|
|
|
// 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 "table/block_based/block_based_table_reader.h"
|
|
|
|
|
|
|
|
#include <algorithm>
|
|
|
|
#include <array>
|
Always verify SST unique IDs on SST file open (#10532)
Summary:
Although we've been tracking SST unique IDs in the DB manifest
unconditionally, checking has been opt-in and with an extra pass at DB::Open
time. This changes the behavior of `verify_sst_unique_id_in_manifest` to
check unique ID against manifest every time an SST file is opened through
table cache (normal DB operations), replacing the explicit pass over files
at DB::Open time. This change also enables the option by default and
removes the "EXPERIMENTAL" designation.
One possible criticism is that the option no longer ensures the integrity
of a DB at Open time. This is far from an all-or-nothing issue. Verifying
the IDs of all SST files hardly ensures all the data in the DB is readable.
(VerifyChecksum is supposed to do that.) Also, with
max_open_files=-1 (default, extremely common), all SST files are
opened at DB::Open time anyway.
Implementation details:
* `VerifySstUniqueIdInManifest()` functions are the extra/explicit pass
that is now removed.
* Unit tests that manipulate/corrupt table properties have to opt out of
this check, because that corrupts the "actual" unique id. (And even for
testing we don't currently have a mechanism to set "no unique id"
in the in-memory file metadata for new files.)
* A lot of other unit test churn relates to (a) default checking on, and
(b) checking on SST open even without DB::Open (e.g. on flush)
* Use `FileMetaData` for more `TableCache` operations (in place of
`FileDescriptor`) so that we have access to the unique_id whenever
we might need to open an SST file. **There is the possibility of
performance impact because we can no longer use the more
localized `fd` part of an `FdWithKeyRange` but instead follow the
`file_metadata` pointer. However, this change (possible regression)
is only done for `GetMemoryUsageByTableReaders`.**
* Removed a completely unnecessary constructor overload of
`TableReaderOptions`
Possible follow-up:
* Verification only happens when opening through table cache. Are there
more places where this should happen?
* Improve error message when there is a file size mismatch vs. manifest
(FIXME added in the appropriate place).
* I'm not sure there's a justification for `FileDescriptor` to be distinct from
`FileMetaData`.
* I'm skeptical that `FdWithKeyRange` really still makes sense for
optimizing some data locality by duplicating some data in memory, but I
could be wrong.
* An unnecessary overload of NewTableReader was recently added, in
the public API nonetheless (though unusable there). It should be cleaned
up to put most things under `TableReaderOptions`.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10532
Test Plan:
updated unit tests
Performance test showing no significant difference (just noise I think):
`./db_bench -benchmarks=readwhilewriting[-X10] -num=3000000 -disable_wal=1 -bloom_bits=8 -write_buffer_size=1000000 -target_file_size_base=1000000`
Before: readwhilewriting [AVG 10 runs] : 68702 (± 6932) ops/sec
After: readwhilewriting [AVG 10 runs] : 68239 (± 7198) ops/sec
Reviewed By: jay-zhuang
Differential Revision: D38765551
Pulled By: pdillinger
fbshipit-source-id: a827a708155f12344ab2a5c16e7701c7636da4c2
2 years ago
|
|
|
#include <atomic>
|
|
|
|
#include <cstdint>
|
|
|
|
#include <limits>
|
|
|
|
#include <memory>
|
|
|
|
#include <string>
|
|
|
|
#include <unordered_set>
|
|
|
|
#include <utility>
|
|
|
|
#include <vector>
|
|
|
|
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
#include "block_cache.h"
|
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
|
|
|
#include "cache/cache_entry_roles.h"
|
New stable, fixed-length cache keys (#9126)
Summary:
This change standardizes on a new 16-byte cache key format for
block cache (incl compressed and secondary) and persistent cache (but
not table cache and row cache).
The goal is a really fast cache key with practically ideal stability and
uniqueness properties without external dependencies (e.g. from FileSystem).
A fixed key size of 16 bytes should enable future optimizations to the
concurrent hash table for block cache, which is a heavy CPU user /
bottleneck, but there appears to be measurable performance improvement
even with no changes to LRUCache.
This change replaces a lot of disjointed and ugly code handling cache
keys with calls to a simple, clean new internal API (cache_key.h).
(Preserving the old cache key logic under an option would be very ugly
and likely negate the performance gain of the new approach. Complete
replacement carries some inherent risk, but I think that's acceptable
with sufficient analysis and testing.)
The scheme for encoding new cache keys is complicated but explained
in cache_key.cc.
Also: EndianSwapValue is moved to math.h to be next to other bit
operations. (Explains some new include "math.h".) ReverseBits operation
added and unit tests added to hash_test for both.
Fixes https://github.com/facebook/rocksdb/issues/7405 (presuming a root cause)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9126
Test Plan:
### Basic correctness
Several tests needed updates to work with the new functionality, mostly
because we are no longer relying on filesystem for stable cache keys
so table builders & readers need more context info to agree on cache
keys. This functionality is so core, a huge number of existing tests
exercise the cache key functionality.
### Performance
Create db with
`TEST_TMPDIR=/dev/shm ./db_bench -bloom_bits=10 -benchmarks=fillrandom -num=3000000 -partition_index_and_filters`
And test performance with
`TEST_TMPDIR=/dev/shm ./db_bench -readonly -use_existing_db -bloom_bits=10 -benchmarks=readrandom -num=3000000 -duration=30 -cache_index_and_filter_blocks -cache_size=250000 -threads=4`
using DEBUG_LEVEL=0 and simultaneous before & after runs.
Before ops/sec, avg over 100 runs: 121924
After ops/sec, avg over 100 runs: 125385 (+2.8%)
### Collision probability
I have built a tool, ./cache_bench -stress_cache_key to broadly simulate host-wide cache activity
over many months, by making some pessimistic simplifying assumptions:
* Every generated file has a cache entry for every byte offset in the file (contiguous range of cache keys)
* All of every file is cached for its entire lifetime
We use a simple table with skewed address assignment and replacement on address collision
to simulate files coming & going, with quite a variance (super-Poisson) in ages. Some output
with `./cache_bench -stress_cache_key -sck_keep_bits=40`:
```
Total cache or DBs size: 32TiB Writing 925.926 MiB/s or 76.2939TiB/day
Multiply by 9.22337e+18 to correct for simulation losses (but still assume whole file cached)
```
These come from default settings of 2.5M files per day of 32 MB each, and
`-sck_keep_bits=40` means that to represent a single file, we are only keeping 40 bits of
the 128-bit cache key. With file size of 2\*\*25 contiguous keys (pessimistic), our simulation
is about 2\*\*(128-40-25) or about 9 billion billion times more prone to collision than reality.
More default assumptions, relatively pessimistic:
* 100 DBs in same process (doesn't matter much)
* Re-open DB in same process (new session ID related to old session ID) on average
every 100 files generated
* Restart process (all new session IDs unrelated to old) 24 times per day
After enough data, we get a result at the end:
```
(keep 40 bits) 17 collisions after 2 x 90 days, est 10.5882 days between (9.76592e+19 corrected)
```
If we believe the (pessimistic) simulation and the mathematical generalization, we would need to run a billion machines all for 97 billion days to expect a cache key collision. To help verify that our generalization ("corrected") is robust, we can make our simulation more precise with `-sck_keep_bits=41` and `42`, which takes more running time to get enough data:
```
(keep 41 bits) 16 collisions after 4 x 90 days, est 22.5 days between (1.03763e+20 corrected)
(keep 42 bits) 19 collisions after 10 x 90 days, est 47.3684 days between (1.09224e+20 corrected)
```
The generalized prediction still holds. With the `-sck_randomize` option, we can see that we are beating "random" cache keys (except offsets still non-randomized) by a modest amount (roughly 20x less collision prone than random), which should make us reasonably comfortable even in "degenerate" cases:
```
197 collisions after 1 x 90 days, est 0.456853 days between (4.21372e+18 corrected)
```
I've run other tests to validate other conditions behave as expected, never behaving "worse than random" unless we start chopping off structured data.
Reviewed By: zhichao-cao
Differential Revision: D33171746
Pulled By: pdillinger
fbshipit-source-id: f16a57e369ed37be5e7e33525ace848d0537c88f
3 years ago
|
|
|
#include "cache/cache_key.h"
|
Improve / clean up meta block code & integrity (#9163)
Summary:
* Checksums are now checked on meta blocks unless specifically
suppressed or not applicable (e.g. plain table). (Was other way around.)
This means a number of cases that were not checking checksums now are,
including direct read TableProperties in Version::GetTableProperties
(fixed in meta_blocks ReadTableProperties), reading any block from
PersistentCache (fixed in BlockFetcher), read TableProperties in
SstFileDumper (ldb/sst_dump/BackupEngine) before table reader open,
maybe more.
* For that to work, I moved the global_seqno+TableProperties checksum
logic to the shared table/ code, because that is used by many utilies
such as SstFileDumper.
* Also for that to work, we have to know when we're dealing with a block
that has a checksum (trailer), so added that capability to Footer based
on magic number, and from there BlockFetcher.
* Knowledge of trailer presence has also fixed a problem where other
table formats were reading blocks including bytes for a non-existant
trailer--and awkwardly kind-of not using them, e.g. no shared code
checking checksums. (BlockFetcher compression type was populated
incorrectly.) Now we only read what is needed.
* Minimized code duplication and differing/incompatible/awkward
abstractions in meta_blocks.{cc,h} (e.g. SeekTo in metaindex block
without parsing block handle)
* Moved some meta block handling code from table_properties*.*
* Moved some code specific to block-based table from shared table/ code
to BlockBasedTable class. The checksum stuff means we can't completely
separate it, but things that don't need to be in shared table/ code
should not be.
* Use unique_ptr rather than raw ptr in more places. (Note: you can
std::move from unique_ptr to shared_ptr.)
Without enhancements to GetPropertiesOfAllTablesTest (see below),
net reduction of roughly 100 lines of code.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9163
Test Plan:
existing tests and
* Enhanced DBTablePropertiesTest.GetPropertiesOfAllTablesTest to verify that
checksums are now checked on direct read of table properties by TableCache
(new test would fail before this change)
* Also enhanced DBTablePropertiesTest.GetPropertiesOfAllTablesTest to test
putting table properties under old meta name
* Also generally enhanced that same test to actually test what it was
supposed to be testing already, by kicking things out of table cache when
we don't want them there.
Reviewed By: ajkr, mrambacher
Differential Revision: D32514757
Pulled By: pdillinger
fbshipit-source-id: 507964b9311d186ae8d1131182290cbd97a99fa9
3 years ago
|
|
|
#include "db/compaction/compaction_picker.h"
|
|
|
|
#include "db/dbformat.h"
|
Introduce FullMergeV2 (eliminate memcpy from merge operators)
Summary:
This diff update the code to pin the merge operator operands while the merge operation is done, so that we can eliminate the memcpy cost, to do that we need a new public API for FullMerge that replace the std::deque<std::string> with std::vector<Slice>
This diff is stacked on top of D56493 and D56511
In this diff we
- Update FullMergeV2 arguments to be encapsulated in MergeOperationInput and MergeOperationOutput which will make it easier to add new arguments in the future
- Replace std::deque<std::string> with std::vector<Slice> to pass operands
- Replace MergeContext std::deque with std::vector (based on a simple benchmark I ran https://gist.github.com/IslamAbdelRahman/78fc86c9ab9f52b1df791e58943fb187)
- Allow FullMergeV2 output to be an existing operand
```
[Everything in Memtable | 10K operands | 10 KB each | 1 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --merge_keys=10000 --num=10000 --disable_auto_compactions --value_size=10240 --write_buffer_size=1000000000
[FullMergeV2]
readseq : 0.607 micros/op 1648235 ops/sec; 16121.2 MB/s
readseq : 0.478 micros/op 2091546 ops/sec; 20457.2 MB/s
readseq : 0.252 micros/op 3972081 ops/sec; 38850.5 MB/s
readseq : 0.237 micros/op 4218328 ops/sec; 41259.0 MB/s
readseq : 0.247 micros/op 4043927 ops/sec; 39553.2 MB/s
[master]
readseq : 3.935 micros/op 254140 ops/sec; 2485.7 MB/s
readseq : 3.722 micros/op 268657 ops/sec; 2627.7 MB/s
readseq : 3.149 micros/op 317605 ops/sec; 3106.5 MB/s
readseq : 3.125 micros/op 320024 ops/sec; 3130.1 MB/s
readseq : 4.075 micros/op 245374 ops/sec; 2400.0 MB/s
```
```
[Everything in Memtable | 10K operands | 10 KB each | 10 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --merge_keys=1000 --num=10000 --disable_auto_compactions --value_size=10240 --write_buffer_size=1000000000
[FullMergeV2]
readseq : 3.472 micros/op 288018 ops/sec; 2817.1 MB/s
readseq : 2.304 micros/op 434027 ops/sec; 4245.2 MB/s
readseq : 1.163 micros/op 859845 ops/sec; 8410.0 MB/s
readseq : 1.192 micros/op 838926 ops/sec; 8205.4 MB/s
readseq : 1.250 micros/op 800000 ops/sec; 7824.7 MB/s
[master]
readseq : 24.025 micros/op 41623 ops/sec; 407.1 MB/s
readseq : 18.489 micros/op 54086 ops/sec; 529.0 MB/s
readseq : 18.693 micros/op 53495 ops/sec; 523.2 MB/s
readseq : 23.621 micros/op 42335 ops/sec; 414.1 MB/s
readseq : 18.775 micros/op 53262 ops/sec; 521.0 MB/s
```
```
[Everything in Block cache | 10K operands | 10 KB each | 1 operand per key]
[FullMergeV2]
$ DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --num=100000 --db="/dev/shm/merge-random-10K-10KB" --cache_size=1000000000 --use_existing_db --disable_auto_compactions
readseq : 14.741 micros/op 67837 ops/sec; 663.5 MB/s
readseq : 1.029 micros/op 971446 ops/sec; 9501.6 MB/s
readseq : 0.974 micros/op 1026229 ops/sec; 10037.4 MB/s
readseq : 0.965 micros/op 1036080 ops/sec; 10133.8 MB/s
readseq : 0.943 micros/op 1060657 ops/sec; 10374.2 MB/s
[master]
readseq : 16.735 micros/op 59755 ops/sec; 584.5 MB/s
readseq : 3.029 micros/op 330151 ops/sec; 3229.2 MB/s
readseq : 3.136 micros/op 318883 ops/sec; 3119.0 MB/s
readseq : 3.065 micros/op 326245 ops/sec; 3191.0 MB/s
readseq : 3.014 micros/op 331813 ops/sec; 3245.4 MB/s
```
```
[Everything in Block cache | 10K operands | 10 KB each | 10 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --num=100000 --db="/dev/shm/merge-random-10-operands-10K-10KB" --cache_size=1000000000 --use_existing_db --disable_auto_compactions
[FullMergeV2]
readseq : 24.325 micros/op 41109 ops/sec; 402.1 MB/s
readseq : 1.470 micros/op 680272 ops/sec; 6653.7 MB/s
readseq : 1.231 micros/op 812347 ops/sec; 7945.5 MB/s
readseq : 1.091 micros/op 916590 ops/sec; 8965.1 MB/s
readseq : 1.109 micros/op 901713 ops/sec; 8819.6 MB/s
[master]
readseq : 27.257 micros/op 36687 ops/sec; 358.8 MB/s
readseq : 4.443 micros/op 225073 ops/sec; 2201.4 MB/s
readseq : 5.830 micros/op 171526 ops/sec; 1677.7 MB/s
readseq : 4.173 micros/op 239635 ops/sec; 2343.8 MB/s
readseq : 4.150 micros/op 240963 ops/sec; 2356.8 MB/s
```
Test Plan: COMPILE_WITH_ASAN=1 make check -j64
Reviewers: yhchiang, andrewkr, sdong
Reviewed By: sdong
Subscribers: lovro, andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D57075
8 years ago
|
|
|
#include "db/pinned_iterators_manager.h"
|
|
|
|
#include "file/file_prefetch_buffer.h"
|
|
|
|
#include "file/file_util.h"
|
|
|
|
#include "file/random_access_file_reader.h"
|
|
|
|
#include "logging/logging.h"
|
|
|
|
#include "monitoring/perf_context_imp.h"
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
#include "parsed_full_filter_block.h"
|
|
|
|
#include "port/lang.h"
|
|
|
|
#include "rocksdb/cache.h"
|
|
|
|
#include "rocksdb/comparator.h"
|
|
|
|
#include "rocksdb/convenience.h"
|
|
|
|
#include "rocksdb/env.h"
|
Introduce a new storage specific Env API (#5761)
Summary:
The current Env API encompasses both storage/file operations, as well as OS related operations. Most of the APIs return a Status, which does not have enough metadata about an error, such as whether its retry-able or not, scope (i.e fault domain) of the error etc., that may be required in order to properly handle a storage error. The file APIs also do not provide enough control over the IO SLA, such as timeout, prioritization, hinting about placement and redundancy etc.
This PR separates out the file/storage APIs from Env into a new FileSystem class. The APIs are updated to return an IOStatus with metadata about the error, as well as to take an IOOptions structure as input in order to allow more control over the IO.
The user can set both ```options.env``` and ```options.file_system``` to specify that RocksDB should use the former for OS related operations and the latter for storage operations. Internally, a ```CompositeEnvWrapper``` has been introduced that inherits from ```Env``` and redirects individual methods to either an ```Env``` implementation or the ```FileSystem``` as appropriate. When options are sanitized during ```DB::Open```, ```options.env``` is replaced with a newly allocated ```CompositeEnvWrapper``` instance if both env and file_system have been specified. This way, the rest of the RocksDB code can continue to function as before.
This PR also ports PosixEnv to the new API by splitting it into two - PosixEnv and PosixFileSystem. PosixEnv is defined as a sub-class of CompositeEnvWrapper, and threading/time functions are overridden with Posix specific implementations in order to avoid an extra level of indirection.
The ```CompositeEnvWrapper``` translates ```IOStatus``` return code to ```Status```, and sets the severity to ```kSoftError``` if the io_status is retryable. The error handling code in RocksDB can then recover the DB automatically.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5761
Differential Revision: D18868376
Pulled By: anand1976
fbshipit-source-id: 39efe18a162ea746fabac6360ff529baba48486f
5 years ago
|
|
|
#include "rocksdb/file_system.h"
|
|
|
|
#include "rocksdb/filter_policy.h"
|
|
|
|
#include "rocksdb/iterator.h"
|
|
|
|
#include "rocksdb/options.h"
|
|
|
|
#include "rocksdb/snapshot.h"
|
|
|
|
#include "rocksdb/statistics.h"
|
|
|
|
#include "rocksdb/system_clock.h"
|
|
|
|
#include "rocksdb/table.h"
|
|
|
|
#include "rocksdb/table_properties.h"
|
|
|
|
#include "rocksdb/trace_record.h"
|
|
|
|
#include "table/block_based/binary_search_index_reader.h"
|
|
|
|
#include "table/block_based/block.h"
|
|
|
|
#include "table/block_based/block_based_table_factory.h"
|
|
|
|
#include "table/block_based/block_based_table_iterator.h"
|
|
|
|
#include "table/block_based/block_prefix_index.h"
|
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
|
|
|
#include "table/block_based/block_type.h"
|
|
|
|
#include "table/block_based/filter_block.h"
|
|
|
|
#include "table/block_based/filter_policy_internal.h"
|
|
|
|
#include "table/block_based/full_filter_block.h"
|
|
|
|
#include "table/block_based/hash_index_reader.h"
|
|
|
|
#include "table/block_based/partitioned_filter_block.h"
|
|
|
|
#include "table/block_based/partitioned_index_reader.h"
|
|
|
|
#include "table/block_fetcher.h"
|
|
|
|
#include "table/format.h"
|
|
|
|
#include "table/get_context.h"
|
|
|
|
#include "table/internal_iterator.h"
|
|
|
|
#include "table/meta_blocks.h"
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
6 years ago
|
|
|
#include "table/multiget_context.h"
|
|
|
|
#include "table/persistent_cache_helper.h"
|
|
|
|
#include "table/persistent_cache_options.h"
|
|
|
|
#include "table/sst_file_writer_collectors.h"
|
|
|
|
#include "table/two_level_iterator.h"
|
|
|
|
#include "test_util/sync_point.h"
|
|
|
|
#include "util/coding.h"
|
|
|
|
#include "util/crc32c.h"
|
|
|
|
#include "util/stop_watch.h"
|
|
|
|
#include "util/string_util.h"
|
|
|
|
|
Multi file concurrency in MultiGet using coroutines and async IO (#9968)
Summary:
This PR implements a coroutine version of batched MultiGet in order to concurrently read from multiple SST files in a level using async IO, thus reducing the latency of the MultiGet. The API from the user perspective is still synchronous and single threaded, with the RocksDB part of the processing happening in the context of the caller's thread. In Version::MultiGet, the decision is made whether to call synchronous or coroutine code.
A good way to review this PR is to review the first 4 commits in order - de773b3, 70c2f70, 10b50e1, and 377a597 - before reviewing the rest.
TODO:
1. Figure out how to build it in CircleCI (requires some dependencies to be installed)
2. Do some stress testing with coroutines enabled
No regression in synchronous MultiGet between this branch and main -
```
./db_bench -use_existing_db=true --db=/data/mysql/rocksdb/prefix_scan -benchmarks="readseq,multireadrandom" -key_size=32 -value_size=512 -num=5000000 -batch_size=64 -multiread_batched=true -use_direct_reads=false -duration=60 -ops_between_duration_checks=1 -readonly=true -adaptive_readahead=true -threads=16 -cache_size=10485760000 -async_io=false -multiread_stride=40000 -statistics
```
Branch - ```multireadrandom : 4.025 micros/op 3975111 ops/sec 60.001 seconds 238509056 operations; 2062.3 MB/s (14767808 of 14767808 found)```
Main - ```multireadrandom : 3.987 micros/op 4013216 ops/sec 60.001 seconds 240795392 operations; 2082.1 MB/s (15231040 of 15231040 found)```
More benchmarks in various scenarios are given below. The measurements were taken with ```async_io=false``` (no coroutines) and ```async_io=true``` (use coroutines). For an IO bound workload (with every key requiring an IO), the coroutines version shows a clear benefit, being ~2.6X faster. For CPU bound workloads, the coroutines version has ~6-15% higher CPU utilization, depending on how many keys overlap an SST file.
1. Single thread IO bound workload on remote storage with sparse MultiGet batch keys (~1 key overlap/file) -
No coroutines - ```multireadrandom : 831.774 micros/op 1202 ops/sec 60.001 seconds 72136 operations; 0.6 MB/s (72136 of 72136 found)```
Using coroutines - ```multireadrandom : 318.742 micros/op 3137 ops/sec 60.003 seconds 188248 operations; 1.6 MB/s (188248 of 188248 found)```
2. Single thread CPU bound workload (all data cached) with ~1 key overlap/file -
No coroutines - ```multireadrandom : 4.127 micros/op 242322 ops/sec 60.000 seconds 14539384 operations; 125.7 MB/s (14539384 of 14539384 found)```
Using coroutines - ```multireadrandom : 4.741 micros/op 210935 ops/sec 60.000 seconds 12656176 operations; 109.4 MB/s (12656176 of 12656176 found)```
3. Single thread CPU bound workload with ~2 key overlap/file -
No coroutines - ```multireadrandom : 3.717 micros/op 269000 ops/sec 60.000 seconds 16140024 operations; 139.6 MB/s (16140024 of 16140024 found)```
Using coroutines - ```multireadrandom : 4.146 micros/op 241204 ops/sec 60.000 seconds 14472296 operations; 125.1 MB/s (14472296 of 14472296 found)```
4. CPU bound multi-threaded (16 threads) with ~4 key overlap/file -
No coroutines - ```multireadrandom : 4.534 micros/op 3528792 ops/sec 60.000 seconds 211728728 operations; 1830.7 MB/s (12737024 of 12737024 found) ```
Using coroutines - ```multireadrandom : 4.872 micros/op 3283812 ops/sec 60.000 seconds 197030096 operations; 1703.6 MB/s (12548032 of 12548032 found) ```
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9968
Reviewed By: akankshamahajan15
Differential Revision: D36348563
Pulled By: anand1976
fbshipit-source-id: c0ce85a505fd26ebfbb09786cbd7f25202038696
3 years ago
|
|
|
namespace ROCKSDB_NAMESPACE {
|
|
|
|
namespace {
|
|
|
|
|
|
|
|
CacheAllocationPtr CopyBufferToHeap(MemoryAllocator* allocator, Slice& buf) {
|
|
|
|
CacheAllocationPtr heap_buf;
|
|
|
|
heap_buf = AllocateBlock(buf.size(), allocator);
|
|
|
|
memcpy(heap_buf.get(), buf.data(), buf.size());
|
|
|
|
return heap_buf;
|
|
|
|
}
|
|
|
|
} // namespace
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
|
|
|
|
// Explicitly instantiate templates for each "blocklike" type we use (and
|
|
|
|
// before implicit specialization).
|
|
|
|
// This makes it possible to keep the template definitions in the .cc file.
|
|
|
|
#define INSTANTIATE_RETRIEVE_BLOCK(T) \
|
|
|
|
template Status BlockBasedTable::RetrieveBlock<T>( \
|
|
|
|
FilePrefetchBuffer * prefetch_buffer, const ReadOptions& ro, \
|
|
|
|
const BlockHandle& handle, const UncompressionDict& uncompression_dict, \
|
|
|
|
CachableEntry<T>* out_parsed_block, GetContext* get_context, \
|
|
|
|
BlockCacheLookupContext* lookup_context, bool for_compaction, \
|
HyperClockCache support for SecondaryCache, with refactoring (#11301)
Summary:
Internally refactors SecondaryCache integration out of LRUCache specifically and into a wrapper/adapter class that works with various Cache implementations. Notably, this relies on separating the notion of async lookup handles from other cache handles, so that HyperClockCache doesn't have to deal with the problem of allocating handles from the hash table for lookups that might fail anyway, and might be on the same key without support for coalescing. (LRUCache's hash table can incorporate previously allocated handles thanks to its pointer indirection.) Specifically, I'm worried about the case in which hundreds of threads try to access the same block and probing in the hash table degrades to linear search on the pile of entries with the same key.
This change is a big step in the direction of supporting stacked SecondaryCaches, but there are obstacles to completing that. Especially, there is no SecondaryCache hook for evictions to pass from one to the next. It has been proposed that evictions be transmitted simply as the persisted data (as in SaveToCallback), but given the current structure provided by the CacheItemHelpers, that would require an extra copy of the block data, because there's intentionally no way to ask for a contiguous Slice of the data (to allow for flexibility in storage). `AsyncLookupHandle` and the re-worked `WaitAll()` should be essentially prepared for stacked SecondaryCaches, but several "TODO with stacked secondaries" issues remain in various places.
It could be argued that the stacking instead be done as a SecondaryCache adapter that wraps two (or more) SecondaryCaches, but at least with the current API that would require an extra heap allocation on SecondaryCache Lookup for a wrapper SecondaryCacheResultHandle that can transfer a Lookup between secondaries. We could also consider trying to unify the Cache and SecondaryCache APIs, though that might be difficult if `AsyncLookupHandle` is kept a fixed struct.
## cache.h (public API)
Moves `secondary_cache` option from LRUCacheOptions to ShardedCacheOptions so that it is applicable to HyperClockCache.
## advanced_cache.h (advanced public API)
* Add `Cache::CreateStandalone()` so that the SecondaryCache support wrapper can use it.
* Add `SetEvictionCallback()` / `eviction_callback_` so that the SecondaryCache support wrapper can use it. Only a single callback is supported for efficiency. If there is ever a need for more than one, hopefully that can be handled with a broadcast callback wrapper.
These are essentially the two "extra" pieces of `Cache` for pulling out specific SecondaryCache support from the `Cache` implementation. I think it's a good trade-off as these are reasonable, limited, and reusable "cut points" into the `Cache` implementations.
* Remove async capability from standard `Lookup()` (getting rid of awkward restrictions on pending Handles) and add `AsyncLookupHandle` and `StartAsyncLookup()`. As noted in the comments, the full struct of `AsyncLookupHandle` is exposed so that it can be stack allocated, for efficiency, though more data is being copied around than before, which could impact performance. (Lookup info -> AsyncLookupHandle -> Handle vs. Lookup info -> Handle)
I could foresee a future in which a Cache internally saves a pointer to the AsyncLookupHandle, which means it's dangerous to allow it to be copyable or even movable. It also means it's not compatible with std::vector (which I don't like requiring as an API parameter anyway), so `WaitAll()` expects any contiguous array of AsyncLookupHandles. I believe this is best for common case efficiency, while behaving well in other cases also. For example, `WaitAll()` has no effect on default-constructed AsyncLookupHandles, which look like a completed cache miss.
## cacheable_entry.h
A couple of functions are obsolete because Cache::Handle can no longer be pending.
## cache.cc
Provides default implementations for new or revamped Cache functions, especially appropriate for non-blocking caches.
## secondary_cache_adapter.{h,cc}
The full details of the Cache wrapper adding SecondaryCache support. Essentially replicates the SecondaryCache handling that was in LRUCache, but obviously refactored. There is a bit of logic duplication, where Lookup() is essentially a manually optimized version of StartAsyncLookup() and Wait(), but it's roughly a dozen lines of code.
## sharded_cache.h, typed_cache.h, charged_cache.{h,cc}, sim_cache.cc
Simply updated for Cache API changes.
## lru_cache.{h,cc}
Carefully remove SecondaryCache logic, implement `CreateStandalone` and eviction handler functionality.
## clock_cache.{h,cc}
Expose existing `CreateStandalone` functionality, add eviction handler functionality. Light refactoring.
## block_based_table_reader*
Mostly re-worked the only usage of async Lookup, which is in BlockBasedTable::MultiGet. Used arrays in place of autovector in some places for efficiency. Simplified some logic by not trying to process some cache results before they're all ready.
Created new function `BlockBasedTable::GetCachePriority()` to reduce some pre-existing code duplication (and avoid making it worse).
Fixed at least one small bug from the prior confusing mixture of async and sync Lookups. In MaybeReadBlockAndLoadToCache(), called by RetrieveBlock(), called by MultiGet() with wait=false, is_cache_hit for the block_cache_tracer entry would not be set to true if the handle was pending after Lookup and before Wait.
## Intended follow-up work
* Figure out if there are any missing stats or block_cache_tracer work in refactored BlockBasedTable::MultiGet
* Stacked secondary caches (see above discussion)
* See if we can make up for the small MultiGet performance regression.
* Study more performance with SecondaryCache
* Items evicted from over-full LRUCache in Release were not being demoted to SecondaryCache, and still aren't to minimize unit test churn. Ideally they would be demoted, but it's an exceptional case so not a big deal.
* Use CreateStandalone for cache reservations (save unnecessary hash table operations). Not a big deal, but worthy cleanup.
* Somehow I got the contract for SecondaryCache::Insert wrong in #10945. (Doesn't take ownership!) That API comment needs to be fixed, but didn't want to mingle that in here.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11301
Test Plan:
## Unit tests
Generally updated to include HCC in SecondaryCache tests, though HyperClockCache has some different, less strict behaviors that leads to some tests not really being set up to work with it. Some of the tests remain disabled with it, but I think we have good coverage without them.
## Crash/stress test
Updated to use the new combination.
## Performance
First, let's check for regression on caches without secondary cache configured. Adding support for the eviction callback is likely to have a tiny effect, but it shouldn't be worrisome. LRUCache could benefit slightly from less logic around SecondaryCache handling. We can test with cache_bench default settings, built with DEBUG_LEVEL=0 and PORTABLE=0.
```
(while :; do base/cache_bench --cache_type=hyper_clock_cache | grep Rough; done) | awk '{ sum += $9; count++; print $0; print "Average: " int(sum / count) }'
```
**Before** this and #11299 (which could also have a small effect), running for about an hour, before & after running concurrently for each cache type:
HyperClockCache: 3168662 (average parallel ops/sec)
LRUCache: 2940127
**After** this and #11299, running for about an hour:
HyperClockCache: 3164862 (average parallel ops/sec) (0.12% slower)
LRUCache: 2940928 (0.03% faster)
This is an acceptable difference IMHO.
Next, let's consider essentially the worst case of new CPU overhead affecting overall performance. MultiGet uses the async lookup interface regardless of whether SecondaryCache or folly are used. We can configure a benchmark where all block cache queries are for data blocks, and all are hits.
Create DB and test (before and after tests running simultaneously):
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16
TEST_TMPDIR=/dev/shm base/db_bench -benchmarks=multireadrandom[-X30] -readonly -multiread_batched -batch_size=32 -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16
```
**Before**:
multireadrandom [AVG 30 runs] : 3444202 (± 57049) ops/sec; 240.9 (± 4.0) MB/sec
multireadrandom [MEDIAN 30 runs] : 3514443 ops/sec; 245.8 MB/sec
**After**:
multireadrandom [AVG 30 runs] : 3291022 (± 58851) ops/sec; 230.2 (± 4.1) MB/sec
multireadrandom [MEDIAN 30 runs] : 3366179 ops/sec; 235.4 MB/sec
So that's roughly a 3% regression, on kind of a *worst case* test of MultiGet CPU. Similar story with HyperClockCache:
**Before**:
multireadrandom [AVG 30 runs] : 3933777 (± 41840) ops/sec; 275.1 (± 2.9) MB/sec
multireadrandom [MEDIAN 30 runs] : 3970667 ops/sec; 277.7 MB/sec
**After**:
multireadrandom [AVG 30 runs] : 3755338 (± 30391) ops/sec; 262.6 (± 2.1) MB/sec
multireadrandom [MEDIAN 30 runs] : 3785696 ops/sec; 264.8 MB/sec
Roughly a 4-5% regression. Not ideal, but not the whole story, fortunately.
Let's also look at Get() in db_bench:
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X30] -readonly -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16
```
**Before**:
readrandom [AVG 30 runs] : 2198685 (± 13412) ops/sec; 153.8 (± 0.9) MB/sec
readrandom [MEDIAN 30 runs] : 2209498 ops/sec; 154.5 MB/sec
**After**:
readrandom [AVG 30 runs] : 2292814 (± 43508) ops/sec; 160.3 (± 3.0) MB/sec
readrandom [MEDIAN 30 runs] : 2365181 ops/sec; 165.4 MB/sec
That's showing roughly a 4% improvement, perhaps because of the secondary cache code that is no longer part of LRUCache. But weirdly, HyperClockCache is also showing 2-3% improvement:
**Before**:
readrandom [AVG 30 runs] : 2272333 (± 9992) ops/sec; 158.9 (± 0.7) MB/sec
readrandom [MEDIAN 30 runs] : 2273239 ops/sec; 159.0 MB/sec
**After**:
readrandom [AVG 30 runs] : 2332407 (± 11252) ops/sec; 163.1 (± 0.8) MB/sec
readrandom [MEDIAN 30 runs] : 2335329 ops/sec; 163.3 MB/sec
Reviewed By: ltamasi
Differential Revision: D44177044
Pulled By: pdillinger
fbshipit-source-id: e808e48ff3fe2f792a79841ba617be98e48689f5
2 years ago
|
|
|
bool use_cache, bool async_read) const;
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
|
|
|
|
INSTANTIATE_RETRIEVE_BLOCK(ParsedFullFilterBlock);
|
|
|
|
INSTANTIATE_RETRIEVE_BLOCK(UncompressionDict);
|
|
|
|
INSTANTIATE_RETRIEVE_BLOCK(Block_kData);
|
|
|
|
INSTANTIATE_RETRIEVE_BLOCK(Block_kIndex);
|
|
|
|
INSTANTIATE_RETRIEVE_BLOCK(Block_kFilterPartitionIndex);
|
|
|
|
INSTANTIATE_RETRIEVE_BLOCK(Block_kRangeDeletion);
|
|
|
|
INSTANTIATE_RETRIEVE_BLOCK(Block_kMetaIndex);
|
|
|
|
|
Multi file concurrency in MultiGet using coroutines and async IO (#9968)
Summary:
This PR implements a coroutine version of batched MultiGet in order to concurrently read from multiple SST files in a level using async IO, thus reducing the latency of the MultiGet. The API from the user perspective is still synchronous and single threaded, with the RocksDB part of the processing happening in the context of the caller's thread. In Version::MultiGet, the decision is made whether to call synchronous or coroutine code.
A good way to review this PR is to review the first 4 commits in order - de773b3, 70c2f70, 10b50e1, and 377a597 - before reviewing the rest.
TODO:
1. Figure out how to build it in CircleCI (requires some dependencies to be installed)
2. Do some stress testing with coroutines enabled
No regression in synchronous MultiGet between this branch and main -
```
./db_bench -use_existing_db=true --db=/data/mysql/rocksdb/prefix_scan -benchmarks="readseq,multireadrandom" -key_size=32 -value_size=512 -num=5000000 -batch_size=64 -multiread_batched=true -use_direct_reads=false -duration=60 -ops_between_duration_checks=1 -readonly=true -adaptive_readahead=true -threads=16 -cache_size=10485760000 -async_io=false -multiread_stride=40000 -statistics
```
Branch - ```multireadrandom : 4.025 micros/op 3975111 ops/sec 60.001 seconds 238509056 operations; 2062.3 MB/s (14767808 of 14767808 found)```
Main - ```multireadrandom : 3.987 micros/op 4013216 ops/sec 60.001 seconds 240795392 operations; 2082.1 MB/s (15231040 of 15231040 found)```
More benchmarks in various scenarios are given below. The measurements were taken with ```async_io=false``` (no coroutines) and ```async_io=true``` (use coroutines). For an IO bound workload (with every key requiring an IO), the coroutines version shows a clear benefit, being ~2.6X faster. For CPU bound workloads, the coroutines version has ~6-15% higher CPU utilization, depending on how many keys overlap an SST file.
1. Single thread IO bound workload on remote storage with sparse MultiGet batch keys (~1 key overlap/file) -
No coroutines - ```multireadrandom : 831.774 micros/op 1202 ops/sec 60.001 seconds 72136 operations; 0.6 MB/s (72136 of 72136 found)```
Using coroutines - ```multireadrandom : 318.742 micros/op 3137 ops/sec 60.003 seconds 188248 operations; 1.6 MB/s (188248 of 188248 found)```
2. Single thread CPU bound workload (all data cached) with ~1 key overlap/file -
No coroutines - ```multireadrandom : 4.127 micros/op 242322 ops/sec 60.000 seconds 14539384 operations; 125.7 MB/s (14539384 of 14539384 found)```
Using coroutines - ```multireadrandom : 4.741 micros/op 210935 ops/sec 60.000 seconds 12656176 operations; 109.4 MB/s (12656176 of 12656176 found)```
3. Single thread CPU bound workload with ~2 key overlap/file -
No coroutines - ```multireadrandom : 3.717 micros/op 269000 ops/sec 60.000 seconds 16140024 operations; 139.6 MB/s (16140024 of 16140024 found)```
Using coroutines - ```multireadrandom : 4.146 micros/op 241204 ops/sec 60.000 seconds 14472296 operations; 125.1 MB/s (14472296 of 14472296 found)```
4. CPU bound multi-threaded (16 threads) with ~4 key overlap/file -
No coroutines - ```multireadrandom : 4.534 micros/op 3528792 ops/sec 60.000 seconds 211728728 operations; 1830.7 MB/s (12737024 of 12737024 found) ```
Using coroutines - ```multireadrandom : 4.872 micros/op 3283812 ops/sec 60.000 seconds 197030096 operations; 1703.6 MB/s (12548032 of 12548032 found) ```
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9968
Reviewed By: akankshamahajan15
Differential Revision: D36348563
Pulled By: anand1976
fbshipit-source-id: c0ce85a505fd26ebfbb09786cbd7f25202038696
3 years ago
|
|
|
} // namespace ROCKSDB_NAMESPACE
|
|
|
|
|
|
|
|
// Generate the regular and coroutine versions of some methods by
|
|
|
|
// including block_based_table_reader_sync_and_async.h twice
|
|
|
|
// Macros in the header will expand differently based on whether
|
|
|
|
// WITH_COROUTINES or WITHOUT_COROUTINES is defined
|
|
|
|
// clang-format off
|
|
|
|
#define WITHOUT_COROUTINES
|
|
|
|
#include "table/block_based/block_based_table_reader_sync_and_async.h"
|
|
|
|
#undef WITHOUT_COROUTINES
|
|
|
|
#define WITH_COROUTINES
|
|
|
|
#include "table/block_based/block_based_table_reader_sync_and_async.h"
|
|
|
|
#undef WITH_COROUTINES
|
Multi file concurrency in MultiGet using coroutines and async IO (#9968)
Summary:
This PR implements a coroutine version of batched MultiGet in order to concurrently read from multiple SST files in a level using async IO, thus reducing the latency of the MultiGet. The API from the user perspective is still synchronous and single threaded, with the RocksDB part of the processing happening in the context of the caller's thread. In Version::MultiGet, the decision is made whether to call synchronous or coroutine code.
A good way to review this PR is to review the first 4 commits in order - de773b3, 70c2f70, 10b50e1, and 377a597 - before reviewing the rest.
TODO:
1. Figure out how to build it in CircleCI (requires some dependencies to be installed)
2. Do some stress testing with coroutines enabled
No regression in synchronous MultiGet between this branch and main -
```
./db_bench -use_existing_db=true --db=/data/mysql/rocksdb/prefix_scan -benchmarks="readseq,multireadrandom" -key_size=32 -value_size=512 -num=5000000 -batch_size=64 -multiread_batched=true -use_direct_reads=false -duration=60 -ops_between_duration_checks=1 -readonly=true -adaptive_readahead=true -threads=16 -cache_size=10485760000 -async_io=false -multiread_stride=40000 -statistics
```
Branch - ```multireadrandom : 4.025 micros/op 3975111 ops/sec 60.001 seconds 238509056 operations; 2062.3 MB/s (14767808 of 14767808 found)```
Main - ```multireadrandom : 3.987 micros/op 4013216 ops/sec 60.001 seconds 240795392 operations; 2082.1 MB/s (15231040 of 15231040 found)```
More benchmarks in various scenarios are given below. The measurements were taken with ```async_io=false``` (no coroutines) and ```async_io=true``` (use coroutines). For an IO bound workload (with every key requiring an IO), the coroutines version shows a clear benefit, being ~2.6X faster. For CPU bound workloads, the coroutines version has ~6-15% higher CPU utilization, depending on how many keys overlap an SST file.
