/sec = 107176 Operation latency (ns): Count: 32000000 Average: 9257.9421 StdDev: 122412.04 Min: 134 Median: 3623.0493 Max: 56918500 Percentiles: P50: 3623.05 P75: 10288.02 P99: 30219.35 P99.9: 683522.04 P99.99: 7302791.63 ``` New: (add USE_FOLLY=1) ``` Complete in 16.674 s; Rough parallel ops/sec = 1919135 (+21%) Thread ops/sec = 135487 Operation latency (ns): Count: 32000000 Average: 7304.9294 StdDev: 108530.28 Min: 132 Median: 3777.6012 Max: 91030902 Percentiles: P50: 3777.60 P75: 10169.89 P99: 24504.51 P99.9: 59721.59 P99.99: 1861151.83 ``` Reviewed By: anand1976 Differential Revision: D37182983 Pulled By: pdillinger fbshipit-source-id: a17eb05f25b832b6a2c1356f5c657e831a5af8d1
2 years ago
DMutexLock l(mutex_);
table_.clear();
for (auto& handle : list_) {
UnsetInCache(&handle, &context);
}
}
Cleanup(context);
}
class ClockCache final : public ShardedCache {
public:
ClockCache(size_t capacity, int num_shard_bits, bool strict_capacity_limit,
CacheMetadataChargePolicy metadata_charge_policy)
: ShardedCache(capacity, num_shard_bits, strict_capacity_limit) {
int num_shards = 1 << num_shard_bits;
shards_ = new ClockCacheShard[num_shards];
for (int i = 0; i < num_shards; i++) {
shards_[i].set_metadata_charge_policy(metadata_charge_policy);
}
SetCapacity(capacity);
SetStrictCapacityLimit(strict_capacity_limit);
}
~ClockCache() override { delete[] shards_; }
const char* Name() const override { return "ClockCache"; }
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
CacheShard* GetShard(uint32_t shard) override {
return reinterpret_cast<CacheShard*>(&shards_[shard]);
}
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
const CacheShard* GetShard(uint32_t shard) const override {
return reinterpret_cast<CacheShard*>(&shards_[shard]);
}
void* Value(Handle* handle) override {
return reinterpret_cast<const CacheHandle*>(handle)->value;
}
size_t GetCharge(Handle* handle) const override {
return reinterpret_cast<const CacheHandle*>(handle)->charge;
}
uint32_t GetHash(Handle* handle) const override {
return reinterpret_cast<const CacheHandle*>(handle)->hash;
}
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
DeleterFn GetDeleter(Handle* handle) const override {
return reinterpret_cast<const CacheHandle*>(handle)->deleter;
}
Fix use-after-free threading bug in ClockCache (#8261) Summary: In testing for https://github.com/facebook/rocksdb/issues/8225 I found cache_bench would crash with -use_clock_cache, as well as db_bench -use_clock_cache, but not single-threaded. Smaller cache size hits failure much faster. ASAN reported the failuer as calling malloc_usable_size on the `key` pointer of a ClockCache handle after it was reportedly freed. On detailed inspection I found this bad sequence of operations for a cache entry: state=InCache=1,refs=1 [thread 1] Start ClockCacheShard::Unref (from Release, no mutex) [thread 1] Decrement ref count state=InCache=1,refs=0 [thread 1] Suspend before CalcTotalCharge (no mutex) [thread 2] Start UnsetInCache (from Insert, mutex held) [thread 2] clear InCache bit state=InCache=0,refs=0 [thread 2] Calls RecycleHandle (based on pre-updated state) [thread 2] Returns to Insert which calls Cleanup which deletes `key` [thread 1] Resume ClockCacheShard::Unref [thread 1] Read `key` in CalcTotalCharge To fix this, I've added a field to the handle to store the metadata charge so that we can efficiently remember everything we need from the handle in Unref. We must not read from the handle again if we decrement the count to zero with InCache=1, which means we don't own the entry and someone else could eject/overwrite it immediately. Note before this change, on amd64 sizeof(Handle) == 56 even though there are only 48 bytes of data. Grouping together the uint32_t fields would cut it down to 48, but I've added another uint32_t, which takes it back up to 56. Not a big deal. Also fixed DisownData to cooperate with ASAN as in LRUCache. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8261 Test Plan: Manual + adding use_clock_cache to db_crashtest.py Base performance ./cache_bench -use_clock_cache Complete in 17.060 s; QPS = 2458513 New performance ./cache_bench -use_clock_cache Complete in 17.052 s; QPS = 2459695 Any difference is easily buried in small noise. Crash test shows still more bug(s) in ClockCache, so I'm expecting to disable ClockCache from production code in a follow-up PR (if we can't find and fix the bug(s)) Reviewed By: mrambacher Differential Revision: D28207358 Pulled By: pdillinger fbshipit-source-id: aa7a9322afc6f18f30e462c75dbbe4a1206eb294
4 years ago
void DisownData() override {
// Leak data only if that won't generate an ASAN/valgrind warning
if (!kMustFreeHeapAllocations) {
shards_ = nullptr;
}
Fix use-after-free threading bug in ClockCache (#8261) Summary: In testing for https://github.com/facebook/rocksdb/issues/8225 I found cache_bench would crash with -use_clock_cache, as well as db_bench -use_clock_cache, but not single-threaded. Smaller cache size hits failure much faster. ASAN reported the failuer as calling malloc_usable_size on the `key` pointer of a ClockCache handle after it was reportedly freed. On detailed inspection I found this bad sequence of operations for a cache entry: state=InCache=1,refs=1 [thread 1] Start ClockCacheShard::Unref (from Release, no mutex) [thread 1] Decrement ref count state=InCache=1,refs=0 [thread 1] Suspend before CalcTotalCharge (no mutex) [thread 2] Start UnsetInCache (from Insert, mutex held) [thread 2] clear InCache bit state=InCache=0,refs=0 [thread 2] Calls RecycleHandle (based on pre-updated state) [thread 2] Returns to Insert which calls Cleanup which deletes `key` [thread 1] Resume ClockCacheShard::Unref [thread 1] Read `key` in CalcTotalCharge To fix this, I've added a field to the handle to store the metadata charge so that we can efficiently remember everything we need from the handle in Unref. We must not read from the handle again if we decrement the count to zero with InCache=1, which means we don't own the entry and someone else could eject/overwrite it immediately. Note before this change, on amd64 sizeof(Handle) == 56 even though there are only 48 bytes of data. Grouping together the uint32_t fields would cut it down to 48, but I've added another uint32_t, which takes it back up to 56. Not a big deal. Also fixed DisownData to cooperate with ASAN as in LRUCache. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8261 Test Plan: Manual + adding use_clock_cache to db_crashtest.py Base performance ./cache_bench -use_clock_cache Complete in 17.060 s; QPS = 2458513 New performance ./cache_bench -use_clock_cache Complete in 17.052 s; QPS = 2459695 Any difference is easily buried in small noise. Crash test shows still more bug(s) in ClockCache, so I'm expecting to disable ClockCache from production code in a follow-up PR (if we can't find and fix the bug(s)) Reviewed By: mrambacher Differential Revision: D28207358 Pulled By: pdillinger fbshipit-source-id: aa7a9322afc6f18f30e462c75dbbe4a1206eb294
4 years ago
}
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
void WaitAll(std::vector<Handle*>& /*handles*/) override {}
private:
ClockCacheShard* shards_;
};
} // end anonymous namespace
std::shared_ptr<Cache> NewClockCache(
size_t capacity, int num_shard_bits, bool strict_capacity_limit,
CacheMetadataChargePolicy metadata_charge_policy) {
if (num_shard_bits < 0) {
num_shard_bits = GetDefaultCacheShardBits(capacity);
}
return std::make_shared<ClockCache>(
capacity, num_shard_bits, strict_capacity_limit, metadata_charge_policy);
}
} // namespace ROCKSDB_NAMESPACE
#endif // SUPPORT_CLOCK_CACHE