1. Single thread IO bound workload on remote storage with sparse MultiGet batch keys (~1 key overlap/file) -
No coroutines - ```multireadrandom : 831.774 micros/op 1202 ops/sec 60.001 seconds 72136 operations; 0.6 MB/s (72136 of 72136 found)```
Using coroutines - ```multireadrandom : 318.742 micros/op 3137 ops/sec 60.003 seconds 188248 operations; 1.6 MB/s (188248 of 188248 found)```
2. Single thread CPU bound workload (all data cached) with ~1 key overlap/file -
No coroutines - ```multireadrandom : 4.127 micros/op 242322 ops/sec 60.000 seconds 14539384 operations; 125.7 MB/s (14539384 of 14539384 found)```
Using coroutines - ```multireadrandom : 4.741 micros/op 210935 ops/sec 60.000 seconds 12656176 operations; 109.4 MB/s (12656176 of 12656176 found)```
3. Single thread CPU bound workload with ~2 key overlap/file -
No coroutines - ```multireadrandom : 3.717 micros/op 269000 ops/sec 60.000 seconds 16140024 operations; 139.6 MB/s (16140024 of 16140024 found)```
Using coroutines - ```multireadrandom : 4.146 micros/op 241204 ops/sec 60.000 seconds 14472296 operations; 125.1 MB/s (14472296 of 14472296 found)```
4. CPU bound multi-threaded (16 threads) with ~4 key overlap/file -
No coroutines - ```multireadrandom : 4.534 micros/op 3528792 ops/sec 60.000 seconds 211728728 operations; 1830.7 MB/s (12737024 of 12737024 found) ```
Using coroutines - ```multireadrandom : 4.872 micros/op 3283812 ops/sec 60.000 seconds 197030096 operations; 1703.6 MB/s (12548032 of 12548032 found) ```
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9968
Reviewed By: akankshamahajan15
Differential Revision: D36348563
Pulled By: anand1976
fbshipit-source-id: c0ce85a505fd26ebfbb09786cbd7f25202038696
3 years ago
|
|
|
// clang-format on
|
|
|
|
|
|
|
|
namespace ROCKSDB_NAMESPACE {
|
|
|
|
|
|
|
|
extern const uint64_t kBlockBasedTableMagicNumber;
|
|
|
|
extern const std::string kHashIndexPrefixesBlock;
|
|
|
|
extern const std::string kHashIndexPrefixesMetadataBlock;
|
|
|
|
|
|
|
|
BlockBasedTable::~BlockBasedTable() { delete rep_; }
|
|
|
|
|
|
|
|
namespace {
|
|
|
|
// Read the block identified by "handle" from "file".
|
|
|
|
// The only relevant option is options.verify_checksums for now.
|
|
|
|
// On failure return non-OK.
|
|
|
|
// On success fill *result and return OK - caller owns *result
|
|
|
|
// @param uncompression_dict Data for presetting the compression library's
|
Shared dictionary compression using reference block
Summary:
This adds a new metablock containing a shared dictionary that is used
to compress all data blocks in the SST file. The size of the shared dictionary
is configurable in CompressionOptions and defaults to 0. It's currently only
used for zlib/lz4/lz4hc, but the block will be stored in the SST regardless of
the compression type if the user chooses a nonzero dictionary size.
During compaction, computes the dictionary by randomly sampling the first
output file in each subcompaction. It pre-computes the intervals to sample
by assuming the output file will have the maximum allowable length. In case
the file is smaller, some of the pre-computed sampling intervals can be beyond
end-of-file, in which case we skip over those samples and the dictionary will
be a bit smaller. After the dictionary is generated using the first file in a
subcompaction, it is loaded into the compression library before writing each
block in each subsequent file of that subcompaction.
On the read path, gets the dictionary from the metablock, if it exists. Then,
loads that dictionary into the compression library before reading each block.
Test Plan: new unit test
Reviewers: yhchiang, IslamAbdelRahman, cyan, sdong
Reviewed By: sdong
Subscribers: andrewkr, yoshinorim, kradhakrishnan, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D52287
9 years ago
|
|
|
// dictionary.
|
|
|
|
template <typename TBlocklike>
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
Status ReadAndParseBlockFromFile(
|
|
|
|
RandomAccessFileReader* file, FilePrefetchBuffer* prefetch_buffer,
|
|
|
|
const Footer& footer, const ReadOptions& options, const BlockHandle& handle,
|
|
|
|
std::unique_ptr<TBlocklike>* result, const ImmutableOptions& ioptions,
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
BlockCreateContext& create_context, bool maybe_compressed,
|
|
|
|
const UncompressionDict& uncompression_dict,
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
const PersistentCacheOptions& cache_options,
|
|
|
|
MemoryAllocator* memory_allocator, bool for_compaction, bool async_read) {
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
assert(result);
|
|
|
|
|
|
|
|
BlockContents contents;
|
|
|
|
BlockFetcher block_fetcher(
|
|
|
|
file, prefetch_buffer, footer, options, handle, &contents, ioptions,
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
/*do_uncompress*/ maybe_compressed, maybe_compressed,
|
|
|
|
TBlocklike::kBlockType, uncompression_dict, cache_options,
|
|
|
|
memory_allocator, nullptr, for_compaction);
|
|
|
|
Status s;
|
|
|
|
// If prefetch_buffer is not allocated, it will fallback to synchronous
|
|
|
|
// reading of block contents.
|
|
|
|
if (async_read && prefetch_buffer != nullptr) {
|
|
|
|
s = block_fetcher.ReadAsyncBlockContents();
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
s = block_fetcher.ReadBlockContents();
|
|
|
|
}
|
|
|
|
if (s.ok()) {
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
create_context.Create(result, std::move(contents));
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
// For hash based index, return false if table_properties->prefix_extractor_name
|
|
|
|
// and prefix_extractor both exist and match, otherwise true.
|
|
|
|
inline bool PrefixExtractorChangedHelper(
|
|
|
|
const TableProperties* table_properties,
|
|
|
|
const SliceTransform* prefix_extractor) {
|
|
|
|
// BlockBasedTableOptions::kHashSearch requires prefix_extractor to be set.
|
|
|
|
// Turn off hash index in prefix_extractor is not set; if prefix_extractor
|
|
|
|
// is set but prefix_extractor_block is not set, also disable hash index
|
|
|
|
if (prefix_extractor == nullptr || table_properties == nullptr ||
|
|
|
|
table_properties->prefix_extractor_name.empty()) {
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
// prefix_extractor and prefix_extractor_block are both non-empty
|
|
|
|
if (table_properties->prefix_extractor_name != prefix_extractor->AsString()) {
|
|
|
|
return true;
|
|
|
|
} else {
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
template <typename TBlocklike>
|
|
|
|
uint32_t GetBlockNumRestarts(const TBlocklike& block) {
|
|
|
|
if constexpr (std::is_convertible_v<const TBlocklike&, const Block&>) {
|
|
|
|
const Block& b = block;
|
|
|
|
return b.NumRestarts();
|
|
|
|
} else {
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
} // namespace
|
|
|
|
|
|
|
|
void BlockBasedTable::UpdateCacheHitMetrics(BlockType block_type,
|
|
|
|
GetContext* get_context,
|
|
|
|
size_t usage) const {
|
|
|
|
Statistics* const statistics = rep_->ioptions.stats;
|
|
|
|
|
|
|
|
PERF_COUNTER_ADD(block_cache_hit_count, 1);
|
|
|
|
PERF_COUNTER_BY_LEVEL_ADD(block_cache_hit_count, 1,
|
|
|
|
static_cast<uint32_t>(rep_->level));
|
|
|
|
|
|
|
|
if (get_context) {
|
|
|
|
++get_context->get_context_stats_.num_cache_hit;
|
|
|
|
get_context->get_context_stats_.num_cache_bytes_read += usage;
|
|
|
|
} else {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_HIT);
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_BYTES_READ, usage);
|
|
|
|
}
|
|
|
|
|
|
|
|
switch (block_type) {
|
|
|
|
case BlockType::kFilter:
|
|
|
|
case BlockType::kFilterPartitionIndex:
|
|
|
|
PERF_COUNTER_ADD(block_cache_filter_hit_count, 1);
|
|
|
|
|
|
|
|
if (get_context) {
|
|
|
|
++get_context->get_context_stats_.num_cache_filter_hit;
|
|
|
|
} else {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_FILTER_HIT);
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
|
|
|
|
case BlockType::kCompressionDictionary:
|
|
|
|
// TODO: introduce perf counter for compression dictionary hit count
|
|
|
|
if (get_context) {
|
|
|
|
++get_context->get_context_stats_.num_cache_compression_dict_hit;
|
|
|
|
} else {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_COMPRESSION_DICT_HIT);
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
|
|
|
|
case BlockType::kIndex:
|
|
|
|
PERF_COUNTER_ADD(block_cache_index_hit_count, 1);
|
|
|
|
|
|
|
|
if (get_context) {
|
|
|
|
++get_context->get_context_stats_.num_cache_index_hit;
|
|
|
|
} else {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_INDEX_HIT);
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
|
|
|
|
default:
|
|
|
|
// TODO: introduce dedicated tickers/statistics/counters
|
|
|
|
// for range tombstones
|
|
|
|
if (get_context) {
|
|
|
|
++get_context->get_context_stats_.num_cache_data_hit;
|
|
|
|
} else {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_DATA_HIT);
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void BlockBasedTable::UpdateCacheMissMetrics(BlockType block_type,
|
|
|
|
GetContext* get_context) const {
|
|
|
|
Statistics* const statistics = rep_->ioptions.stats;
|
|
|
|
|
|
|
|
// TODO: introduce aggregate (not per-level) block cache miss count
|
|
|
|
PERF_COUNTER_BY_LEVEL_ADD(block_cache_miss_count, 1,
|
|
|
|
static_cast<uint32_t>(rep_->level));
|
|
|
|
|
|
|
|
if (get_context) {
|
|
|
|
++get_context->get_context_stats_.num_cache_miss;
|
|
|
|
} else {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_MISS);
|
|
|
|
}
|
|
|
|
|
|
|
|
// TODO: introduce perf counters for misses per block type
|
|
|
|
switch (block_type) {
|
|
|
|
case BlockType::kFilter:
|
|
|
|
case BlockType::kFilterPartitionIndex:
|
|
|
|
if (get_context) {
|
|
|
|
++get_context->get_context_stats_.num_cache_filter_miss;
|
|
|
|
} else {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_FILTER_MISS);
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
|
|
|
|
case BlockType::kCompressionDictionary:
|
|
|
|
if (get_context) {
|
|
|
|
++get_context->get_context_stats_.num_cache_compression_dict_miss;
|
|
|
|
} else {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_COMPRESSION_DICT_MISS);
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
|
|
|
|
case BlockType::kIndex:
|
|
|
|
if (get_context) {
|
|
|
|
++get_context->get_context_stats_.num_cache_index_miss;
|
|
|
|
} else {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_INDEX_MISS);
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
|
|
|
|
default:
|
|
|
|
// TODO: introduce dedicated tickers/statistics/counters
|
|
|
|
// for range tombstones
|
|
|
|
if (get_context) {
|
|
|
|
++get_context->get_context_stats_.num_cache_data_miss;
|
|
|
|
} else {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_DATA_MISS);
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void BlockBasedTable::UpdateCacheInsertionMetrics(
|
|
|
|
BlockType block_type, GetContext* get_context, size_t usage, bool redundant,
|
|
|
|
Statistics* const statistics) {
|
|
|
|
// TODO: introduce perf counters for block cache insertions
|
|
|
|
if (get_context) {
|
|
|
|
++get_context->get_context_stats_.num_cache_add;
|
|
|
|
if (redundant) {
|
|
|
|
++get_context->get_context_stats_.num_cache_add_redundant;
|
|
|
|
}
|
|
|
|
get_context->get_context_stats_.num_cache_bytes_write += usage;
|
|
|
|
} else {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_ADD);
|
|
|
|
if (redundant) {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_ADD_REDUNDANT);
|
|
|
|
}
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_BYTES_WRITE, usage);
|
|
|
|
}
|
|
|
|
|
|
|
|
switch (block_type) {
|
|
|
|
case BlockType::kFilter:
|
|
|
|
case BlockType::kFilterPartitionIndex:
|
|
|
|
if (get_context) {
|
|
|
|
++get_context->get_context_stats_.num_cache_filter_add;
|
|
|
|
if (redundant) {
|
|
|
|
++get_context->get_context_stats_.num_cache_filter_add_redundant;
|
|
|
|
}
|
|
|
|
get_context->get_context_stats_.num_cache_filter_bytes_insert += usage;
|
|
|
|
} else {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_FILTER_ADD);
|
|
|
|
if (redundant) {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_FILTER_ADD_REDUNDANT);
|
|
|
|
}
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_FILTER_BYTES_INSERT, usage);
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
|
|
|
|
case BlockType::kCompressionDictionary:
|
|
|
|
if (get_context) {
|
|
|
|
++get_context->get_context_stats_.num_cache_compression_dict_add;
|
|
|
|
if (redundant) {
|
|
|
|
++get_context->get_context_stats_
|
|
|
|
.num_cache_compression_dict_add_redundant;
|
|
|
|
}
|
|
|
|
get_context->get_context_stats_
|
|
|
|
.num_cache_compression_dict_bytes_insert += usage;
|
|
|
|
} else {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_COMPRESSION_DICT_ADD);
|
|
|
|
if (redundant) {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_COMPRESSION_DICT_ADD_REDUNDANT);
|
|
|
|
}
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_COMPRESSION_DICT_BYTES_INSERT,
|
|
|
|
usage);
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
|
|
|
|
case BlockType::kIndex:
|
|
|
|
if (get_context) {
|
|
|
|
++get_context->get_context_stats_.num_cache_index_add;
|
|
|
|
if (redundant) {
|
|
|
|
++get_context->get_context_stats_.num_cache_index_add_redundant;
|
|
|
|
}
|
|
|
|
get_context->get_context_stats_.num_cache_index_bytes_insert += usage;
|
|
|
|
} else {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_INDEX_ADD);
|
|
|
|
if (redundant) {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_INDEX_ADD_REDUNDANT);
|
|
|
|
}
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_INDEX_BYTES_INSERT, usage);
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
|
|
|
|
default:
|
|
|
|
// TODO: introduce dedicated tickers/statistics/counters
|
|
|
|
// for range tombstones
|
|
|
|
if (get_context) {
|
|
|
|
++get_context->get_context_stats_.num_cache_data_add;
|
|
|
|
if (redundant) {
|
|
|
|
++get_context->get_context_stats_.num_cache_data_add_redundant;
|
|
|
|
}
|
|
|
|
get_context->get_context_stats_.num_cache_data_bytes_insert += usage;
|
|
|
|
} else {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_DATA_ADD);
|
|
|
|
if (redundant) {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_DATA_ADD_REDUNDANT);
|
|
|
|
}
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_DATA_BYTES_INSERT, usage);
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
namespace {
|
|
|
|
// Return True if table_properties has `user_prop_name` has a `true` value
|
|
|
|
// or it doesn't contain this property (for backward compatible).
|
|
|
|
bool IsFeatureSupported(const TableProperties& table_properties,
|
|
|
|
const std::string& user_prop_name, Logger* info_log) {
|
|
|
|
auto& props = table_properties.user_collected_properties;
|
|
|
|
auto pos = props.find(user_prop_name);
|
|
|
|
// Older version doesn't have this value set. Skip this check.
|
|
|
|
if (pos != props.end()) {
|
|
|
|
if (pos->second == kPropFalse) {
|
|
|
|
return false;
|
|
|
|
} else if (pos->second != kPropTrue) {
|
|
|
|
ROCKS_LOG_WARN(info_log, "Property %s has invalidate value %s",
|
|
|
|
user_prop_name.c_str(), pos->second.c_str());
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Caller has to ensure seqno is not nullptr.
|
|
|
|
Status GetGlobalSequenceNumber(const TableProperties& table_properties,
|
|
|
|
SequenceNumber largest_seqno,
|
|
|
|
SequenceNumber* seqno) {
|
|
|
|
const auto& props = table_properties.user_collected_properties;
|
|
|
|
const auto version_pos = props.find(ExternalSstFilePropertyNames::kVersion);
|
|
|
|
const auto seqno_pos = props.find(ExternalSstFilePropertyNames::kGlobalSeqno);
|
|
|
|
|
|
|
|
*seqno = kDisableGlobalSequenceNumber;
|
|
|
|
if (version_pos == props.end()) {
|
|
|
|
if (seqno_pos != props.end()) {
|
|
|
|
std::array<char, 200> msg_buf;
|
|
|
|
// This is not an external sst file, global_seqno is not supported.
|
|
|
|
snprintf(
|
|
|
|
msg_buf.data(), msg_buf.max_size(),
|
|
|
|
"A non-external sst file have global seqno property with value %s",
|
|
|
|
seqno_pos->second.c_str());
|
|
|
|
return Status::Corruption(msg_buf.data());
|
|
|
|
}
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
|
|
|
uint32_t version = DecodeFixed32(version_pos->second.c_str());
|
|
|
|
if (version < 2) {
|
|
|
|
if (seqno_pos != props.end() || version != 1) {
|
|
|
|
std::array<char, 200> msg_buf;
|
|
|
|
// This is a v1 external sst file, global_seqno is not supported.
|
|
|
|
snprintf(msg_buf.data(), msg_buf.max_size(),
|
|
|
|
"An external sst file with version %u have global seqno "
|
|
|
|
"property with value %s",
|
|
|
|
version, seqno_pos->second.c_str());
|
|
|
|
return Status::Corruption(msg_buf.data());
|
|
|
|
}
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
|
|
|
// Since we have a plan to deprecate global_seqno, we do not return failure
|
|
|
|
// if seqno_pos == props.end(). We rely on version_pos to detect whether the
|
|
|
|
// SST is external.
|
|
|
|
SequenceNumber global_seqno(0);
|
|
|
|
if (seqno_pos != props.end()) {
|
|
|
|
global_seqno = DecodeFixed64(seqno_pos->second.c_str());
|
|
|
|
}
|
|
|
|
// SstTableReader open table reader with kMaxSequenceNumber as largest_seqno
|
|
|
|
// to denote it is unknown.
|
|
|
|
if (largest_seqno < kMaxSequenceNumber) {
|
|
|
|
if (global_seqno == 0) {
|
|
|
|
global_seqno = largest_seqno;
|
|
|
|
}
|
|
|
|
if (global_seqno != largest_seqno) {
|
|
|
|
std::array<char, 200> msg_buf;
|
|
|
|
snprintf(
|
|
|
|
msg_buf.data(), msg_buf.max_size(),
|
|
|
|
"An external sst file with version %u have global seqno property "
|
|
|
|
"with value %s, while largest seqno in the file is %llu",
|
|
|
|
version, seqno_pos->second.c_str(),
|
|
|
|
static_cast<unsigned long long>(largest_seqno));
|
|
|
|
return Status::Corruption(msg_buf.data());
|
|
|
|
}
|
|
|
|
}
|
|
|
|
*seqno = global_seqno;
|
|
|
|
|
|
|
|
if (global_seqno > kMaxSequenceNumber) {
|
|
|
|
std::array<char, 200> msg_buf;
|
|
|
|
snprintf(msg_buf.data(), msg_buf.max_size(),
|
|
|
|
"An external sst file with version %u have global seqno property "
|
|
|
|
"with value %llu, which is greater than kMaxSequenceNumber",
|
|
|
|
version, static_cast<unsigned long long>(global_seqno));
|
|
|
|
return Status::Corruption(msg_buf.data());
|
|
|
|
}
|
|
|
|
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
} // namespace
|
|
|
|
|
New stable, fixed-length cache keys (#9126)
Summary:
This change standardizes on a new 16-byte cache key format for
block cache (incl compressed and secondary) and persistent cache (but
not table cache and row cache).
The goal is a really fast cache key with practically ideal stability and
uniqueness properties without external dependencies (e.g. from FileSystem).
A fixed key size of 16 bytes should enable future optimizations to the
concurrent hash table for block cache, which is a heavy CPU user /
bottleneck, but there appears to be measurable performance improvement
even with no changes to LRUCache.
This change replaces a lot of disjointed and ugly code handling cache
keys with calls to a simple, clean new internal API (cache_key.h).
(Preserving the old cache key logic under an option would be very ugly
and likely negate the performance gain of the new approach. Complete
replacement carries some inherent risk, but I think that's acceptable
with sufficient analysis and testing.)
The scheme for encoding new cache keys is complicated but explained
in cache_key.cc.
Also: EndianSwapValue is moved to math.h to be next to other bit
operations. (Explains some new include "math.h".) ReverseBits operation
added and unit tests added to hash_test for both.
Fixes https://github.com/facebook/rocksdb/issues/7405 (presuming a root cause)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9126
Test Plan:
### Basic correctness
Several tests needed updates to work with the new functionality, mostly
because we are no longer relying on filesystem for stable cache keys
so table builders & readers need more context info to agree on cache
keys. This functionality is so core, a huge number of existing tests
exercise the cache key functionality.
### Performance
Create db with
`TEST_TMPDIR=/dev/shm ./db_bench -bloom_bits=10 -benchmarks=fillrandom -num=3000000 -partition_index_and_filters`
And test performance with
`TEST_TMPDIR=/dev/shm ./db_bench -readonly -use_existing_db -bloom_bits=10 -benchmarks=readrandom -num=3000000 -duration=30 -cache_index_and_filter_blocks -cache_size=250000 -threads=4`
using DEBUG_LEVEL=0 and simultaneous before & after runs.
Before ops/sec, avg over 100 runs: 121924
After ops/sec, avg over 100 runs: 125385 (+2.8%)
### Collision probability
I have built a tool, ./cache_bench -stress_cache_key to broadly simulate host-wide cache activity
over many months, by making some pessimistic simplifying assumptions:
* Every generated file has a cache entry for every byte offset in the file (contiguous range of cache keys)
* All of every file is cached for its entire lifetime
We use a simple table with skewed address assignment and replacement on address collision
to simulate files coming & going, with quite a variance (super-Poisson) in ages. Some output
with `./cache_bench -stress_cache_key -sck_keep_bits=40`:
```
Total cache or DBs size: 32TiB Writing 925.926 MiB/s or 76.2939TiB/day
Multiply by 9.22337e+18 to correct for simulation losses (but still assume whole file cached)
```
These come from default settings of 2.5M files per day of 32 MB each, and
`-sck_keep_bits=40` means that to represent a single file, we are only keeping 40 bits of
the 128-bit cache key. With file size of 2\*\*25 contiguous keys (pessimistic), our simulation
is about 2\*\*(128-40-25) or about 9 billion billion times more prone to collision than reality.
More default assumptions, relatively pessimistic:
* 100 DBs in same process (doesn't matter much)
* Re-open DB in same process (new session ID related to old session ID) on average
every 100 files generated
* Restart process (all new session IDs unrelated to old) 24 times per day
After enough data, we get a result at the end:
```
(keep 40 bits) 17 collisions after 2 x 90 days, est 10.5882 days between (9.76592e+19 corrected)
```
If we believe the (pessimistic) simulation and the mathematical generalization, we would need to run a billion machines all for 97 billion days to expect a cache key collision. To help verify that our generalization ("corrected") is robust, we can make our simulation more precise with `-sck_keep_bits=41` and `42`, which takes more running time to get enough data:
```
(keep 41 bits) 16 collisions after 4 x 90 days, est 22.5 days between (1.03763e+20 corrected)
(keep 42 bits) 19 collisions after 10 x 90 days, est 47.3684 days between (1.09224e+20 corrected)
```
The generalized prediction still holds. With the `-sck_randomize` option, we can see that we are beating "random" cache keys (except offsets still non-randomized) by a modest amount (roughly 20x less collision prone than random), which should make us reasonably comfortable even in "degenerate" cases:
```
197 collisions after 1 x 90 days, est 0.456853 days between (4.21372e+18 corrected)
```
I've run other tests to validate other conditions behave as expected, never behaving "worse than random" unless we start chopping off structured data.
Reviewed By: zhichao-cao
Differential Revision: D33171746
Pulled By: pdillinger
fbshipit-source-id: f16a57e369ed37be5e7e33525ace848d0537c88f
3 years ago
|
|
|
void BlockBasedTable::SetupBaseCacheKey(const TableProperties* properties,
|
|
|
|
const std::string& cur_db_session_id,
|
|
|
|
uint64_t cur_file_number,
|
|
|
|
OffsetableCacheKey* out_base_cache_key,
|
|
|
|
bool* out_is_stable) {
|
|
|
|
// Use a stable cache key if sufficient data is in table properties
|
|
|
|
std::string db_session_id;
|
|
|
|
uint64_t file_num;
|
|
|
|
std::string db_id;
|
|
|
|
if (properties && !properties->db_session_id.empty() &&
|
|
|
|
properties->orig_file_number > 0) {
|
|
|
|
// (Newer SST file case)
|
|
|
|
// We must have both properties to get a stable unique id because
|
|
|
|
// CreateColumnFamilyWithImport or IngestExternalFiles can change the
|
|
|
|
// file numbers on a file.
|
|
|
|
db_session_id = properties->db_session_id;
|
|
|
|
file_num = properties->orig_file_number;
|
|
|
|
// Less critical, populated in earlier release than above
|
|
|
|
db_id = properties->db_id;
|
|
|
|
if (out_is_stable) {
|
|
|
|
*out_is_stable = true;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
// (Old SST file case)
|
|
|
|
// We use (unique) cache keys based on current identifiers. These are at
|
|
|
|
// least stable across table file close and re-open, but not across
|
|
|
|
// different DBs nor DB close and re-open.
|
|
|
|
db_session_id = cur_db_session_id;
|
|
|
|
file_num = cur_file_number;
|
|
|
|
// Plumbing through the DB ID to here would be annoying, and of limited
|
|
|
|
// value because of the case of VersionSet::Recover opening some table
|
|
|
|
// files and later setting the DB ID. So we just rely on uniqueness
|
|
|
|
// level provided by session ID.
|
|
|
|
db_id = "unknown";
|
|
|
|
if (out_is_stable) {
|
|
|
|
*out_is_stable = false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Too many tests to update to get these working
|
|
|
|
// assert(file_num > 0);
|
|
|
|
// assert(!db_session_id.empty());
|
|
|
|
// assert(!db_id.empty());
|
|
|
|
|
|
|
|
// Minimum block size is 5 bytes; therefore we can trim off two lower bits
|
|
|
|
// from offsets. See GetCacheKey.
|
Derive cache keys from SST unique IDs (#10394)
Summary:
... so that cache keys can be derived from DB manifest data
before reading the file from storage--so that every part of the file
can potentially go in a persistent cache.
See updated comments in cache_key.cc for technical details. Importantly,
the new cache key encoding uses some fancy but efficient math to pack
data into the cache key without depending on the sizes of the various
pieces. This simplifies some existing code creating cache keys, like
cache warming before the file size is known.
This should provide us an essentially permanent mapping between SST
unique IDs and base cache keys, with the ability to "upgrade" SST
unique IDs (and thus cache keys) with new SST format_versions.
These cache keys are of similar, perhaps indistinguishable quality to
the previous generation. Before this change (see "corrected" days
between collision):
```
./cache_bench -stress_cache_key -sck_keep_bits=43
18 collisions after 2 x 90 days, est 10 days between (1.15292e+19 corrected)
```
After this change (keep 43 bits, up through 50, to validate "trajectory"
is ok on "corrected" days between collision):
```
19 collisions after 3 x 90 days, est 14.2105 days between (1.63836e+19 corrected)
16 collisions after 5 x 90 days, est 28.125 days between (1.6213e+19 corrected)
15 collisions after 7 x 90 days, est 42 days between (1.21057e+19 corrected)
15 collisions after 17 x 90 days, est 102 days between (1.46997e+19 corrected)
15 collisions after 49 x 90 days, est 294 days between (2.11849e+19 corrected)
15 collisions after 62 x 90 days, est 372 days between (1.34027e+19 corrected)
15 collisions after 53 x 90 days, est 318 days between (5.72858e+18 corrected)
15 collisions after 309 x 90 days, est 1854 days between (1.66994e+19 corrected)
```
However, the change does modify (probably weaken) the "guaranteed unique" promise from this
> SST files generated in a single process are guaranteed to have unique cache keys, unless/until number session ids * max file number = 2**86
to this (see https://github.com/facebook/rocksdb/issues/10388)
> With the DB id limitation, we only have nice guaranteed unique cache keys for files generated in a single process until biggest session_id_counter and offset_in_file reach combined 64 bits
I don't think this is a practical concern, though.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10394
Test Plan: unit tests updated, see simulation results above
Reviewed By: jay-zhuang
Differential Revision: D38667529
Pulled By: pdillinger
fbshipit-source-id: 49af3fe7f47e5b61162809a78b76c769fd519fba
2 years ago
|
|
|
*out_base_cache_key = OffsetableCacheKey(db_id, db_session_id, file_num);
|
New stable, fixed-length cache keys (#9126)
Summary:
This change standardizes on a new 16-byte cache key format for
block cache (incl compressed and secondary) and persistent cache (but
not table cache and row cache).
The goal is a really fast cache key with practically ideal stability and
uniqueness properties without external dependencies (e.g. from FileSystem).
A fixed key size of 16 bytes should enable future optimizations to the
concurrent hash table for block cache, which is a heavy CPU user /
bottleneck, but there appears to be measurable performance improvement
even with no changes to LRUCache.
This change replaces a lot of disjointed and ugly code handling cache
keys with calls to a simple, clean new internal API (cache_key.h).
(Preserving the old cache key logic under an option would be very ugly
and likely negate the performance gain of the new approach. Complete
replacement carries some inherent risk, but I think that's acceptable
with sufficient analysis and testing.)
The scheme for encoding new cache keys is complicated but explained
in cache_key.cc.
Also: EndianSwapValue is moved to math.h to be next to other bit
operations. (Explains some new include "math.h".) ReverseBits operation
added and unit tests added to hash_test for both.
Fixes https://github.com/facebook/rocksdb/issues/7405 (presuming a root cause)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9126
Test Plan:
### Basic correctness
Several tests needed updates to work with the new functionality, mostly
because we are no longer relying on filesystem for stable cache keys
so table builders & readers need more context info to agree on cache
keys. This functionality is so core, a huge number of existing tests
exercise the cache key functionality.
### Performance
Create db with
`TEST_TMPDIR=/dev/shm ./db_bench -bloom_bits=10 -benchmarks=fillrandom -num=3000000 -partition_index_and_filters`
And test performance with
`TEST_TMPDIR=/dev/shm ./db_bench -readonly -use_existing_db -bloom_bits=10 -benchmarks=readrandom -num=3000000 -duration=30 -cache_index_and_filter_blocks -cache_size=250000 -threads=4`
using DEBUG_LEVEL=0 and simultaneous before & after runs.
Before ops/sec, avg over 100 runs: 121924
After ops/sec, avg over 100 runs: 125385 (+2.8%)
### Collision probability
I have built a tool, ./cache_bench -stress_cache_key to broadly simulate host-wide cache activity
over many months, by making some pessimistic simplifying assumptions:
* Every generated file has a cache entry for every byte offset in the file (contiguous range of cache keys)
* All of every file is cached for its entire lifetime
We use a simple table with skewed address assignment and replacement on address collision
to simulate files coming & going, with quite a variance (super-Poisson) in ages. Some output
with `./cache_bench -stress_cache_key -sck_keep_bits=40`:
```
Total cache or DBs size: 32TiB Writing 925.926 MiB/s or 76.2939TiB/day
Multiply by 9.22337e+18 to correct for simulation losses (but still assume whole file cached)
```
These come from default settings of 2.5M files per day of 32 MB each, and
`-sck_keep_bits=40` means that to represent a single file, we are only keeping 40 bits of
the 128-bit cache key. With file size of 2\*\*25 contiguous keys (pessimistic), our simulation
is about 2\*\*(128-40-25) or about 9 billion billion times more prone to collision than reality.
More default assumptions, relatively pessimistic:
* 100 DBs in same process (doesn't matter much)
* Re-open DB in same process (new session ID related to old session ID) on average
every 100 files generated
* Restart process (all new session IDs unrelated to old) 24 times per day
After enough data, we get a result at the end:
```
(keep 40 bits) 17 collisions after 2 x 90 days, est 10.5882 days between (9.76592e+19 corrected)
```
If we believe the (pessimistic) simulation and the mathematical generalization, we would need to run a billion machines all for 97 billion days to expect a cache key collision. To help verify that our generalization ("corrected") is robust, we can make our simulation more precise with `-sck_keep_bits=41` and `42`, which takes more running time to get enough data:
```
(keep 41 bits) 16 collisions after 4 x 90 days, est 22.5 days between (1.03763e+20 corrected)
(keep 42 bits) 19 collisions after 10 x 90 days, est 47.3684 days between (1.09224e+20 corrected)
```
The generalized prediction still holds. With the `-sck_randomize` option, we can see that we are beating "random" cache keys (except offsets still non-randomized) by a modest amount (roughly 20x less collision prone than random), which should make us reasonably comfortable even in "degenerate" cases:
```
197 collisions after 1 x 90 days, est 0.456853 days between (4.21372e+18 corrected)
```
I've run other tests to validate other conditions behave as expected, never behaving "worse than random" unless we start chopping off structured data.
Reviewed By: zhichao-cao
Differential Revision: D33171746
Pulled By: pdillinger
fbshipit-source-id: f16a57e369ed37be5e7e33525ace848d0537c88f
3 years ago
|
|
|
}
|
|
|
|
|
|
|
|
CacheKey BlockBasedTable::GetCacheKey(const OffsetableCacheKey& base_cache_key,
|
|
|
|
const BlockHandle& handle) {
|
|
|
|
// Minimum block size is 5 bytes; therefore we can trim off two lower bits
|
|
|
|
// from offet.
|
|
|
|
return base_cache_key.WithOffset(handle.offset() >> 2);
|
|
|
|
}
|
|
|
|
|
|
|
|
Status BlockBasedTable::Open(
|
|
|
|
const ReadOptions& read_options, const ImmutableOptions& ioptions,
|
|
|
|
const EnvOptions& env_options, const BlockBasedTableOptions& table_options,
|
|
|
|
const InternalKeyComparator& internal_comparator,
|
|
|
|
std::unique_ptr<RandomAccessFileReader>&& file, uint64_t file_size,
|
|
|
|
uint8_t block_protection_bytes_per_key,
|
Record and use the tail size to prefetch table tail (#11406)
Summary:
**Context:**
We prefetch the tail part of a SST file (i.e, the blocks after data blocks till the end of the file) during each SST file open in hope to prefetch all the stuff at once ahead of time for later read e.g, footer, meta index, filter/index etc. The existing approach to estimate the tail size to prefetch is through `TailPrefetchStats` heuristics introduced in https://github.com/facebook/rocksdb/pull/4156, which has caused small reads in unlucky case (e.g, small read into the tail buffer during table open in thread 1 under the same BlockBasedTableFactory object can make thread 2's tail prefetching use a small size that it shouldn't) and is hard to debug. Therefore we decide to record the exact tail size and use it directly to prefetch tail of the SST instead of relying heuristics.
**Summary:**
- Obtain and record in manifest the tail size in `BlockBasedTableBuilder::Finish()`
- For backward compatibility, we fall back to TailPrefetchStats and last to simple heuristics that the tail size is a linear portion of the file size - see PR conversation for more.
- Make`tail_start_offset` part of the table properties and deduct tail size to record in manifest for external files (e.g, file ingestion, import CF) and db repair (with no access to manifest).
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11406
Test Plan:
1. New UT
2. db bench
Note: db bench on /tmp/ where direct read is supported is too slow to finish and the default pinning setting in db bench is not helpful to profile # sst read of Get. Therefore I hacked the following to obtain the following comparison.
```
diff --git a/table/block_based/block_based_table_reader.cc b/table/block_based/block_based_table_reader.cc
index bd5669f0f..791484c1f 100644
--- a/table/block_based/block_based_table_reader.cc
+++ b/table/block_based/block_based_table_reader.cc
@@ -838,7 +838,7 @@ Status BlockBasedTable::PrefetchTail(
&tail_prefetch_size);
// Try file system prefetch
- if (!file->use_direct_io() && !force_direct_prefetch) {
+ if (false && !file->use_direct_io() && !force_direct_prefetch) {
if (!file->Prefetch(prefetch_off, prefetch_len, ro.rate_limiter_priority)
.IsNotSupported()) {
prefetch_buffer->reset(new FilePrefetchBuffer(
diff --git a/tools/db_bench_tool.cc b/tools/db_bench_tool.cc
index ea40f5fa0..39a0ac385 100644
--- a/tools/db_bench_tool.cc
+++ b/tools/db_bench_tool.cc
@@ -4191,6 +4191,8 @@ class Benchmark {
std::shared_ptr<TableFactory>(NewCuckooTableFactory(table_options));
} else {
BlockBasedTableOptions block_based_options;
+ block_based_options.metadata_cache_options.partition_pinning =
+ PinningTier::kAll;
block_based_options.checksum =
static_cast<ChecksumType>(FLAGS_checksum_type);
if (FLAGS_use_hash_search) {
```
Create DB
```
./db_bench --bloom_bits=3 --use_existing_db=1 --seed=1682546046158958 --partition_index_and_filters=1 --statistics=1 -db=/dev/shm/testdb/ -benchmarks=readrandom -key_size=3200 -value_size=512 -num=1000000 -write_buffer_size=6550000 -disable_auto_compactions=false -target_file_size_base=6550000 -compression_type=none
```
ReadRandom
```
./db_bench --bloom_bits=3 --use_existing_db=1 --seed=1682546046158958 --partition_index_and_filters=1 --statistics=1 -db=/dev/shm/testdb/ -benchmarks=readrandom -key_size=3200 -value_size=512 -num=1000000 -write_buffer_size=6550000 -disable_auto_compactions=false -target_file_size_base=6550000 -compression_type=none
```
(a) Existing (Use TailPrefetchStats for tail size + use seperate prefetch buffer in PartitionedFilter/IndexReader::CacheDependencies())
```
rocksdb.table.open.prefetch.tail.hit COUNT : 3395
rocksdb.sst.read.micros P50 : 5.655570 P95 : 9.931396 P99 : 14.845454 P100 : 585.000000 COUNT : 999905 SUM : 6590614
```
(b) This PR (Record tail size + use the same tail buffer in PartitionedFilter/IndexReader::CacheDependencies())
```
rocksdb.table.open.prefetch.tail.hit COUNT : 14257
rocksdb.sst.read.micros P50 : 5.173347 P95 : 9.015017 P99 : 12.912610 P100 : 228.000000 COUNT : 998547 SUM : 5976540
```
As we can see, we increase the prefetch tail hit count and decrease SST read count with this PR
3. Test backward compatibility by stepping through reading with post-PR code on a db generated pre-PR.
Reviewed By: pdillinger
Differential Revision: D45413346
Pulled By: hx235
fbshipit-source-id: 7d5e36a60a72477218f79905168d688452a4c064
2 years ago
|
|
|
std::unique_ptr<TableReader>* table_reader, uint64_t tail_size,
|
|
|
|
std::shared_ptr<CacheReservationManager> table_reader_cache_res_mgr,
|
|
|
|
const std::shared_ptr<const SliceTransform>& prefix_extractor,
|
|
|
|
const bool prefetch_index_and_filter_in_cache, const bool skip_filters,
|
|
|
|
const int level, const bool immortal_table,
|
|
|
|
const SequenceNumber largest_seqno, const bool force_direct_prefetch,
|
|
|
|
TailPrefetchStats* tail_prefetch_stats,
|
|
|
|
BlockCacheTracer* const block_cache_tracer,
|
|
|
|
size_t max_file_size_for_l0_meta_pin, const std::string& cur_db_session_id,
|
Add support to strip / pad timestamp when creating / reading a block based table (#11495)
Summary:
Add support to strip timestamp in block based table builder and pad timestamp in block based table reader.
On the write path, use the per column family option `AdvancedColumnFamilyOptions.persist_user_defined_timestamps` to indicate whether user-defined timestamps should be stripped for all block based tables created for the column family.
On the read path, added a per table `TableReadOption.user_defined_timestamps_persisted` to flag whether the user keys in the table contains user defined timestamps.
This patch is mostly passing the related flags down to the block building/parsing level with the exception of handling the `first_internal_key` in `IndexValue`, which is included in the `IndexBuilder` level. The value part of range deletion entries should have a similar handling, I haven't decided where to best fit this piece of logic, I will do it in a follow up.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11495
Test Plan:
Existing test `BlockBasedTableReaderTest` is parameterized to run with:
1) different UDT test modes: kNone, kNormal, kStripUserDefinedTimestamp
2) all four index types, when index type is `kTwoLevelIndexSearch`, also enables partitioned filters
3) parallel vs non-parallel compression
4) enable/disable compression dictionary.
Also added tests for API `BlockBasedTableReader::NewIterator`.
`PartitionedFilterBlockTest` is parameterized to run with different UDT test modes:kNone, kNormal, kStripUserDefinedTimestamp.
```
make all check
./block_based_table_reader_test
./partitioned_filter_block_test
```
Reviewed By: ltamasi
Differential Revision: D46344577
Pulled By: jowlyzhang
fbshipit-source-id: 93ac8542b19319d1298712b8bed908c8831ba675
1 year ago
|
|
|
uint64_t cur_file_num, UniqueId64x2 expected_unique_id,
|
|
|
|
const bool user_defined_timestamps_persisted) {
|
|
|
|
table_reader->reset();
|
|
|
|
|
|
|
|
Status s;
|
|
|
|
Footer footer;
|
|
|
|
std::unique_ptr<FilePrefetchBuffer> prefetch_buffer;
|
|
|
|
|
|
|
|
// From read_options, retain deadline, io_timeout, rate_limiter_priority, and
|
|
|
|
// verify_checksums. In future, we may retain more options.
|
|
|
|
ReadOptions ro;
|
|
|
|
ro.deadline = read_options.deadline;
|
|
|
|
ro.io_timeout = read_options.io_timeout;
|
Set Read rate limiter priority dynamically and pass it to FS (#9996)
Summary:
### Context:
Background compactions and flush generate large reads and writes, and can be long running, especially for universal compaction. In some cases, this can impact foreground reads and writes by users.
### Solution
User, Flush, and Compaction reads share some code path. For this task, we update the rate_limiter_priority in ReadOptions for code paths (e.g. FindTable (mainly in BlockBasedTable::Open()) and various iterators), and eventually update the rate_limiter_priority in IOOptions for FSRandomAccessFile.
**This PR is for the Read path.** The **Read:** dynamic priority for different state are listed as follows:
| State | Normal | Delayed | Stalled |
| ----- | ------ | ------- | ------- |
| Flush (verification read in BuildTable()) | IO_USER | IO_USER | IO_USER |
| Compaction | IO_LOW | IO_USER | IO_USER |
| User | User provided | User provided | User provided |
We will respect the read_options that the user provided and will not set it.
The only sst read for Flush is the verification read in BuildTable(). It claims to be "regard as user read".
**Details**
1. Set read_options.rate_limiter_priority dynamically:
- User: Do not update the read_options. Use the read_options that the user provided.
- Compaction: Update read_options in CompactionJob::ProcessKeyValueCompaction().
- Flush: Update read_options in BuildTable().
2. Pass the rate limiter priority to FSRandomAccessFile functions:
- After calling the FindTable(), read_options is passed through GetTableReader(table_cache.cc), BlockBasedTableFactory::NewTableReader(block_based_table_factory.cc), and BlockBasedTable::Open(). The Open() needs some updates for the ReadOptions variable and the updates are also needed for the called functions, including PrefetchTail(), PrepareIOOptions(), ReadFooterFromFile(), ReadMetaIndexblock(), ReadPropertiesBlock(), PrefetchIndexAndFilterBlocks(), and ReadRangeDelBlock().
- In RandomAccessFileReader, the functions to be updated include Read(), MultiRead(), ReadAsync(), and Prefetch().
- Update the downstream functions of NewIndexIterator(), NewDataBlockIterator(), and BlockBasedTableIterator().
### Test Plans
Add unit tests.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9996
Reviewed By: anand1976
Differential Revision: D36452483
Pulled By: gitbw95
fbshipit-source-id: 60978204a4f849bb9261cb78d9bc1cb56d6008cf
3 years ago
|
|
|
ro.rate_limiter_priority = read_options.rate_limiter_priority;
|
|
|
|
ro.verify_checksums = read_options.verify_checksums;
|
|
|
|
ro.io_activity = read_options.io_activity;
|
|
|
|
|
|
|
|
// prefetch both index and filters, down to all partitions
|
|
|
|
const bool prefetch_all = prefetch_index_and_filter_in_cache || level == 0;
|
|
|
|
const bool preload_all = !table_options.cache_index_and_filter_blocks;
|
|
|
|
|
|
|
|
if (!ioptions.allow_mmap_reads) {
|
|
|
|
s = PrefetchTail(ro, file.get(), file_size, force_direct_prefetch,
|
|
|
|
tail_prefetch_stats, prefetch_all, preload_all,
|
Record and use the tail size to prefetch table tail (#11406)
Summary:
**Context:**
We prefetch the tail part of a SST file (i.e, the blocks after data blocks till the end of the file) during each SST file open in hope to prefetch all the stuff at once ahead of time for later read e.g, footer, meta index, filter/index etc. The existing approach to estimate the tail size to prefetch is through `TailPrefetchStats` heuristics introduced in https://github.com/facebook/rocksdb/pull/4156, which has caused small reads in unlucky case (e.g, small read into the tail buffer during table open in thread 1 under the same BlockBasedTableFactory object can make thread 2's tail prefetching use a small size that it shouldn't) and is hard to debug. Therefore we decide to record the exact tail size and use it directly to prefetch tail of the SST instead of relying heuristics.
**Summary:**
- Obtain and record in manifest the tail size in `BlockBasedTableBuilder::Finish()`
- For backward compatibility, we fall back to TailPrefetchStats and last to simple heuristics that the tail size is a linear portion of the file size - see PR conversation for more.
- Make`tail_start_offset` part of the table properties and deduct tail size to record in manifest for external files (e.g, file ingestion, import CF) and db repair (with no access to manifest).
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11406
Test Plan:
1. New UT
2. db bench
Note: db bench on /tmp/ where direct read is supported is too slow to finish and the default pinning setting in db bench is not helpful to profile # sst read of Get. Therefore I hacked the following to obtain the following comparison.
```
diff --git a/table/block_based/block_based_table_reader.cc b/table/block_based/block_based_table_reader.cc
index bd5669f0f..791484c1f 100644
--- a/table/block_based/block_based_table_reader.cc
+++ b/table/block_based/block_based_table_reader.cc
@@ -838,7 +838,7 @@ Status BlockBasedTable::PrefetchTail(
&tail_prefetch_size);
// Try file system prefetch
- if (!file->use_direct_io() && !force_direct_prefetch) {
+ if (false && !file->use_direct_io() && !force_direct_prefetch) {
if (!file->Prefetch(prefetch_off, prefetch_len, ro.rate_limiter_priority)
.IsNotSupported()) {
prefetch_buffer->reset(new FilePrefetchBuffer(
diff --git a/tools/db_bench_tool.cc b/tools/db_bench_tool.cc
index ea40f5fa0..39a0ac385 100644
--- a/tools/db_bench_tool.cc
+++ b/tools/db_bench_tool.cc
@@ -4191,6 +4191,8 @@ class Benchmark {
std::shared_ptr<TableFactory>(NewCuckooTableFactory(table_options));
} else {
BlockBasedTableOptions block_based_options;
+ block_based_options.metadata_cache_options.partition_pinning =
+ PinningTier::kAll;
block_based_options.checksum =
static_cast<ChecksumType>(FLAGS_checksum_type);
if (FLAGS_use_hash_search) {
```
Create DB
```
./db_bench --bloom_bits=3 --use_existing_db=1 --seed=1682546046158958 --partition_index_and_filters=1 --statistics=1 -db=/dev/shm/testdb/ -benchmarks=readrandom -key_size=3200 -value_size=512 -num=1000000 -write_buffer_size=6550000 -disable_auto_compactions=false -target_file_size_base=6550000 -compression_type=none
```
ReadRandom
```
./db_bench --bloom_bits=3 --use_existing_db=1 --seed=1682546046158958 --partition_index_and_filters=1 --statistics=1 -db=/dev/shm/testdb/ -benchmarks=readrandom -key_size=3200 -value_size=512 -num=1000000 -write_buffer_size=6550000 -disable_auto_compactions=false -target_file_size_base=6550000 -compression_type=none
```
(a) Existing (Use TailPrefetchStats for tail size + use seperate prefetch buffer in PartitionedFilter/IndexReader::CacheDependencies())
```
rocksdb.table.open.prefetch.tail.hit COUNT : 3395
rocksdb.sst.read.micros P50 : 5.655570 P95 : 9.931396 P99 : 14.845454 P100 : 585.000000 COUNT : 999905 SUM : 6590614
```
(b) This PR (Record tail size + use the same tail buffer in PartitionedFilter/IndexReader::CacheDependencies())
```
rocksdb.table.open.prefetch.tail.hit COUNT : 14257
rocksdb.sst.read.micros P50 : 5.173347 P95 : 9.015017 P99 : 12.912610 P100 : 228.000000 COUNT : 998547 SUM : 5976540
```
As we can see, we increase the prefetch tail hit count and decrease SST read count with this PR
3. Test backward compatibility by stepping through reading with post-PR code on a db generated pre-PR.
Reviewed By: pdillinger
Differential Revision: D45413346
Pulled By: hx235
fbshipit-source-id: 7d5e36a60a72477218f79905168d688452a4c064
2 years ago
|
|
|
&prefetch_buffer, ioptions.stats, tail_size,
|
|
|
|
ioptions.logger);
|
|
|
|
// Return error in prefetch path to users.
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
// Should not prefetch for mmap mode.
|
|
|
|
prefetch_buffer.reset(new FilePrefetchBuffer(
|
|
|
|
0 /* readahead_size */, 0 /* max_readahead_size */, false /* enable */,
|
|
|
|
true /* track_min_offset */));
|
|
|
|
}
|
|
|
|
|
|
|
|
// Read in the following order:
|
|
|
|
// 1. Footer
|
|
|
|
// 2. [metaindex block]
|
|
|
|
// 3. [meta block: properties]
|
|
|
|
// 4. [meta block: range deletion tombstone]
|
|
|
|
// 5. [meta block: compression dictionary]
|
|
|
|
// 6. [meta block: index]
|
|
|
|
// 7. [meta block: filter]
|
|
|
|
IOOptions opts;
|
|
|
|
s = file->PrepareIOOptions(ro, opts);
|
|
|
|
if (s.ok()) {
|
|
|
|
s = ReadFooterFromFile(opts, file.get(), *ioptions.fs,
|
|
|
|
prefetch_buffer.get(), file_size, &footer,
|
|
|
|
kBlockBasedTableMagicNumber);
|
|
|
|
}
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
if (!IsSupportedFormatVersion(footer.format_version())) {
|
|
|
|
return Status::Corruption(
|
|
|
|
"Unknown Footer version. Maybe this file was created with newer "
|
|
|
|
"version of RocksDB?");
|
|
|
|
}
|
|
|
|
|
|
|
|
BlockCacheLookupContext lookup_context{TableReaderCaller::kPrefetch};
|
Add support to strip / pad timestamp when creating / reading a block based table (#11495)
Summary:
Add support to strip timestamp in block based table builder and pad timestamp in block based table reader.
On the write path, use the per column family option `AdvancedColumnFamilyOptions.persist_user_defined_timestamps` to indicate whether user-defined timestamps should be stripped for all block based tables created for the column family.
On the read path, added a per table `TableReadOption.user_defined_timestamps_persisted` to flag whether the user keys in the table contains user defined timestamps.
This patch is mostly passing the related flags down to the block building/parsing level with the exception of handling the `first_internal_key` in `IndexValue`, which is included in the `IndexBuilder` level. The value part of range deletion entries should have a similar handling, I haven't decided where to best fit this piece of logic, I will do it in a follow up.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11495
Test Plan:
Existing test `BlockBasedTableReaderTest` is parameterized to run with:
1) different UDT test modes: kNone, kNormal, kStripUserDefinedTimestamp
2) all four index types, when index type is `kTwoLevelIndexSearch`, also enables partitioned filters
3) parallel vs non-parallel compression
4) enable/disable compression dictionary.
Also added tests for API `BlockBasedTableReader::NewIterator`.
`PartitionedFilterBlockTest` is parameterized to run with different UDT test modes:kNone, kNormal, kStripUserDefinedTimestamp.
```
make all check
./block_based_table_reader_test
./partitioned_filter_block_test
```
Reviewed By: ltamasi
Differential Revision: D46344577
Pulled By: jowlyzhang
fbshipit-source-id: 93ac8542b19319d1298712b8bed908c8831ba675
1 year ago
|
|
|
Rep* rep = new BlockBasedTable::Rep(
|
|
|
|
ioptions, env_options, table_options, internal_comparator, skip_filters,
|
|
|
|
file_size, level, immortal_table, user_defined_timestamps_persisted);
|
|
|
|
rep->file = std::move(file);
|
|
|
|
rep->footer = footer;
|
|
|
|
|
|
|
|
// For fully portable/stable cache keys, we need to read the properties
|
|
|
|
// block before setting up cache keys. TODO: consider setting up a bootstrap
|
|
|
|
// cache key for PersistentCache to use for metaindex and properties blocks.
|
|
|
|
rep->persistent_cache_options = PersistentCacheOptions();
|
|
|
|
|
|
|
|
// Meta-blocks are not dictionary compressed. Explicitly set the dictionary
|
|
|
|
// handle to null, otherwise it may be seen as uninitialized during the below
|
|
|
|
// meta-block reads.
|
|
|
|
rep->compression_dict_handle = BlockHandle::NullBlockHandle();
|
|
|
|
|
|
|
|
rep->create_context.protection_bytes_per_key = block_protection_bytes_per_key;
|
|
|
|
// Read metaindex
|
|
|
|
std::unique_ptr<BlockBasedTable> new_table(
|
|
|
|
new BlockBasedTable(rep, block_cache_tracer));
|
|
|
|
std::unique_ptr<Block> metaindex;
|
|
|
|
std::unique_ptr<InternalIterator> metaindex_iter;
|
|
|
|
s = new_table->ReadMetaIndexBlock(ro, prefetch_buffer.get(), &metaindex,
|
|
|
|
&metaindex_iter);
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
// Populates table_properties and some fields that depend on it,
|
|
|
|
// such as index_type.
|
|
|
|
s = new_table->ReadPropertiesBlock(ro, prefetch_buffer.get(),
|
|
|
|
metaindex_iter.get(), largest_seqno);
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
Always verify SST unique IDs on SST file open (#10532)
Summary:
Although we've been tracking SST unique IDs in the DB manifest
unconditionally, checking has been opt-in and with an extra pass at DB::Open
time. This changes the behavior of `verify_sst_unique_id_in_manifest` to
check unique ID against manifest every time an SST file is opened through
table cache (normal DB operations), replacing the explicit pass over files
at DB::Open time. This change also enables the option by default and
removes the "EXPERIMENTAL" designation.
One possible criticism is that the option no longer ensures the integrity
of a DB at Open time. This is far from an all-or-nothing issue. Verifying
the IDs of all SST files hardly ensures all the data in the DB is readable.
(VerifyChecksum is supposed to do that.) Also, with
max_open_files=-1 (default, extremely common), all SST files are
opened at DB::Open time anyway.
Implementation details:
* `VerifySstUniqueIdInManifest()` functions are the extra/explicit pass
that is now removed.
* Unit tests that manipulate/corrupt table properties have to opt out of
this check, because that corrupts the "actual" unique id. (And even for
testing we don't currently have a mechanism to set "no unique id"
in the in-memory file metadata for new files.)
* A lot of other unit test churn relates to (a) default checking on, and
(b) checking on SST open even without DB::Open (e.g. on flush)
* Use `FileMetaData` for more `TableCache` operations (in place of
`FileDescriptor`) so that we have access to the unique_id whenever
we might need to open an SST file. **There is the possibility of
performance impact because we can no longer use the more
localized `fd` part of an `FdWithKeyRange` but instead follow the
`file_metadata` pointer. However, this change (possible regression)
is only done for `GetMemoryUsageByTableReaders`.**
* Removed a completely unnecessary constructor overload of
`TableReaderOptions`
Possible follow-up:
* Verification only happens when opening through table cache. Are there
more places where this should happen?
* Improve error message when there is a file size mismatch vs. manifest
(FIXME added in the appropriate place).
* I'm not sure there's a justification for `FileDescriptor` to be distinct from
`FileMetaData`.
* I'm skeptical that `FdWithKeyRange` really still makes sense for
optimizing some data locality by duplicating some data in memory, but I
could be wrong.
* An unnecessary overload of NewTableReader was recently added, in
the public API nonetheless (though unusable there). It should be cleaned
up to put most things under `TableReaderOptions`.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10532
Test Plan:
updated unit tests
Performance test showing no significant difference (just noise I think):
`./db_bench -benchmarks=readwhilewriting[-X10] -num=3000000 -disable_wal=1 -bloom_bits=8 -write_buffer_size=1000000 -target_file_size_base=1000000`
Before: readwhilewriting [AVG 10 runs] : 68702 (± 6932) ops/sec
After: readwhilewriting [AVG 10 runs] : 68239 (± 7198) ops/sec
Reviewed By: jay-zhuang
Differential Revision: D38765551
Pulled By: pdillinger
fbshipit-source-id: a827a708155f12344ab2a5c16e7701c7636da4c2
2 years ago
|
|
|
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
// Populate BlockCreateContext
|
|
|
|
bool blocks_definitely_zstd_compressed =
|
|
|
|
rep->table_properties &&
|
|
|
|
(rep->table_properties->compression_name ==
|
|
|
|
CompressionTypeToString(kZSTD) ||
|
|
|
|
rep->table_properties->compression_name ==
|
|
|
|
CompressionTypeToString(kZSTDNotFinalCompression));
|
|
|
|
rep->create_context = BlockCreateContext(
|
|
|
|
&rep->table_options, rep->ioptions.stats,
|
|
|
|
blocks_definitely_zstd_compressed, block_protection_bytes_per_key,
|
|
|
|
rep->internal_comparator.user_comparator(), rep->index_value_is_full,
|
|
|
|
rep->index_has_first_key);
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
|
Always verify SST unique IDs on SST file open (#10532)
Summary:
Although we've been tracking SST unique IDs in the DB manifest
unconditionally, checking has been opt-in and with an extra pass at DB::Open
time. This changes the behavior of `verify_sst_unique_id_in_manifest` to
check unique ID against manifest every time an SST file is opened through
table cache (normal DB operations), replacing the explicit pass over files
at DB::Open time. This change also enables the option by default and
removes the "EXPERIMENTAL" designation.
One possible criticism is that the option no longer ensures the integrity
of a DB at Open time. This is far from an all-or-nothing issue. Verifying
the IDs of all SST files hardly ensures all the data in the DB is readable.
(VerifyChecksum is supposed to do that.) Also, with
max_open_files=-1 (default, extremely common), all SST files are
opened at DB::Open time anyway.
Implementation details:
* `VerifySstUniqueIdInManifest()` functions are the extra/explicit pass
that is now removed.
* Unit tests that manipulate/corrupt table properties have to opt out of
this check, because that corrupts the "actual" unique id. (And even for
testing we don't currently have a mechanism to set "no unique id"
in the in-memory file metadata for new files.)
* A lot of other unit test churn relates to (a) default checking on, and
(b) checking on SST open even without DB::Open (e.g. on flush)
* Use `FileMetaData` for more `TableCache` operations (in place of
`FileDescriptor`) so that we have access to the unique_id whenever
we might need to open an SST file. **There is the possibility of
performance impact because we can no longer use the more
localized `fd` part of an `FdWithKeyRange` but instead follow the
`file_metadata` pointer. However, this change (possible regression)
is only done for `GetMemoryUsageByTableReaders`.**
* Removed a completely unnecessary constructor overload of
`TableReaderOptions`
Possible follow-up:
* Verification only happens when opening through table cache. Are there
more places where this should happen?
* Improve error message when there is a file size mismatch vs. manifest
(FIXME added in the appropriate place).
* I'm not sure there's a justification for `FileDescriptor` to be distinct from
`FileMetaData`.
* I'm skeptical that `FdWithKeyRange` really still makes sense for
optimizing some data locality by duplicating some data in memory, but I
could be wrong.
* An unnecessary overload of NewTableReader was recently added, in
the public API nonetheless (though unusable there). It should be cleaned
up to put most things under `TableReaderOptions`.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10532
Test Plan:
updated unit tests
Performance test showing no significant difference (just noise I think):
`./db_bench -benchmarks=readwhilewriting[-X10] -num=3000000 -disable_wal=1 -bloom_bits=8 -write_buffer_size=1000000 -target_file_size_base=1000000`
Before: readwhilewriting [AVG 10 runs] : 68702 (± 6932) ops/sec
After: readwhilewriting [AVG 10 runs] : 68239 (± 7198) ops/sec
Reviewed By: jay-zhuang
Differential Revision: D38765551
Pulled By: pdillinger
fbshipit-source-id: a827a708155f12344ab2a5c16e7701c7636da4c2
2 years ago
|
|
|
// Check expected unique id if provided
|
|
|
|
if (expected_unique_id != kNullUniqueId64x2) {
|
|
|
|
auto props = rep->table_properties;
|
|
|
|
if (!props) {
|
|
|
|
return Status::Corruption("Missing table properties on file " +
|
|
|
|
std::to_string(cur_file_num) +
|
|
|
|
" with known unique ID");
|
|
|
|
}
|
|
|
|
UniqueId64x2 actual_unique_id{};
|
|
|
|
s = GetSstInternalUniqueId(props->db_id, props->db_session_id,
|
|
|
|
props->orig_file_number, &actual_unique_id,
|
|
|
|
/*force*/ true);
|
|
|
|
assert(s.ok()); // because force=true
|
|
|
|
if (expected_unique_id != actual_unique_id) {
|
|
|
|
return Status::Corruption(
|
|
|
|
"Mismatch in unique ID on table file " +
|
|
|
|
std::to_string(cur_file_num) +
|
|
|
|
". Expected: " + InternalUniqueIdToHumanString(&expected_unique_id) +
|
|
|
|
" Actual: " + InternalUniqueIdToHumanString(&actual_unique_id));
|
|
|
|
}
|
|
|
|
TEST_SYNC_POINT_CALLBACK("BlockBasedTable::Open::PassedVerifyUniqueId",
|
|
|
|
&actual_unique_id);
|
|
|
|
} else {
|
|
|
|
TEST_SYNC_POINT_CALLBACK("BlockBasedTable::Open::SkippedVerifyUniqueId",
|
|
|
|
nullptr);
|
|
|
|
if (ioptions.verify_sst_unique_id_in_manifest && ioptions.logger) {
|
|
|
|
// A crude but isolated way of reporting unverified files. This should not
|
|
|
|
// be an ongoing concern so doesn't deserve a place in Statistics IMHO.
|
|
|
|
static std::atomic<uint64_t> unverified_count{0};
|
|
|
|
auto prev_count =
|
|
|
|
unverified_count.fetch_add(1, std::memory_order_relaxed);
|
|
|
|
if (prev_count == 0) {
|
|
|
|
ROCKS_LOG_WARN(
|
|
|
|
ioptions.logger,
|
|
|
|
"At least one SST file opened without unique ID to verify: %" PRIu64
|
|
|
|
".sst",
|
|
|
|
cur_file_num);
|
|
|
|
} else if (prev_count % 1000 == 0) {
|
|
|
|
ROCKS_LOG_WARN(
|
|
|
|
ioptions.logger,
|
|
|
|
"Another ~1000 SST files opened without unique ID to verify");
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Set up prefix extracto as needed
|
|
|
|
bool force_null_table_prefix_extractor = false;
|
|
|
|
TEST_SYNC_POINT_CALLBACK(
|
|
|
|
"BlockBasedTable::Open::ForceNullTablePrefixExtractor",
|
|
|
|
&force_null_table_prefix_extractor);
|
|
|
|
if (force_null_table_prefix_extractor) {
|
|
|
|
assert(!rep->table_prefix_extractor);
|
|
|
|
} else if (!PrefixExtractorChangedHelper(rep->table_properties.get(),
|
|
|
|
prefix_extractor.get())) {
|
|
|
|
// Establish fast path for unchanged prefix_extractor
|
|
|
|
rep->table_prefix_extractor = prefix_extractor;
|
|
|
|
} else {
|
|
|
|
// Current prefix_extractor doesn't match table
|
|
|
|
if (rep->table_properties) {
|
|
|
|
//**TODO: If/When the DBOptions has a registry in it, the ConfigOptions
|
|
|
|
// will need to use it
|
|
|
|
ConfigOptions config_options;
|
|
|
|
Status st = SliceTransform::CreateFromString(
|
|
|
|
config_options, rep->table_properties->prefix_extractor_name,
|
|
|
|
&(rep->table_prefix_extractor));
|
|
|
|
if (!st.ok()) {
|
|
|
|
//**TODO: Should this be error be returned or swallowed?
|
|
|
|
ROCKS_LOG_ERROR(rep->ioptions.logger,
|
|
|
|
"Failed to create prefix extractor[%s]: %s",
|
|
|
|
rep->table_properties->prefix_extractor_name.c_str(),
|
|
|
|
st.ToString().c_str());
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
New stable, fixed-length cache keys (#9126)
Summary:
This change standardizes on a new 16-byte cache key format for
block cache (incl compressed and secondary) and persistent cache (but
not table cache and row cache).
The goal is a really fast cache key with practically ideal stability and
uniqueness properties without external dependencies (e.g. from FileSystem).
A fixed key size of 16 bytes should enable future optimizations to the
concurrent hash table for block cache, which is a heavy CPU user /
bottleneck, but there appears to be measurable performance improvement
even with no changes to LRUCache.
This change replaces a lot of disjointed and ugly code handling cache
keys with calls to a simple, clean new internal API (cache_key.h).
(Preserving the old cache key logic under an option would be very ugly
and likely negate the performance gain of the new approach. Complete
replacement carries some inherent risk, but I think that's acceptable
with sufficient analysis and testing.)
The scheme for encoding new cache keys is complicated but explained
in cache_key.cc.
Also: EndianSwapValue is moved to math.h to be next to other bit
operations. (Explains some new include "math.h".) ReverseBits operation
added and unit tests added to hash_test for both.
Fixes https://github.com/facebook/rocksdb/issues/7405 (presuming a root cause)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9126
Test Plan:
### Basic correctness
Several tests needed updates to work with the new functionality, mostly
because we are no longer relying on filesystem for stable cache keys
so table builders & readers need more context info to agree on cache
keys. This functionality is so core, a huge number of existing tests
exercise the cache key functionality.
### Performance
Create db with
`TEST_TMPDIR=/dev/shm ./db_bench -bloom_bits=10 -benchmarks=fillrandom -num=3000000 -partition_index_and_filters`
And test performance with
`TEST_TMPDIR=/dev/shm ./db_bench -readonly -use_existing_db -bloom_bits=10 -benchmarks=readrandom -num=3000000 -duration=30 -cache_index_and_filter_blocks -cache_size=250000 -threads=4`
using DEBUG_LEVEL=0 and simultaneous before & after runs.
Before ops/sec, avg over 100 runs: 121924
After ops/sec, avg over 100 runs: 125385 (+2.8%)
### Collision probability
I have built a tool, ./cache_bench -stress_cache_key to broadly simulate host-wide cache activity
over many months, by making some pessimistic simplifying assumptions:
* Every generated file has a cache entry for every byte offset in the file (contiguous range of cache keys)
* All of every file is cached for its entire lifetime
We use a simple table with skewed address assignment and replacement on address collision
to simulate files coming & going, with quite a variance (super-Poisson) in ages. Some output
with `./cache_bench -stress_cache_key -sck_keep_bits=40`:
```
Total cache or DBs size: 32TiB Writing 925.926 MiB/s or 76.2939TiB/day
Multiply by 9.22337e+18 to correct for simulation losses (but still assume whole file cached)
```
These come from default settings of 2.5M files per day of 32 MB each, and
`-sck_keep_bits=40` means that to represent a single file, we are only keeping 40 bits of
the 128-bit cache key. With file size of 2\*\*25 contiguous keys (pessimistic), our simulation
is about 2\*\*(128-40-25) or about 9 billion billion times more prone to collision than reality.
More default assumptions, relatively pessimistic:
* 100 DBs in same process (doesn't matter much)
* Re-open DB in same process (new session ID related to old session ID) on average
every 100 files generated
* Restart process (all new session IDs unrelated to old) 24 times per day
After enough data, we get a result at the end:
```
(keep 40 bits) 17 collisions after 2 x 90 days, est 10.5882 days between (9.76592e+19 corrected)
```
If we believe the (pessimistic) simulation and the mathematical generalization, we would need to run a billion machines all for 97 billion days to expect a cache key collision. To help verify that our generalization ("corrected") is robust, we can make our simulation more precise with `-sck_keep_bits=41` and `42`, which takes more running time to get enough data:
```
(keep 41 bits) 16 collisions after 4 x 90 days, est 22.5 days between (1.03763e+20 corrected)
(keep 42 bits) 19 collisions after 10 x 90 days, est 47.3684 days between (1.09224e+20 corrected)
```
The generalized prediction still holds. With the `-sck_randomize` option, we can see that we are beating "random" cache keys (except offsets still non-randomized) by a modest amount (roughly 20x less collision prone than random), which should make us reasonably comfortable even in "degenerate" cases:
```
197 collisions after 1 x 90 days, est 0.456853 days between (4.21372e+18 corrected)
```
I've run other tests to validate other conditions behave as expected, never behaving "worse than random" unless we start chopping off structured data.
Reviewed By: zhichao-cao
Differential Revision: D33171746
Pulled By: pdillinger
fbshipit-source-id: f16a57e369ed37be5e7e33525ace848d0537c88f
3 years ago
|
|
|
// With properties loaded, we can set up portable/stable cache keys
|
|
|
|
SetupBaseCacheKey(rep->table_properties.get(), cur_db_session_id,
|
Derive cache keys from SST unique IDs (#10394)
Summary:
... so that cache keys can be derived from DB manifest data
before reading the file from storage--so that every part of the file
can potentially go in a persistent cache.
See updated comments in cache_key.cc for technical details. Importantly,
the new cache key encoding uses some fancy but efficient math to pack
data into the cache key without depending on the sizes of the various
pieces. This simplifies some existing code creating cache keys, like
cache warming before the file size is known.
This should provide us an essentially permanent mapping between SST
unique IDs and base cache keys, with the ability to "upgrade" SST
unique IDs (and thus cache keys) with new SST format_versions.
These cache keys are of similar, perhaps indistinguishable quality to
the previous generation. Before this change (see "corrected" days
between collision):
```
./cache_bench -stress_cache_key -sck_keep_bits=43
18 collisions after 2 x 90 days, est 10 days between (1.15292e+19 corrected)
```
After this change (keep 43 bits, up through 50, to validate "trajectory"
is ok on "corrected" days between collision):
```
19 collisions after 3 x 90 days, est 14.2105 days between (1.63836e+19 corrected)
16 collisions after 5 x 90 days, est 28.125 days between (1.6213e+19 corrected)
15 collisions after 7 x 90 days, est 42 days between (1.21057e+19 corrected)
15 collisions after 17 x 90 days, est 102 days between (1.46997e+19 corrected)
15 collisions after 49 x 90 days, est 294 days between (2.11849e+19 corrected)
15 collisions after 62 x 90 days, est 372 days between (1.34027e+19 corrected)
15 collisions after 53 x 90 days, est 318 days between (5.72858e+18 corrected)
15 collisions after 309 x 90 days, est 1854 days between (1.66994e+19 corrected)
```
However, the change does modify (probably weaken) the "guaranteed unique" promise from this
> SST files generated in a single process are guaranteed to have unique cache keys, unless/until number session ids * max file number = 2**86
to this (see https://github.com/facebook/rocksdb/issues/10388)
> With the DB id limitation, we only have nice guaranteed unique cache keys for files generated in a single process until biggest session_id_counter and offset_in_file reach combined 64 bits
I don't think this is a practical concern, though.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10394
Test Plan: unit tests updated, see simulation results above
Reviewed By: jay-zhuang
Differential Revision: D38667529
Pulled By: pdillinger
fbshipit-source-id: 49af3fe7f47e5b61162809a78b76c769fd519fba
2 years ago
|
|
|
cur_file_num, &rep->base_cache_key);
|
New stable, fixed-length cache keys (#9126)
Summary:
This change standardizes on a new 16-byte cache key format for
block cache (incl compressed and secondary) and persistent cache (but
not table cache and row cache).
The goal is a really fast cache key with practically ideal stability and
uniqueness properties without external dependencies (e.g. from FileSystem).
A fixed key size of 16 bytes should enable future optimizations to the
concurrent hash table for block cache, which is a heavy CPU user /
bottleneck, but there appears to be measurable performance improvement
even with no changes to LRUCache.
This change replaces a lot of disjointed and ugly code handling cache
keys with calls to a simple, clean new internal API (cache_key.h).
(Preserving the old cache key logic under an option would be very ugly
and likely negate the performance gain of the new approach. Complete
replacement carries some inherent risk, but I think that's acceptable
with sufficient analysis and testing.)
The scheme for encoding new cache keys is complicated but explained
in cache_key.cc.
Also: EndianSwapValue is moved to math.h to be next to other bit
operations. (Explains some new include "math.h".) ReverseBits operation
added and unit tests added to hash_test for both.
Fixes https://github.com/facebook/rocksdb/issues/7405 (presuming a root cause)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9126
Test Plan:
### Basic correctness
Several tests needed updates to work with the new functionality, mostly
because we are no longer relying on filesystem for stable cache keys
so table builders & readers need more context info to agree on cache
keys. This functionality is so core, a huge number of existing tests
exercise the cache key functionality.
### Performance
Create db with
`TEST_TMPDIR=/dev/shm ./db_bench -bloom_bits=10 -benchmarks=fillrandom -num=3000000 -partition_index_and_filters`
And test performance with
`TEST_TMPDIR=/dev/shm ./db_bench -readonly -use_existing_db -bloom_bits=10 -benchmarks=readrandom -num=3000000 -duration=30 -cache_index_and_filter_blocks -cache_size=250000 -threads=4`
using DEBUG_LEVEL=0 and simultaneous before & after runs.
Before ops/sec, avg over 100 runs: 121924
After ops/sec, avg over 100 runs: 125385 (+2.8%)
### Collision probability
I have built a tool, ./cache_bench -stress_cache_key to broadly simulate host-wide cache activity
over many months, by making some pessimistic simplifying assumptions:
* Every generated file has a cache entry for every byte offset in the file (contiguous range of cache keys)
* All of every file is cached for its entire lifetime
We use a simple table with skewed address assignment and replacement on address collision
to simulate files coming & going, with quite a variance (super-Poisson) in ages. Some output
with `./cache_bench -stress_cache_key -sck_keep_bits=40`:
```
Total cache or DBs size: 32TiB Writing 925.926 MiB/s or 76.2939TiB/day
Multiply by 9.22337e+18 to correct for simulation losses (but still assume whole file cached)
```
These come from default settings of 2.5M files per day of 32 MB each, and
`-sck_keep_bits=40` means that to represent a single file, we are only keeping 40 bits of
the 128-bit cache key. With file size of 2\*\*25 contiguous keys (pessimistic), our simulation
is about 2\*\*(128-40-25) or about 9 billion billion times more prone to collision than reality.
More default assumptions, relatively pessimistic:
* 100 DBs in same process (doesn't matter much)
* Re-open DB in same process (new session ID related to old session ID) on average
every 100 files generated
* Restart process (all new session IDs unrelated to old) 24 times per day
After enough data, we get a result at the end:
```
(keep 40 bits) 17 collisions after 2 x 90 days, est 10.5882 days between (9.76592e+19 corrected)
```
If we believe the (pessimistic) simulation and the mathematical generalization, we would need to run a billion machines all for 97 billion days to expect a cache key collision. To help verify that our generalization ("corrected") is robust, we can make our simulation more precise with `-sck_keep_bits=41` and `42`, which takes more running time to get enough data:
```
(keep 41 bits) 16 collisions after 4 x 90 days, est 22.5 days between (1.03763e+20 corrected)
(keep 42 bits) 19 collisions after 10 x 90 days, est 47.3684 days between (1.09224e+20 corrected)
```
The generalized prediction still holds. With the `-sck_randomize` option, we can see that we are beating "random" cache keys (except offsets still non-randomized) by a modest amount (roughly 20x less collision prone than random), which should make us reasonably comfortable even in "degenerate" cases:
```
197 collisions after 1 x 90 days, est 0.456853 days between (4.21372e+18 corrected)
```
I've run other tests to validate other conditions behave as expected, never behaving "worse than random" unless we start chopping off structured data.
Reviewed By: zhichao-cao
Differential Revision: D33171746
Pulled By: pdillinger
fbshipit-source-id: f16a57e369ed37be5e7e33525ace848d0537c88f
3 years ago
|
|
|
|
|
|
|
rep->persistent_cache_options =
|
|
|
|
PersistentCacheOptions(rep->table_options.persistent_cache,
|
|
|
|
rep->base_cache_key, rep->ioptions.stats);
|
|
|
|
|
Add support to strip / pad timestamp when creating / reading a block based table (#11495)
Summary:
Add support to strip timestamp in block based table builder and pad timestamp in block based table reader.
On the write path, use the per column family option `AdvancedColumnFamilyOptions.persist_user_defined_timestamps` to indicate whether user-defined timestamps should be stripped for all block based tables created for the column family.
On the read path, added a per table `TableReadOption.user_defined_timestamps_persisted` to flag whether the user keys in the table contains user defined timestamps.
This patch is mostly passing the related flags down to the block building/parsing level with the exception of handling the `first_internal_key` in `IndexValue`, which is included in the `IndexBuilder` level. The value part of range deletion entries should have a similar handling, I haven't decided where to best fit this piece of logic, I will do it in a follow up.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11495
Test Plan:
Existing test `BlockBasedTableReaderTest` is parameterized to run with:
1) different UDT test modes: kNone, kNormal, kStripUserDefinedTimestamp
2) all four index types, when index type is `kTwoLevelIndexSearch`, also enables partitioned filters
3) parallel vs non-parallel compression
4) enable/disable compression dictionary.
Also added tests for API `BlockBasedTableReader::NewIterator`.
`PartitionedFilterBlockTest` is parameterized to run with different UDT test modes:kNone, kNormal, kStripUserDefinedTimestamp.
```
make all check
./block_based_table_reader_test
./partitioned_filter_block_test
```
Reviewed By: ltamasi
Differential Revision: D46344577
Pulled By: jowlyzhang
fbshipit-source-id: 93ac8542b19319d1298712b8bed908c8831ba675
1 year ago
|
|
|
// TODO(yuzhangyu): handle range deletion entries for UDT in memtable only.
|
|
|
|
s = new_table->ReadRangeDelBlock(ro, prefetch_buffer.get(),
|
|
|
|
metaindex_iter.get(), internal_comparator,
|
|
|
|
&lookup_context);
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
s = new_table->PrefetchIndexAndFilterBlocks(
|
|
|
|
ro, prefetch_buffer.get(), metaindex_iter.get(), new_table.get(),
|
|
|
|
prefetch_all, table_options, level, file_size,
|
|
|
|
max_file_size_for_l0_meta_pin, &lookup_context);
|
|
|
|
|
|
|
|
if (s.ok()) {
|
|
|
|
// Update tail prefetch stats
|
|
|
|
assert(prefetch_buffer.get() != nullptr);
|
|
|
|
if (tail_prefetch_stats != nullptr) {
|
|
|
|
assert(prefetch_buffer->min_offset_read() < file_size);
|
|
|
|
tail_prefetch_stats->RecordEffectiveSize(
|
|
|
|
static_cast<size_t>(file_size) - prefetch_buffer->min_offset_read());
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if (s.ok() && table_reader_cache_res_mgr) {
|
|
|
|
std::size_t mem_usage = new_table->ApproximateMemoryUsage();
|
|
|
|
s = table_reader_cache_res_mgr->MakeCacheReservation(
|
|
|
|
mem_usage, &(rep->table_reader_cache_res_handle));
|
|
|
|
if (s.IsMemoryLimit()) {
|
|
|
|
s = Status::MemoryLimit(
|
|
|
|
"Can't allocate " +
|
|
|
|
kCacheEntryRoleToCamelString[static_cast<std::uint32_t>(
|
|
|
|
CacheEntryRole::kBlockBasedTableReader)] +
|
|
|
|
" due to memory limit based on "
|
|
|
|
"cache capacity for memory allocation");
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if (s.ok()) {
|
|
|
|
*table_reader = std::move(new_table);
|
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
Status BlockBasedTable::PrefetchTail(
|
|
|
|
const ReadOptions& ro, RandomAccessFileReader* file, uint64_t file_size,
|
|
|
|
bool force_direct_prefetch, TailPrefetchStats* tail_prefetch_stats,
|
|
|
|
const bool prefetch_all, const bool preload_all,
|
Record and use the tail size to prefetch table tail (#11406)
Summary:
**Context:**
We prefetch the tail part of a SST file (i.e, the blocks after data blocks till the end of the file) during each SST file open in hope to prefetch all the stuff at once ahead of time for later read e.g, footer, meta index, filter/index etc. The existing approach to estimate the tail size to prefetch is through `TailPrefetchStats` heuristics introduced in https://github.com/facebook/rocksdb/pull/4156, which has caused small reads in unlucky case (e.g, small read into the tail buffer during table open in thread 1 under the same BlockBasedTableFactory object can make thread 2's tail prefetching use a small size that it shouldn't) and is hard to debug. Therefore we decide to record the exact tail size and use it directly to prefetch tail of the SST instead of relying heuristics.
**Summary:**
- Obtain and record in manifest the tail size in `BlockBasedTableBuilder::Finish()`
- For backward compatibility, we fall back to TailPrefetchStats and last to simple heuristics that the tail size is a linear portion of the file size - see PR conversation for more.
- Make`tail_start_offset` part of the table properties and deduct tail size to record in manifest for external files (e.g, file ingestion, import CF) and db repair (with no access to manifest).
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11406
Test Plan:
1. New UT
2. db bench
Note: db bench on /tmp/ where direct read is supported is too slow to finish and the default pinning setting in db bench is not helpful to profile # sst read of Get. Therefore I hacked the following to obtain the following comparison.
```
diff --git a/table/block_based/block_based_table_reader.cc b/table/block_based/block_based_table_reader.cc
index bd5669f0f..791484c1f 100644
--- a/table/block_based/block_based_table_reader.cc
+++ b/table/block_based/block_based_table_reader.cc
@@ -838,7 +838,7 @@ Status BlockBasedTable::PrefetchTail(
&tail_prefetch_size);
// Try file system prefetch
- if (!file->use_direct_io() && !force_direct_prefetch) {
+ if (false && !file->use_direct_io() && !force_direct_prefetch) {
if (!file->Prefetch(prefetch_off, prefetch_len, ro.rate_limiter_priority)
.IsNotSupported()) {
prefetch_buffer->reset(new FilePrefetchBuffer(
diff --git a/tools/db_bench_tool.cc b/tools/db_bench_tool.cc
index ea40f5fa0..39a0ac385 100644
--- a/tools/db_bench_tool.cc
+++ b/tools/db_bench_tool.cc
@@ -4191,6 +4191,8 @@ class Benchmark {
std::shared_ptr<TableFactory>(NewCuckooTableFactory(table_options));
} else {
BlockBasedTableOptions block_based_options;
+ block_based_options.metadata_cache_options.partition_pinning =
+ PinningTier::kAll;
block_based_options.checksum =
static_cast<ChecksumType>(FLAGS_checksum_type);
if (FLAGS_use_hash_search) {
```
Create DB
```
./db_bench --bloom_bits=3 --use_existing_db=1 --seed=1682546046158958 --partition_index_and_filters=1 --statistics=1 -db=/dev/shm/testdb/ -benchmarks=readrandom -key_size=3200 -value_size=512 -num=1000000 -write_buffer_size=6550000 -disable_auto_compactions=false -target_file_size_base=6550000 -compression_type=none
```
ReadRandom
```
./db_bench --bloom_bits=3 --use_existing_db=1 --seed=1682546046158958 --partition_index_and_filters=1 --statistics=1 -db=/dev/shm/testdb/ -benchmarks=readrandom -key_size=3200 -value_size=512 -num=1000000 -write_buffer_size=6550000 -disable_auto_compactions=false -target_file_size_base=6550000 -compression_type=none
```
(a) Existing (Use TailPrefetchStats for tail size + use seperate prefetch buffer in PartitionedFilter/IndexReader::CacheDependencies())
```
rocksdb.table.open.prefetch.tail.hit COUNT : 3395
rocksdb.sst.read.micros P50 : 5.655570 P95 : 9.931396 P99 : 14.845454 P100 : 585.000000 COUNT : 999905 SUM : 6590614
```
(b) This PR (Record tail size + use the same tail buffer in PartitionedFilter/IndexReader::CacheDependencies())
```
rocksdb.table.open.prefetch.tail.hit COUNT : 14257
rocksdb.sst.read.micros P50 : 5.173347 P95 : 9.015017 P99 : 12.912610 P100 : 228.000000 COUNT : 998547 SUM : 5976540
```
As we can see, we increase the prefetch tail hit count and decrease SST read count with this PR
3. Test backward compatibility by stepping through reading with post-PR code on a db generated pre-PR.
Reviewed By: pdillinger
Differential Revision: D45413346
Pulled By: hx235
fbshipit-source-id: 7d5e36a60a72477218f79905168d688452a4c064
2 years ago
|
|
|
std::unique_ptr<FilePrefetchBuffer>* prefetch_buffer, Statistics* stats,
|
|
|
|
uint64_t tail_size, Logger* const logger) {
|
|
|
|
assert(tail_size <= file_size);
|
|
|
|
|
|
|
|
size_t tail_prefetch_size = 0;
|
Record and use the tail size to prefetch table tail (#11406)
Summary:
**Context:**
We prefetch the tail part of a SST file (i.e, the blocks after data blocks till the end of the file) during each SST file open in hope to prefetch all the stuff at once ahead of time for later read e.g, footer, meta index, filter/index etc. The existing approach to estimate the tail size to prefetch is through `TailPrefetchStats` heuristics introduced in https://github.com/facebook/rocksdb/pull/4156, which has caused small reads in unlucky case (e.g, small read into the tail buffer during table open in thread 1 under the same BlockBasedTableFactory object can make thread 2's tail prefetching use a small size that it shouldn't) and is hard to debug. Therefore we decide to record the exact tail size and use it directly to prefetch tail of the SST instead of relying heuristics.
**Summary:**
- Obtain and record in manifest the tail size in `BlockBasedTableBuilder::Finish()`
- For backward compatibility, we fall back to TailPrefetchStats and last to simple heuristics that the tail size is a linear portion of the file size - see PR conversation for more.
- Make`tail_start_offset` part of the table properties and deduct tail size to record in manifest for external files (e.g, file ingestion, import CF) and db repair (with no access to manifest).
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11406
Test Plan:
1. New UT
2. db bench
Note: db bench on /tmp/ where direct read is supported is too slow to finish and the default pinning setting in db bench is not helpful to profile # sst read of Get. Therefore I hacked the following to obtain the following comparison.
```
diff --git a/table/block_based/block_based_table_reader.cc b/table/block_based/block_based_table_reader.cc
index bd5669f0f..791484c1f 100644
--- a/table/block_based/block_based_table_reader.cc
+++ b/table/block_based/block_based_table_reader.cc
@@ -838,7 +838,7 @@ Status BlockBasedTable::PrefetchTail(
&tail_prefetch_size);
// Try file system prefetch
- if (!file->use_direct_io() && !force_direct_prefetch) {
+ if (false && !file->use_direct_io() && !force_direct_prefetch) {
if (!file->Prefetch(prefetch_off, prefetch_len, ro.rate_limiter_priority)
.IsNotSupported()) {
prefetch_buffer->reset(new FilePrefetchBuffer(
diff --git a/tools/db_bench_tool.cc b/tools/db_bench_tool.cc
index ea40f5fa0..39a0ac385 100644
--- a/tools/db_bench_tool.cc
+++ b/tools/db_bench_tool.cc
@@ -4191,6 +4191,8 @@ class Benchmark {
std::shared_ptr<TableFactory>(NewCuckooTableFactory(table_options));
} else {
BlockBasedTableOptions block_based_options;
+ block_based_options.metadata_cache_options.partition_pinning =
+ PinningTier::kAll;
block_based_options.checksum =
static_cast<ChecksumType>(FLAGS_checksum_type);
if (FLAGS_use_hash_search) {
```
Create DB
```
./db_bench --bloom_bits=3 --use_existing_db=1 --seed=1682546046158958 --partition_index_and_filters=1 --statistics=1 -db=/dev/shm/testdb/ -benchmarks=readrandom -key_size=3200 -value_size=512 -num=1000000 -write_buffer_size=6550000 -disable_auto_compactions=false -target_file_size_base=6550000 -compression_type=none
```
ReadRandom
```
./db_bench --bloom_bits=3 --use_existing_db=1 --seed=1682546046158958 --partition_index_and_filters=1 --statistics=1 -db=/dev/shm/testdb/ -benchmarks=readrandom -key_size=3200 -value_size=512 -num=1000000 -write_buffer_size=6550000 -disable_auto_compactions=false -target_file_size_base=6550000 -compression_type=none
```
(a) Existing (Use TailPrefetchStats for tail size + use seperate prefetch buffer in PartitionedFilter/IndexReader::CacheDependencies())
```
rocksdb.table.open.prefetch.tail.hit COUNT : 3395
rocksdb.sst.read.micros P50 : 5.655570 P95 : 9.931396 P99 : 14.845454 P100 : 585.000000 COUNT : 999905 SUM : 6590614
```
(b) This PR (Record tail size + use the same tail buffer in PartitionedFilter/IndexReader::CacheDependencies())
```
rocksdb.table.open.prefetch.tail.hit COUNT : 14257
rocksdb.sst.read.micros P50 : 5.173347 P95 : 9.015017 P99 : 12.912610 P100 : 228.000000 COUNT : 998547 SUM : 5976540
```
As we can see, we increase the prefetch tail hit count and decrease SST read count with this PR
3. Test backward compatibility by stepping through reading with post-PR code on a db generated pre-PR.
Reviewed By: pdillinger
Differential Revision: D45413346
Pulled By: hx235
fbshipit-source-id: 7d5e36a60a72477218f79905168d688452a4c064
2 years ago
|
|
|
if (tail_size != 0) {
|
|
|
|
tail_prefetch_size = tail_size;
|
|
|
|
} else {
|
|
|
|
if (tail_prefetch_stats != nullptr) {
|
|
|
|
// Multiple threads may get a 0 (no history) when running in parallel,
|
|
|
|
// but it will get cleared after the first of them finishes.
|
|
|
|
tail_prefetch_size = tail_prefetch_stats->GetSuggestedPrefetchSize();
|
|
|
|
}
|
|
|
|
if (tail_prefetch_size == 0) {
|
|
|
|
// Before read footer, readahead backwards to prefetch data. Do more
|
|
|
|
// readahead if we're going to read index/filter.
|
|
|
|
// TODO: This may incorrectly select small readahead in case partitioned
|
|
|
|
// index/filter is enabled and top-level partition pinning is enabled.
|
|
|
|
// That's because we need to issue readahead before we read the
|
|
|
|
// properties, at which point we don't yet know the index type.
|
|
|
|
tail_prefetch_size = prefetch_all || preload_all ? 512 * 1024 : 4 * 1024;
|
Record and use the tail size to prefetch table tail (#11406)
Summary:
**Context:**
We prefetch the tail part of a SST file (i.e, the blocks after data blocks till the end of the file) during each SST file open in hope to prefetch all the stuff at once ahead of time for later read e.g, footer, meta index, filter/index etc. The existing approach to estimate the tail size to prefetch is through `TailPrefetchStats` heuristics introduced in https://github.com/facebook/rocksdb/pull/4156, which has caused small reads in unlucky case (e.g, small read into the tail buffer during table open in thread 1 under the same BlockBasedTableFactory object can make thread 2's tail prefetching use a small size that it shouldn't) and is hard to debug. Therefore we decide to record the exact tail size and use it directly to prefetch tail of the SST instead of relying heuristics.
**Summary:**
- Obtain and record in manifest the tail size in `BlockBasedTableBuilder::Finish()`
- For backward compatibility, we fall back to TailPrefetchStats and last to simple heuristics that the tail size is a linear portion of the file size - see PR conversation for more.
- Make`tail_start_offset` part of the table properties and deduct tail size to record in manifest for external files (e.g, file ingestion, import CF) and db repair (with no access to manifest).
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11406
Test Plan:
1. New UT
2. db bench
Note: db bench on /tmp/ where direct read is supported is too slow to finish and the default pinning setting in db bench is not helpful to profile # sst read of Get. Therefore I hacked the following to obtain the following comparison.
```
diff --git a/table/block_based/block_based_table_reader.cc b/table/block_based/block_based_table_reader.cc
index bd5669f0f..791484c1f 100644
--- a/table/block_based/block_based_table_reader.cc
+++ b/table/block_based/block_based_table_reader.cc
@@ -838,7 +838,7 @@ Status BlockBasedTable::PrefetchTail(
&tail_prefetch_size);
// Try file system prefetch
- if (!file->use_direct_io() && !force_direct_prefetch) {
+ if (false && !file->use_direct_io() && !force_direct_prefetch) {
if (!file->Prefetch(prefetch_off, prefetch_len, ro.rate_limiter_priority)
.IsNotSupported()) {
prefetch_buffer->reset(new FilePrefetchBuffer(
diff --git a/tools/db_bench_tool.cc b/tools/db_bench_tool.cc
index ea40f5fa0..39a0ac385 100644
--- a/tools/db_bench_tool.cc
+++ b/tools/db_bench_tool.cc
@@ -4191,6 +4191,8 @@ class Benchmark {
std::shared_ptr<TableFactory>(NewCuckooTableFactory(table_options));
} else {
BlockBasedTableOptions block_based_options;
+ block_based_options.metadata_cache_options.partition_pinning =
+ PinningTier::kAll;
block_based_options.checksum =
static_cast<ChecksumType>(FLAGS_checksum_type);
if (FLAGS_use_hash_search) {
```
Create DB
```
./db_bench --bloom_bits=3 --use_existing_db=1 --seed=1682546046158958 --partition_index_and_filters=1 --statistics=1 -db=/dev/shm/testdb/ -benchmarks=readrandom -key_size=3200 -value_size=512 -num=1000000 -write_buffer_size=6550000 -disable_auto_compactions=false -target_file_size_base=6550000 -compression_type=none
```
ReadRandom
```
./db_bench --bloom_bits=3 --use_existing_db=1 --seed=1682546046158958 --partition_index_and_filters=1 --statistics=1 -db=/dev/shm/testdb/ -benchmarks=readrandom -key_size=3200 -value_size=512 -num=1000000 -write_buffer_size=6550000 -disable_auto_compactions=false -target_file_size_base=6550000 -compression_type=none
```
(a) Existing (Use TailPrefetchStats for tail size + use seperate prefetch buffer in PartitionedFilter/IndexReader::CacheDependencies())
```
rocksdb.table.open.prefetch.tail.hit COUNT : 3395
rocksdb.sst.read.micros P50 : 5.655570 P95 : 9.931396 P99 : 14.845454 P100 : 585.000000 COUNT : 999905 SUM : 6590614
```
(b) This PR (Record tail size + use the same tail buffer in PartitionedFilter/IndexReader::CacheDependencies())
```
rocksdb.table.open.prefetch.tail.hit COUNT : 14257
rocksdb.sst.read.micros P50 : 5.173347 P95 : 9.015017 P99 : 12.912610 P100 : 228.000000 COUNT : 998547 SUM : 5976540
```
As we can see, we increase the prefetch tail hit count and decrease SST read count with this PR
3. Test backward compatibility by stepping through reading with post-PR code on a db generated pre-PR.
Reviewed By: pdillinger
Differential Revision: D45413346
Pulled By: hx235
fbshipit-source-id: 7d5e36a60a72477218f79905168d688452a4c064
2 years ago
|
|
|
|
|
|
|
ROCKS_LOG_WARN(logger,
|
|
|
|
"Tail prefetch size %zu is calculated based on heuristics",
|
|
|
|
tail_prefetch_size);
|
|
|
|
} else {
|
|
|
|
ROCKS_LOG_WARN(
|
|
|
|
logger,
|
|
|
|
"Tail prefetch size %zu is calculated based on TailPrefetchStats",
|
|
|
|
tail_prefetch_size);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
size_t prefetch_off;
|
|
|
|
size_t prefetch_len;
|
|
|
|
if (file_size < tail_prefetch_size) {
|
|
|
|
prefetch_off = 0;
|
|
|
|
prefetch_len = static_cast<size_t>(file_size);
|
|
|
|
} else {
|
|
|
|
prefetch_off = static_cast<size_t>(file_size - tail_prefetch_size);
|
|
|
|
prefetch_len = tail_prefetch_size;
|
|
|
|
}
|
|
|
|
TEST_SYNC_POINT_CALLBACK("BlockBasedTable::Open::TailPrefetchLen",
|
|
|
|
&tail_prefetch_size);
|
|
|
|
|
|
|
|
// Try file system prefetch
|
|
|
|
if (!file->use_direct_io() && !force_direct_prefetch) {
|
Set Read rate limiter priority dynamically and pass it to FS (#9996)
Summary:
### Context:
Background compactions and flush generate large reads and writes, and can be long running, especially for universal compaction. In some cases, this can impact foreground reads and writes by users.
### Solution
User, Flush, and Compaction reads share some code path. For this task, we update the rate_limiter_priority in ReadOptions for code paths (e.g. FindTable (mainly in BlockBasedTable::Open()) and various iterators), and eventually update the rate_limiter_priority in IOOptions for FSRandomAccessFile.
**This PR is for the Read path.** The **Read:** dynamic priority for different state are listed as follows:
| State | Normal | Delayed | Stalled |
| ----- | ------ | ------- | ------- |
| Flush (verification read in BuildTable()) | IO_USER | IO_USER | IO_USER |
| Compaction | IO_LOW | IO_USER | IO_USER |
| User | User provided | User provided | User provided |
We will respect the read_options that the user provided and will not set it.
The only sst read for Flush is the verification read in BuildTable(). It claims to be "regard as user read".
**Details**
1. Set read_options.rate_limiter_priority dynamically:
- User: Do not update the read_options. Use the read_options that the user provided.
- Compaction: Update read_options in CompactionJob::ProcessKeyValueCompaction().
- Flush: Update read_options in BuildTable().
2. Pass the rate limiter priority to FSRandomAccessFile functions:
- After calling the FindTable(), read_options is passed through GetTableReader(table_cache.cc), BlockBasedTableFactory::NewTableReader(block_based_table_factory.cc), and BlockBasedTable::Open(). The Open() needs some updates for the ReadOptions variable and the updates are also needed for the called functions, including PrefetchTail(), PrepareIOOptions(), ReadFooterFromFile(), ReadMetaIndexblock(), ReadPropertiesBlock(), PrefetchIndexAndFilterBlocks(), and ReadRangeDelBlock().
- In RandomAccessFileReader, the functions to be updated include Read(), MultiRead(), ReadAsync(), and Prefetch().
- Update the downstream functions of NewIndexIterator(), NewDataBlockIterator(), and BlockBasedTableIterator().
### Test Plans
Add unit tests.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9996
Reviewed By: anand1976
Differential Revision: D36452483
Pulled By: gitbw95
fbshipit-source-id: 60978204a4f849bb9261cb78d9bc1cb56d6008cf
3 years ago
|
|
|
if (!file->Prefetch(prefetch_off, prefetch_len, ro.rate_limiter_priority)
|
|
|
|
.IsNotSupported()) {
|
|
|
|
prefetch_buffer->reset(new FilePrefetchBuffer(
|
|
|
|
0 /* readahead_size */, 0 /* max_readahead_size */,
|
|
|
|
false /* enable */, true /* track_min_offset */));
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Use `FilePrefetchBuffer`
|
Add new stat rocksdb.table.open.prefetch.tail.read.bytes, rocksdb.table.open.prefetch.tail.{miss|hit} (#11265)
Summary:
**Context/Summary:**
We are adding new stats to measure behavior of prefetched tail size and look up into this buffer
The stat collection is done in FilePrefetchBuffer but only for prefetched tail buffer during table open for now using FilePrefetchBuffer enum. It's cleaner than the alternative of implementing in upper-level call places of FilePrefetchBuffer for table open. It also has the benefit of extensible to other types of FilePrefetchBuffer if needed. See db bench for perf regression concern.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11265
Test Plan:
**- Piggyback on existing test**
**- rocksdb.table.open.prefetch.tail.miss is harder to UT so I manually set prefetch tail read bytes to be small and run db bench.**
```
./db_bench -db=/tmp/testdb -statistics=true -benchmarks="fillseq" -key_size=32 -value_size=512 -num=5000 -write_buffer_size=655 -target_file_size_base=655 -disable_auto_compactions=false -compression_type=none -bloom_bits=3 -use_direct_reads=true
```
```
rocksdb.table.open.prefetch.tail.read.bytes P50 : 4096.000000 P95 : 4096.000000 P99 : 4096.000000 P100 : 4096.000000 COUNT : 225 SUM : 921600
rocksdb.table.open.prefetch.tail.miss COUNT : 91
rocksdb.table.open.prefetch.tail.hit COUNT : 1034
```
**- No perf regression observed in db_bench**
SETUP command: create same db with ~900 files for pre-change/post-change.
```
./db_bench -db=/tmp/testdb -benchmarks="fillseq" -key_size=32 -value_size=512 -num=500000 -write_buffer_size=655360 -disable_auto_compactions=true -target_file_size_base=16777216 -compression_type=none
```
TEST command 60 runs or til convergence: as suggested by anand1976 and akankshamahajan15, vary `seek_nexts` and `async_io` in testing.
```
./db_bench -use_existing_db=true -db=/tmp/testdb -statistics=false -cache_size=0 -cache_index_and_filter_blocks=false -benchmarks=seekrandom[-X60] -num=50000 -seek_nexts={10, 500, 1000} -async_io={0|1} -use_direct_reads=true
```
async io = 0, direct io read = true
| seek_nexts = 10, 30 runs | seek_nexts = 500, 12 runs | seek_nexts = 1000, 6 runs
-- | -- | -- | --
pre-post change | 4776 (± 28) ops/sec; 24.8 (± 0.1) MB/sec | 288 (± 1) ops/sec; 74.8 (± 0.4) MB/sec | 145 (± 4) ops/sec; 75.6 (± 2.2) MB/sec
post-change | 4790 (± 32) ops/sec; 24.9 (± 0.2) MB/sec | 288 (± 3) ops/sec; 74.7 (± 0.8) MB/sec | 143 (± 3) ops/sec; 74.5 (± 1.6) MB/sec
async io = 1, direct io read = true
| seek_nexts = 10, 54 runs | seek_nexts = 500, 6 runs | seek_nexts = 1000, 4 runs
-- | -- | -- | --
pre-post change | 3350 (± 36) ops/sec; 17.4 (± 0.2) MB/sec | 264 (± 0) ops/sec; 68.7 (± 0.2) MB/sec | 138 (± 1) ops/sec; 71.8 (± 1.0) MB/sec
post-change | 3358 (± 27) ops/sec; 17.4 (± 0.1) MB/sec | 263 (± 2) ops/sec; 68.3 (± 0.8) MB/sec | 139 (± 1) ops/sec; 72.6 (± 0.6) MB/sec
Reviewed By: ajkr
Differential Revision: D43781467
Pulled By: hx235
fbshipit-source-id: a706a18472a8edb2b952bac3af40eec803537f2a
2 years ago
|
|
|
prefetch_buffer->reset(new FilePrefetchBuffer(
|
|
|
|
0 /* readahead_size */, 0 /* max_readahead_size */, true /* enable */,
|
|
|
|
true /* track_min_offset */, false /* implicit_auto_readahead */,
|
|
|
|
0 /* num_file_reads */, 0 /* num_file_reads_for_auto_readahead */,
|
|
|
|
nullptr /* fs */, nullptr /* clock */, stats,
|
|
|
|
FilePrefetchBufferUsage::kTableOpenPrefetchTail));
|
Set Read rate limiter priority dynamically and pass it to FS (#9996)
Summary:
### Context:
Background compactions and flush generate large reads and writes, and can be long running, especially for universal compaction. In some cases, this can impact foreground reads and writes by users.
### Solution
User, Flush, and Compaction reads share some code path. For this task, we update the rate_limiter_priority in ReadOptions for code paths (e.g. FindTable (mainly in BlockBasedTable::Open()) and various iterators), and eventually update the rate_limiter_priority in IOOptions for FSRandomAccessFile.
**This PR is for the Read path.** The **Read:** dynamic priority for different state are listed as follows:
| State | Normal | Delayed | Stalled |
| ----- | ------ | ------- | ------- |
| Flush (verification read in BuildTable()) | IO_USER | IO_USER | IO_USER |
| Compaction | IO_LOW | IO_USER | IO_USER |
| User | User provided | User provided | User provided |
We will respect the read_options that the user provided and will not set it.
The only sst read for Flush is the verification read in BuildTable(). It claims to be "regard as user read".
**Details**
1. Set read_options.rate_limiter_priority dynamically:
- User: Do not update the read_options. Use the read_options that the user provided.
- Compaction: Update read_options in CompactionJob::ProcessKeyValueCompaction().
- Flush: Update read_options in BuildTable().
2. Pass the rate limiter priority to FSRandomAccessFile functions:
- After calling the FindTable(), read_options is passed through GetTableReader(table_cache.cc), BlockBasedTableFactory::NewTableReader(block_based_table_factory.cc), and BlockBasedTable::Open(). The Open() needs some updates for the ReadOptions variable and the updates are also needed for the called functions, including PrefetchTail(), PrepareIOOptions(), ReadFooterFromFile(), ReadMetaIndexblock(), ReadPropertiesBlock(), PrefetchIndexAndFilterBlocks(), and ReadRangeDelBlock().
- In RandomAccessFileReader, the functions to be updated include Read(), MultiRead(), ReadAsync(), and Prefetch().
- Update the downstream functions of NewIndexIterator(), NewDataBlockIterator(), and BlockBasedTableIterator().
### Test Plans
Add unit tests.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9996
Reviewed By: anand1976
Differential Revision: D36452483
Pulled By: gitbw95
fbshipit-source-id: 60978204a4f849bb9261cb78d9bc1cb56d6008cf
3 years ago
|
|
|
|
|
|
|
IOOptions opts;
|
|
|
|
Status s = file->PrepareIOOptions(ro, opts);
|
|
|
|
if (s.ok()) {
|
|
|
|
s = (*prefetch_buffer)
|
|
|
|
->Prefetch(opts, file, prefetch_off, prefetch_len,
|
|
|
|
ro.rate_limiter_priority);
|
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
Status BlockBasedTable::ReadPropertiesBlock(
|
|
|
|
const ReadOptions& ro, FilePrefetchBuffer* prefetch_buffer,
|
|
|
|
InternalIterator* meta_iter, const SequenceNumber largest_seqno) {
|
|
|
|
Status s;
|
Improve / clean up meta block code & integrity (#9163)
Summary:
* Checksums are now checked on meta blocks unless specifically
suppressed or not applicable (e.g. plain table). (Was other way around.)
This means a number of cases that were not checking checksums now are,
including direct read TableProperties in Version::GetTableProperties
(fixed in meta_blocks ReadTableProperties), reading any block from
PersistentCache (fixed in BlockFetcher), read TableProperties in
SstFileDumper (ldb/sst_dump/BackupEngine) before table reader open,
maybe more.
* For that to work, I moved the global_seqno+TableProperties checksum
logic to the shared table/ code, because that is used by many utilies
such as SstFileDumper.
* Also for that to work, we have to know when we're dealing with a block
that has a checksum (trailer), so added that capability to Footer based
on magic number, and from there BlockFetcher.
* Knowledge of trailer presence has also fixed a problem where other
table formats were reading blocks including bytes for a non-existant
trailer--and awkwardly kind-of not using them, e.g. no shared code
checking checksums. (BlockFetcher compression type was populated
incorrectly.) Now we only read what is needed.
* Minimized code duplication and differing/incompatible/awkward
abstractions in meta_blocks.{cc,h} (e.g. SeekTo in metaindex block
without parsing block handle)
* Moved some meta block handling code from table_properties*.*
* Moved some code specific to block-based table from shared table/ code
to BlockBasedTable class. The checksum stuff means we can't completely
separate it, but things that don't need to be in shared table/ code
should not be.
* Use unique_ptr rather than raw ptr in more places. (Note: you can
std::move from unique_ptr to shared_ptr.)
Without enhancements to GetPropertiesOfAllTablesTest (see below),
net reduction of roughly 100 lines of code.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9163
Test Plan:
existing tests and
* Enhanced DBTablePropertiesTest.GetPropertiesOfAllTablesTest to verify that
checksums are now checked on direct read of table properties by TableCache
(new test would fail before this change)
* Also enhanced DBTablePropertiesTest.GetPropertiesOfAllTablesTest to test
putting table properties under old meta name
* Also generally enhanced that same test to actually test what it was
supposed to be testing already, by kicking things out of table cache when
we don't want them there.
Reviewed By: ajkr, mrambacher
Differential Revision: D32514757
Pulled By: pdillinger
fbshipit-source-id: 507964b9311d186ae8d1131182290cbd97a99fa9
3 years ago
|
|
|
BlockHandle handle;
|
|
|
|
s = FindOptionalMetaBlock(meta_iter, kPropertiesBlockName, &handle);
|
RocksDB 2.8 to be able to read files generated by 2.6
Summary:
From 2.6 to 2.7, property block name is renamed from rocksdb.stats to rocksdb.properties. Older properties were not able to be loaded. In 2.8, we seem to have added some logic that uses property block without checking null pointers, which create segment faults.
In this patch, we fix it by:
(1) try rocksdb.stats if rocksdb.properties is not found
(2) add some null checking before consuming rep->table_properties
Test Plan: make sure a file generated in 2.7 couldn't be opened now can be opened.
Reviewers: haobo, igor, yhchiang
Reviewed By: igor
CC: ljin, xjin, dhruba, kailiu, leveldb
Differential Revision: https://reviews.facebook.net/D17961
11 years ago
|
|
|
|
|
|
|
if (!s.ok()) {
|
|
|
|
ROCKS_LOG_WARN(rep_->ioptions.logger,
|
|
|
|
"Error when seeking to properties block from file: %s",
|
|
|
|
s.ToString().c_str());
|
Improve / clean up meta block code & integrity (#9163)
Summary:
* Checksums are now checked on meta blocks unless specifically
suppressed or not applicable (e.g. plain table). (Was other way around.)
This means a number of cases that were not checking checksums now are,
including direct read TableProperties in Version::GetTableProperties
(fixed in meta_blocks ReadTableProperties), reading any block from
PersistentCache (fixed in BlockFetcher), read TableProperties in
SstFileDumper (ldb/sst_dump/BackupEngine) before table reader open,
maybe more.
* For that to work, I moved the global_seqno+TableProperties checksum
logic to the shared table/ code, because that is used by many utilies
such as SstFileDumper.
* Also for that to work, we have to know when we're dealing with a block
that has a checksum (trailer), so added that capability to Footer based
on magic number, and from there BlockFetcher.
* Knowledge of trailer presence has also fixed a problem where other
table formats were reading blocks including bytes for a non-existant
trailer--and awkwardly kind-of not using them, e.g. no shared code
checking checksums. (BlockFetcher compression type was populated
incorrectly.) Now we only read what is needed.
* Minimized code duplication and differing/incompatible/awkward
abstractions in meta_blocks.{cc,h} (e.g. SeekTo in metaindex block
without parsing block handle)
* Moved some meta block handling code from table_properties*.*
* Moved some code specific to block-based table from shared table/ code
to BlockBasedTable class. The checksum stuff means we can't completely
separate it, but things that don't need to be in shared table/ code
should not be.
* Use unique_ptr rather than raw ptr in more places. (Note: you can
std::move from unique_ptr to shared_ptr.)
Without enhancements to GetPropertiesOfAllTablesTest (see below),
net reduction of roughly 100 lines of code.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9163
Test Plan:
existing tests and
* Enhanced DBTablePropertiesTest.GetPropertiesOfAllTablesTest to verify that
checksums are now checked on direct read of table properties by TableCache
(new test would fail before this change)
* Also enhanced DBTablePropertiesTest.GetPropertiesOfAllTablesTest to test
putting table properties under old meta name
* Also generally enhanced that same test to actually test what it was
supposed to be testing already, by kicking things out of table cache when
we don't want them there.
Reviewed By: ajkr, mrambacher
Differential Revision: D32514757
Pulled By: pdillinger
fbshipit-source-id: 507964b9311d186ae8d1131182290cbd97a99fa9
3 years ago
|
|
|
} else if (!handle.IsNull()) {
|
|
|
|
s = meta_iter->status();
|
Improve / clean up meta block code & integrity (#9163)
Summary:
* Checksums are now checked on meta blocks unless specifically
suppressed or not applicable (e.g. plain table). (Was other way around.)
This means a number of cases that were not checking checksums now are,
including direct read TableProperties in Version::GetTableProperties
(fixed in meta_blocks ReadTableProperties), reading any block from
PersistentCache (fixed in BlockFetcher), read TableProperties in
SstFileDumper (ldb/sst_dump/BackupEngine) before table reader open,
maybe more.
* For that to work, I moved the global_seqno+TableProperties checksum
logic to the shared table/ code, because that is used by many utilies
such as SstFileDumper.
* Also for that to work, we have to know when we're dealing with a block
that has a checksum (trailer), so added that capability to Footer based
on magic number, and from there BlockFetcher.
* Knowledge of trailer presence has also fixed a problem where other
table formats were reading blocks including bytes for a non-existant
trailer--and awkwardly kind-of not using them, e.g. no shared code
checking checksums. (BlockFetcher compression type was populated
incorrectly.) Now we only read what is needed.
* Minimized code duplication and differing/incompatible/awkward
abstractions in meta_blocks.{cc,h} (e.g. SeekTo in metaindex block
without parsing block handle)
* Moved some meta block handling code from table_properties*.*
* Moved some code specific to block-based table from shared table/ code
to BlockBasedTable class. The checksum stuff means we can't completely
separate it, but things that don't need to be in shared table/ code
should not be.
* Use unique_ptr rather than raw ptr in more places. (Note: you can
std::move from unique_ptr to shared_ptr.)
Without enhancements to GetPropertiesOfAllTablesTest (see below),
net reduction of roughly 100 lines of code.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9163
Test Plan:
existing tests and
* Enhanced DBTablePropertiesTest.GetPropertiesOfAllTablesTest to verify that
checksums are now checked on direct read of table properties by TableCache
(new test would fail before this change)
* Also enhanced DBTablePropertiesTest.GetPropertiesOfAllTablesTest to test
putting table properties under old meta name
* Also generally enhanced that same test to actually test what it was
supposed to be testing already, by kicking things out of table cache when
we don't want them there.
Reviewed By: ajkr, mrambacher
Differential Revision: D32514757
Pulled By: pdillinger
fbshipit-source-id: 507964b9311d186ae8d1131182290cbd97a99fa9
3 years ago
|
|
|
std::unique_ptr<TableProperties> table_properties;
|
|
|
|
if (s.ok()) {
|
Improve / clean up meta block code & integrity (#9163)
Summary:
* Checksums are now checked on meta blocks unless specifically
suppressed or not applicable (e.g. plain table). (Was other way around.)
This means a number of cases that were not checking checksums now are,
including direct read TableProperties in Version::GetTableProperties
(fixed in meta_blocks ReadTableProperties), reading any block from
PersistentCache (fixed in BlockFetcher), read TableProperties in
SstFileDumper (ldb/sst_dump/BackupEngine) before table reader open,
maybe more.
* For that to work, I moved the global_seqno+TableProperties checksum
logic to the shared table/ code, because that is used by many utilies
such as SstFileDumper.
* Also for that to work, we have to know when we're dealing with a block
that has a checksum (trailer), so added that capability to Footer based
on magic number, and from there BlockFetcher.
* Knowledge of trailer presence has also fixed a problem where other
table formats were reading blocks including bytes for a non-existant
trailer--and awkwardly kind-of not using them, e.g. no shared code
checking checksums. (BlockFetcher compression type was populated
incorrectly.) Now we only read what is needed.
* Minimized code duplication and differing/incompatible/awkward
abstractions in meta_blocks.{cc,h} (e.g. SeekTo in metaindex block
without parsing block handle)
* Moved some meta block handling code from table_properties*.*
* Moved some code specific to block-based table from shared table/ code
to BlockBasedTable class. The checksum stuff means we can't completely
separate it, but things that don't need to be in shared table/ code
should not be.
* Use unique_ptr rather than raw ptr in more places. (Note: you can
std::move from unique_ptr to shared_ptr.)
Without enhancements to GetPropertiesOfAllTablesTest (see below),
net reduction of roughly 100 lines of code.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9163
Test Plan:
existing tests and
* Enhanced DBTablePropertiesTest.GetPropertiesOfAllTablesTest to verify that
checksums are now checked on direct read of table properties by TableCache
(new test would fail before this change)
* Also enhanced DBTablePropertiesTest.GetPropertiesOfAllTablesTest to test
putting table properties under old meta name
* Also generally enhanced that same test to actually test what it was
supposed to be testing already, by kicking things out of table cache when
we don't want them there.
Reviewed By: ajkr, mrambacher
Differential Revision: D32514757
Pulled By: pdillinger
fbshipit-source-id: 507964b9311d186ae8d1131182290cbd97a99fa9
3 years ago
|
|
|
s = ReadTablePropertiesHelper(
|
|
|
|
ro, handle, rep_->file.get(), prefetch_buffer, rep_->footer,
|
|
|
|
rep_->ioptions, &table_properties, nullptr /* memory_allocator */);
|
|
|
|
}
|
|
|
|
IGNORE_STATUS_IF_ERROR(s);
|
|
|
|
|
|
|
|
if (!s.ok()) {
|
|
|
|
ROCKS_LOG_WARN(rep_->ioptions.logger,
|
|
|
|
"Encountered error while reading data from properties "
|
|
|
|
"block %s",
|
|
|
|
s.ToString().c_str());
|
|
|
|
} else {
|
|
|
|
assert(table_properties != nullptr);
|
Improve / clean up meta block code & integrity (#9163)
Summary:
* Checksums are now checked on meta blocks unless specifically
suppressed or not applicable (e.g. plain table). (Was other way around.)
This means a number of cases that were not checking checksums now are,
including direct read TableProperties in Version::GetTableProperties
(fixed in meta_blocks ReadTableProperties), reading any block from
PersistentCache (fixed in BlockFetcher), read TableProperties in
SstFileDumper (ldb/sst_dump/BackupEngine) before table reader open,
maybe more.
* For that to work, I moved the global_seqno+TableProperties checksum
logic to the shared table/ code, because that is used by many utilies
such as SstFileDumper.
* Also for that to work, we have to know when we're dealing with a block
that has a checksum (trailer), so added that capability to Footer based
on magic number, and from there BlockFetcher.
* Knowledge of trailer presence has also fixed a problem where other
table formats were reading blocks including bytes for a non-existant
trailer--and awkwardly kind-of not using them, e.g. no shared code
checking checksums. (BlockFetcher compression type was populated
incorrectly.) Now we only read what is needed.
* Minimized code duplication and differing/incompatible/awkward
abstractions in meta_blocks.{cc,h} (e.g. SeekTo in metaindex block
without parsing block handle)
* Moved some meta block handling code from table_properties*.*
* Moved some code specific to block-based table from shared table/ code
to BlockBasedTable class. The checksum stuff means we can't completely
separate it, but things that don't need to be in shared table/ code
should not be.
* Use unique_ptr rather than raw ptr in more places. (Note: you can
std::move from unique_ptr to shared_ptr.)
Without enhancements to GetPropertiesOfAllTablesTest (see below),
net reduction of roughly 100 lines of code.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9163
Test Plan:
existing tests and
* Enhanced DBTablePropertiesTest.GetPropertiesOfAllTablesTest to verify that
checksums are now checked on direct read of table properties by TableCache
(new test would fail before this change)
* Also enhanced DBTablePropertiesTest.GetPropertiesOfAllTablesTest to test
putting table properties under old meta name
* Also generally enhanced that same test to actually test what it was
supposed to be testing already, by kicking things out of table cache when
we don't want them there.
Reviewed By: ajkr, mrambacher
Differential Revision: D32514757
Pulled By: pdillinger
fbshipit-source-id: 507964b9311d186ae8d1131182290cbd97a99fa9
3 years ago
|
|
|
rep_->table_properties = std::move(table_properties);
|
|
|
|
rep_->blocks_maybe_compressed =
|
|
|
|
rep_->table_properties->compression_name !=
|
|
|
|
CompressionTypeToString(kNoCompression);
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
ROCKS_LOG_ERROR(rep_->ioptions.logger,
|
|
|
|
"Cannot find Properties block from file.");
|
|
|
|
}
|
|
|
|
|
|
|
|
// Read the table properties, if provided.
|
|
|
|
if (rep_->table_properties) {
|
|
|
|
rep_->whole_key_filtering &=
|
|
|
|
IsFeatureSupported(*(rep_->table_properties),
|
|
|
|
BlockBasedTablePropertyNames::kWholeKeyFiltering,
|
|
|
|
rep_->ioptions.logger);
|
|
|
|
rep_->prefix_filtering &= IsFeatureSupported(
|
|
|
|
*(rep_->table_properties),
|
|
|
|
BlockBasedTablePropertyNames::kPrefixFiltering, rep_->ioptions.logger);
|
|
|
|
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
rep_->index_key_includes_seq =
|
|
|
|
rep_->table_properties->index_key_is_user_key == 0;
|
|
|
|
rep_->index_value_is_full =
|
|
|
|
rep_->table_properties->index_value_is_delta_encoded == 0;
|
|
|
|
|
|
|
|
// Update index_type with the true type.
|
|
|
|
// If table properties don't contain index type, we assume that the table
|
|
|
|
// is in very old format and has kBinarySearch index type.
|
|
|
|
auto& props = rep_->table_properties->user_collected_properties;
|
|
|
|
auto index_type_pos = props.find(BlockBasedTablePropertyNames::kIndexType);
|
|
|
|
if (index_type_pos != props.end()) {
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
rep_->index_type = static_cast<BlockBasedTableOptions::IndexType>(
|
|
|
|
DecodeFixed32(index_type_pos->second.c_str()));
|
|
|
|
}
|
|
|
|
auto min_ts_pos = props.find("rocksdb.timestamp_min");
|
|
|
|
if (min_ts_pos != props.end()) {
|
|
|
|
rep_->min_timestamp = Slice(min_ts_pos->second);
|
|
|
|
}
|
|
|
|
auto max_ts_pos = props.find("rocksdb.timestamp_max");
|
|
|
|
if (max_ts_pos != props.end()) {
|
|
|
|
rep_->max_timestamp = Slice(max_ts_pos->second);
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
}
|
|
|
|
|
|
|
|
rep_->index_has_first_key =
|
|
|
|
rep_->index_type == BlockBasedTableOptions::kBinarySearchWithFirstKey;
|
|
|
|
|
|
|
|
s = GetGlobalSequenceNumber(*(rep_->table_properties), largest_seqno,
|
|
|
|
&(rep_->global_seqno));
|
|
|
|
if (!s.ok()) {
|
|
|
|
ROCKS_LOG_ERROR(rep_->ioptions.logger, "%s", s.ToString().c_str());
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
Status BlockBasedTable::ReadRangeDelBlock(
|
|
|
|
const ReadOptions& read_options, FilePrefetchBuffer* prefetch_buffer,
|
|
|
|
InternalIterator* meta_iter,
|
|
|
|
const InternalKeyComparator& internal_comparator,
|
|
|
|
BlockCacheLookupContext* lookup_context) {
|
|
|
|
Status s;
|
|
|
|
BlockHandle range_del_handle;
|
|
|
|
s = FindOptionalMetaBlock(meta_iter, kRangeDelBlockName, &range_del_handle);
|
|
|
|
if (!s.ok()) {
|
|
|
|
ROCKS_LOG_WARN(
|
|
|
|
rep_->ioptions.logger,
|
|
|
|
"Error when seeking to range delete tombstones block from file: %s",
|
|
|
|
s.ToString().c_str());
|
Improve / clean up meta block code & integrity (#9163)
Summary:
* Checksums are now checked on meta blocks unless specifically
suppressed or not applicable (e.g. plain table). (Was other way around.)
This means a number of cases that were not checking checksums now are,
including direct read TableProperties in Version::GetTableProperties
(fixed in meta_blocks ReadTableProperties), reading any block from
PersistentCache (fixed in BlockFetcher), read TableProperties in
SstFileDumper (ldb/sst_dump/BackupEngine) before table reader open,
maybe more.
* For that to work, I moved the global_seqno+TableProperties checksum
logic to the shared table/ code, because that is used by many utilies
such as SstFileDumper.
* Also for that to work, we have to know when we're dealing with a block
that has a checksum (trailer), so added that capability to Footer based
on magic number, and from there BlockFetcher.
* Knowledge of trailer presence has also fixed a problem where other
table formats were reading blocks including bytes for a non-existant
trailer--and awkwardly kind-of not using them, e.g. no shared code
checking checksums. (BlockFetcher compression type was populated
incorrectly.) Now we only read what is needed.
* Minimized code duplication and differing/incompatible/awkward
abstractions in meta_blocks.{cc,h} (e.g. SeekTo in metaindex block
without parsing block handle)
* Moved some meta block handling code from table_properties*.*
* Moved some code specific to block-based table from shared table/ code
to BlockBasedTable class. The checksum stuff means we can't completely
separate it, but things that don't need to be in shared table/ code
should not be.
* Use unique_ptr rather than raw ptr in more places. (Note: you can
std::move from unique_ptr to shared_ptr.)
Without enhancements to GetPropertiesOfAllTablesTest (see below),
net reduction of roughly 100 lines of code.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9163
Test Plan:
existing tests and
* Enhanced DBTablePropertiesTest.GetPropertiesOfAllTablesTest to verify that
checksums are now checked on direct read of table properties by TableCache
(new test would fail before this change)
* Also enhanced DBTablePropertiesTest.GetPropertiesOfAllTablesTest to test
putting table properties under old meta name
* Also generally enhanced that same test to actually test what it was
supposed to be testing already, by kicking things out of table cache when
we don't want them there.
Reviewed By: ajkr, mrambacher
Differential Revision: D32514757
Pulled By: pdillinger
fbshipit-source-id: 507964b9311d186ae8d1131182290cbd97a99fa9
3 years ago
|
|
|
} else if (!range_del_handle.IsNull()) {
|
|
|
|
Status tmp_status;
|
|
|
|
std::unique_ptr<InternalIterator> iter(NewDataBlockIterator<DataBlockIter>(
|
|
|
|
read_options, range_del_handle,
|
|
|
|
/*input_iter=*/nullptr, BlockType::kRangeDeletion,
|
|
|
|
/*get_context=*/nullptr, lookup_context, prefetch_buffer,
|
|
|
|
/*for_compaction= */ false, /*async_read= */ false, tmp_status));
|
|
|
|
assert(iter != nullptr);
|
|
|
|
s = iter->status();
|
Cache fragmented range tombstones in BlockBasedTableReader (#4493)
Summary:
This allows tombstone fragmenting to only be performed when the table is opened, and cached for subsequent accesses.
On the same DB used in #4449, running `readrandom` results in the following:
```
readrandom : 0.983 micros/op 1017076 ops/sec; 78.3 MB/s (63103 of 100000 found)
```
Now that Get performance in the presence of range tombstones is reasonable, I also compared the performance between a DB with range tombstones, "expanded" range tombstones (several point tombstones that cover the same keys the equivalent range tombstone would cover, a common workaround for DeleteRange), and no range tombstones. The created DBs had 5 million keys each, and DeleteRange was called at regular intervals (depending on the total number of range tombstones being written) after 4.5 million Puts. The table below summarizes the results of a `readwhilewriting` benchmark (in order to provide somewhat more realistic results):
```
Tombstones? | avg micros/op | stddev micros/op | avg ops/s | stddev ops/s
----------------- | ------------- | ---------------- | ------------ | ------------
None | 0.6186 | 0.04637 | 1,625,252.90 | 124,679.41
500 Expanded | 0.6019 | 0.03628 | 1,666,670.40 | 101,142.65
500 Unexpanded | 0.6435 | 0.03994 | 1,559,979.40 | 104,090.52
1k Expanded | 0.6034 | 0.04349 | 1,665,128.10 | 125,144.57
1k Unexpanded | 0.6261 | 0.03093 | 1,600,457.50 | 79,024.94
5k Expanded | 0.6163 | 0.05926 | 1,636,668.80 | 154,888.85
5k Unexpanded | 0.6402 | 0.04002 | 1,567,804.70 | 100,965.55
10k Expanded | 0.6036 | 0.05105 | 1,667,237.70 | 142,830.36
10k Unexpanded | 0.6128 | 0.02598 | 1,634,633.40 | 72,161.82
25k Expanded | 0.6198 | 0.04542 | 1,620,980.50 | 116,662.93
25k Unexpanded | 0.5478 | 0.0362 | 1,833,059.10 | 121,233.81
50k Expanded | 0.5104 | 0.04347 | 1,973,107.90 | 184,073.49
50k Unexpanded | 0.4528 | 0.03387 | 2,219,034.50 | 170,984.32
```
After a large enough quantity of range tombstones are written, range tombstone Gets can become faster than reading from an equivalent DB with several point tombstones.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4493
Differential Revision: D10842844
Pulled By: abhimadan
fbshipit-source-id: a7d44534f8120e6aabb65779d26c6b9df954c509
6 years ago
|
|
|
if (!s.ok()) {
|
|
|
|
ROCKS_LOG_WARN(
|
|
|
|
rep_->ioptions.logger,
|
Cache fragmented range tombstones in BlockBasedTableReader (#4493)
Summary:
This allows tombstone fragmenting to only be performed when the table is opened, and cached for subsequent accesses.
On the same DB used in #4449, running `readrandom` results in the following:
```
readrandom : 0.983 micros/op 1017076 ops/sec; 78.3 MB/s (63103 of 100000 found)
```
Now that Get performance in the presence of range tombstones is reasonable, I also compared the performance between a DB with range tombstones, "expanded" range tombstones (several point tombstones that cover the same keys the equivalent range tombstone would cover, a common workaround for DeleteRange), and no range tombstones. The created DBs had 5 million keys each, and DeleteRange was called at regular intervals (depending on the total number of range tombstones being written) after 4.5 million Puts. The table below summarizes the results of a `readwhilewriting` benchmark (in order to provide somewhat more realistic results):
```
Tombstones? | avg micros/op | stddev micros/op | avg ops/s | stddev ops/s
----------------- | ------------- | ---------------- | ------------ | ------------
None | 0.6186 | 0.04637 | 1,625,252.90 | 124,679.41
500 Expanded | 0.6019 | 0.03628 | 1,666,670.40 | 101,142.65
500 Unexpanded | 0.6435 | 0.03994 | 1,559,979.40 | 104,090.52
1k Expanded | 0.6034 | 0.04349 | 1,665,128.10 | 125,144.57
1k Unexpanded | 0.6261 | 0.03093 | 1,600,457.50 | 79,024.94
5k Expanded | 0.6163 | 0.05926 | 1,636,668.80 | 154,888.85
5k Unexpanded | 0.6402 | 0.04002 | 1,567,804.70 | 100,965.55
10k Expanded | 0.6036 | 0.05105 | 1,667,237.70 | 142,830.36
10k Unexpanded | 0.6128 | 0.02598 | 1,634,633.40 | 72,161.82
25k Expanded | 0.6198 | 0.04542 | 1,620,980.50 | 116,662.93
25k Unexpanded | 0.5478 | 0.0362 | 1,833,059.10 | 121,233.81
50k Expanded | 0.5104 | 0.04347 | 1,973,107.90 | 184,073.49
50k Unexpanded | 0.4528 | 0.03387 | 2,219,034.50 | 170,984.32
```
After a large enough quantity of range tombstones are written, range tombstone Gets can become faster than reading from an equivalent DB with several point tombstones.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4493
Differential Revision: D10842844
Pulled By: abhimadan
fbshipit-source-id: a7d44534f8120e6aabb65779d26c6b9df954c509
6 years ago
|
|
|
"Encountered error while reading data from range del block %s",
|
|
|
|
s.ToString().c_str());
|
|
|
|
IGNORE_STATUS_IF_ERROR(s);
|
|
|
|
} else {
|
|
|
|
rep_->fragmented_range_dels =
|
|
|
|
std::make_shared<FragmentedRangeTombstoneList>(std::move(iter),
|
|
|
|
internal_comparator);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
Status BlockBasedTable::PrefetchIndexAndFilterBlocks(
|
|
|
|
const ReadOptions& ro, FilePrefetchBuffer* prefetch_buffer,
|
|
|
|
InternalIterator* meta_iter, BlockBasedTable* new_table, bool prefetch_all,
|
|
|
|
const BlockBasedTableOptions& table_options, const int level,
|
|
|
|
size_t file_size, size_t max_file_size_for_l0_meta_pin,
|
|
|
|
BlockCacheLookupContext* lookup_context) {
|
|
|
|
// Find filter handle and filter type
|
|
|
|
if (rep_->filter_policy) {
|
|
|
|
auto name = rep_->filter_policy->CompatibilityName();
|
|
|
|
bool builtin_compatible =
|
|
|
|
strcmp(name, BuiltinFilterPolicy::kCompatibilityName()) == 0;
|
|
|
|
|
|
|
|
for (const auto& [filter_type, prefix] :
|
|
|
|
{std::make_pair(Rep::FilterType::kFullFilter, kFullFilterBlockPrefix),
|
|
|
|
std::make_pair(Rep::FilterType::kPartitionedFilter,
|
|
|
|
kPartitionedFilterBlockPrefix),
|
Remove deprecated block-based filter (#10184)
Summary:
In https://github.com/facebook/rocksdb/issues/9535, release 7.0, we hid the old block-based filter from being created using
the public API, because of its inefficiency. Although we normally maintain read compatibility
on old DBs forever, filters are not required for reading a DB, only for optimizing read
performance. Thus, it should be acceptable to remove this code and the substantial
maintenance burden it carries as useful features are developed and validated (such
as user timestamp).
This change completely removes the code for reading and writing the old block-based
filters, net removing about 1370 lines of code no longer needed. Options removed from
testing / benchmarking tools. The prior existence is only evident in a couple of places:
* `CacheEntryRole::kDeprecatedFilterBlock` - We can update this public API enum in
a major release to minimize source code incompatibilities.
* A warning is logged when an old table file is opened that used the old block-based
filter. This is provided as a courtesy, and would be a pain to unit test, so manual testing
should suffice. Unfortunately, sst_dump does not tell you whether a file uses
block-based filter, and the structure of the code makes it very difficult to fix.
* To detect that case, `kObsoleteFilterBlockPrefix` (renamed from `kFilterBlockPrefix`)
for metaindex is maintained (for now).
Other notes:
* In some cases where numbers are associated with filter configurations, we have had to
update the assigned numbers so that they all correspond to something that exists.
* Fixed potential stat counting bug by assuming `filter_checked = false` for cases
like `filter == nullptr` rather than assuming `filter_checked = true`
* Removed obsolete `block_offset` and `prefix_extractor` parameters from several
functions.
* Removed some unnecessary checks `if (!table_prefix_extractor() && !prefix_extractor)`
because the caller guarantees the prefix extractor exists and is compatible
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10184
Test Plan:
tests updated, manually test new warning in LOG using base version to
generate a DB
Reviewed By: riversand963
Differential Revision: D37212647
Pulled By: pdillinger
fbshipit-source-id: 06ee020d8de3b81260ffc36ad0c1202cbf463a80
2 years ago
|
|
|
std::make_pair(Rep::FilterType::kNoFilter,
|
|
|
|
kObsoleteFilterBlockPrefix)}) {
|
|
|
|
if (builtin_compatible) {
|
|
|
|
// This code is only here to deal with a hiccup in early 7.0.x where
|
|
|
|
// there was an unintentional name change in the SST files metadata.
|
|
|
|
// It should be OK to remove this in the future (late 2022) and just
|
|
|
|
// have the 'else' code.
|
|
|
|
// NOTE: the test:: names below are likely not needed but included
|
|
|
|
// out of caution
|
|
|
|
static const std::unordered_set<std::string> kBuiltinNameAndAliases = {
|
|
|
|
BuiltinFilterPolicy::kCompatibilityName(),
|
|
|
|
test::LegacyBloomFilterPolicy::kClassName(),
|
|
|
|
test::FastLocalBloomFilterPolicy::kClassName(),
|
|
|
|
test::Standard128RibbonFilterPolicy::kClassName(),
|
Remove deprecated block-based filter (#10184)
Summary:
In https://github.com/facebook/rocksdb/issues/9535, release 7.0, we hid the old block-based filter from being created using
the public API, because of its inefficiency. Although we normally maintain read compatibility
on old DBs forever, filters are not required for reading a DB, only for optimizing read
performance. Thus, it should be acceptable to remove this code and the substantial
maintenance burden it carries as useful features are developed and validated (such
as user timestamp).
This change completely removes the code for reading and writing the old block-based
filters, net removing about 1370 lines of code no longer needed. Options removed from
testing / benchmarking tools. The prior existence is only evident in a couple of places:
* `CacheEntryRole::kDeprecatedFilterBlock` - We can update this public API enum in
a major release to minimize source code incompatibilities.
* A warning is logged when an old table file is opened that used the old block-based
filter. This is provided as a courtesy, and would be a pain to unit test, so manual testing
should suffice. Unfortunately, sst_dump does not tell you whether a file uses
block-based filter, and the structure of the code makes it very difficult to fix.
* To detect that case, `kObsoleteFilterBlockPrefix` (renamed from `kFilterBlockPrefix`)
for metaindex is maintained (for now).
Other notes:
* In some cases where numbers are associated with filter configurations, we have had to
update the assigned numbers so that they all correspond to something that exists.
* Fixed potential stat counting bug by assuming `filter_checked = false` for cases
like `filter == nullptr` rather than assuming `filter_checked = true`
* Removed obsolete `block_offset` and `prefix_extractor` parameters from several
functions.
* Removed some unnecessary checks `if (!table_prefix_extractor() && !prefix_extractor)`
because the caller guarantees the prefix extractor exists and is compatible
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10184
Test Plan:
tests updated, manually test new warning in LOG using base version to
generate a DB
Reviewed By: riversand963
Differential Revision: D37212647
Pulled By: pdillinger
fbshipit-source-id: 06ee020d8de3b81260ffc36ad0c1202cbf463a80
2 years ago
|
|
|
"rocksdb.internal.DeprecatedBlockBasedBloomFilter",
|
|
|
|
BloomFilterPolicy::kClassName(),
|
|
|
|
RibbonFilterPolicy::kClassName(),
|
|
|
|
};
|
|
|
|
|
|
|
|
// For efficiency, do a prefix seek and see if the first match is
|
|
|
|
// good.
|
|
|
|
meta_iter->Seek(prefix);
|
|
|
|
if (meta_iter->status().ok() && meta_iter->Valid()) {
|
|
|
|
Slice key = meta_iter->key();
|
|
|
|
if (key.starts_with(prefix)) {
|
|
|
|
key.remove_prefix(prefix.size());
|
|
|
|
if (kBuiltinNameAndAliases.find(key.ToString()) !=
|
|
|
|
kBuiltinNameAndAliases.end()) {
|
|
|
|
Slice v = meta_iter->value();
|
|
|
|
Status s = rep_->filter_handle.DecodeFrom(&v);
|
|
|
|
if (s.ok()) {
|
|
|
|
rep_->filter_type = filter_type;
|
Remove deprecated block-based filter (#10184)
Summary:
In https://github.com/facebook/rocksdb/issues/9535, release 7.0, we hid the old block-based filter from being created using
the public API, because of its inefficiency. Although we normally maintain read compatibility
on old DBs forever, filters are not required for reading a DB, only for optimizing read
performance. Thus, it should be acceptable to remove this code and the substantial
maintenance burden it carries as useful features are developed and validated (such
as user timestamp).
This change completely removes the code for reading and writing the old block-based
filters, net removing about 1370 lines of code no longer needed. Options removed from
testing / benchmarking tools. The prior existence is only evident in a couple of places:
* `CacheEntryRole::kDeprecatedFilterBlock` - We can update this public API enum in
a major release to minimize source code incompatibilities.
* A warning is logged when an old table file is opened that used the old block-based
filter. This is provided as a courtesy, and would be a pain to unit test, so manual testing
should suffice. Unfortunately, sst_dump does not tell you whether a file uses
block-based filter, and the structure of the code makes it very difficult to fix.
* To detect that case, `kObsoleteFilterBlockPrefix` (renamed from `kFilterBlockPrefix`)
for metaindex is maintained (for now).
Other notes:
* In some cases where numbers are associated with filter configurations, we have had to
update the assigned numbers so that they all correspond to something that exists.
* Fixed potential stat counting bug by assuming `filter_checked = false` for cases
like `filter == nullptr` rather than assuming `filter_checked = true`
* Removed obsolete `block_offset` and `prefix_extractor` parameters from several
functions.
* Removed some unnecessary checks `if (!table_prefix_extractor() && !prefix_extractor)`
because the caller guarantees the prefix extractor exists and is compatible
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10184
Test Plan:
tests updated, manually test new warning in LOG using base version to
generate a DB
Reviewed By: riversand963
Differential Revision: D37212647
Pulled By: pdillinger
fbshipit-source-id: 06ee020d8de3b81260ffc36ad0c1202cbf463a80
2 years ago
|
|
|
if (filter_type == Rep::FilterType::kNoFilter) {
|
|
|
|
ROCKS_LOG_WARN(rep_->ioptions.logger,
|
|
|
|
"Detected obsolete filter type in %s. Read "
|
|
|
|
"performance might suffer until DB is fully "
|
|
|
|
"re-compacted.",
|
|
|
|
rep_->file->file_name().c_str());
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
std::string filter_block_key = prefix + name;
|
|
|
|
if (FindMetaBlock(meta_iter, filter_block_key, &rep_->filter_handle)
|
|
|
|
.ok()) {
|
|
|
|
rep_->filter_type = filter_type;
|
Remove deprecated block-based filter (#10184)
Summary:
In https://github.com/facebook/rocksdb/issues/9535, release 7.0, we hid the old block-based filter from being created using
the public API, because of its inefficiency. Although we normally maintain read compatibility
on old DBs forever, filters are not required for reading a DB, only for optimizing read
performance. Thus, it should be acceptable to remove this code and the substantial
maintenance burden it carries as useful features are developed and validated (such
as user timestamp).
This change completely removes the code for reading and writing the old block-based
filters, net removing about 1370 lines of code no longer needed. Options removed from
testing / benchmarking tools. The prior existence is only evident in a couple of places:
* `CacheEntryRole::kDeprecatedFilterBlock` - We can update this public API enum in
a major release to minimize source code incompatibilities.
* A warning is logged when an old table file is opened that used the old block-based
filter. This is provided as a courtesy, and would be a pain to unit test, so manual testing
should suffice. Unfortunately, sst_dump does not tell you whether a file uses
block-based filter, and the structure of the code makes it very difficult to fix.
* To detect that case, `kObsoleteFilterBlockPrefix` (renamed from `kFilterBlockPrefix`)
for metaindex is maintained (for now).
Other notes:
* In some cases where numbers are associated with filter configurations, we have had to
update the assigned numbers so that they all correspond to something that exists.
* Fixed potential stat counting bug by assuming `filter_checked = false` for cases
like `filter == nullptr` rather than assuming `filter_checked = true`
* Removed obsolete `block_offset` and `prefix_extractor` parameters from several
functions.
* Removed some unnecessary checks `if (!table_prefix_extractor() && !prefix_extractor)`
because the caller guarantees the prefix extractor exists and is compatible
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10184
Test Plan:
tests updated, manually test new warning in LOG using base version to
generate a DB
Reviewed By: riversand963
Differential Revision: D37212647
Pulled By: pdillinger
fbshipit-source-id: 06ee020d8de3b81260ffc36ad0c1202cbf463a80
2 years ago
|
|
|
if (filter_type == Rep::FilterType::kNoFilter) {
|
|
|
|
ROCKS_LOG_WARN(
|
|
|
|
rep_->ioptions.logger,
|
|
|
|
"Detected obsolete filter type in %s. Read performance might "
|
|
|
|
"suffer until DB is fully re-compacted.",
|
|
|
|
rep_->file->file_name().c_str());
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
// Partition filters cannot be enabled without partition indexes
|
|
|
|
assert(rep_->filter_type != Rep::FilterType::kPartitionedFilter ||
|
|
|
|
rep_->index_type == BlockBasedTableOptions::kTwoLevelIndexSearch);
|
|
|
|
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
// Find compression dictionary handle
|
|
|
|
Status s = FindOptionalMetaBlock(meta_iter, kCompressionDictBlockName,
|
|
|
|
&rep_->compression_dict_handle);
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
BlockBasedTableOptions::IndexType index_type = rep_->index_type;
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
|
|
|
|
const bool use_cache = table_options.cache_index_and_filter_blocks;
|
|
|
|
|
|
|
|
const bool maybe_flushed =
|
|
|
|
level == 0 && file_size <= max_file_size_for_l0_meta_pin;
|
|
|
|
std::function<bool(PinningTier, PinningTier)> is_pinned =
|
|
|
|
[maybe_flushed, &is_pinned](PinningTier pinning_tier,
|
|
|
|
PinningTier fallback_pinning_tier) {
|
|
|
|
// Fallback to fallback would lead to infinite recursion. Disallow it.
|
|
|
|
assert(fallback_pinning_tier != PinningTier::kFallback);
|
|
|
|
|
|
|
|
switch (pinning_tier) {
|
|
|
|
case PinningTier::kFallback:
|
|
|
|
return is_pinned(fallback_pinning_tier,
|
|
|
|
PinningTier::kNone /* fallback_pinning_tier */);
|
|
|
|
case PinningTier::kNone:
|
|
|
|
return false;
|
|
|
|
case PinningTier::kFlushedAndSimilar:
|
|
|
|
return maybe_flushed;
|
|
|
|
case PinningTier::kAll:
|
|
|
|
return true;
|
|
|
|
};
|
|
|
|
|
|
|
|
// In GCC, this is needed to suppress `control reaches end of non-void
|
|
|
|
// function [-Werror=return-type]`.
|
|
|
|
assert(false);
|
|
|
|
return false;
|
|
|
|
};
|
|
|
|
const bool pin_top_level_index = is_pinned(
|
|
|
|
table_options.metadata_cache_options.top_level_index_pinning,
|
|
|
|
table_options.pin_top_level_index_and_filter ? PinningTier::kAll
|
|
|
|
: PinningTier::kNone);
|
|
|
|
const bool pin_partition =
|
|
|
|
is_pinned(table_options.metadata_cache_options.partition_pinning,
|
|
|
|
table_options.pin_l0_filter_and_index_blocks_in_cache
|
|
|
|
? PinningTier::kFlushedAndSimilar
|
|
|
|
: PinningTier::kNone);
|
|
|
|
const bool pin_unpartitioned =
|
|
|
|
is_pinned(table_options.metadata_cache_options.unpartitioned_pinning,
|
|
|
|
table_options.pin_l0_filter_and_index_blocks_in_cache
|
|
|
|
? PinningTier::kFlushedAndSimilar
|
|
|
|
: PinningTier::kNone);
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
|
|
|
|
// pin the first level of index
|
|
|
|
const bool pin_index =
|
|
|
|
index_type == BlockBasedTableOptions::kTwoLevelIndexSearch
|
|
|
|
? pin_top_level_index
|
|
|
|
: pin_unpartitioned;
|
|
|
|
// prefetch the first level of index
|
|
|
|
// WART: this might be redundant (unnecessary cache hit) if !pin_index,
|
|
|
|
// depending on prepopulate_block_cache option
|
|
|
|
const bool prefetch_index = prefetch_all || pin_index;
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
|
|
|
|
std::unique_ptr<IndexReader> index_reader;
|
|
|
|
s = new_table->CreateIndexReader(ro, prefetch_buffer, meta_iter, use_cache,
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
prefetch_index, pin_index, lookup_context,
|
|
|
|
&index_reader);
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
rep_->index_reader = std::move(index_reader);
|
|
|
|
|
|
|
|
// The partitions of partitioned index are always stored in cache. They
|
|
|
|
// are hence follow the configuration for pin and prefetch regardless of
|
|
|
|
// the value of cache_index_and_filter_blocks
|
|
|
|
if (prefetch_all || pin_partition) {
|
Record and use the tail size to prefetch table tail (#11406)
Summary:
**Context:**
We prefetch the tail part of a SST file (i.e, the blocks after data blocks till the end of the file) during each SST file open in hope to prefetch all the stuff at once ahead of time for later read e.g, footer, meta index, filter/index etc. The existing approach to estimate the tail size to prefetch is through `TailPrefetchStats` heuristics introduced in https://github.com/facebook/rocksdb/pull/4156, which has caused small reads in unlucky case (e.g, small read into the tail buffer during table open in thread 1 under the same BlockBasedTableFactory object can make thread 2's tail prefetching use a small size that it shouldn't) and is hard to debug. Therefore we decide to record the exact tail size and use it directly to prefetch tail of the SST instead of relying heuristics.
**Summary:**
- Obtain and record in manifest the tail size in `BlockBasedTableBuilder::Finish()`
- For backward compatibility, we fall back to TailPrefetchStats and last to simple heuristics that the tail size is a linear portion of the file size - see PR conversation for more.
- Make`tail_start_offset` part of the table properties and deduct tail size to record in manifest for external files (e.g, file ingestion, import CF) and db repair (with no access to manifest).
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11406
Test Plan:
1. New UT
2. db bench
Note: db bench on /tmp/ where direct read is supported is too slow to finish and the default pinning setting in db bench is not helpful to profile # sst read of Get. Therefore I hacked the following to obtain the following comparison.
```
diff --git a/table/block_based/block_based_table_reader.cc b/table/block_based/block_based_table_reader.cc
index bd5669f0f..791484c1f 100644
--- a/table/block_based/block_based_table_reader.cc
+++ b/table/block_based/block_based_table_reader.cc
@@ -838,7 +838,7 @@ Status BlockBasedTable::PrefetchTail(
&tail_prefetch_size);
// Try file system prefetch
- if (!file->use_direct_io() && !force_direct_prefetch) {
+ if (false && !file->use_direct_io() && !force_direct_prefetch) {
if (!file->Prefetch(prefetch_off, prefetch_len, ro.rate_limiter_priority)
.IsNotSupported()) {
prefetch_buffer->reset(new FilePrefetchBuffer(
diff --git a/tools/db_bench_tool.cc b/tools/db_bench_tool.cc
index ea40f5fa0..39a0ac385 100644
--- a/tools/db_bench_tool.cc
+++ b/tools/db_bench_tool.cc
@@ -4191,6 +4191,8 @@ class Benchmark {
std::shared_ptr<TableFactory>(NewCuckooTableFactory(table_options));
} else {
BlockBasedTableOptions block_based_options;
+ block_based_options.metadata_cache_options.partition_pinning =
+ PinningTier::kAll;
block_based_options.checksum =
static_cast<ChecksumType>(FLAGS_checksum_type);
if (FLAGS_use_hash_search) {
```
Create DB
```
./db_bench --bloom_bits=3 --use_existing_db=1 --seed=1682546046158958 --partition_index_and_filters=1 --statistics=1 -db=/dev/shm/testdb/ -benchmarks=readrandom -key_size=3200 -value_size=512 -num=1000000 -write_buffer_size=6550000 -disable_auto_compactions=false -target_file_size_base=6550000 -compression_type=none
```
ReadRandom
```
./db_bench --bloom_bits=3 --use_existing_db=1 --seed=1682546046158958 --partition_index_and_filters=1 --statistics=1 -db=/dev/shm/testdb/ -benchmarks=readrandom -key_size=3200 -value_size=512 -num=1000000 -write_buffer_size=6550000 -disable_auto_compactions=false -target_file_size_base=6550000 -compression_type=none
```
(a) Existing (Use TailPrefetchStats for tail size + use seperate prefetch buffer in PartitionedFilter/IndexReader::CacheDependencies())
```
rocksdb.table.open.prefetch.tail.hit COUNT : 3395
rocksdb.sst.read.micros P50 : 5.655570 P95 : 9.931396 P99 : 14.845454 P100 : 585.000000 COUNT : 999905 SUM : 6590614
```
(b) This PR (Record tail size + use the same tail buffer in PartitionedFilter/IndexReader::CacheDependencies())
```
rocksdb.table.open.prefetch.tail.hit COUNT : 14257
rocksdb.sst.read.micros P50 : 5.173347 P95 : 9.015017 P99 : 12.912610 P100 : 228.000000 COUNT : 998547 SUM : 5976540
```
As we can see, we increase the prefetch tail hit count and decrease SST read count with this PR
3. Test backward compatibility by stepping through reading with post-PR code on a db generated pre-PR.
Reviewed By: pdillinger
Differential Revision: D45413346
Pulled By: hx235
fbshipit-source-id: 7d5e36a60a72477218f79905168d688452a4c064
2 years ago
|
|
|
s = rep_->index_reader->CacheDependencies(ro, pin_partition,
|
|
|
|
prefetch_buffer);
|
|
|
|
}
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
}
|
|
|
|
|
|
|
|
// pin the first level of filter
|
|
|
|
const bool pin_filter =
|
|
|
|
rep_->filter_type == Rep::FilterType::kPartitionedFilter
|
|
|
|
? pin_top_level_index
|
|
|
|
: pin_unpartitioned;
|
|
|
|
// prefetch the first level of filter
|
|
|
|
// WART: this might be redundant (unnecessary cache hit) if !pin_filter,
|
|
|
|
// depending on prepopulate_block_cache option
|
|
|
|
const bool prefetch_filter = prefetch_all || pin_filter;
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
if (rep_->filter_policy) {
|
|
|
|
auto filter = new_table->CreateFilterBlockReader(
|
|
|
|
ro, prefetch_buffer, use_cache, prefetch_filter, pin_filter,
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
lookup_context);
|
|
|
|
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
if (filter) {
|
|
|
|
// Refer to the comment above about paritioned indexes always being cached
|
|
|
|
if (prefetch_all || pin_partition) {
|
Record and use the tail size to prefetch table tail (#11406)
Summary:
**Context:**
We prefetch the tail part of a SST file (i.e, the blocks after data blocks till the end of the file) during each SST file open in hope to prefetch all the stuff at once ahead of time for later read e.g, footer, meta index, filter/index etc. The existing approach to estimate the tail size to prefetch is through `TailPrefetchStats` heuristics introduced in https://github.com/facebook/rocksdb/pull/4156, which has caused small reads in unlucky case (e.g, small read into the tail buffer during table open in thread 1 under the same BlockBasedTableFactory object can make thread 2's tail prefetching use a small size that it shouldn't) and is hard to debug. Therefore we decide to record the exact tail size and use it directly to prefetch tail of the SST instead of relying heuristics.
**Summary:**
- Obtain and record in manifest the tail size in `BlockBasedTableBuilder::Finish()`
- For backward compatibility, we fall back to TailPrefetchStats and last to simple heuristics that the tail size is a linear portion of the file size - see PR conversation for more.
- Make`tail_start_offset` part of the table properties and deduct tail size to record in manifest for external files (e.g, file ingestion, import CF) and db repair (with no access to manifest).
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11406
Test Plan:
1. New UT
2. db bench
Note: db bench on /tmp/ where direct read is supported is too slow to finish and the default pinning setting in db bench is not helpful to profile # sst read of Get. Therefore I hacked the following to obtain the following comparison.
```
diff --git a/table/block_based/block_based_table_reader.cc b/table/block_based/block_based_table_reader.cc
index bd5669f0f..791484c1f 100644
--- a/table/block_based/block_based_table_reader.cc
+++ b/table/block_based/block_based_table_reader.cc
@@ -838,7 +838,7 @@ Status BlockBasedTable::PrefetchTail(
&tail_prefetch_size);
// Try file system prefetch
- if (!file->use_direct_io() && !force_direct_prefetch) {
+ if (false && !file->use_direct_io() && !force_direct_prefetch) {
if (!file->Prefetch(prefetch_off, prefetch_len, ro.rate_limiter_priority)
.IsNotSupported()) {
prefetch_buffer->reset(new FilePrefetchBuffer(
diff --git a/tools/db_bench_tool.cc b/tools/db_bench_tool.cc
index ea40f5fa0..39a0ac385 100644
--- a/tools/db_bench_tool.cc
+++ b/tools/db_bench_tool.cc
@@ -4191,6 +4191,8 @@ class Benchmark {
std::shared_ptr<TableFactory>(NewCuckooTableFactory(table_options));
} else {
BlockBasedTableOptions block_based_options;
+ block_based_options.metadata_cache_options.partition_pinning =
+ PinningTier::kAll;
block_based_options.checksum =
static_cast<ChecksumType>(FLAGS_checksum_type);
if (FLAGS_use_hash_search) {
```
Create DB
```
./db_bench --bloom_bits=3 --use_existing_db=1 --seed=1682546046158958 --partition_index_and_filters=1 --statistics=1 -db=/dev/shm/testdb/ -benchmarks=readrandom -key_size=3200 -value_size=512 -num=1000000 -write_buffer_size=6550000 -disable_auto_compactions=false -target_file_size_base=6550000 -compression_type=none
```
ReadRandom
```
./db_bench --bloom_bits=3 --use_existing_db=1 --seed=1682546046158958 --partition_index_and_filters=1 --statistics=1 -db=/dev/shm/testdb/ -benchmarks=readrandom -key_size=3200 -value_size=512 -num=1000000 -write_buffer_size=6550000 -disable_auto_compactions=false -target_file_size_base=6550000 -compression_type=none
```
(a) Existing (Use TailPrefetchStats for tail size + use seperate prefetch buffer in PartitionedFilter/IndexReader::CacheDependencies())
```
rocksdb.table.open.prefetch.tail.hit COUNT : 3395
rocksdb.sst.read.micros P50 : 5.655570 P95 : 9.931396 P99 : 14.845454 P100 : 585.000000 COUNT : 999905 SUM : 6590614
```
(b) This PR (Record tail size + use the same tail buffer in PartitionedFilter/IndexReader::CacheDependencies())
```
rocksdb.table.open.prefetch.tail.hit COUNT : 14257
rocksdb.sst.read.micros P50 : 5.173347 P95 : 9.015017 P99 : 12.912610 P100 : 228.000000 COUNT : 998547 SUM : 5976540
```
As we can see, we increase the prefetch tail hit count and decrease SST read count with this PR
3. Test backward compatibility by stepping through reading with post-PR code on a db generated pre-PR.
Reviewed By: pdillinger
Differential Revision: D45413346
Pulled By: hx235
fbshipit-source-id: 7d5e36a60a72477218f79905168d688452a4c064
2 years ago
|
|
|
s = filter->CacheDependencies(ro, pin_partition, prefetch_buffer);
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
}
|
|
|
|
rep_->filter = std::move(filter);
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!rep_->compression_dict_handle.IsNull()) {
|
|
|
|
std::unique_ptr<UncompressionDictReader> uncompression_dict_reader;
|
|
|
|
s = UncompressionDictReader::Create(
|
|
|
|
this, ro, prefetch_buffer, use_cache, prefetch_all || pin_unpartitioned,
|
|
|
|
pin_unpartitioned, lookup_context, &uncompression_dict_reader);
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
|
|
|
|
rep_->uncompression_dict_reader = std::move(uncompression_dict_reader);
|
|
|
|
}
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
|
|
|
|
assert(s.ok());
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
void BlockBasedTable::SetupForCompaction() {
|
|
|
|
switch (rep_->ioptions.access_hint_on_compaction_start) {
|
|
|
|
case Options::NONE:
|
|
|
|
break;
|
|
|
|
case Options::NORMAL:
|
Introduce a new storage specific Env API (#5761)
Summary:
The current Env API encompasses both storage/file operations, as well as OS related operations. Most of the APIs return a Status, which does not have enough metadata about an error, such as whether its retry-able or not, scope (i.e fault domain) of the error etc., that may be required in order to properly handle a storage error. The file APIs also do not provide enough control over the IO SLA, such as timeout, prioritization, hinting about placement and redundancy etc.
This PR separates out the file/storage APIs from Env into a new FileSystem class. The APIs are updated to return an IOStatus with metadata about the error, as well as to take an IOOptions structure as input in order to allow more control over the IO.
The user can set both ```options.env``` and ```options.file_system``` to specify that RocksDB should use the former for OS related operations and the latter for storage operations. Internally, a ```CompositeEnvWrapper``` has been introduced that inherits from ```Env``` and redirects individual methods to either an ```Env``` implementation or the ```FileSystem``` as appropriate. When options are sanitized during ```DB::Open```, ```options.env``` is replaced with a newly allocated ```CompositeEnvWrapper``` instance if both env and file_system have been specified. This way, the rest of the RocksDB code can continue to function as before.
This PR also ports PosixEnv to the new API by splitting it into two - PosixEnv and PosixFileSystem. PosixEnv is defined as a sub-class of CompositeEnvWrapper, and threading/time functions are overridden with Posix specific implementations in order to avoid an extra level of indirection.
The ```CompositeEnvWrapper``` translates ```IOStatus``` return code to ```Status```, and sets the severity to ```kSoftError``` if the io_status is retryable. The error handling code in RocksDB can then recover the DB automatically.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5761
Differential Revision: D18868376
Pulled By: anand1976
fbshipit-source-id: 39efe18a162ea746fabac6360ff529baba48486f
5 years ago
|
|
|
rep_->file->file()->Hint(FSRandomAccessFile::kNormal);
|
|
|
|
break;
|
|
|
|
case Options::SEQUENTIAL:
|
Introduce a new storage specific Env API (#5761)
Summary:
The current Env API encompasses both storage/file operations, as well as OS related operations. Most of the APIs return a Status, which does not have enough metadata about an error, such as whether its retry-able or not, scope (i.e fault domain) of the error etc., that may be required in order to properly handle a storage error. The file APIs also do not provide enough control over the IO SLA, such as timeout, prioritization, hinting about placement and redundancy etc.
This PR separates out the file/storage APIs from Env into a new FileSystem class. The APIs are updated to return an IOStatus with metadata about the error, as well as to take an IOOptions structure as input in order to allow more control over the IO.
The user can set both ```options.env``` and ```options.file_system``` to specify that RocksDB should use the former for OS related operations and the latter for storage operations. Internally, a ```CompositeEnvWrapper``` has been introduced that inherits from ```Env``` and redirects individual methods to either an ```Env``` implementation or the ```FileSystem``` as appropriate. When options are sanitized during ```DB::Open```, ```options.env``` is replaced with a newly allocated ```CompositeEnvWrapper``` instance if both env and file_system have been specified. This way, the rest of the RocksDB code can continue to function as before.
This PR also ports PosixEnv to the new API by splitting it into two - PosixEnv and PosixFileSystem. PosixEnv is defined as a sub-class of CompositeEnvWrapper, and threading/time functions are overridden with Posix specific implementations in order to avoid an extra level of indirection.
The ```CompositeEnvWrapper``` translates ```IOStatus``` return code to ```Status```, and sets the severity to ```kSoftError``` if the io_status is retryable. The error handling code in RocksDB can then recover the DB automatically.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5761
Differential Revision: D18868376
Pulled By: anand1976
fbshipit-source-id: 39efe18a162ea746fabac6360ff529baba48486f
5 years ago
|
|
|
rep_->file->file()->Hint(FSRandomAccessFile::kSequential);
|
|
|
|
break;
|
|
|
|
case Options::WILLNEED:
|
Introduce a new storage specific Env API (#5761)
Summary:
The current Env API encompasses both storage/file operations, as well as OS related operations. Most of the APIs return a Status, which does not have enough metadata about an error, such as whether its retry-able or not, scope (i.e fault domain) of the error etc., that may be required in order to properly handle a storage error. The file APIs also do not provide enough control over the IO SLA, such as timeout, prioritization, hinting about placement and redundancy etc.
This PR separates out the file/storage APIs from Env into a new FileSystem class. The APIs are updated to return an IOStatus with metadata about the error, as well as to take an IOOptions structure as input in order to allow more control over the IO.
The user can set both ```options.env``` and ```options.file_system``` to specify that RocksDB should use the former for OS related operations and the latter for storage operations. Internally, a ```CompositeEnvWrapper``` has been introduced that inherits from ```Env``` and redirects individual methods to either an ```Env``` implementation or the ```FileSystem``` as appropriate. When options are sanitized during ```DB::Open```, ```options.env``` is replaced with a newly allocated ```CompositeEnvWrapper``` instance if both env and file_system have been specified. This way, the rest of the RocksDB code can continue to function as before.
This PR also ports PosixEnv to the new API by splitting it into two - PosixEnv and PosixFileSystem. PosixEnv is defined as a sub-class of CompositeEnvWrapper, and threading/time functions are overridden with Posix specific implementations in order to avoid an extra level of indirection.
The ```CompositeEnvWrapper``` translates ```IOStatus``` return code to ```Status```, and sets the severity to ```kSoftError``` if the io_status is retryable. The error handling code in RocksDB can then recover the DB automatically.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5761
Differential Revision: D18868376
Pulled By: anand1976
fbshipit-source-id: 39efe18a162ea746fabac6360ff529baba48486f
5 years ago
|
|
|
rep_->file->file()->Hint(FSRandomAccessFile::kWillNeed);
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
assert(false);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
std::shared_ptr<const TableProperties> BlockBasedTable::GetTableProperties()
|
|
|
|
const {
|
|
|
|
return rep_->table_properties;
|
|
|
|
}
|
|
|
|
|
|
|
|
size_t BlockBasedTable::ApproximateMemoryUsage() const {
|
|
|
|
size_t usage = 0;
|
|
|
|
if (rep_) {
|
|
|
|
usage += rep_->ApproximateMemoryUsage();
|
|
|
|
} else {
|
|
|
|
return usage;
|
|
|
|
}
|
|
|
|
if (rep_->filter) {
|
|
|
|
usage += rep_->filter->ApproximateMemoryUsage();
|
|
|
|
}
|
|
|
|
if (rep_->index_reader) {
|
|
|
|
usage += rep_->index_reader->ApproximateMemoryUsage();
|
|
|
|
}
|
|
|
|
if (rep_->uncompression_dict_reader) {
|
|
|
|
usage += rep_->uncompression_dict_reader->ApproximateMemoryUsage();
|
|
|
|
}
|
|
|
|
if (rep_->table_properties) {
|
|
|
|
usage += rep_->table_properties->ApproximateMemoryUsage();
|
|
|
|
}
|
|
|
|
return usage;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Load the meta-index-block from the file. On success, return the loaded
|
|
|
|
// metaindex
|
|
|
|
// block and its iterator.
|
|
|
|
Status BlockBasedTable::ReadMetaIndexBlock(
|
|
|
|
const ReadOptions& ro, FilePrefetchBuffer* prefetch_buffer,
|
|
|
|
std::unique_ptr<Block>* metaindex_block,
|
|
|
|
std::unique_ptr<InternalIterator>* iter) {
|
|
|
|
// TODO(sanjay): Skip this if footer.metaindex_handle() size indicates
|
|
|
|
// it is an empty block.
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
std::unique_ptr<Block_kMetaIndex> metaindex;
|
|
|
|
Status s = ReadAndParseBlockFromFile(
|
|
|
|
rep_->file.get(), prefetch_buffer, rep_->footer, ro,
|
|
|
|
rep_->footer.metaindex_handle(), &metaindex, rep_->ioptions,
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
rep_->create_context, true /*maybe_compressed*/,
|
|
|
|
UncompressionDict::GetEmptyDict(), rep_->persistent_cache_options,
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
GetMemoryAllocator(rep_->table_options), false /* for_compaction */,
|
|
|
|
false /* async_read */);
|
|
|
|
|
|
|
|
if (!s.ok()) {
|
|
|
|
ROCKS_LOG_ERROR(rep_->ioptions.logger,
|
|
|
|
"Encountered error while reading data from properties"
|
|
|
|
" block %s",
|
|
|
|
s.ToString().c_str());
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
*metaindex_block = std::move(metaindex);
|
|
|
|
// meta block uses bytewise comparator.
|
|
|
|
iter->reset(metaindex_block->get()->NewMetaIterator());
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
template <typename TBlocklike>
|
HyperClockCache support for SecondaryCache, with refactoring (#11301)
Summary:
Internally refactors SecondaryCache integration out of LRUCache specifically and into a wrapper/adapter class that works with various Cache implementations. Notably, this relies on separating the notion of async lookup handles from other cache handles, so that HyperClockCache doesn't have to deal with the problem of allocating handles from the hash table for lookups that might fail anyway, and might be on the same key without support for coalescing. (LRUCache's hash table can incorporate previously allocated handles thanks to its pointer indirection.) Specifically, I'm worried about the case in which hundreds of threads try to access the same block and probing in the hash table degrades to linear search on the pile of entries with the same key.
This change is a big step in the direction of supporting stacked SecondaryCaches, but there are obstacles to completing that. Especially, there is no SecondaryCache hook for evictions to pass from one to the next. It has been proposed that evictions be transmitted simply as the persisted data (as in SaveToCallback), but given the current structure provided by the CacheItemHelpers, that would require an extra copy of the block data, because there's intentionally no way to ask for a contiguous Slice of the data (to allow for flexibility in storage). `AsyncLookupHandle` and the re-worked `WaitAll()` should be essentially prepared for stacked SecondaryCaches, but several "TODO with stacked secondaries" issues remain in various places.
It could be argued that the stacking instead be done as a SecondaryCache adapter that wraps two (or more) SecondaryCaches, but at least with the current API that would require an extra heap allocation on SecondaryCache Lookup for a wrapper SecondaryCacheResultHandle that can transfer a Lookup between secondaries. We could also consider trying to unify the Cache and SecondaryCache APIs, though that might be difficult if `AsyncLookupHandle` is kept a fixed struct.
## cache.h (public API)
Moves `secondary_cache` option from LRUCacheOptions to ShardedCacheOptions so that it is applicable to HyperClockCache.
## advanced_cache.h (advanced public API)
* Add `Cache::CreateStandalone()` so that the SecondaryCache support wrapper can use it.
* Add `SetEvictionCallback()` / `eviction_callback_` so that the SecondaryCache support wrapper can use it. Only a single callback is supported for efficiency. If there is ever a need for more than one, hopefully that can be handled with a broadcast callback wrapper.
These are essentially the two "extra" pieces of `Cache` for pulling out specific SecondaryCache support from the `Cache` implementation. I think it's a good trade-off as these are reasonable, limited, and reusable "cut points" into the `Cache` implementations.
* Remove async capability from standard `Lookup()` (getting rid of awkward restrictions on pending Handles) and add `AsyncLookupHandle` and `StartAsyncLookup()`. As noted in the comments, the full struct of `AsyncLookupHandle` is exposed so that it can be stack allocated, for efficiency, though more data is being copied around than before, which could impact performance. (Lookup info -> AsyncLookupHandle -> Handle vs. Lookup info -> Handle)
I could foresee a future in which a Cache internally saves a pointer to the AsyncLookupHandle, which means it's dangerous to allow it to be copyable or even movable. It also means it's not compatible with std::vector (which I don't like requiring as an API parameter anyway), so `WaitAll()` expects any contiguous array of AsyncLookupHandles. I believe this is best for common case efficiency, while behaving well in other cases also. For example, `WaitAll()` has no effect on default-constructed AsyncLookupHandles, which look like a completed cache miss.
## cacheable_entry.h
A couple of functions are obsolete because Cache::Handle can no longer be pending.
## cache.cc
Provides default implementations for new or revamped Cache functions, especially appropriate for non-blocking caches.
## secondary_cache_adapter.{h,cc}
The full details of the Cache wrapper adding SecondaryCache support. Essentially replicates the SecondaryCache handling that was in LRUCache, but obviously refactored. There is a bit of logic duplication, where Lookup() is essentially a manually optimized version of StartAsyncLookup() and Wait(), but it's roughly a dozen lines of code.
## sharded_cache.h, typed_cache.h, charged_cache.{h,cc}, sim_cache.cc
Simply updated for Cache API changes.
## lru_cache.{h,cc}
Carefully remove SecondaryCache logic, implement `CreateStandalone` and eviction handler functionality.
## clock_cache.{h,cc}
Expose existing `CreateStandalone` functionality, add eviction handler functionality. Light refactoring.
## block_based_table_reader*
Mostly re-worked the only usage of async Lookup, which is in BlockBasedTable::MultiGet. Used arrays in place of autovector in some places for efficiency. Simplified some logic by not trying to process some cache results before they're all ready.
Created new function `BlockBasedTable::GetCachePriority()` to reduce some pre-existing code duplication (and avoid making it worse).
Fixed at least one small bug from the prior confusing mixture of async and sync Lookups. In MaybeReadBlockAndLoadToCache(), called by RetrieveBlock(), called by MultiGet() with wait=false, is_cache_hit for the block_cache_tracer entry would not be set to true if the handle was pending after Lookup and before Wait.
## Intended follow-up work
* Figure out if there are any missing stats or block_cache_tracer work in refactored BlockBasedTable::MultiGet
* Stacked secondary caches (see above discussion)
* See if we can make up for the small MultiGet performance regression.
* Study more performance with SecondaryCache
* Items evicted from over-full LRUCache in Release were not being demoted to SecondaryCache, and still aren't to minimize unit test churn. Ideally they would be demoted, but it's an exceptional case so not a big deal.
* Use CreateStandalone for cache reservations (save unnecessary hash table operations). Not a big deal, but worthy cleanup.
* Somehow I got the contract for SecondaryCache::Insert wrong in #10945. (Doesn't take ownership!) That API comment needs to be fixed, but didn't want to mingle that in here.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11301
Test Plan:
## Unit tests
Generally updated to include HCC in SecondaryCache tests, though HyperClockCache has some different, less strict behaviors that leads to some tests not really being set up to work with it. Some of the tests remain disabled with it, but I think we have good coverage without them.
## Crash/stress test
Updated to use the new combination.
## Performance
First, let's check for regression on caches without secondary cache configured. Adding support for the eviction callback is likely to have a tiny effect, but it shouldn't be worrisome. LRUCache could benefit slightly from less logic around SecondaryCache handling. We can test with cache_bench default settings, built with DEBUG_LEVEL=0 and PORTABLE=0.
```
(while :; do base/cache_bench --cache_type=hyper_clock_cache | grep Rough; done) | awk '{ sum += $9; count++; print $0; print "Average: " int(sum / count) }'
```
**Before** this and #11299 (which could also have a small effect), running for about an hour, before & after running concurrently for each cache type:
HyperClockCache: 3168662 (average parallel ops/sec)
LRUCache: 2940127
**After** this and #11299, running for about an hour:
HyperClockCache: 3164862 (average parallel ops/sec) (0.12% slower)
LRUCache: 2940928 (0.03% faster)
This is an acceptable difference IMHO.
Next, let's consider essentially the worst case of new CPU overhead affecting overall performance. MultiGet uses the async lookup interface regardless of whether SecondaryCache or folly are used. We can configure a benchmark where all block cache queries are for data blocks, and all are hits.
Create DB and test (before and after tests running simultaneously):
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16
TEST_TMPDIR=/dev/shm base/db_bench -benchmarks=multireadrandom[-X30] -readonly -multiread_batched -batch_size=32 -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16
```
**Before**:
multireadrandom [AVG 30 runs] : 3444202 (± 57049) ops/sec; 240.9 (± 4.0) MB/sec
multireadrandom [MEDIAN 30 runs] : 3514443 ops/sec; 245.8 MB/sec
**After**:
multireadrandom [AVG 30 runs] : 3291022 (± 58851) ops/sec; 230.2 (± 4.1) MB/sec
multireadrandom [MEDIAN 30 runs] : 3366179 ops/sec; 235.4 MB/sec
So that's roughly a 3% regression, on kind of a *worst case* test of MultiGet CPU. Similar story with HyperClockCache:
**Before**:
multireadrandom [AVG 30 runs] : 3933777 (± 41840) ops/sec; 275.1 (± 2.9) MB/sec
multireadrandom [MEDIAN 30 runs] : 3970667 ops/sec; 277.7 MB/sec
**After**:
multireadrandom [AVG 30 runs] : 3755338 (± 30391) ops/sec; 262.6 (± 2.1) MB/sec
multireadrandom [MEDIAN 30 runs] : 3785696 ops/sec; 264.8 MB/sec
Roughly a 4-5% regression. Not ideal, but not the whole story, fortunately.
Let's also look at Get() in db_bench:
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X30] -readonly -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16
```
**Before**:
readrandom [AVG 30 runs] : 2198685 (± 13412) ops/sec; 153.8 (± 0.9) MB/sec
readrandom [MEDIAN 30 runs] : 2209498 ops/sec; 154.5 MB/sec
**After**:
readrandom [AVG 30 runs] : 2292814 (± 43508) ops/sec; 160.3 (± 3.0) MB/sec
readrandom [MEDIAN 30 runs] : 2365181 ops/sec; 165.4 MB/sec
That's showing roughly a 4% improvement, perhaps because of the secondary cache code that is no longer part of LRUCache. But weirdly, HyperClockCache is also showing 2-3% improvement:
**Before**:
readrandom [AVG 30 runs] : 2272333 (± 9992) ops/sec; 158.9 (± 0.7) MB/sec
readrandom [MEDIAN 30 runs] : 2273239 ops/sec; 159.0 MB/sec
**After**:
readrandom [AVG 30 runs] : 2332407 (± 11252) ops/sec; 163.1 (± 0.8) MB/sec
readrandom [MEDIAN 30 runs] : 2335329 ops/sec; 163.3 MB/sec
Reviewed By: ltamasi
Differential Revision: D44177044
Pulled By: pdillinger
fbshipit-source-id: e808e48ff3fe2f792a79841ba617be98e48689f5
2 years ago
|
|
|
Cache::Priority BlockBasedTable::GetCachePriority() const {
|
|
|
|
// Here we treat the legacy name "...index_and_filter_blocks..." to mean all
|
|
|
|
// metadata blocks that might go into block cache, EXCEPT only those needed
|
|
|
|
// for the read path (Get, etc.). TableProperties should not be needed on the
|
|
|
|
// read path (prefix extractor setting is an O(1) size special case that we
|
|
|
|
// are working not to require from TableProperties), so it is not given
|
|
|
|
// high-priority treatment if it should go into BlockCache.
|
HyperClockCache support for SecondaryCache, with refactoring (#11301)
Summary:
Internally refactors SecondaryCache integration out of LRUCache specifically and into a wrapper/adapter class that works with various Cache implementations. Notably, this relies on separating the notion of async lookup handles from other cache handles, so that HyperClockCache doesn't have to deal with the problem of allocating handles from the hash table for lookups that might fail anyway, and might be on the same key without support for coalescing. (LRUCache's hash table can incorporate previously allocated handles thanks to its pointer indirection.) Specifically, I'm worried about the case in which hundreds of threads try to access the same block and probing in the hash table degrades to linear search on the pile of entries with the same key.
This change is a big step in the direction of supporting stacked SecondaryCaches, but there are obstacles to completing that. Especially, there is no SecondaryCache hook for evictions to pass from one to the next. It has been proposed that evictions be transmitted simply as the persisted data (as in SaveToCallback), but given the current structure provided by the CacheItemHelpers, that would require an extra copy of the block data, because there's intentionally no way to ask for a contiguous Slice of the data (to allow for flexibility in storage). `AsyncLookupHandle` and the re-worked `WaitAll()` should be essentially prepared for stacked SecondaryCaches, but several "TODO with stacked secondaries" issues remain in various places.
It could be argued that the stacking instead be done as a SecondaryCache adapter that wraps two (or more) SecondaryCaches, but at least with the current API that would require an extra heap allocation on SecondaryCache Lookup for a wrapper SecondaryCacheResultHandle that can transfer a Lookup between secondaries. We could also consider trying to unify the Cache and SecondaryCache APIs, though that might be difficult if `AsyncLookupHandle` is kept a fixed struct.
## cache.h (public API)
Moves `secondary_cache` option from LRUCacheOptions to ShardedCacheOptions so that it is applicable to HyperClockCache.
## advanced_cache.h (advanced public API)
* Add `Cache::CreateStandalone()` so that the SecondaryCache support wrapper can use it.
* Add `SetEvictionCallback()` / `eviction_callback_` so that the SecondaryCache support wrapper can use it. Only a single callback is supported for efficiency. If there is ever a need for more than one, hopefully that can be handled with a broadcast callback wrapper.
These are essentially the two "extra" pieces of `Cache` for pulling out specific SecondaryCache support from the `Cache` implementation. I think it's a good trade-off as these are reasonable, limited, and reusable "cut points" into the `Cache` implementations.
* Remove async capability from standard `Lookup()` (getting rid of awkward restrictions on pending Handles) and add `AsyncLookupHandle` and `StartAsyncLookup()`. As noted in the comments, the full struct of `AsyncLookupHandle` is exposed so that it can be stack allocated, for efficiency, though more data is being copied around than before, which could impact performance. (Lookup info -> AsyncLookupHandle -> Handle vs. Lookup info -> Handle)
I could foresee a future in which a Cache internally saves a pointer to the AsyncLookupHandle, which means it's dangerous to allow it to be copyable or even movable. It also means it's not compatible with std::vector (which I don't like requiring as an API parameter anyway), so `WaitAll()` expects any contiguous array of AsyncLookupHandles. I believe this is best for common case efficiency, while behaving well in other cases also. For example, `WaitAll()` has no effect on default-constructed AsyncLookupHandles, which look like a completed cache miss.
## cacheable_entry.h
A couple of functions are obsolete because Cache::Handle can no longer be pending.
## cache.cc
Provides default implementations for new or revamped Cache functions, especially appropriate for non-blocking caches.
## secondary_cache_adapter.{h,cc}
The full details of the Cache wrapper adding SecondaryCache support. Essentially replicates the SecondaryCache handling that was in LRUCache, but obviously refactored. There is a bit of logic duplication, where Lookup() is essentially a manually optimized version of StartAsyncLookup() and Wait(), but it's roughly a dozen lines of code.
## sharded_cache.h, typed_cache.h, charged_cache.{h,cc}, sim_cache.cc
Simply updated for Cache API changes.
## lru_cache.{h,cc}
Carefully remove SecondaryCache logic, implement `CreateStandalone` and eviction handler functionality.
## clock_cache.{h,cc}
Expose existing `CreateStandalone` functionality, add eviction handler functionality. Light refactoring.
## block_based_table_reader*
Mostly re-worked the only usage of async Lookup, which is in BlockBasedTable::MultiGet. Used arrays in place of autovector in some places for efficiency. Simplified some logic by not trying to process some cache results before they're all ready.
Created new function `BlockBasedTable::GetCachePriority()` to reduce some pre-existing code duplication (and avoid making it worse).
Fixed at least one small bug from the prior confusing mixture of async and sync Lookups. In MaybeReadBlockAndLoadToCache(), called by RetrieveBlock(), called by MultiGet() with wait=false, is_cache_hit for the block_cache_tracer entry would not be set to true if the handle was pending after Lookup and before Wait.
## Intended follow-up work
* Figure out if there are any missing stats or block_cache_tracer work in refactored BlockBasedTable::MultiGet
* Stacked secondary caches (see above discussion)
* See if we can make up for the small MultiGet performance regression.
* Study more performance with SecondaryCache
* Items evicted from over-full LRUCache in Release were not being demoted to SecondaryCache, and still aren't to minimize unit test churn. Ideally they would be demoted, but it's an exceptional case so not a big deal.
* Use CreateStandalone for cache reservations (save unnecessary hash table operations). Not a big deal, but worthy cleanup.
* Somehow I got the contract for SecondaryCache::Insert wrong in #10945. (Doesn't take ownership!) That API comment needs to be fixed, but didn't want to mingle that in here.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11301
Test Plan:
## Unit tests
Generally updated to include HCC in SecondaryCache tests, though HyperClockCache has some different, less strict behaviors that leads to some tests not really being set up to work with it. Some of the tests remain disabled with it, but I think we have good coverage without them.
## Crash/stress test
Updated to use the new combination.
## Performance
First, let's check for regression on caches without secondary cache configured. Adding support for the eviction callback is likely to have a tiny effect, but it shouldn't be worrisome. LRUCache could benefit slightly from less logic around SecondaryCache handling. We can test with cache_bench default settings, built with DEBUG_LEVEL=0 and PORTABLE=0.
```
(while :; do base/cache_bench --cache_type=hyper_clock_cache | grep Rough; done) | awk '{ sum += $9; count++; print $0; print "Average: " int(sum / count) }'
```
**Before** this and #11299 (which could also have a small effect), running for about an hour, before & after running concurrently for each cache type:
HyperClockCache: 3168662 (average parallel ops/sec)
LRUCache: 2940127
**After** this and #11299, running for about an hour:
HyperClockCache: 3164862 (average parallel ops/sec) (0.12% slower)
LRUCache: 2940928 (0.03% faster)
This is an acceptable difference IMHO.
Next, let's consider essentially the worst case of new CPU overhead affecting overall performance. MultiGet uses the async lookup interface regardless of whether SecondaryCache or folly are used. We can configure a benchmark where all block cache queries are for data blocks, and all are hits.
Create DB and test (before and after tests running simultaneously):
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16
TEST_TMPDIR=/dev/shm base/db_bench -benchmarks=multireadrandom[-X30] -readonly -multiread_batched -batch_size=32 -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16
```
**Before**:
multireadrandom [AVG 30 runs] : 3444202 (± 57049) ops/sec; 240.9 (± 4.0) MB/sec
multireadrandom [MEDIAN 30 runs] : 3514443 ops/sec; 245.8 MB/sec
**After**:
multireadrandom [AVG 30 runs] : 3291022 (± 58851) ops/sec; 230.2 (± 4.1) MB/sec
multireadrandom [MEDIAN 30 runs] : 3366179 ops/sec; 235.4 MB/sec
So that's roughly a 3% regression, on kind of a *worst case* test of MultiGet CPU. Similar story with HyperClockCache:
**Before**:
multireadrandom [AVG 30 runs] : 3933777 (± 41840) ops/sec; 275.1 (± 2.9) MB/sec
multireadrandom [MEDIAN 30 runs] : 3970667 ops/sec; 277.7 MB/sec
**After**:
multireadrandom [AVG 30 runs] : 3755338 (± 30391) ops/sec; 262.6 (± 2.1) MB/sec
multireadrandom [MEDIAN 30 runs] : 3785696 ops/sec; 264.8 MB/sec
Roughly a 4-5% regression. Not ideal, but not the whole story, fortunately.
Let's also look at Get() in db_bench:
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X30] -readonly -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16
```
**Before**:
readrandom [AVG 30 runs] : 2198685 (± 13412) ops/sec; 153.8 (± 0.9) MB/sec
readrandom [MEDIAN 30 runs] : 2209498 ops/sec; 154.5 MB/sec
**After**:
readrandom [AVG 30 runs] : 2292814 (± 43508) ops/sec; 160.3 (± 3.0) MB/sec
readrandom [MEDIAN 30 runs] : 2365181 ops/sec; 165.4 MB/sec
That's showing roughly a 4% improvement, perhaps because of the secondary cache code that is no longer part of LRUCache. But weirdly, HyperClockCache is also showing 2-3% improvement:
**Before**:
readrandom [AVG 30 runs] : 2272333 (± 9992) ops/sec; 158.9 (± 0.7) MB/sec
readrandom [MEDIAN 30 runs] : 2273239 ops/sec; 159.0 MB/sec
**After**:
readrandom [AVG 30 runs] : 2332407 (± 11252) ops/sec; 163.1 (± 0.8) MB/sec
readrandom [MEDIAN 30 runs] : 2335329 ops/sec; 163.3 MB/sec
Reviewed By: ltamasi
Differential Revision: D44177044
Pulled By: pdillinger
fbshipit-source-id: e808e48ff3fe2f792a79841ba617be98e48689f5
2 years ago
|
|
|
if constexpr (TBlocklike::kBlockType == BlockType::kData ||
|
|
|
|
TBlocklike::kBlockType == BlockType::kProperties) {
|
|
|
|
return Cache::Priority::LOW;
|
|
|
|
} else if (rep_->table_options
|
|
|
|
.cache_index_and_filter_blocks_with_high_priority) {
|
|
|
|
return Cache::Priority::HIGH;
|
|
|
|
} else {
|
|
|
|
return Cache::Priority::LOW;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
template <typename TBlocklike>
|
|
|
|
WithBlocklikeCheck<Status, TBlocklike> BlockBasedTable::GetDataBlockFromCache(
|
|
|
|
const Slice& cache_key, BlockCacheInterface<TBlocklike> block_cache,
|
|
|
|
CachableEntry<TBlocklike>* out_parsed_block,
|
|
|
|
GetContext* get_context) const {
|
|
|
|
assert(out_parsed_block);
|
|
|
|
assert(out_parsed_block->IsEmpty());
|
|
|
|
|
|
|
|
Status s;
|
|
|
|
Statistics* statistics = rep_->ioptions.statistics.get();
|
|
|
|
|
|
|
|
// Lookup uncompressed cache first
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
if (block_cache) {
|
New stable, fixed-length cache keys (#9126)
Summary:
This change standardizes on a new 16-byte cache key format for
block cache (incl compressed and secondary) and persistent cache (but
not table cache and row cache).
The goal is a really fast cache key with practically ideal stability and
uniqueness properties without external dependencies (e.g. from FileSystem).
A fixed key size of 16 bytes should enable future optimizations to the
concurrent hash table for block cache, which is a heavy CPU user /
bottleneck, but there appears to be measurable performance improvement
even with no changes to LRUCache.
This change replaces a lot of disjointed and ugly code handling cache
keys with calls to a simple, clean new internal API (cache_key.h).
(Preserving the old cache key logic under an option would be very ugly
and likely negate the performance gain of the new approach. Complete
replacement carries some inherent risk, but I think that's acceptable
with sufficient analysis and testing.)
The scheme for encoding new cache keys is complicated but explained
in cache_key.cc.
Also: EndianSwapValue is moved to math.h to be next to other bit
operations. (Explains some new include "math.h".) ReverseBits operation
added and unit tests added to hash_test for both.
Fixes https://github.com/facebook/rocksdb/issues/7405 (presuming a root cause)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9126
Test Plan:
### Basic correctness
Several tests needed updates to work with the new functionality, mostly
because we are no longer relying on filesystem for stable cache keys
so table builders & readers need more context info to agree on cache
keys. This functionality is so core, a huge number of existing tests
exercise the cache key functionality.
### Performance
Create db with
`TEST_TMPDIR=/dev/shm ./db_bench -bloom_bits=10 -benchmarks=fillrandom -num=3000000 -partition_index_and_filters`
And test performance with
`TEST_TMPDIR=/dev/shm ./db_bench -readonly -use_existing_db -bloom_bits=10 -benchmarks=readrandom -num=3000000 -duration=30 -cache_index_and_filter_blocks -cache_size=250000 -threads=4`
using DEBUG_LEVEL=0 and simultaneous before & after runs.
Before ops/sec, avg over 100 runs: 121924
After ops/sec, avg over 100 runs: 125385 (+2.8%)
### Collision probability
I have built a tool, ./cache_bench -stress_cache_key to broadly simulate host-wide cache activity
over many months, by making some pessimistic simplifying assumptions:
* Every generated file has a cache entry for every byte offset in the file (contiguous range of cache keys)
* All of every file is cached for its entire lifetime
We use a simple table with skewed address assignment and replacement on address collision
to simulate files coming & going, with quite a variance (super-Poisson) in ages. Some output
with `./cache_bench -stress_cache_key -sck_keep_bits=40`:
```
Total cache or DBs size: 32TiB Writing 925.926 MiB/s or 76.2939TiB/day
Multiply by 9.22337e+18 to correct for simulation losses (but still assume whole file cached)
```
These come from default settings of 2.5M files per day of 32 MB each, and
`-sck_keep_bits=40` means that to represent a single file, we are only keeping 40 bits of
the 128-bit cache key. With file size of 2\*\*25 contiguous keys (pessimistic), our simulation
is about 2\*\*(128-40-25) or about 9 billion billion times more prone to collision than reality.
More default assumptions, relatively pessimistic:
* 100 DBs in same process (doesn't matter much)
* Re-open DB in same process (new session ID related to old session ID) on average
every 100 files generated
* Restart process (all new session IDs unrelated to old) 24 times per day
After enough data, we get a result at the end:
```
(keep 40 bits) 17 collisions after 2 x 90 days, est 10.5882 days between (9.76592e+19 corrected)
```
If we believe the (pessimistic) simulation and the mathematical generalization, we would need to run a billion machines all for 97 billion days to expect a cache key collision. To help verify that our generalization ("corrected") is robust, we can make our simulation more precise with `-sck_keep_bits=41` and `42`, which takes more running time to get enough data:
```
(keep 41 bits) 16 collisions after 4 x 90 days, est 22.5 days between (1.03763e+20 corrected)
(keep 42 bits) 19 collisions after 10 x 90 days, est 47.3684 days between (1.09224e+20 corrected)
```
The generalized prediction still holds. With the `-sck_randomize` option, we can see that we are beating "random" cache keys (except offsets still non-randomized) by a modest amount (roughly 20x less collision prone than random), which should make us reasonably comfortable even in "degenerate" cases:
```
197 collisions after 1 x 90 days, est 0.456853 days between (4.21372e+18 corrected)
```
I've run other tests to validate other conditions behave as expected, never behaving "worse than random" unless we start chopping off structured data.
Reviewed By: zhichao-cao
Differential Revision: D33171746
Pulled By: pdillinger
fbshipit-source-id: f16a57e369ed37be5e7e33525ace848d0537c88f
3 years ago
|
|
|
assert(!cache_key.empty());
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
auto cache_handle = block_cache.LookupFull(
|
HyperClockCache support for SecondaryCache, with refactoring (#11301)
Summary:
Internally refactors SecondaryCache integration out of LRUCache specifically and into a wrapper/adapter class that works with various Cache implementations. Notably, this relies on separating the notion of async lookup handles from other cache handles, so that HyperClockCache doesn't have to deal with the problem of allocating handles from the hash table for lookups that might fail anyway, and might be on the same key without support for coalescing. (LRUCache's hash table can incorporate previously allocated handles thanks to its pointer indirection.) Specifically, I'm worried about the case in which hundreds of threads try to access the same block and probing in the hash table degrades to linear search on the pile of entries with the same key.
This change is a big step in the direction of supporting stacked SecondaryCaches, but there are obstacles to completing that. Especially, there is no SecondaryCache hook for evictions to pass from one to the next. It has been proposed that evictions be transmitted simply as the persisted data (as in SaveToCallback), but given the current structure provided by the CacheItemHelpers, that would require an extra copy of the block data, because there's intentionally no way to ask for a contiguous Slice of the data (to allow for flexibility in storage). `AsyncLookupHandle` and the re-worked `WaitAll()` should be essentially prepared for stacked SecondaryCaches, but several "TODO with stacked secondaries" issues remain in various places.
It could be argued that the stacking instead be done as a SecondaryCache adapter that wraps two (or more) SecondaryCaches, but at least with the current API that would require an extra heap allocation on SecondaryCache Lookup for a wrapper SecondaryCacheResultHandle that can transfer a Lookup between secondaries. We could also consider trying to unify the Cache and SecondaryCache APIs, though that might be difficult if `AsyncLookupHandle` is kept a fixed struct.
## cache.h (public API)
Moves `secondary_cache` option from LRUCacheOptions to ShardedCacheOptions so that it is applicable to HyperClockCache.
## advanced_cache.h (advanced public API)
* Add `Cache::CreateStandalone()` so that the SecondaryCache support wrapper can use it.
* Add `SetEvictionCallback()` / `eviction_callback_` so that the SecondaryCache support wrapper can use it. Only a single callback is supported for efficiency. If there is ever a need for more than one, hopefully that can be handled with a broadcast callback wrapper.
These are essentially the two "extra" pieces of `Cache` for pulling out specific SecondaryCache support from the `Cache` implementation. I think it's a good trade-off as these are reasonable, limited, and reusable "cut points" into the `Cache` implementations.
* Remove async capability from standard `Lookup()` (getting rid of awkward restrictions on pending Handles) and add `AsyncLookupHandle` and `StartAsyncLookup()`. As noted in the comments, the full struct of `AsyncLookupHandle` is exposed so that it can be stack allocated, for efficiency, though more data is being copied around than before, which could impact performance. (Lookup info -> AsyncLookupHandle -> Handle vs. Lookup info -> Handle)
I could foresee a future in which a Cache internally saves a pointer to the AsyncLookupHandle, which means it's dangerous to allow it to be copyable or even movable. It also means it's not compatible with std::vector (which I don't like requiring as an API parameter anyway), so `WaitAll()` expects any contiguous array of AsyncLookupHandles. I believe this is best for common case efficiency, while behaving well in other cases also. For example, `WaitAll()` has no effect on default-constructed AsyncLookupHandles, which look like a completed cache miss.
## cacheable_entry.h
A couple of functions are obsolete because Cache::Handle can no longer be pending.
## cache.cc
Provides default implementations for new or revamped Cache functions, especially appropriate for non-blocking caches.
## secondary_cache_adapter.{h,cc}
The full details of the Cache wrapper adding SecondaryCache support. Essentially replicates the SecondaryCache handling that was in LRUCache, but obviously refactored. There is a bit of logic duplication, where Lookup() is essentially a manually optimized version of StartAsyncLookup() and Wait(), but it's roughly a dozen lines of code.
## sharded_cache.h, typed_cache.h, charged_cache.{h,cc}, sim_cache.cc
Simply updated for Cache API changes.
## lru_cache.{h,cc}
Carefully remove SecondaryCache logic, implement `CreateStandalone` and eviction handler functionality.
## clock_cache.{h,cc}
Expose existing `CreateStandalone` functionality, add eviction handler functionality. Light refactoring.
## block_based_table_reader*
Mostly re-worked the only usage of async Lookup, which is in BlockBasedTable::MultiGet. Used arrays in place of autovector in some places for efficiency. Simplified some logic by not trying to process some cache results before they're all ready.
Created new function `BlockBasedTable::GetCachePriority()` to reduce some pre-existing code duplication (and avoid making it worse).
Fixed at least one small bug from the prior confusing mixture of async and sync Lookups. In MaybeReadBlockAndLoadToCache(), called by RetrieveBlock(), called by MultiGet() with wait=false, is_cache_hit for the block_cache_tracer entry would not be set to true if the handle was pending after Lookup and before Wait.
## Intended follow-up work
* Figure out if there are any missing stats or block_cache_tracer work in refactored BlockBasedTable::MultiGet
* Stacked secondary caches (see above discussion)
* See if we can make up for the small MultiGet performance regression.
* Study more performance with SecondaryCache
* Items evicted from over-full LRUCache in Release were not being demoted to SecondaryCache, and still aren't to minimize unit test churn. Ideally they would be demoted, but it's an exceptional case so not a big deal.
* Use CreateStandalone for cache reservations (save unnecessary hash table operations). Not a big deal, but worthy cleanup.
* Somehow I got the contract for SecondaryCache::Insert wrong in #10945. (Doesn't take ownership!) That API comment needs to be fixed, but didn't want to mingle that in here.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11301
Test Plan:
## Unit tests
Generally updated to include HCC in SecondaryCache tests, though HyperClockCache has some different, less strict behaviors that leads to some tests not really being set up to work with it. Some of the tests remain disabled with it, but I think we have good coverage without them.
## Crash/stress test
Updated to use the new combination.
## Performance
First, let's check for regression on caches without secondary cache configured. Adding support for the eviction callback is likely to have a tiny effect, but it shouldn't be worrisome. LRUCache could benefit slightly from less logic around SecondaryCache handling. We can test with cache_bench default settings, built with DEBUG_LEVEL=0 and PORTABLE=0.
```
(while :; do base/cache_bench --cache_type=hyper_clock_cache | grep Rough; done) | awk '{ sum += $9; count++; print $0; print "Average: " int(sum / count) }'
```
**Before** this and #11299 (which could also have a small effect), running for about an hour, before & after running concurrently for each cache type:
HyperClockCache: 3168662 (average parallel ops/sec)
LRUCache: 2940127
**After** this and #11299, running for about an hour:
HyperClockCache: 3164862 (average parallel ops/sec) (0.12% slower)
LRUCache: 2940928 (0.03% faster)
This is an acceptable difference IMHO.
Next, let's consider essentially the worst case of new CPU overhead affecting overall performance. MultiGet uses the async lookup interface regardless of whether SecondaryCache or folly are used. We can configure a benchmark where all block cache queries are for data blocks, and all are hits.
Create DB and test (before and after tests running simultaneously):
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16
TEST_TMPDIR=/dev/shm base/db_bench -benchmarks=multireadrandom[-X30] -readonly -multiread_batched -batch_size=32 -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16
```
**Before**:
multireadrandom [AVG 30 runs] : 3444202 (± 57049) ops/sec; 240.9 (± 4.0) MB/sec
multireadrandom [MEDIAN 30 runs] : 3514443 ops/sec; 245.8 MB/sec
**After**:
multireadrandom [AVG 30 runs] : 3291022 (± 58851) ops/sec; 230.2 (± 4.1) MB/sec
multireadrandom [MEDIAN 30 runs] : 3366179 ops/sec; 235.4 MB/sec
So that's roughly a 3% regression, on kind of a *worst case* test of MultiGet CPU. Similar story with HyperClockCache:
**Before**:
multireadrandom [AVG 30 runs] : 3933777 (± 41840) ops/sec; 275.1 (± 2.9) MB/sec
multireadrandom [MEDIAN 30 runs] : 3970667 ops/sec; 277.7 MB/sec
**After**:
multireadrandom [AVG 30 runs] : 3755338 (± 30391) ops/sec; 262.6 (± 2.1) MB/sec
multireadrandom [MEDIAN 30 runs] : 3785696 ops/sec; 264.8 MB/sec
Roughly a 4-5% regression. Not ideal, but not the whole story, fortunately.
Let's also look at Get() in db_bench:
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X30] -readonly -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16
```
**Before**:
readrandom [AVG 30 runs] : 2198685 (± 13412) ops/sec; 153.8 (± 0.9) MB/sec
readrandom [MEDIAN 30 runs] : 2209498 ops/sec; 154.5 MB/sec
**After**:
readrandom [AVG 30 runs] : 2292814 (± 43508) ops/sec; 160.3 (± 3.0) MB/sec
readrandom [MEDIAN 30 runs] : 2365181 ops/sec; 165.4 MB/sec
That's showing roughly a 4% improvement, perhaps because of the secondary cache code that is no longer part of LRUCache. But weirdly, HyperClockCache is also showing 2-3% improvement:
**Before**:
readrandom [AVG 30 runs] : 2272333 (± 9992) ops/sec; 158.9 (± 0.7) MB/sec
readrandom [MEDIAN 30 runs] : 2273239 ops/sec; 159.0 MB/sec
**After**:
readrandom [AVG 30 runs] : 2332407 (± 11252) ops/sec; 163.1 (± 0.8) MB/sec
readrandom [MEDIAN 30 runs] : 2335329 ops/sec; 163.3 MB/sec
Reviewed By: ltamasi
Differential Revision: D44177044
Pulled By: pdillinger
fbshipit-source-id: e808e48ff3fe2f792a79841ba617be98e48689f5
2 years ago
|
|
|
cache_key, &rep_->create_context, GetCachePriority<TBlocklike>(),
|
|
|
|
statistics, rep_->ioptions.lowest_used_cache_tier);
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
|
|
|
|
// Avoid updating metrics here if the handle is not complete yet. This
|
|
|
|
// happens with MultiGet and secondary cache. So update the metrics only
|
|
|
|
// if its a miss, or a hit and value is ready
|
|
|
|
if (!cache_handle) {
|
|
|
|
UpdateCacheMissMetrics(TBlocklike::kBlockType, get_context);
|
|
|
|
} else {
|
|
|
|
TBlocklike* value = block_cache.Value(cache_handle);
|
|
|
|
if (value) {
|
|
|
|
UpdateCacheHitMetrics(TBlocklike::kBlockType, get_context,
|
|
|
|
block_cache.get()->GetUsage(cache_handle));
|
|
|
|
}
|
|
|
|
out_parsed_block->SetCachedValue(value, block_cache.get(), cache_handle);
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// If not found, search from the compressed block cache.
|
Refactor to avoid confusing "raw block" (#10408)
Summary:
We have a lot of confusing code because of mixed, sometimes
completely opposite uses of of the term "raw block" or "raw contents",
sometimes within the same source file. For example, in `BlockBasedTableBuilder`,
`raw_block_contents` and `raw_size` generally referred to uncompressed block
contents and size, while `WriteRawBlock` referred to writing a block that
is already compressed if it is going to be. Meanwhile, in
`BlockBasedTable`, `raw_block_contents` either referred to a (maybe
compressed) block with trailer, or a maybe compressed block maybe
without trailer. (Note: left as follow-up work to use C++ typing to
better sort out the various kinds of BlockContents.)
This change primarily tries to apply some consistent terminology around
the kinds of block representations, avoiding the unclear "raw". (Any
meaning of "raw" assumes some bias toward the storage layer or toward
the logical data layer.) Preferred terminology:
* **Serialized block** - bytes that go into storage. For block-based table
(usually the case) this includes the block trailer. WART: block `size` may or
may not include the trailer; need to be clear about whether it does or not.
* **Maybe compressed block** - like a serialized block, but without the
trailer (or no promise of including a trailer). Must be accompanied by a
CompressionType.
* **Uncompressed block** - "payload" bytes that are either stored with no
compression, used as input to compression function, or result of
decompression function.
* **Parsed block** - an in-memory form of a block in block cache, as it is
used by the table reader. Different C++ types are used depending on the
block type (see block_like_traits.h).
Other refactorings:
* Misc corrections/improvements of internal API comments
* Remove a few misleading / unhelpful / redundant comments.
* Use move semantics in some places to simplify contracts
* Use better parameter names to indicate which parameters are used for
outputs
* Remove some extraneous `extern`
* Various clean-ups to `CacheDumperImpl` (mostly unnecessary code)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10408
Test Plan: existing tests
Reviewed By: akankshamahajan15
Differential Revision: D38172617
Pulled By: pdillinger
fbshipit-source-id: ccb99299f324ac5ca46996d34c5089621a4f260c
2 years ago
|
|
|
assert(out_parsed_block->IsEmpty());
|
|
|
|
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
template <typename TBlocklike>
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
WithBlocklikeCheck<Status, TBlocklike> BlockBasedTable::PutDataBlockToCache(
|
|
|
|
const Slice& cache_key, BlockCacheInterface<TBlocklike> block_cache,
|
Refactor to avoid confusing "raw block" (#10408)
Summary:
We have a lot of confusing code because of mixed, sometimes
completely opposite uses of of the term "raw block" or "raw contents",
sometimes within the same source file. For example, in `BlockBasedTableBuilder`,
`raw_block_contents` and `raw_size` generally referred to uncompressed block
contents and size, while `WriteRawBlock` referred to writing a block that
is already compressed if it is going to be. Meanwhile, in
`BlockBasedTable`, `raw_block_contents` either referred to a (maybe
compressed) block with trailer, or a maybe compressed block maybe
without trailer. (Note: left as follow-up work to use C++ typing to
better sort out the various kinds of BlockContents.)
This change primarily tries to apply some consistent terminology around
the kinds of block representations, avoiding the unclear "raw". (Any
meaning of "raw" assumes some bias toward the storage layer or toward
the logical data layer.) Preferred terminology:
* **Serialized block** - bytes that go into storage. For block-based table
(usually the case) this includes the block trailer. WART: block `size` may or
may not include the trailer; need to be clear about whether it does or not.
* **Maybe compressed block** - like a serialized block, but without the
trailer (or no promise of including a trailer). Must be accompanied by a
CompressionType.
* **Uncompressed block** - "payload" bytes that are either stored with no
compression, used as input to compression function, or result of
decompression function.
* **Parsed block** - an in-memory form of a block in block cache, as it is
used by the table reader. Different C++ types are used depending on the
block type (see block_like_traits.h).
Other refactorings:
* Misc corrections/improvements of internal API comments
* Remove a few misleading / unhelpful / redundant comments.
* Use move semantics in some places to simplify contracts
* Use better parameter names to indicate which parameters are used for
outputs
* Remove some extraneous `extern`
* Various clean-ups to `CacheDumperImpl` (mostly unnecessary code)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10408
Test Plan: existing tests
Reviewed By: akankshamahajan15
Differential Revision: D38172617
Pulled By: pdillinger
fbshipit-source-id: ccb99299f324ac5ca46996d34c5089621a4f260c
2 years ago
|
|
|
CachableEntry<TBlocklike>* out_parsed_block, BlockContents&& block_contents,
|
|
|
|
CompressionType block_comp_type,
|
|
|
|
const UncompressionDict& uncompression_dict,
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
MemoryAllocator* memory_allocator, GetContext* get_context) const {
|
|
|
|
const ImmutableOptions& ioptions = rep_->ioptions;
|
|
|
|
const uint32_t format_version = rep_->table_options.format_version;
|
Refactor to avoid confusing "raw block" (#10408)
Summary:
We have a lot of confusing code because of mixed, sometimes
completely opposite uses of of the term "raw block" or "raw contents",
sometimes within the same source file. For example, in `BlockBasedTableBuilder`,
`raw_block_contents` and `raw_size` generally referred to uncompressed block
contents and size, while `WriteRawBlock` referred to writing a block that
is already compressed if it is going to be. Meanwhile, in
`BlockBasedTable`, `raw_block_contents` either referred to a (maybe
compressed) block with trailer, or a maybe compressed block maybe
without trailer. (Note: left as follow-up work to use C++ typing to
better sort out the various kinds of BlockContents.)
This change primarily tries to apply some consistent terminology around
the kinds of block representations, avoiding the unclear "raw". (Any
meaning of "raw" assumes some bias toward the storage layer or toward
the logical data layer.) Preferred terminology:
* **Serialized block** - bytes that go into storage. For block-based table
(usually the case) this includes the block trailer. WART: block `size` may or
may not include the trailer; need to be clear about whether it does or not.
* **Maybe compressed block** - like a serialized block, but without the
trailer (or no promise of including a trailer). Must be accompanied by a
CompressionType.
* **Uncompressed block** - "payload" bytes that are either stored with no
compression, used as input to compression function, or result of
decompression function.
* **Parsed block** - an in-memory form of a block in block cache, as it is
used by the table reader. Different C++ types are used depending on the
block type (see block_like_traits.h).
Other refactorings:
* Misc corrections/improvements of internal API comments
* Remove a few misleading / unhelpful / redundant comments.
* Use move semantics in some places to simplify contracts
* Use better parameter names to indicate which parameters are used for
outputs
* Remove some extraneous `extern`
* Various clean-ups to `CacheDumperImpl` (mostly unnecessary code)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10408
Test Plan: existing tests
Reviewed By: akankshamahajan15
Differential Revision: D38172617
Pulled By: pdillinger
fbshipit-source-id: ccb99299f324ac5ca46996d34c5089621a4f260c
2 years ago
|
|
|
assert(out_parsed_block);
|
|
|
|
assert(out_parsed_block->IsEmpty());
|
|
|
|
|
|
|
|
Status s;
|
|
|
|
Statistics* statistics = ioptions.stats;
|
|
|
|
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
std::unique_ptr<TBlocklike> block_holder;
|
Refactor to avoid confusing "raw block" (#10408)
Summary:
We have a lot of confusing code because of mixed, sometimes
completely opposite uses of of the term "raw block" or "raw contents",
sometimes within the same source file. For example, in `BlockBasedTableBuilder`,
`raw_block_contents` and `raw_size` generally referred to uncompressed block
contents and size, while `WriteRawBlock` referred to writing a block that
is already compressed if it is going to be. Meanwhile, in
`BlockBasedTable`, `raw_block_contents` either referred to a (maybe
compressed) block with trailer, or a maybe compressed block maybe
without trailer. (Note: left as follow-up work to use C++ typing to
better sort out the various kinds of BlockContents.)
This change primarily tries to apply some consistent terminology around
the kinds of block representations, avoiding the unclear "raw". (Any
meaning of "raw" assumes some bias toward the storage layer or toward
the logical data layer.) Preferred terminology:
* **Serialized block** - bytes that go into storage. For block-based table
(usually the case) this includes the block trailer. WART: block `size` may or
may not include the trailer; need to be clear about whether it does or not.
* **Maybe compressed block** - like a serialized block, but without the
trailer (or no promise of including a trailer). Must be accompanied by a
CompressionType.
* **Uncompressed block** - "payload" bytes that are either stored with no
compression, used as input to compression function, or result of
decompression function.
* **Parsed block** - an in-memory form of a block in block cache, as it is
used by the table reader. Different C++ types are used depending on the
block type (see block_like_traits.h).
Other refactorings:
* Misc corrections/improvements of internal API comments
* Remove a few misleading / unhelpful / redundant comments.
* Use move semantics in some places to simplify contracts
* Use better parameter names to indicate which parameters are used for
outputs
* Remove some extraneous `extern`
* Various clean-ups to `CacheDumperImpl` (mostly unnecessary code)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10408
Test Plan: existing tests
Reviewed By: akankshamahajan15
Differential Revision: D38172617
Pulled By: pdillinger
fbshipit-source-id: ccb99299f324ac5ca46996d34c5089621a4f260c
2 years ago
|
|
|
if (block_comp_type != kNoCompression) {
|
|
|
|
// Retrieve the uncompressed contents into a new buffer
|
|
|
|
BlockContents uncompressed_block_contents;
|
Refactor to avoid confusing "raw block" (#10408)
Summary:
We have a lot of confusing code because of mixed, sometimes
completely opposite uses of of the term "raw block" or "raw contents",
sometimes within the same source file. For example, in `BlockBasedTableBuilder`,
`raw_block_contents` and `raw_size` generally referred to uncompressed block
contents and size, while `WriteRawBlock` referred to writing a block that
is already compressed if it is going to be. Meanwhile, in
`BlockBasedTable`, `raw_block_contents` either referred to a (maybe
compressed) block with trailer, or a maybe compressed block maybe
without trailer. (Note: left as follow-up work to use C++ typing to
better sort out the various kinds of BlockContents.)
This change primarily tries to apply some consistent terminology around
the kinds of block representations, avoiding the unclear "raw". (Any
meaning of "raw" assumes some bias toward the storage layer or toward
the logical data layer.) Preferred terminology:
* **Serialized block** - bytes that go into storage. For block-based table
(usually the case) this includes the block trailer. WART: block `size` may or
may not include the trailer; need to be clear about whether it does or not.
* **Maybe compressed block** - like a serialized block, but without the
trailer (or no promise of including a trailer). Must be accompanied by a
CompressionType.
* **Uncompressed block** - "payload" bytes that are either stored with no
compression, used as input to compression function, or result of
decompression function.
* **Parsed block** - an in-memory form of a block in block cache, as it is
used by the table reader. Different C++ types are used depending on the
block type (see block_like_traits.h).
Other refactorings:
* Misc corrections/improvements of internal API comments
* Remove a few misleading / unhelpful / redundant comments.
* Use move semantics in some places to simplify contracts
* Use better parameter names to indicate which parameters are used for
outputs
* Remove some extraneous `extern`
* Various clean-ups to `CacheDumperImpl` (mostly unnecessary code)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10408
Test Plan: existing tests
Reviewed By: akankshamahajan15
Differential Revision: D38172617
Pulled By: pdillinger
fbshipit-source-id: ccb99299f324ac5ca46996d34c5089621a4f260c
2 years ago
|
|
|
UncompressionContext context(block_comp_type);
|
|
|
|
UncompressionInfo info(context, uncompression_dict, block_comp_type);
|
|
|
|
s = UncompressBlockData(info, block_contents.data.data(),
|
|
|
|
block_contents.data.size(),
|
|
|
|
&uncompressed_block_contents, format_version,
|
|
|
|
ioptions, memory_allocator);
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
rep_->create_context.Create(&block_holder,
|
|
|
|
std::move(uncompressed_block_contents));
|
|
|
|
} else {
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
rep_->create_context.Create(&block_holder, std::move(block_contents));
|
|
|
|
}
|
|
|
|
|
|
|
|
// insert into uncompressed block cache
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
if (block_cache && block_holder->own_bytes()) {
|
|
|
|
size_t charge = block_holder->ApproximateMemoryUsage();
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
BlockCacheTypedHandle<TBlocklike>* cache_handle = nullptr;
|
|
|
|
s = block_cache.InsertFull(cache_key, block_holder.get(), charge,
|
HyperClockCache support for SecondaryCache, with refactoring (#11301)
Summary:
Internally refactors SecondaryCache integration out of LRUCache specifically and into a wrapper/adapter class that works with various Cache implementations. Notably, this relies on separating the notion of async lookup handles from other cache handles, so that HyperClockCache doesn't have to deal with the problem of allocating handles from the hash table for lookups that might fail anyway, and might be on the same key without support for coalescing. (LRUCache's hash table can incorporate previously allocated handles thanks to its pointer indirection.) Specifically, I'm worried about the case in which hundreds of threads try to access the same block and probing in the hash table degrades to linear search on the pile of entries with the same key.
This change is a big step in the direction of supporting stacked SecondaryCaches, but there are obstacles to completing that. Especially, there is no SecondaryCache hook for evictions to pass from one to the next. It has been proposed that evictions be transmitted simply as the persisted data (as in SaveToCallback), but given the current structure provided by the CacheItemHelpers, that would require an extra copy of the block data, because there's intentionally no way to ask for a contiguous Slice of the data (to allow for flexibility in storage). `AsyncLookupHandle` and the re-worked `WaitAll()` should be essentially prepared for stacked SecondaryCaches, but several "TODO with stacked secondaries" issues remain in various places.
It could be argued that the stacking instead be done as a SecondaryCache adapter that wraps two (or more) SecondaryCaches, but at least with the current API that would require an extra heap allocation on SecondaryCache Lookup for a wrapper SecondaryCacheResultHandle that can transfer a Lookup between secondaries. We could also consider trying to unify the Cache and SecondaryCache APIs, though that might be difficult if `AsyncLookupHandle` is kept a fixed struct.
## cache.h (public API)
Moves `secondary_cache` option from LRUCacheOptions to ShardedCacheOptions so that it is applicable to HyperClockCache.
## advanced_cache.h (advanced public API)
* Add `Cache::CreateStandalone()` so that the SecondaryCache support wrapper can use it.
* Add `SetEvictionCallback()` / `eviction_callback_` so that the SecondaryCache support wrapper can use it. Only a single callback is supported for efficiency. If there is ever a need for more than one, hopefully that can be handled with a broadcast callback wrapper.
These are essentially the two "extra" pieces of `Cache` for pulling out specific SecondaryCache support from the `Cache` implementation. I think it's a good trade-off as these are reasonable, limited, and reusable "cut points" into the `Cache` implementations.
* Remove async capability from standard `Lookup()` (getting rid of awkward restrictions on pending Handles) and add `AsyncLookupHandle` and `StartAsyncLookup()`. As noted in the comments, the full struct of `AsyncLookupHandle` is exposed so that it can be stack allocated, for efficiency, though more data is being copied around than before, which could impact performance. (Lookup info -> AsyncLookupHandle -> Handle vs. Lookup info -> Handle)
I could foresee a future in which a Cache internally saves a pointer to the AsyncLookupHandle, which means it's dangerous to allow it to be copyable or even movable. It also means it's not compatible with std::vector (which I don't like requiring as an API parameter anyway), so `WaitAll()` expects any contiguous array of AsyncLookupHandles. I believe this is best for common case efficiency, while behaving well in other cases also. For example, `WaitAll()` has no effect on default-constructed AsyncLookupHandles, which look like a completed cache miss.
## cacheable_entry.h
A couple of functions are obsolete because Cache::Handle can no longer be pending.
## cache.cc
Provides default implementations for new or revamped Cache functions, especially appropriate for non-blocking caches.
## secondary_cache_adapter.{h,cc}
The full details of the Cache wrapper adding SecondaryCache support. Essentially replicates the SecondaryCache handling that was in LRUCache, but obviously refactored. There is a bit of logic duplication, where Lookup() is essentially a manually optimized version of StartAsyncLookup() and Wait(), but it's roughly a dozen lines of code.
## sharded_cache.h, typed_cache.h, charged_cache.{h,cc}, sim_cache.cc
Simply updated for Cache API changes.
## lru_cache.{h,cc}
Carefully remove SecondaryCache logic, implement `CreateStandalone` and eviction handler functionality.
## clock_cache.{h,cc}
Expose existing `CreateStandalone` functionality, add eviction handler functionality. Light refactoring.
## block_based_table_reader*
Mostly re-worked the only usage of async Lookup, which is in BlockBasedTable::MultiGet. Used arrays in place of autovector in some places for efficiency. Simplified some logic by not trying to process some cache results before they're all ready.
Created new function `BlockBasedTable::GetCachePriority()` to reduce some pre-existing code duplication (and avoid making it worse).
Fixed at least one small bug from the prior confusing mixture of async and sync Lookups. In MaybeReadBlockAndLoadToCache(), called by RetrieveBlock(), called by MultiGet() with wait=false, is_cache_hit for the block_cache_tracer entry would not be set to true if the handle was pending after Lookup and before Wait.
## Intended follow-up work
* Figure out if there are any missing stats or block_cache_tracer work in refactored BlockBasedTable::MultiGet
* Stacked secondary caches (see above discussion)
* See if we can make up for the small MultiGet performance regression.
* Study more performance with SecondaryCache
* Items evicted from over-full LRUCache in Release were not being demoted to SecondaryCache, and still aren't to minimize unit test churn. Ideally they would be demoted, but it's an exceptional case so not a big deal.
* Use CreateStandalone for cache reservations (save unnecessary hash table operations). Not a big deal, but worthy cleanup.
* Somehow I got the contract for SecondaryCache::Insert wrong in #10945. (Doesn't take ownership!) That API comment needs to be fixed, but didn't want to mingle that in here.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11301
Test Plan:
## Unit tests
Generally updated to include HCC in SecondaryCache tests, though HyperClockCache has some different, less strict behaviors that leads to some tests not really being set up to work with it. Some of the tests remain disabled with it, but I think we have good coverage without them.
## Crash/stress test
Updated to use the new combination.
## Performance
First, let's check for regression on caches without secondary cache configured. Adding support for the eviction callback is likely to have a tiny effect, but it shouldn't be worrisome. LRUCache could benefit slightly from less logic around SecondaryCache handling. We can test with cache_bench default settings, built with DEBUG_LEVEL=0 and PORTABLE=0.
```
(while :; do base/cache_bench --cache_type=hyper_clock_cache | grep Rough; done) | awk '{ sum += $9; count++; print $0; print "Average: " int(sum / count) }'
```
**Before** this and #11299 (which could also have a small effect), running for about an hour, before & after running concurrently for each cache type:
HyperClockCache: 3168662 (average parallel ops/sec)
LRUCache: 2940127
**After** this and #11299, running for about an hour:
HyperClockCache: 3164862 (average parallel ops/sec) (0.12% slower)
LRUCache: 2940928 (0.03% faster)
This is an acceptable difference IMHO.
Next, let's consider essentially the worst case of new CPU overhead affecting overall performance. MultiGet uses the async lookup interface regardless of whether SecondaryCache or folly are used. We can configure a benchmark where all block cache queries are for data blocks, and all are hits.
Create DB and test (before and after tests running simultaneously):
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16
TEST_TMPDIR=/dev/shm base/db_bench -benchmarks=multireadrandom[-X30] -readonly -multiread_batched -batch_size=32 -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16
```
**Before**:
multireadrandom [AVG 30 runs] : 3444202 (± 57049) ops/sec; 240.9 (± 4.0) MB/sec
multireadrandom [MEDIAN 30 runs] : 3514443 ops/sec; 245.8 MB/sec
**After**:
multireadrandom [AVG 30 runs] : 3291022 (± 58851) ops/sec; 230.2 (± 4.1) MB/sec
multireadrandom [MEDIAN 30 runs] : 3366179 ops/sec; 235.4 MB/sec
So that's roughly a 3% regression, on kind of a *worst case* test of MultiGet CPU. Similar story with HyperClockCache:
**Before**:
multireadrandom [AVG 30 runs] : 3933777 (± 41840) ops/sec; 275.1 (± 2.9) MB/sec
multireadrandom [MEDIAN 30 runs] : 3970667 ops/sec; 277.7 MB/sec
**After**:
multireadrandom [AVG 30 runs] : 3755338 (± 30391) ops/sec; 262.6 (± 2.1) MB/sec
multireadrandom [MEDIAN 30 runs] : 3785696 ops/sec; 264.8 MB/sec
Roughly a 4-5% regression. Not ideal, but not the whole story, fortunately.
Let's also look at Get() in db_bench:
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X30] -readonly -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16
```
**Before**:
readrandom [AVG 30 runs] : 2198685 (± 13412) ops/sec; 153.8 (± 0.9) MB/sec
readrandom [MEDIAN 30 runs] : 2209498 ops/sec; 154.5 MB/sec
**After**:
readrandom [AVG 30 runs] : 2292814 (± 43508) ops/sec; 160.3 (± 3.0) MB/sec
readrandom [MEDIAN 30 runs] : 2365181 ops/sec; 165.4 MB/sec
That's showing roughly a 4% improvement, perhaps because of the secondary cache code that is no longer part of LRUCache. But weirdly, HyperClockCache is also showing 2-3% improvement:
**Before**:
readrandom [AVG 30 runs] : 2272333 (± 9992) ops/sec; 158.9 (± 0.7) MB/sec
readrandom [MEDIAN 30 runs] : 2273239 ops/sec; 159.0 MB/sec
**After**:
readrandom [AVG 30 runs] : 2332407 (± 11252) ops/sec; 163.1 (± 0.8) MB/sec
readrandom [MEDIAN 30 runs] : 2335329 ops/sec; 163.3 MB/sec
Reviewed By: ltamasi
Differential Revision: D44177044
Pulled By: pdillinger
fbshipit-source-id: e808e48ff3fe2f792a79841ba617be98e48689f5
2 years ago
|
|
|
&cache_handle, GetCachePriority<TBlocklike>(),
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
rep_->ioptions.lowest_used_cache_tier);
|
|
|
|
|
|
|
|
if (s.ok()) {
|
|
|
|
assert(cache_handle != nullptr);
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
out_parsed_block->SetCachedValue(block_holder.release(),
|
|
|
|
block_cache.get(), cache_handle);
|
|
|
|
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
UpdateCacheInsertionMetrics(TBlocklike::kBlockType, get_context, charge,
|
|
|
|
s.IsOkOverwritten(), rep_->ioptions.stats);
|
|
|
|
} else {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_ADD_FAILURES);
|
|
|
|
}
|
|
|
|
} else {
|
Refactor to avoid confusing "raw block" (#10408)
Summary:
We have a lot of confusing code because of mixed, sometimes
completely opposite uses of of the term "raw block" or "raw contents",
sometimes within the same source file. For example, in `BlockBasedTableBuilder`,
`raw_block_contents` and `raw_size` generally referred to uncompressed block
contents and size, while `WriteRawBlock` referred to writing a block that
is already compressed if it is going to be. Meanwhile, in
`BlockBasedTable`, `raw_block_contents` either referred to a (maybe
compressed) block with trailer, or a maybe compressed block maybe
without trailer. (Note: left as follow-up work to use C++ typing to
better sort out the various kinds of BlockContents.)
This change primarily tries to apply some consistent terminology around
the kinds of block representations, avoiding the unclear "raw". (Any
meaning of "raw" assumes some bias toward the storage layer or toward
the logical data layer.) Preferred terminology:
* **Serialized block** - bytes that go into storage. For block-based table
(usually the case) this includes the block trailer. WART: block `size` may or
may not include the trailer; need to be clear about whether it does or not.
* **Maybe compressed block** - like a serialized block, but without the
trailer (or no promise of including a trailer). Must be accompanied by a
CompressionType.
* **Uncompressed block** - "payload" bytes that are either stored with no
compression, used as input to compression function, or result of
decompression function.
* **Parsed block** - an in-memory form of a block in block cache, as it is
used by the table reader. Different C++ types are used depending on the
block type (see block_like_traits.h).
Other refactorings:
* Misc corrections/improvements of internal API comments
* Remove a few misleading / unhelpful / redundant comments.
* Use move semantics in some places to simplify contracts
* Use better parameter names to indicate which parameters are used for
outputs
* Remove some extraneous `extern`
* Various clean-ups to `CacheDumperImpl` (mostly unnecessary code)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10408
Test Plan: existing tests
Reviewed By: akankshamahajan15
Differential Revision: D38172617
Pulled By: pdillinger
fbshipit-source-id: ccb99299f324ac5ca46996d34c5089621a4f260c
2 years ago
|
|
|
out_parsed_block->SetOwnedValue(std::move(block_holder));
|
|
|
|
}
|
|
|
|
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
std::unique_ptr<FilterBlockReader> BlockBasedTable::CreateFilterBlockReader(
|
|
|
|
const ReadOptions& ro, FilePrefetchBuffer* prefetch_buffer, bool use_cache,
|
|
|
|
bool prefetch, bool pin, BlockCacheLookupContext* lookup_context) {
|
|
|
|
auto& rep = rep_;
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
auto filter_type = rep->filter_type;
|
|
|
|
if (filter_type == Rep::FilterType::kNoFilter) {
|
|
|
|
return std::unique_ptr<FilterBlockReader>();
|
|
|
|
}
|
|
|
|
|
|
|
|
assert(rep->filter_policy);
|
|
|
|
|
|
|
|
switch (filter_type) {
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
case Rep::FilterType::kPartitionedFilter:
|
|
|
|
return PartitionedFilterBlockReader::Create(
|
|
|
|
this, ro, prefetch_buffer, use_cache, prefetch, pin, lookup_context);
|
|
|
|
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
case Rep::FilterType::kFullFilter:
|
|
|
|
return FullFilterBlockReader::Create(this, ro, prefetch_buffer, use_cache,
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
prefetch, pin, lookup_context);
|
|
|
|
|
|
|
|
default:
|
|
|
|
// filter_type is either kNoFilter (exited the function at the first if),
|
|
|
|
// or it must be covered in this switch block
|
|
|
|
assert(false);
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
return std::unique_ptr<FilterBlockReader>();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// disable_prefix_seek should be set to true when prefix_extractor found in SST
|
|
|
|
// differs from the one in mutable_cf_options and index type is HashBasedIndex
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
InternalIteratorBase<IndexValue>* BlockBasedTable::NewIndexIterator(
|
|
|
|
const ReadOptions& read_options, bool disable_prefix_seek,
|
|
|
|
IndexBlockIter* input_iter, GetContext* get_context,
|
|
|
|
BlockCacheLookupContext* lookup_context) const {
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
assert(rep_ != nullptr);
|
|
|
|
assert(rep_->index_reader != nullptr);
|
|
|
|
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
// We don't return pinned data from index blocks, so no need
|
|
|
|
// to set `block_contents_pinned`.
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
return rep_->index_reader->NewIterator(read_options, disable_prefix_seek,
|
|
|
|
input_iter, get_context,
|
|
|
|
lookup_context);
|
|
|
|
}
|
|
|
|
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
// TODO?
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
template <>
|
|
|
|
DataBlockIter* BlockBasedTable::InitBlockIterator<DataBlockIter>(
|
|
|
|
const Rep* rep, Block* block, BlockType block_type,
|
|
|
|
DataBlockIter* input_iter, bool block_contents_pinned) {
|
|
|
|
return block->NewDataIterator(rep->internal_comparator.user_comparator(),
|
|
|
|
rep->get_global_seqno(block_type), input_iter,
|
Add support to strip / pad timestamp when creating / reading a block based table (#11495)
Summary:
Add support to strip timestamp in block based table builder and pad timestamp in block based table reader.
On the write path, use the per column family option `AdvancedColumnFamilyOptions.persist_user_defined_timestamps` to indicate whether user-defined timestamps should be stripped for all block based tables created for the column family.
On the read path, added a per table `TableReadOption.user_defined_timestamps_persisted` to flag whether the user keys in the table contains user defined timestamps.
This patch is mostly passing the related flags down to the block building/parsing level with the exception of handling the `first_internal_key` in `IndexValue`, which is included in the `IndexBuilder` level. The value part of range deletion entries should have a similar handling, I haven't decided where to best fit this piece of logic, I will do it in a follow up.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11495
Test Plan:
Existing test `BlockBasedTableReaderTest` is parameterized to run with:
1) different UDT test modes: kNone, kNormal, kStripUserDefinedTimestamp
2) all four index types, when index type is `kTwoLevelIndexSearch`, also enables partitioned filters
3) parallel vs non-parallel compression
4) enable/disable compression dictionary.
Also added tests for API `BlockBasedTableReader::NewIterator`.
`PartitionedFilterBlockTest` is parameterized to run with different UDT test modes:kNone, kNormal, kStripUserDefinedTimestamp.
```
make all check
./block_based_table_reader_test
./partitioned_filter_block_test
```
Reviewed By: ltamasi
Differential Revision: D46344577
Pulled By: jowlyzhang
fbshipit-source-id: 93ac8542b19319d1298712b8bed908c8831ba675
1 year ago
|
|
|
rep->ioptions.stats, block_contents_pinned,
|
|
|
|
rep->user_defined_timestamps_persisted);
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
}
|
|
|
|
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
// TODO?
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
template <>
|
|
|
|
IndexBlockIter* BlockBasedTable::InitBlockIterator<IndexBlockIter>(
|
|
|
|
const Rep* rep, Block* block, BlockType block_type,
|
|
|
|
IndexBlockIter* input_iter, bool block_contents_pinned) {
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
return block->NewIndexIterator(
|
|
|
|
rep->internal_comparator.user_comparator(),
|
|
|
|
rep->get_global_seqno(block_type), input_iter, rep->ioptions.stats,
|
|
|
|
/* total_order_seek */ true, rep->index_has_first_key,
|
|
|
|
rep->index_key_includes_seq, rep->index_value_is_full,
|
Add support to strip / pad timestamp when creating / reading a block based table (#11495)
Summary:
Add support to strip timestamp in block based table builder and pad timestamp in block based table reader.
On the write path, use the per column family option `AdvancedColumnFamilyOptions.persist_user_defined_timestamps` to indicate whether user-defined timestamps should be stripped for all block based tables created for the column family.
On the read path, added a per table `TableReadOption.user_defined_timestamps_persisted` to flag whether the user keys in the table contains user defined timestamps.
This patch is mostly passing the related flags down to the block building/parsing level with the exception of handling the `first_internal_key` in `IndexValue`, which is included in the `IndexBuilder` level. The value part of range deletion entries should have a similar handling, I haven't decided where to best fit this piece of logic, I will do it in a follow up.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11495
Test Plan:
Existing test `BlockBasedTableReaderTest` is parameterized to run with:
1) different UDT test modes: kNone, kNormal, kStripUserDefinedTimestamp
2) all four index types, when index type is `kTwoLevelIndexSearch`, also enables partitioned filters
3) parallel vs non-parallel compression
4) enable/disable compression dictionary.
Also added tests for API `BlockBasedTableReader::NewIterator`.
`PartitionedFilterBlockTest` is parameterized to run with different UDT test modes:kNone, kNormal, kStripUserDefinedTimestamp.
```
make all check
./block_based_table_reader_test
./partitioned_filter_block_test
```
Reviewed By: ltamasi
Differential Revision: D46344577
Pulled By: jowlyzhang
fbshipit-source-id: 93ac8542b19319d1298712b8bed908c8831ba675
1 year ago
|
|
|
block_contents_pinned, rep->user_defined_timestamps_persisted);
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
}
|
|
|
|
|
|
|
|
// If contents is nullptr, this function looks up the block caches for the
|
|
|
|
// data block referenced by handle, and read the block from disk if necessary.
|
|
|
|
// If contents is non-null, it skips the cache lookup and disk read, since
|
|
|
|
// the caller has already read it. In both cases, if ro.fill_cache is true,
|
|
|
|
// it inserts the block into the block cache.
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
template <typename TBlocklike>
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
WithBlocklikeCheck<Status, TBlocklike>
|
|
|
|
BlockBasedTable::MaybeReadBlockAndLoadToCache(
|
|
|
|
FilePrefetchBuffer* prefetch_buffer, const ReadOptions& ro,
|
|
|
|
const BlockHandle& handle, const UncompressionDict& uncompression_dict,
|
HyperClockCache support for SecondaryCache, with refactoring (#11301)
Summary:
Internally refactors SecondaryCache integration out of LRUCache specifically and into a wrapper/adapter class that works with various Cache implementations. Notably, this relies on separating the notion of async lookup handles from other cache handles, so that HyperClockCache doesn't have to deal with the problem of allocating handles from the hash table for lookups that might fail anyway, and might be on the same key without support for coalescing. (LRUCache's hash table can incorporate previously allocated handles thanks to its pointer indirection.) Specifically, I'm worried about the case in which hundreds of threads try to access the same block and probing in the hash table degrades to linear search on the pile of entries with the same key.
This change is a big step in the direction of supporting stacked SecondaryCaches, but there are obstacles to completing that. Especially, there is no SecondaryCache hook for evictions to pass from one to the next. It has been proposed that evictions be transmitted simply as the persisted data (as in SaveToCallback), but given the current structure provided by the CacheItemHelpers, that would require an extra copy of the block data, because there's intentionally no way to ask for a contiguous Slice of the data (to allow for flexibility in storage). `AsyncLookupHandle` and the re-worked `WaitAll()` should be essentially prepared for stacked SecondaryCaches, but several "TODO with stacked secondaries" issues remain in various places.
It could be argued that the stacking instead be done as a SecondaryCache adapter that wraps two (or more) SecondaryCaches, but at least with the current API that would require an extra heap allocation on SecondaryCache Lookup for a wrapper SecondaryCacheResultHandle that can transfer a Lookup between secondaries. We could also consider trying to unify the Cache and SecondaryCache APIs, though that might be difficult if `AsyncLookupHandle` is kept a fixed struct.
## cache.h (public API)
Moves `secondary_cache` option from LRUCacheOptions to ShardedCacheOptions so that it is applicable to HyperClockCache.
## advanced_cache.h (advanced public API)
* Add `Cache::CreateStandalone()` so that the SecondaryCache support wrapper can use it.
* Add `SetEvictionCallback()` / `eviction_callback_` so that the SecondaryCache support wrapper can use it. Only a single callback is supported for efficiency. If there is ever a need for more than one, hopefully that can be handled with a broadcast callback wrapper.
These are essentially the two "extra" pieces of `Cache` for pulling out specific SecondaryCache support from the `Cache` implementation. I think it's a good trade-off as these are reasonable, limited, and reusable "cut points" into the `Cache` implementations.
* Remove async capability from standard `Lookup()` (getting rid of awkward restrictions on pending Handles) and add `AsyncLookupHandle` and `StartAsyncLookup()`. As noted in the comments, the full struct of `AsyncLookupHandle` is exposed so that it can be stack allocated, for efficiency, though more data is being copied around than before, which could impact performance. (Lookup info -> AsyncLookupHandle -> Handle vs. Lookup info -> Handle)
I could foresee a future in which a Cache internally saves a pointer to the AsyncLookupHandle, which means it's dangerous to allow it to be copyable or even movable. It also means it's not compatible with std::vector (which I don't like requiring as an API parameter anyway), so `WaitAll()` expects any contiguous array of AsyncLookupHandles. I believe this is best for common case efficiency, while behaving well in other cases also. For example, `WaitAll()` has no effect on default-constructed AsyncLookupHandles, which look like a completed cache miss.
## cacheable_entry.h
A couple of functions are obsolete because Cache::Handle can no longer be pending.
## cache.cc
Provides default implementations for new or revamped Cache functions, especially appropriate for non-blocking caches.
## secondary_cache_adapter.{h,cc}
The full details of the Cache wrapper adding SecondaryCache support. Essentially replicates the SecondaryCache handling that was in LRUCache, but obviously refactored. There is a bit of logic duplication, where Lookup() is essentially a manually optimized version of StartAsyncLookup() and Wait(), but it's roughly a dozen lines of code.
## sharded_cache.h, typed_cache.h, charged_cache.{h,cc}, sim_cache.cc
Simply updated for Cache API changes.
## lru_cache.{h,cc}
Carefully remove SecondaryCache logic, implement `CreateStandalone` and eviction handler functionality.
## clock_cache.{h,cc}
Expose existing `CreateStandalone` functionality, add eviction handler functionality. Light refactoring.
## block_based_table_reader*
Mostly re-worked the only usage of async Lookup, which is in BlockBasedTable::MultiGet. Used arrays in place of autovector in some places for efficiency. Simplified some logic by not trying to process some cache results before they're all ready.
Created new function `BlockBasedTable::GetCachePriority()` to reduce some pre-existing code duplication (and avoid making it worse).
Fixed at least one small bug from the prior confusing mixture of async and sync Lookups. In MaybeReadBlockAndLoadToCache(), called by RetrieveBlock(), called by MultiGet() with wait=false, is_cache_hit for the block_cache_tracer entry would not be set to true if the handle was pending after Lookup and before Wait.
## Intended follow-up work
* Figure out if there are any missing stats or block_cache_tracer work in refactored BlockBasedTable::MultiGet
* Stacked secondary caches (see above discussion)
* See if we can make up for the small MultiGet performance regression.
* Study more performance with SecondaryCache
* Items evicted from over-full LRUCache in Release were not being demoted to SecondaryCache, and still aren't to minimize unit test churn. Ideally they would be demoted, but it's an exceptional case so not a big deal.
* Use CreateStandalone for cache reservations (save unnecessary hash table operations). Not a big deal, but worthy cleanup.
* Somehow I got the contract for SecondaryCache::Insert wrong in #10945. (Doesn't take ownership!) That API comment needs to be fixed, but didn't want to mingle that in here.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11301
Test Plan:
## Unit tests
Generally updated to include HCC in SecondaryCache tests, though HyperClockCache has some different, less strict behaviors that leads to some tests not really being set up to work with it. Some of the tests remain disabled with it, but I think we have good coverage without them.
## Crash/stress test
Updated to use the new combination.
## Performance
First, let's check for regression on caches without secondary cache configured. Adding support for the eviction callback is likely to have a tiny effect, but it shouldn't be worrisome. LRUCache could benefit slightly from less logic around SecondaryCache handling. We can test with cache_bench default settings, built with DEBUG_LEVEL=0 and PORTABLE=0.
```
(while :; do base/cache_bench --cache_type=hyper_clock_cache | grep Rough; done) | awk '{ sum += $9; count++; print $0; print "Average: " int(sum / count) }'
```
**Before** this and #11299 (which could also have a small effect), running for about an hour, before & after running concurrently for each cache type:
HyperClockCache: 3168662 (average parallel ops/sec)
LRUCache: 2940127
**After** this and #11299, running for about an hour:
HyperClockCache: 3164862 (average parallel ops/sec) (0.12% slower)
LRUCache: 2940928 (0.03% faster)
This is an acceptable difference IMHO.
Next, let's consider essentially the worst case of new CPU overhead affecting overall performance. MultiGet uses the async lookup interface regardless of whether SecondaryCache or folly are used. We can configure a benchmark where all block cache queries are for data blocks, and all are hits.
Create DB and test (before and after tests running simultaneously):
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16
TEST_TMPDIR=/dev/shm base/db_bench -benchmarks=multireadrandom[-X30] -readonly -multiread_batched -batch_size=32 -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16
```
**Before**:
multireadrandom [AVG 30 runs] : 3444202 (± 57049) ops/sec; 240.9 (± 4.0) MB/sec
multireadrandom [MEDIAN 30 runs] : 3514443 ops/sec; 245.8 MB/sec
**After**:
multireadrandom [AVG 30 runs] : 3291022 (± 58851) ops/sec; 230.2 (± 4.1) MB/sec
multireadrandom [MEDIAN 30 runs] : 3366179 ops/sec; 235.4 MB/sec
So that's roughly a 3% regression, on kind of a *worst case* test of MultiGet CPU. Similar story with HyperClockCache:
**Before**:
multireadrandom [AVG 30 runs] : 3933777 (± 41840) ops/sec; 275.1 (± 2.9) MB/sec
multireadrandom [MEDIAN 30 runs] : 3970667 ops/sec; 277.7 MB/sec
**After**:
multireadrandom [AVG 30 runs] : 3755338 (± 30391) ops/sec; 262.6 (± 2.1) MB/sec
multireadrandom [MEDIAN 30 runs] : 3785696 ops/sec; 264.8 MB/sec
Roughly a 4-5% regression. Not ideal, but not the whole story, fortunately.
Let's also look at Get() in db_bench:
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X30] -readonly -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16
```
**Before**:
readrandom [AVG 30 runs] : 2198685 (± 13412) ops/sec; 153.8 (± 0.9) MB/sec
readrandom [MEDIAN 30 runs] : 2209498 ops/sec; 154.5 MB/sec
**After**:
readrandom [AVG 30 runs] : 2292814 (± 43508) ops/sec; 160.3 (± 3.0) MB/sec
readrandom [MEDIAN 30 runs] : 2365181 ops/sec; 165.4 MB/sec
That's showing roughly a 4% improvement, perhaps because of the secondary cache code that is no longer part of LRUCache. But weirdly, HyperClockCache is also showing 2-3% improvement:
**Before**:
readrandom [AVG 30 runs] : 2272333 (± 9992) ops/sec; 158.9 (± 0.7) MB/sec
readrandom [MEDIAN 30 runs] : 2273239 ops/sec; 159.0 MB/sec
**After**:
readrandom [AVG 30 runs] : 2332407 (± 11252) ops/sec; 163.1 (± 0.8) MB/sec
readrandom [MEDIAN 30 runs] : 2335329 ops/sec; 163.3 MB/sec
Reviewed By: ltamasi
Differential Revision: D44177044
Pulled By: pdillinger
fbshipit-source-id: e808e48ff3fe2f792a79841ba617be98e48689f5
2 years ago
|
|
|
bool for_compaction, CachableEntry<TBlocklike>* out_parsed_block,
|
|
|
|
GetContext* get_context, BlockCacheLookupContext* lookup_context,
|
|
|
|
BlockContents* contents, bool async_read) const {
|
Refactor to avoid confusing "raw block" (#10408)
Summary:
We have a lot of confusing code because of mixed, sometimes
completely opposite uses of of the term "raw block" or "raw contents",
sometimes within the same source file. For example, in `BlockBasedTableBuilder`,
`raw_block_contents` and `raw_size` generally referred to uncompressed block
contents and size, while `WriteRawBlock` referred to writing a block that
is already compressed if it is going to be. Meanwhile, in
`BlockBasedTable`, `raw_block_contents` either referred to a (maybe
compressed) block with trailer, or a maybe compressed block maybe
without trailer. (Note: left as follow-up work to use C++ typing to
better sort out the various kinds of BlockContents.)
This change primarily tries to apply some consistent terminology around
the kinds of block representations, avoiding the unclear "raw". (Any
meaning of "raw" assumes some bias toward the storage layer or toward
the logical data layer.) Preferred terminology:
* **Serialized block** - bytes that go into storage. For block-based table
(usually the case) this includes the block trailer. WART: block `size` may or
may not include the trailer; need to be clear about whether it does or not.
* **Maybe compressed block** - like a serialized block, but without the
trailer (or no promise of including a trailer). Must be accompanied by a
CompressionType.
* **Uncompressed block** - "payload" bytes that are either stored with no
compression, used as input to compression function, or result of
decompression function.
* **Parsed block** - an in-memory form of a block in block cache, as it is
used by the table reader. Different C++ types are used depending on the
block type (see block_like_traits.h).
Other refactorings:
* Misc corrections/improvements of internal API comments
* Remove a few misleading / unhelpful / redundant comments.
* Use move semantics in some places to simplify contracts
* Use better parameter names to indicate which parameters are used for
outputs
* Remove some extraneous `extern`
* Various clean-ups to `CacheDumperImpl` (mostly unnecessary code)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10408
Test Plan: existing tests
Reviewed By: akankshamahajan15
Differential Revision: D38172617
Pulled By: pdillinger
fbshipit-source-id: ccb99299f324ac5ca46996d34c5089621a4f260c
2 years ago
|
|
|
assert(out_parsed_block != nullptr);
|
|
|
|
const bool no_io = (ro.read_tier == kBlockCacheTier);
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
BlockCacheInterface<TBlocklike> block_cache{
|
|
|
|
rep_->table_options.block_cache.get()};
|
|
|
|
|
|
|
|
// First, try to get the block from the cache
|
|
|
|
//
|
|
|
|
// If either block cache is enabled, we'll try to read from it.
|
|
|
|
Status s;
|
New stable, fixed-length cache keys (#9126)
Summary:
This change standardizes on a new 16-byte cache key format for
block cache (incl compressed and secondary) and persistent cache (but
not table cache and row cache).
The goal is a really fast cache key with practically ideal stability and
uniqueness properties without external dependencies (e.g. from FileSystem).
A fixed key size of 16 bytes should enable future optimizations to the
concurrent hash table for block cache, which is a heavy CPU user /
bottleneck, but there appears to be measurable performance improvement
even with no changes to LRUCache.
This change replaces a lot of disjointed and ugly code handling cache
keys with calls to a simple, clean new internal API (cache_key.h).
(Preserving the old cache key logic under an option would be very ugly
and likely negate the performance gain of the new approach. Complete
replacement carries some inherent risk, but I think that's acceptable
with sufficient analysis and testing.)
The scheme for encoding new cache keys is complicated but explained
in cache_key.cc.
Also: EndianSwapValue is moved to math.h to be next to other bit
operations. (Explains some new include "math.h".) ReverseBits operation
added and unit tests added to hash_test for both.
Fixes https://github.com/facebook/rocksdb/issues/7405 (presuming a root cause)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9126
Test Plan:
### Basic correctness
Several tests needed updates to work with the new functionality, mostly
because we are no longer relying on filesystem for stable cache keys
so table builders & readers need more context info to agree on cache
keys. This functionality is so core, a huge number of existing tests
exercise the cache key functionality.
### Performance
Create db with
`TEST_TMPDIR=/dev/shm ./db_bench -bloom_bits=10 -benchmarks=fillrandom -num=3000000 -partition_index_and_filters`
And test performance with
`TEST_TMPDIR=/dev/shm ./db_bench -readonly -use_existing_db -bloom_bits=10 -benchmarks=readrandom -num=3000000 -duration=30 -cache_index_and_filter_blocks -cache_size=250000 -threads=4`
using DEBUG_LEVEL=0 and simultaneous before & after runs.
Before ops/sec, avg over 100 runs: 121924
After ops/sec, avg over 100 runs: 125385 (+2.8%)
### Collision probability
I have built a tool, ./cache_bench -stress_cache_key to broadly simulate host-wide cache activity
over many months, by making some pessimistic simplifying assumptions:
* Every generated file has a cache entry for every byte offset in the file (contiguous range of cache keys)
* All of every file is cached for its entire lifetime
We use a simple table with skewed address assignment and replacement on address collision
to simulate files coming & going, with quite a variance (super-Poisson) in ages. Some output
with `./cache_bench -stress_cache_key -sck_keep_bits=40`:
```
Total cache or DBs size: 32TiB Writing 925.926 MiB/s or 76.2939TiB/day
Multiply by 9.22337e+18 to correct for simulation losses (but still assume whole file cached)
```
These come from default settings of 2.5M files per day of 32 MB each, and
`-sck_keep_bits=40` means that to represent a single file, we are only keeping 40 bits of
the 128-bit cache key. With file size of 2\*\*25 contiguous keys (pessimistic), our simulation
is about 2\*\*(128-40-25) or about 9 billion billion times more prone to collision than reality.
More default assumptions, relatively pessimistic:
* 100 DBs in same process (doesn't matter much)
* Re-open DB in same process (new session ID related to old session ID) on average
every 100 files generated
* Restart process (all new session IDs unrelated to old) 24 times per day
After enough data, we get a result at the end:
```
(keep 40 bits) 17 collisions after 2 x 90 days, est 10.5882 days between (9.76592e+19 corrected)
```
If we believe the (pessimistic) simulation and the mathematical generalization, we would need to run a billion machines all for 97 billion days to expect a cache key collision. To help verify that our generalization ("corrected") is robust, we can make our simulation more precise with `-sck_keep_bits=41` and `42`, which takes more running time to get enough data:
```
(keep 41 bits) 16 collisions after 4 x 90 days, est 22.5 days between (1.03763e+20 corrected)
(keep 42 bits) 19 collisions after 10 x 90 days, est 47.3684 days between (1.09224e+20 corrected)
```
The generalized prediction still holds. With the `-sck_randomize` option, we can see that we are beating "random" cache keys (except offsets still non-randomized) by a modest amount (roughly 20x less collision prone than random), which should make us reasonably comfortable even in "degenerate" cases:
```
197 collisions after 1 x 90 days, est 0.456853 days between (4.21372e+18 corrected)
```
I've run other tests to validate other conditions behave as expected, never behaving "worse than random" unless we start chopping off structured data.
Reviewed By: zhichao-cao
Differential Revision: D33171746
Pulled By: pdillinger
fbshipit-source-id: f16a57e369ed37be5e7e33525ace848d0537c88f
3 years ago
|
|
|
CacheKey key_data;
|
|
|
|
Slice key;
|
|
|
|
bool is_cache_hit = false;
|
|
|
|
if (block_cache) {
|
|
|
|
// create key for block cache
|
New stable, fixed-length cache keys (#9126)
Summary:
This change standardizes on a new 16-byte cache key format for
block cache (incl compressed and secondary) and persistent cache (but
not table cache and row cache).
The goal is a really fast cache key with practically ideal stability and
uniqueness properties without external dependencies (e.g. from FileSystem).
A fixed key size of 16 bytes should enable future optimizations to the
concurrent hash table for block cache, which is a heavy CPU user /
bottleneck, but there appears to be measurable performance improvement
even with no changes to LRUCache.
This change replaces a lot of disjointed and ugly code handling cache
keys with calls to a simple, clean new internal API (cache_key.h).
(Preserving the old cache key logic under an option would be very ugly
and likely negate the performance gain of the new approach. Complete
replacement carries some inherent risk, but I think that's acceptable
with sufficient analysis and testing.)
The scheme for encoding new cache keys is complicated but explained
in cache_key.cc.
Also: EndianSwapValue is moved to math.h to be next to other bit
operations. (Explains some new include "math.h".) ReverseBits operation
added and unit tests added to hash_test for both.
Fixes https://github.com/facebook/rocksdb/issues/7405 (presuming a root cause)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9126
Test Plan:
### Basic correctness
Several tests needed updates to work with the new functionality, mostly
because we are no longer relying on filesystem for stable cache keys
so table builders & readers need more context info to agree on cache
keys. This functionality is so core, a huge number of existing tests
exercise the cache key functionality.
### Performance
Create db with
`TEST_TMPDIR=/dev/shm ./db_bench -bloom_bits=10 -benchmarks=fillrandom -num=3000000 -partition_index_and_filters`
And test performance with
`TEST_TMPDIR=/dev/shm ./db_bench -readonly -use_existing_db -bloom_bits=10 -benchmarks=readrandom -num=3000000 -duration=30 -cache_index_and_filter_blocks -cache_size=250000 -threads=4`
using DEBUG_LEVEL=0 and simultaneous before & after runs.
Before ops/sec, avg over 100 runs: 121924
After ops/sec, avg over 100 runs: 125385 (+2.8%)
### Collision probability
I have built a tool, ./cache_bench -stress_cache_key to broadly simulate host-wide cache activity
over many months, by making some pessimistic simplifying assumptions:
* Every generated file has a cache entry for every byte offset in the file (contiguous range of cache keys)
* All of every file is cached for its entire lifetime
We use a simple table with skewed address assignment and replacement on address collision
to simulate files coming & going, with quite a variance (super-Poisson) in ages. Some output
with `./cache_bench -stress_cache_key -sck_keep_bits=40`:
```
Total cache or DBs size: 32TiB Writing 925.926 MiB/s or 76.2939TiB/day
Multiply by 9.22337e+18 to correct for simulation losses (but still assume whole file cached)
```
These come from default settings of 2.5M files per day of 32 MB each, and
`-sck_keep_bits=40` means that to represent a single file, we are only keeping 40 bits of
the 128-bit cache key. With file size of 2\*\*25 contiguous keys (pessimistic), our simulation
is about 2\*\*(128-40-25) or about 9 billion billion times more prone to collision than reality.
More default assumptions, relatively pessimistic:
* 100 DBs in same process (doesn't matter much)
* Re-open DB in same process (new session ID related to old session ID) on average
every 100 files generated
* Restart process (all new session IDs unrelated to old) 24 times per day
After enough data, we get a result at the end:
```
(keep 40 bits) 17 collisions after 2 x 90 days, est 10.5882 days between (9.76592e+19 corrected)
```
If we believe the (pessimistic) simulation and the mathematical generalization, we would need to run a billion machines all for 97 billion days to expect a cache key collision. To help verify that our generalization ("corrected") is robust, we can make our simulation more precise with `-sck_keep_bits=41` and `42`, which takes more running time to get enough data:
```
(keep 41 bits) 16 collisions after 4 x 90 days, est 22.5 days between (1.03763e+20 corrected)
(keep 42 bits) 19 collisions after 10 x 90 days, est 47.3684 days between (1.09224e+20 corrected)
```
The generalized prediction still holds. With the `-sck_randomize` option, we can see that we are beating "random" cache keys (except offsets still non-randomized) by a modest amount (roughly 20x less collision prone than random), which should make us reasonably comfortable even in "degenerate" cases:
```
197 collisions after 1 x 90 days, est 0.456853 days between (4.21372e+18 corrected)
```
I've run other tests to validate other conditions behave as expected, never behaving "worse than random" unless we start chopping off structured data.
Reviewed By: zhichao-cao
Differential Revision: D33171746
Pulled By: pdillinger
fbshipit-source-id: f16a57e369ed37be5e7e33525ace848d0537c88f
3 years ago
|
|
|
key_data = GetCacheKey(rep_->base_cache_key, handle);
|
|
|
|
key = key_data.AsSlice();
|
|
|
|
|
|
|
|
if (!contents) {
|
HyperClockCache support for SecondaryCache, with refactoring (#11301)
Summary:
Internally refactors SecondaryCache integration out of LRUCache specifically and into a wrapper/adapter class that works with various Cache implementations. Notably, this relies on separating the notion of async lookup handles from other cache handles, so that HyperClockCache doesn't have to deal with the problem of allocating handles from the hash table for lookups that might fail anyway, and might be on the same key without support for coalescing. (LRUCache's hash table can incorporate previously allocated handles thanks to its pointer indirection.) Specifically, I'm worried about the case in which hundreds of threads try to access the same block and probing in the hash table degrades to linear search on the pile of entries with the same key.
This change is a big step in the direction of supporting stacked SecondaryCaches, but there are obstacles to completing that. Especially, there is no SecondaryCache hook for evictions to pass from one to the next. It has been proposed that evictions be transmitted simply as the persisted data (as in SaveToCallback), but given the current structure provided by the CacheItemHelpers, that would require an extra copy of the block data, because there's intentionally no way to ask for a contiguous Slice of the data (to allow for flexibility in storage). `AsyncLookupHandle` and the re-worked `WaitAll()` should be essentially prepared for stacked SecondaryCaches, but several "TODO with stacked secondaries" issues remain in various places.
It could be argued that the stacking instead be done as a SecondaryCache adapter that wraps two (or more) SecondaryCaches, but at least with the current API that would require an extra heap allocation on SecondaryCache Lookup for a wrapper SecondaryCacheResultHandle that can transfer a Lookup between secondaries. We could also consider trying to unify the Cache and SecondaryCache APIs, though that might be difficult if `AsyncLookupHandle` is kept a fixed struct.
## cache.h (public API)
Moves `secondary_cache` option from LRUCacheOptions to ShardedCacheOptions so that it is applicable to HyperClockCache.
## advanced_cache.h (advanced public API)
* Add `Cache::CreateStandalone()` so that the SecondaryCache support wrapper can use it.
* Add `SetEvictionCallback()` / `eviction_callback_` so that the SecondaryCache support wrapper can use it. Only a single callback is supported for efficiency. If there is ever a need for more than one, hopefully that can be handled with a broadcast callback wrapper.
These are essentially the two "extra" pieces of `Cache` for pulling out specific SecondaryCache support from the `Cache` implementation. I think it's a good trade-off as these are reasonable, limited, and reusable "cut points" into the `Cache` implementations.
* Remove async capability from standard `Lookup()` (getting rid of awkward restrictions on pending Handles) and add `AsyncLookupHandle` and `StartAsyncLookup()`. As noted in the comments, the full struct of `AsyncLookupHandle` is exposed so that it can be stack allocated, for efficiency, though more data is being copied around than before, which could impact performance. (Lookup info -> AsyncLookupHandle -> Handle vs. Lookup info -> Handle)
I could foresee a future in which a Cache internally saves a pointer to the AsyncLookupHandle, which means it's dangerous to allow it to be copyable or even movable. It also means it's not compatible with std::vector (which I don't like requiring as an API parameter anyway), so `WaitAll()` expects any contiguous array of AsyncLookupHandles. I believe this is best for common case efficiency, while behaving well in other cases also. For example, `WaitAll()` has no effect on default-constructed AsyncLookupHandles, which look like a completed cache miss.
## cacheable_entry.h
A couple of functions are obsolete because Cache::Handle can no longer be pending.
## cache.cc
Provides default implementations for new or revamped Cache functions, especially appropriate for non-blocking caches.
## secondary_cache_adapter.{h,cc}
The full details of the Cache wrapper adding SecondaryCache support. Essentially replicates the SecondaryCache handling that was in LRUCache, but obviously refactored. There is a bit of logic duplication, where Lookup() is essentially a manually optimized version of StartAsyncLookup() and Wait(), but it's roughly a dozen lines of code.
## sharded_cache.h, typed_cache.h, charged_cache.{h,cc}, sim_cache.cc
Simply updated for Cache API changes.
## lru_cache.{h,cc}
Carefully remove SecondaryCache logic, implement `CreateStandalone` and eviction handler functionality.
## clock_cache.{h,cc}
Expose existing `CreateStandalone` functionality, add eviction handler functionality. Light refactoring.
## block_based_table_reader*
Mostly re-worked the only usage of async Lookup, which is in BlockBasedTable::MultiGet. Used arrays in place of autovector in some places for efficiency. Simplified some logic by not trying to process some cache results before they're all ready.
Created new function `BlockBasedTable::GetCachePriority()` to reduce some pre-existing code duplication (and avoid making it worse).
Fixed at least one small bug from the prior confusing mixture of async and sync Lookups. In MaybeReadBlockAndLoadToCache(), called by RetrieveBlock(), called by MultiGet() with wait=false, is_cache_hit for the block_cache_tracer entry would not be set to true if the handle was pending after Lookup and before Wait.
## Intended follow-up work
* Figure out if there are any missing stats or block_cache_tracer work in refactored BlockBasedTable::MultiGet
* Stacked secondary caches (see above discussion)
* See if we can make up for the small MultiGet performance regression.
* Study more performance with SecondaryCache
* Items evicted from over-full LRUCache in Release were not being demoted to SecondaryCache, and still aren't to minimize unit test churn. Ideally they would be demoted, but it's an exceptional case so not a big deal.
* Use CreateStandalone for cache reservations (save unnecessary hash table operations). Not a big deal, but worthy cleanup.
* Somehow I got the contract for SecondaryCache::Insert wrong in #10945. (Doesn't take ownership!) That API comment needs to be fixed, but didn't want to mingle that in here.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11301
Test Plan:
## Unit tests
Generally updated to include HCC in SecondaryCache tests, though HyperClockCache has some different, less strict behaviors that leads to some tests not really being set up to work with it. Some of the tests remain disabled with it, but I think we have good coverage without them.
## Crash/stress test
Updated to use the new combination.
## Performance
First, let's check for regression on caches without secondary cache configured. Adding support for the eviction callback is likely to have a tiny effect, but it shouldn't be worrisome. LRUCache could benefit slightly from less logic around SecondaryCache handling. We can test with cache_bench default settings, built with DEBUG_LEVEL=0 and PORTABLE=0.
```
(while :; do base/cache_bench --cache_type=hyper_clock_cache | grep Rough; done) | awk '{ sum += $9; count++; print $0; print "Average: " int(sum / count) }'
```
**Before** this and #11299 (which could also have a small effect), running for about an hour, before & after running concurrently for each cache type:
HyperClockCache: 3168662 (average parallel ops/sec)
LRUCache: 2940127
**After** this and #11299, running for about an hour:
HyperClockCache: 3164862 (average parallel ops/sec) (0.12% slower)
LRUCache: 2940928 (0.03% faster)
This is an acceptable difference IMHO.
Next, let's consider essentially the worst case of new CPU overhead affecting overall performance. MultiGet uses the async lookup interface regardless of whether SecondaryCache or folly are used. We can configure a benchmark where all block cache queries are for data blocks, and all are hits.
Create DB and test (before and after tests running simultaneously):
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16
TEST_TMPDIR=/dev/shm base/db_bench -benchmarks=multireadrandom[-X30] -readonly -multiread_batched -batch_size=32 -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16
```
**Before**:
multireadrandom [AVG 30 runs] : 3444202 (± 57049) ops/sec; 240.9 (± 4.0) MB/sec
multireadrandom [MEDIAN 30 runs] : 3514443 ops/sec; 245.8 MB/sec
**After**:
multireadrandom [AVG 30 runs] : 3291022 (± 58851) ops/sec; 230.2 (± 4.1) MB/sec
multireadrandom [MEDIAN 30 runs] : 3366179 ops/sec; 235.4 MB/sec
So that's roughly a 3% regression, on kind of a *worst case* test of MultiGet CPU. Similar story with HyperClockCache:
**Before**:
multireadrandom [AVG 30 runs] : 3933777 (± 41840) ops/sec; 275.1 (± 2.9) MB/sec
multireadrandom [MEDIAN 30 runs] : 3970667 ops/sec; 277.7 MB/sec
**After**:
multireadrandom [AVG 30 runs] : 3755338 (± 30391) ops/sec; 262.6 (± 2.1) MB/sec
multireadrandom [MEDIAN 30 runs] : 3785696 ops/sec; 264.8 MB/sec
Roughly a 4-5% regression. Not ideal, but not the whole story, fortunately.
Let's also look at Get() in db_bench:
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X30] -readonly -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16
```
**Before**:
readrandom [AVG 30 runs] : 2198685 (± 13412) ops/sec; 153.8 (± 0.9) MB/sec
readrandom [MEDIAN 30 runs] : 2209498 ops/sec; 154.5 MB/sec
**After**:
readrandom [AVG 30 runs] : 2292814 (± 43508) ops/sec; 160.3 (± 3.0) MB/sec
readrandom [MEDIAN 30 runs] : 2365181 ops/sec; 165.4 MB/sec
That's showing roughly a 4% improvement, perhaps because of the secondary cache code that is no longer part of LRUCache. But weirdly, HyperClockCache is also showing 2-3% improvement:
**Before**:
readrandom [AVG 30 runs] : 2272333 (± 9992) ops/sec; 158.9 (± 0.7) MB/sec
readrandom [MEDIAN 30 runs] : 2273239 ops/sec; 159.0 MB/sec
**After**:
readrandom [AVG 30 runs] : 2332407 (± 11252) ops/sec; 163.1 (± 0.8) MB/sec
readrandom [MEDIAN 30 runs] : 2335329 ops/sec; 163.3 MB/sec
Reviewed By: ltamasi
Differential Revision: D44177044
Pulled By: pdillinger
fbshipit-source-id: e808e48ff3fe2f792a79841ba617be98e48689f5
2 years ago
|
|
|
s = GetDataBlockFromCache(key, block_cache, out_parsed_block,
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
get_context);
|
|
|
|
// Value could still be null at this point, so check the cache handle
|
|
|
|
// and update the read pattern for prefetching
|
Refactor to avoid confusing "raw block" (#10408)
Summary:
We have a lot of confusing code because of mixed, sometimes
completely opposite uses of of the term "raw block" or "raw contents",
sometimes within the same source file. For example, in `BlockBasedTableBuilder`,
`raw_block_contents` and `raw_size` generally referred to uncompressed block
contents and size, while `WriteRawBlock` referred to writing a block that
is already compressed if it is going to be. Meanwhile, in
`BlockBasedTable`, `raw_block_contents` either referred to a (maybe
compressed) block with trailer, or a maybe compressed block maybe
without trailer. (Note: left as follow-up work to use C++ typing to
better sort out the various kinds of BlockContents.)
This change primarily tries to apply some consistent terminology around
the kinds of block representations, avoiding the unclear "raw". (Any
meaning of "raw" assumes some bias toward the storage layer or toward
the logical data layer.) Preferred terminology:
* **Serialized block** - bytes that go into storage. For block-based table
(usually the case) this includes the block trailer. WART: block `size` may or
may not include the trailer; need to be clear about whether it does or not.
* **Maybe compressed block** - like a serialized block, but without the
trailer (or no promise of including a trailer). Must be accompanied by a
CompressionType.
* **Uncompressed block** - "payload" bytes that are either stored with no
compression, used as input to compression function, or result of
decompression function.
* **Parsed block** - an in-memory form of a block in block cache, as it is
used by the table reader. Different C++ types are used depending on the
block type (see block_like_traits.h).
Other refactorings:
* Misc corrections/improvements of internal API comments
* Remove a few misleading / unhelpful / redundant comments.
* Use move semantics in some places to simplify contracts
* Use better parameter names to indicate which parameters are used for
outputs
* Remove some extraneous `extern`
* Various clean-ups to `CacheDumperImpl` (mostly unnecessary code)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10408
Test Plan: existing tests
Reviewed By: akankshamahajan15
Differential Revision: D38172617
Pulled By: pdillinger
fbshipit-source-id: ccb99299f324ac5ca46996d34c5089621a4f260c
2 years ago
|
|
|
if (out_parsed_block->GetValue() || out_parsed_block->GetCacheHandle()) {
|
|
|
|
// TODO(haoyu): Differentiate cache hit on uncompressed block cache and
|
|
|
|
// compressed block cache.
|
|
|
|
is_cache_hit = true;
|
|
|
|
if (prefetch_buffer) {
|
|
|
|
// Update the block details so that PrefetchBuffer can use the read
|
|
|
|
// pattern to determine if reads are sequential or not for
|
|
|
|
// prefetching. It should also take in account blocks read from cache.
|
|
|
|
prefetch_buffer->UpdateReadPattern(
|
|
|
|
handle.offset(), BlockSizeWithTrailer(handle),
|
|
|
|
ro.adaptive_readahead /*decrease_readahead_size*/);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Can't find the block from the cache. If I/O is allowed, read from the
|
|
|
|
// file.
|
Refactor to avoid confusing "raw block" (#10408)
Summary:
We have a lot of confusing code because of mixed, sometimes
completely opposite uses of of the term "raw block" or "raw contents",
sometimes within the same source file. For example, in `BlockBasedTableBuilder`,
`raw_block_contents` and `raw_size` generally referred to uncompressed block
contents and size, while `WriteRawBlock` referred to writing a block that
is already compressed if it is going to be. Meanwhile, in
`BlockBasedTable`, `raw_block_contents` either referred to a (maybe
compressed) block with trailer, or a maybe compressed block maybe
without trailer. (Note: left as follow-up work to use C++ typing to
better sort out the various kinds of BlockContents.)
This change primarily tries to apply some consistent terminology around
the kinds of block representations, avoiding the unclear "raw". (Any
meaning of "raw" assumes some bias toward the storage layer or toward
the logical data layer.) Preferred terminology:
* **Serialized block** - bytes that go into storage. For block-based table
(usually the case) this includes the block trailer. WART: block `size` may or
may not include the trailer; need to be clear about whether it does or not.
* **Maybe compressed block** - like a serialized block, but without the
trailer (or no promise of including a trailer). Must be accompanied by a
CompressionType.
* **Uncompressed block** - "payload" bytes that are either stored with no
compression, used as input to compression function, or result of
decompression function.
* **Parsed block** - an in-memory form of a block in block cache, as it is
used by the table reader. Different C++ types are used depending on the
block type (see block_like_traits.h).
Other refactorings:
* Misc corrections/improvements of internal API comments
* Remove a few misleading / unhelpful / redundant comments.
* Use move semantics in some places to simplify contracts
* Use better parameter names to indicate which parameters are used for
outputs
* Remove some extraneous `extern`
* Various clean-ups to `CacheDumperImpl` (mostly unnecessary code)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10408
Test Plan: existing tests
Reviewed By: akankshamahajan15
Differential Revision: D38172617
Pulled By: pdillinger
fbshipit-source-id: ccb99299f324ac5ca46996d34c5089621a4f260c
2 years ago
|
|
|
if (out_parsed_block->GetValue() == nullptr &&
|
|
|
|
out_parsed_block->GetCacheHandle() == nullptr && !no_io &&
|
|
|
|
ro.fill_cache) {
|
|
|
|
Statistics* statistics = rep_->ioptions.stats;
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
const bool maybe_compressed =
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
TBlocklike::kBlockType != BlockType::kFilter &&
|
|
|
|
TBlocklike::kBlockType != BlockType::kCompressionDictionary &&
|
|
|
|
rep_->blocks_maybe_compressed;
|
|
|
|
const bool do_uncompress = maybe_compressed;
|
Refactor to avoid confusing "raw block" (#10408)
Summary:
We have a lot of confusing code because of mixed, sometimes
completely opposite uses of of the term "raw block" or "raw contents",
sometimes within the same source file. For example, in `BlockBasedTableBuilder`,
`raw_block_contents` and `raw_size` generally referred to uncompressed block
contents and size, while `WriteRawBlock` referred to writing a block that
is already compressed if it is going to be. Meanwhile, in
`BlockBasedTable`, `raw_block_contents` either referred to a (maybe
compressed) block with trailer, or a maybe compressed block maybe
without trailer. (Note: left as follow-up work to use C++ typing to
better sort out the various kinds of BlockContents.)
This change primarily tries to apply some consistent terminology around
the kinds of block representations, avoiding the unclear "raw". (Any
meaning of "raw" assumes some bias toward the storage layer or toward
the logical data layer.) Preferred terminology:
* **Serialized block** - bytes that go into storage. For block-based table
(usually the case) this includes the block trailer. WART: block `size` may or
may not include the trailer; need to be clear about whether it does or not.
* **Maybe compressed block** - like a serialized block, but without the
trailer (or no promise of including a trailer). Must be accompanied by a
CompressionType.
* **Uncompressed block** - "payload" bytes that are either stored with no
compression, used as input to compression function, or result of
decompression function.
* **Parsed block** - an in-memory form of a block in block cache, as it is
used by the table reader. Different C++ types are used depending on the
block type (see block_like_traits.h).
Other refactorings:
* Misc corrections/improvements of internal API comments
* Remove a few misleading / unhelpful / redundant comments.
* Use move semantics in some places to simplify contracts
* Use better parameter names to indicate which parameters are used for
outputs
* Remove some extraneous `extern`
* Various clean-ups to `CacheDumperImpl` (mostly unnecessary code)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10408
Test Plan: existing tests
Reviewed By: akankshamahajan15
Differential Revision: D38172617
Pulled By: pdillinger
fbshipit-source-id: ccb99299f324ac5ca46996d34c5089621a4f260c
2 years ago
|
|
|
CompressionType contents_comp_type;
|
|
|
|
// Maybe serialized or uncompressed
|
|
|
|
BlockContents tmp_contents;
|
|
|
|
if (!contents) {
|
|
|
|
Histograms histogram = for_compaction ? READ_BLOCK_COMPACTION_MICROS
|
|
|
|
: READ_BLOCK_GET_MICROS;
|
|
|
|
StopWatch sw(rep_->ioptions.clock, statistics, histogram);
|
|
|
|
BlockFetcher block_fetcher(
|
|
|
|
rep_->file.get(), prefetch_buffer, rep_->footer, ro, handle,
|
Refactor to avoid confusing "raw block" (#10408)
Summary:
We have a lot of confusing code because of mixed, sometimes
completely opposite uses of of the term "raw block" or "raw contents",
sometimes within the same source file. For example, in `BlockBasedTableBuilder`,
`raw_block_contents` and `raw_size` generally referred to uncompressed block
contents and size, while `WriteRawBlock` referred to writing a block that
is already compressed if it is going to be. Meanwhile, in
`BlockBasedTable`, `raw_block_contents` either referred to a (maybe
compressed) block with trailer, or a maybe compressed block maybe
without trailer. (Note: left as follow-up work to use C++ typing to
better sort out the various kinds of BlockContents.)
This change primarily tries to apply some consistent terminology around
the kinds of block representations, avoiding the unclear "raw". (Any
meaning of "raw" assumes some bias toward the storage layer or toward
the logical data layer.) Preferred terminology:
* **Serialized block** - bytes that go into storage. For block-based table
(usually the case) this includes the block trailer. WART: block `size` may or
may not include the trailer; need to be clear about whether it does or not.
* **Maybe compressed block** - like a serialized block, but without the
trailer (or no promise of including a trailer). Must be accompanied by a
CompressionType.
* **Uncompressed block** - "payload" bytes that are either stored with no
compression, used as input to compression function, or result of
decompression function.
* **Parsed block** - an in-memory form of a block in block cache, as it is
used by the table reader. Different C++ types are used depending on the
block type (see block_like_traits.h).
Other refactorings:
* Misc corrections/improvements of internal API comments
* Remove a few misleading / unhelpful / redundant comments.
* Use move semantics in some places to simplify contracts
* Use better parameter names to indicate which parameters are used for
outputs
* Remove some extraneous `extern`
* Various clean-ups to `CacheDumperImpl` (mostly unnecessary code)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10408
Test Plan: existing tests
Reviewed By: akankshamahajan15
Differential Revision: D38172617
Pulled By: pdillinger
fbshipit-source-id: ccb99299f324ac5ca46996d34c5089621a4f260c
2 years ago
|
|
|
&tmp_contents, rep_->ioptions, do_uncompress, maybe_compressed,
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
TBlocklike::kBlockType, uncompression_dict,
|
|
|
|
rep_->persistent_cache_options,
|
|
|
|
GetMemoryAllocator(rep_->table_options),
|
|
|
|
/*allocator=*/nullptr);
|
|
|
|
|
|
|
|
// If prefetch_buffer is not allocated, it will fallback to synchronous
|
|
|
|
// reading of block contents.
|
|
|
|
if (async_read && prefetch_buffer != nullptr) {
|
|
|
|
s = block_fetcher.ReadAsyncBlockContents();
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
s = block_fetcher.ReadBlockContents();
|
|
|
|
}
|
|
|
|
|
Refactor to avoid confusing "raw block" (#10408)
Summary:
We have a lot of confusing code because of mixed, sometimes
completely opposite uses of of the term "raw block" or "raw contents",
sometimes within the same source file. For example, in `BlockBasedTableBuilder`,
`raw_block_contents` and `raw_size` generally referred to uncompressed block
contents and size, while `WriteRawBlock` referred to writing a block that
is already compressed if it is going to be. Meanwhile, in
`BlockBasedTable`, `raw_block_contents` either referred to a (maybe
compressed) block with trailer, or a maybe compressed block maybe
without trailer. (Note: left as follow-up work to use C++ typing to
better sort out the various kinds of BlockContents.)
This change primarily tries to apply some consistent terminology around
the kinds of block representations, avoiding the unclear "raw". (Any
meaning of "raw" assumes some bias toward the storage layer or toward
the logical data layer.) Preferred terminology:
* **Serialized block** - bytes that go into storage. For block-based table
(usually the case) this includes the block trailer. WART: block `size` may or
may not include the trailer; need to be clear about whether it does or not.
* **Maybe compressed block** - like a serialized block, but without the
trailer (or no promise of including a trailer). Must be accompanied by a
CompressionType.
* **Uncompressed block** - "payload" bytes that are either stored with no
compression, used as input to compression function, or result of
decompression function.
* **Parsed block** - an in-memory form of a block in block cache, as it is
used by the table reader. Different C++ types are used depending on the
block type (see block_like_traits.h).
Other refactorings:
* Misc corrections/improvements of internal API comments
* Remove a few misleading / unhelpful / redundant comments.
* Use move semantics in some places to simplify contracts
* Use better parameter names to indicate which parameters are used for
outputs
* Remove some extraneous `extern`
* Various clean-ups to `CacheDumperImpl` (mostly unnecessary code)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10408
Test Plan: existing tests
Reviewed By: akankshamahajan15
Differential Revision: D38172617
Pulled By: pdillinger
fbshipit-source-id: ccb99299f324ac5ca46996d34c5089621a4f260c
2 years ago
|
|
|
contents_comp_type = block_fetcher.get_compression_type();
|
|
|
|
contents = &tmp_contents;
|
|
|
|
if (get_context) {
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
switch (TBlocklike::kBlockType) {
|
|
|
|
case BlockType::kIndex:
|
|
|
|
++get_context->get_context_stats_.num_index_read;
|
|
|
|
break;
|
|
|
|
case BlockType::kFilter:
|
|
|
|
case BlockType::kFilterPartitionIndex:
|
|
|
|
++get_context->get_context_stats_.num_filter_read;
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
} else {
|
Refactor to avoid confusing "raw block" (#10408)
Summary:
We have a lot of confusing code because of mixed, sometimes
completely opposite uses of of the term "raw block" or "raw contents",
sometimes within the same source file. For example, in `BlockBasedTableBuilder`,
`raw_block_contents` and `raw_size` generally referred to uncompressed block
contents and size, while `WriteRawBlock` referred to writing a block that
is already compressed if it is going to be. Meanwhile, in
`BlockBasedTable`, `raw_block_contents` either referred to a (maybe
compressed) block with trailer, or a maybe compressed block maybe
without trailer. (Note: left as follow-up work to use C++ typing to
better sort out the various kinds of BlockContents.)
This change primarily tries to apply some consistent terminology around
the kinds of block representations, avoiding the unclear "raw". (Any
meaning of "raw" assumes some bias toward the storage layer or toward
the logical data layer.) Preferred terminology:
* **Serialized block** - bytes that go into storage. For block-based table
(usually the case) this includes the block trailer. WART: block `size` may or
may not include the trailer; need to be clear about whether it does or not.
* **Maybe compressed block** - like a serialized block, but without the
trailer (or no promise of including a trailer). Must be accompanied by a
CompressionType.
* **Uncompressed block** - "payload" bytes that are either stored with no
compression, used as input to compression function, or result of
decompression function.
* **Parsed block** - an in-memory form of a block in block cache, as it is
used by the table reader. Different C++ types are used depending on the
block type (see block_like_traits.h).
Other refactorings:
* Misc corrections/improvements of internal API comments
* Remove a few misleading / unhelpful / redundant comments.
* Use move semantics in some places to simplify contracts
* Use better parameter names to indicate which parameters are used for
outputs
* Remove some extraneous `extern`
* Various clean-ups to `CacheDumperImpl` (mostly unnecessary code)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10408
Test Plan: existing tests
Reviewed By: akankshamahajan15
Differential Revision: D38172617
Pulled By: pdillinger
fbshipit-source-id: ccb99299f324ac5ca46996d34c5089621a4f260c
2 years ago
|
|
|
contents_comp_type = GetBlockCompressionType(*contents);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (s.ok()) {
|
|
|
|
// If filling cache is allowed and a cache is configured, try to put the
|
|
|
|
// block to the cache.
|
|
|
|
s = PutDataBlockToCache(
|
|
|
|
key, block_cache, out_parsed_block, std::move(*contents),
|
|
|
|
contents_comp_type, uncompression_dict,
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
GetMemoryAllocator(rep_->table_options), get_context);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
Refactor block cache tracing w/improved MultiGet (#11339)
Summary:
After https://github.com/facebook/rocksdb/issues/11301, I wasn't sure whether I had regressed block cache tracing with MultiGet. Demo PR https://github.com/facebook/rocksdb/issues/11330 shows the flawed state of tracing MultiGet before my change, and based on the unit test, there was essentially no change in tracing behavior with https://github.com/facebook/rocksdb/issues/11301. This change is to leave that code and behavior better than I found it.
This change is not intended to change any production behaviors except when block cache tracing is active, though might improve general read path efficiency by disabling some related tracking when such tracing is disabled.
More detail on production code:
* Refactoring to consolidate the construction of BlockCacheTraceRecord, and other related functionality, in block-based table reader, though it's somewhat awkward to preserve an optimization to avoid copying Slices into temporary strings in BlockCacheLookupContext.
* Accurately track cache hits and misses (etc.) for each data block accessed by a MultiGet(). (Previously reported hits as misses.)
* Reduced repeated checking of `block_cache_tracer_` state (by creating lookup_context only when active) for efficiency and to reduce the risk of corner case bugs where tracing is enabled or disabled for different parts of a read op. (See a TODO below)
* Improved estimate calculation for num_keys_in_block (see code comment)
Possible follow-up:
* `XXX:` use_cache=true means double cache query? (possible double-query of block cache when allow_mmap_reads=true)
* `TODO:` need more than one lookup_context here to track individual filter and index partition hits and misses
* `TODO:` optimize more state checks of `block_cache_tracer_` down to `lookup_context != nullptr`
* Pre-existing `XXX:` There appear to be 'break' statements above that bypass this writing of the block cache trace record
* Expand test coverage (see below)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11339
Test Plan:
* Added a basic unit test for block cache tracing MultiGet, for now just covering one data block with two keys.
* Added HitMissCountingCache to independently verify that the actual block cache trace and expected block cache trace also agree with the actual number of cache hits / misses (nothing missing or mislabeled). For now only used with MultiGet test.
* Better testing of num_keys_in_block, for now just with MultiGet
* Misc improvements to table_test to improve clarity, such as making it clear that certain keys are auto-inserted at the start of every test.
Performance test:
Testing multireadrandom as in https://github.com/facebook/rocksdb/issues/11301, except averaging over distinct runs rather than [-X30] which doesn't seem to sufficiently reset after each run to work as an independent test run.
Base with revert of 11301: 3148926 ops/sec
Base: 3019146 ops/sec
New: 2999529 ops/sec
Possibly a tiny MultiGet CPU regression with this change. We are now always allocating an additional vector for the LookupContexts. I'm still contemplating options to try to correct the regression in https://github.com/facebook/rocksdb/issues/11301.
Testing readrandom:
Base with revert of 11301: 2311988
Base: 2281726
New: 2299722
Possibly a tiny Get CPU improvement with this change. We are now avoiding some unnecessary LookupContext population.
Reviewed By: akankshamahajan15
Differential Revision: D44557845
Pulled By: pdillinger
fbshipit-source-id: b841691799d2a48fb59cc8880dc7cbb1e107ae3d
2 years ago
|
|
|
// TODO: optimize so that lookup_context != nullptr implies the others
|
|
|
|
if (block_cache_tracer_ && block_cache_tracer_->is_tracing_enabled() &&
|
|
|
|
lookup_context) {
|
Refactor block cache tracing w/improved MultiGet (#11339)
Summary:
After https://github.com/facebook/rocksdb/issues/11301, I wasn't sure whether I had regressed block cache tracing with MultiGet. Demo PR https://github.com/facebook/rocksdb/issues/11330 shows the flawed state of tracing MultiGet before my change, and based on the unit test, there was essentially no change in tracing behavior with https://github.com/facebook/rocksdb/issues/11301. This change is to leave that code and behavior better than I found it.
This change is not intended to change any production behaviors except when block cache tracing is active, though might improve general read path efficiency by disabling some related tracking when such tracing is disabled.
More detail on production code:
* Refactoring to consolidate the construction of BlockCacheTraceRecord, and other related functionality, in block-based table reader, though it's somewhat awkward to preserve an optimization to avoid copying Slices into temporary strings in BlockCacheLookupContext.
* Accurately track cache hits and misses (etc.) for each data block accessed by a MultiGet(). (Previously reported hits as misses.)
* Reduced repeated checking of `block_cache_tracer_` state (by creating lookup_context only when active) for efficiency and to reduce the risk of corner case bugs where tracing is enabled or disabled for different parts of a read op. (See a TODO below)
* Improved estimate calculation for num_keys_in_block (see code comment)
Possible follow-up:
* `XXX:` use_cache=true means double cache query? (possible double-query of block cache when allow_mmap_reads=true)
* `TODO:` need more than one lookup_context here to track individual filter and index partition hits and misses
* `TODO:` optimize more state checks of `block_cache_tracer_` down to `lookup_context != nullptr`
* Pre-existing `XXX:` There appear to be 'break' statements above that bypass this writing of the block cache trace record
* Expand test coverage (see below)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11339
Test Plan:
* Added a basic unit test for block cache tracing MultiGet, for now just covering one data block with two keys.
* Added HitMissCountingCache to independently verify that the actual block cache trace and expected block cache trace also agree with the actual number of cache hits / misses (nothing missing or mislabeled). For now only used with MultiGet test.
* Better testing of num_keys_in_block, for now just with MultiGet
* Misc improvements to table_test to improve clarity, such as making it clear that certain keys are auto-inserted at the start of every test.
Performance test:
Testing multireadrandom as in https://github.com/facebook/rocksdb/issues/11301, except averaging over distinct runs rather than [-X30] which doesn't seem to sufficiently reset after each run to work as an independent test run.
Base with revert of 11301: 3148926 ops/sec
Base: 3019146 ops/sec
New: 2999529 ops/sec
Possibly a tiny MultiGet CPU regression with this change. We are now always allocating an additional vector for the LookupContexts. I'm still contemplating options to try to correct the regression in https://github.com/facebook/rocksdb/issues/11301.
Testing readrandom:
Base with revert of 11301: 2311988
Base: 2281726
New: 2299722
Possibly a tiny Get CPU improvement with this change. We are now avoiding some unnecessary LookupContext population.
Reviewed By: akankshamahajan15
Differential Revision: D44557845
Pulled By: pdillinger
fbshipit-source-id: b841691799d2a48fb59cc8880dc7cbb1e107ae3d
2 years ago
|
|
|
SaveLookupContextOrTraceRecord(
|
|
|
|
key, is_cache_hit, ro, out_parsed_block->GetValue(), lookup_context);
|
|
|
|
}
|
|
|
|
|
Refactor to avoid confusing "raw block" (#10408)
Summary:
We have a lot of confusing code because of mixed, sometimes
completely opposite uses of of the term "raw block" or "raw contents",
sometimes within the same source file. For example, in `BlockBasedTableBuilder`,
`raw_block_contents` and `raw_size` generally referred to uncompressed block
contents and size, while `WriteRawBlock` referred to writing a block that
is already compressed if it is going to be. Meanwhile, in
`BlockBasedTable`, `raw_block_contents` either referred to a (maybe
compressed) block with trailer, or a maybe compressed block maybe
without trailer. (Note: left as follow-up work to use C++ typing to
better sort out the various kinds of BlockContents.)
This change primarily tries to apply some consistent terminology around
the kinds of block representations, avoiding the unclear "raw". (Any
meaning of "raw" assumes some bias toward the storage layer or toward
the logical data layer.) Preferred terminology:
* **Serialized block** - bytes that go into storage. For block-based table
(usually the case) this includes the block trailer. WART: block `size` may or
may not include the trailer; need to be clear about whether it does or not.
* **Maybe compressed block** - like a serialized block, but without the
trailer (or no promise of including a trailer). Must be accompanied by a
CompressionType.
* **Uncompressed block** - "payload" bytes that are either stored with no
compression, used as input to compression function, or result of
decompression function.
* **Parsed block** - an in-memory form of a block in block cache, as it is
used by the table reader. Different C++ types are used depending on the
block type (see block_like_traits.h).
Other refactorings:
* Misc corrections/improvements of internal API comments
* Remove a few misleading / unhelpful / redundant comments.
* Use move semantics in some places to simplify contracts
* Use better parameter names to indicate which parameters are used for
outputs
* Remove some extraneous `extern`
* Various clean-ups to `CacheDumperImpl` (mostly unnecessary code)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10408
Test Plan: existing tests
Reviewed By: akankshamahajan15
Differential Revision: D38172617
Pulled By: pdillinger
fbshipit-source-id: ccb99299f324ac5ca46996d34c5089621a4f260c
2 years ago
|
|
|
assert(s.ok() || out_parsed_block->GetValue() == nullptr);
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
Refactor block cache tracing w/improved MultiGet (#11339)
Summary:
After https://github.com/facebook/rocksdb/issues/11301, I wasn't sure whether I had regressed block cache tracing with MultiGet. Demo PR https://github.com/facebook/rocksdb/issues/11330 shows the flawed state of tracing MultiGet before my change, and based on the unit test, there was essentially no change in tracing behavior with https://github.com/facebook/rocksdb/issues/11301. This change is to leave that code and behavior better than I found it.
This change is not intended to change any production behaviors except when block cache tracing is active, though might improve general read path efficiency by disabling some related tracking when such tracing is disabled.
More detail on production code:
* Refactoring to consolidate the construction of BlockCacheTraceRecord, and other related functionality, in block-based table reader, though it's somewhat awkward to preserve an optimization to avoid copying Slices into temporary strings in BlockCacheLookupContext.
* Accurately track cache hits and misses (etc.) for each data block accessed by a MultiGet(). (Previously reported hits as misses.)
* Reduced repeated checking of `block_cache_tracer_` state (by creating lookup_context only when active) for efficiency and to reduce the risk of corner case bugs where tracing is enabled or disabled for different parts of a read op. (See a TODO below)
* Improved estimate calculation for num_keys_in_block (see code comment)
Possible follow-up:
* `XXX:` use_cache=true means double cache query? (possible double-query of block cache when allow_mmap_reads=true)
* `TODO:` need more than one lookup_context here to track individual filter and index partition hits and misses
* `TODO:` optimize more state checks of `block_cache_tracer_` down to `lookup_context != nullptr`
* Pre-existing `XXX:` There appear to be 'break' statements above that bypass this writing of the block cache trace record
* Expand test coverage (see below)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11339
Test Plan:
* Added a basic unit test for block cache tracing MultiGet, for now just covering one data block with two keys.
* Added HitMissCountingCache to independently verify that the actual block cache trace and expected block cache trace also agree with the actual number of cache hits / misses (nothing missing or mislabeled). For now only used with MultiGet test.
* Better testing of num_keys_in_block, for now just with MultiGet
* Misc improvements to table_test to improve clarity, such as making it clear that certain keys are auto-inserted at the start of every test.
Performance test:
Testing multireadrandom as in https://github.com/facebook/rocksdb/issues/11301, except averaging over distinct runs rather than [-X30] which doesn't seem to sufficiently reset after each run to work as an independent test run.
Base with revert of 11301: 3148926 ops/sec
Base: 3019146 ops/sec
New: 2999529 ops/sec
Possibly a tiny MultiGet CPU regression with this change. We are now always allocating an additional vector for the LookupContexts. I'm still contemplating options to try to correct the regression in https://github.com/facebook/rocksdb/issues/11301.
Testing readrandom:
Base with revert of 11301: 2311988
Base: 2281726
New: 2299722
Possibly a tiny Get CPU improvement with this change. We are now avoiding some unnecessary LookupContext population.
Reviewed By: akankshamahajan15
Differential Revision: D44557845
Pulled By: pdillinger
fbshipit-source-id: b841691799d2a48fb59cc8880dc7cbb1e107ae3d
2 years ago
|
|
|
template <typename TBlocklike>
|
|
|
|
WithBlocklikeCheck<void, TBlocklike>
|
|
|
|
BlockBasedTable::SaveLookupContextOrTraceRecord(
|
|
|
|
const Slice& block_key, bool is_cache_hit, const ReadOptions& ro,
|
|
|
|
const TBlocklike* parsed_block_value,
|
|
|
|
BlockCacheLookupContext* lookup_context) const {
|
|
|
|
assert(lookup_context);
|
|
|
|
size_t usage = 0;
|
|
|
|
uint64_t nkeys = 0;
|
|
|
|
if (parsed_block_value) {
|
|
|
|
// Approximate the number of keys in the block using restarts.
|
|
|
|
int interval = rep_->table_options.block_restart_interval;
|
|
|
|
nkeys = interval * GetBlockNumRestarts(*parsed_block_value);
|
|
|
|
// On average, the last restart should be just over half utilized.
|
|
|
|
// Specifically, 1..N should be N/2 + 0.5. For example, 7 -> 4, 8 -> 4.5.
|
|
|
|
// Use the get_id to alternate between rounding up vs. down.
|
|
|
|
if (nkeys > 0) {
|
|
|
|
bool rounding = static_cast<int>(lookup_context->get_id) & 1;
|
|
|
|
nkeys -= (interval - rounding) / 2;
|
|
|
|
}
|
|
|
|
usage = parsed_block_value->ApproximateMemoryUsage();
|
|
|
|
}
|
|
|
|
TraceType trace_block_type = TraceType::kTraceMax;
|
|
|
|
switch (TBlocklike::kBlockType) {
|
|
|
|
case BlockType::kData:
|
|
|
|
trace_block_type = TraceType::kBlockTraceDataBlock;
|
|
|
|
break;
|
|
|
|
case BlockType::kFilter:
|
|
|
|
case BlockType::kFilterPartitionIndex:
|
|
|
|
trace_block_type = TraceType::kBlockTraceFilterBlock;
|
|
|
|
break;
|
|
|
|
case BlockType::kCompressionDictionary:
|
|
|
|
trace_block_type = TraceType::kBlockTraceUncompressionDictBlock;
|
|
|
|
break;
|
|
|
|
case BlockType::kRangeDeletion:
|
|
|
|
trace_block_type = TraceType::kBlockTraceRangeDeletionBlock;
|
|
|
|
break;
|
|
|
|
case BlockType::kIndex:
|
|
|
|
trace_block_type = TraceType::kBlockTraceIndexBlock;
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
// This cannot happen.
|
|
|
|
assert(false);
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
const bool no_io = ro.read_tier == kBlockCacheTier;
|
|
|
|
bool no_insert = no_io || !ro.fill_cache;
|
|
|
|
if (BlockCacheTraceHelper::IsGetOrMultiGetOnDataBlock(
|
|
|
|
trace_block_type, lookup_context->caller)) {
|
|
|
|
// Make a copy of the block key here since it will be logged later.
|
|
|
|
lookup_context->FillLookupContext(is_cache_hit, no_insert, trace_block_type,
|
|
|
|
/*block_size=*/usage,
|
|
|
|
block_key.ToString(), nkeys);
|
|
|
|
|
|
|
|
// Defer logging the access to Get() and MultiGet() to trace additional
|
|
|
|
// information, e.g., referenced_key
|
|
|
|
} else {
|
|
|
|
// Avoid making copy of block_key if it doesn't need to be saved in
|
|
|
|
// BlockCacheLookupContext
|
|
|
|
lookup_context->FillLookupContext(is_cache_hit, no_insert, trace_block_type,
|
|
|
|
/*block_size=*/usage,
|
|
|
|
/*block_key=*/{}, nkeys);
|
|
|
|
|
|
|
|
// Fill in default values for irrelevant/unknown fields
|
|
|
|
FinishTraceRecord(*lookup_context, block_key,
|
|
|
|
lookup_context->referenced_key,
|
|
|
|
/*does_referenced_key_exist*/ false,
|
|
|
|
/*referenced_data_size*/ 0);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void BlockBasedTable::FinishTraceRecord(
|
|
|
|
const BlockCacheLookupContext& lookup_context, const Slice& block_key,
|
|
|
|
const Slice& referenced_key, bool does_referenced_key_exist,
|
|
|
|
uint64_t referenced_data_size) const {
|
|
|
|
// Avoid making copy of referenced_key if it doesn't need to be saved in
|
|
|
|
// BlockCacheLookupContext
|
|
|
|
BlockCacheTraceRecord access_record(
|
|
|
|
rep_->ioptions.clock->NowMicros(),
|
|
|
|
/*block_key=*/"", lookup_context.block_type, lookup_context.block_size,
|
|
|
|
rep_->cf_id_for_tracing(),
|
|
|
|
/*cf_name=*/"", rep_->level_for_tracing(), rep_->sst_number_for_tracing(),
|
|
|
|
lookup_context.caller, lookup_context.is_cache_hit,
|
|
|
|
lookup_context.no_insert, lookup_context.get_id,
|
|
|
|
lookup_context.get_from_user_specified_snapshot,
|
|
|
|
/*referenced_key=*/"", referenced_data_size,
|
|
|
|
lookup_context.num_keys_in_block, does_referenced_key_exist);
|
|
|
|
// TODO: Should handle status here?
|
|
|
|
block_cache_tracer_
|
|
|
|
->WriteBlockAccess(access_record, block_key, rep_->cf_name_for_tracing(),
|
|
|
|
referenced_key)
|
|
|
|
.PermitUncheckedError();
|
|
|
|
}
|
|
|
|
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
template <typename TBlocklike /*, auto*/>
|
|
|
|
WithBlocklikeCheck<Status, TBlocklike> BlockBasedTable::RetrieveBlock(
|
|
|
|
FilePrefetchBuffer* prefetch_buffer, const ReadOptions& ro,
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
const BlockHandle& handle, const UncompressionDict& uncompression_dict,
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
CachableEntry<TBlocklike>* out_parsed_block, GetContext* get_context,
|
|
|
|
BlockCacheLookupContext* lookup_context, bool for_compaction,
|
HyperClockCache support for SecondaryCache, with refactoring (#11301)
Summary:
Internally refactors SecondaryCache integration out of LRUCache specifically and into a wrapper/adapter class that works with various Cache implementations. Notably, this relies on separating the notion of async lookup handles from other cache handles, so that HyperClockCache doesn't have to deal with the problem of allocating handles from the hash table for lookups that might fail anyway, and might be on the same key without support for coalescing. (LRUCache's hash table can incorporate previously allocated handles thanks to its pointer indirection.) Specifically, I'm worried about the case in which hundreds of threads try to access the same block and probing in the hash table degrades to linear search on the pile of entries with the same key.
This change is a big step in the direction of supporting stacked SecondaryCaches, but there are obstacles to completing that. Especially, there is no SecondaryCache hook for evictions to pass from one to the next. It has been proposed that evictions be transmitted simply as the persisted data (as in SaveToCallback), but given the current structure provided by the CacheItemHelpers, that would require an extra copy of the block data, because there's intentionally no way to ask for a contiguous Slice of the data (to allow for flexibility in storage). `AsyncLookupHandle` and the re-worked `WaitAll()` should be essentially prepared for stacked SecondaryCaches, but several "TODO with stacked secondaries" issues remain in various places.
It could be argued that the stacking instead be done as a SecondaryCache adapter that wraps two (or more) SecondaryCaches, but at least with the current API that would require an extra heap allocation on SecondaryCache Lookup for a wrapper SecondaryCacheResultHandle that can transfer a Lookup between secondaries. We could also consider trying to unify the Cache and SecondaryCache APIs, though that might be difficult if `AsyncLookupHandle` is kept a fixed struct.
## cache.h (public API)
Moves `secondary_cache` option from LRUCacheOptions to ShardedCacheOptions so that it is applicable to HyperClockCache.
## advanced_cache.h (advanced public API)
* Add `Cache::CreateStandalone()` so that the SecondaryCache support wrapper can use it.
* Add `SetEvictionCallback()` / `eviction_callback_` so that the SecondaryCache support wrapper can use it. Only a single callback is supported for efficiency. If there is ever a need for more than one, hopefully that can be handled with a broadcast callback wrapper.
These are essentially the two "extra" pieces of `Cache` for pulling out specific SecondaryCache support from the `Cache` implementation. I think it's a good trade-off as these are reasonable, limited, and reusable "cut points" into the `Cache` implementations.
* Remove async capability from standard `Lookup()` (getting rid of awkward restrictions on pending Handles) and add `AsyncLookupHandle` and `StartAsyncLookup()`. As noted in the comments, the full struct of `AsyncLookupHandle` is exposed so that it can be stack allocated, for efficiency, though more data is being copied around than before, which could impact performance. (Lookup info -> AsyncLookupHandle -> Handle vs. Lookup info -> Handle)
I could foresee a future in which a Cache internally saves a pointer to the AsyncLookupHandle, which means it's dangerous to allow it to be copyable or even movable. It also means it's not compatible with std::vector (which I don't like requiring as an API parameter anyway), so `WaitAll()` expects any contiguous array of AsyncLookupHandles. I believe this is best for common case efficiency, while behaving well in other cases also. For example, `WaitAll()` has no effect on default-constructed AsyncLookupHandles, which look like a completed cache miss.
## cacheable_entry.h
A couple of functions are obsolete because Cache::Handle can no longer be pending.
## cache.cc
Provides default implementations for new or revamped Cache functions, especially appropriate for non-blocking caches.
## secondary_cache_adapter.{h,cc}
The full details of the Cache wrapper adding SecondaryCache support. Essentially replicates the SecondaryCache handling that was in LRUCache, but obviously refactored. There is a bit of logic duplication, where Lookup() is essentially a manually optimized version of StartAsyncLookup() and Wait(), but it's roughly a dozen lines of code.
## sharded_cache.h, typed_cache.h, charged_cache.{h,cc}, sim_cache.cc
Simply updated for Cache API changes.
## lru_cache.{h,cc}
Carefully remove SecondaryCache logic, implement `CreateStandalone` and eviction handler functionality.
## clock_cache.{h,cc}
Expose existing `CreateStandalone` functionality, add eviction handler functionality. Light refactoring.
## block_based_table_reader*
Mostly re-worked the only usage of async Lookup, which is in BlockBasedTable::MultiGet. Used arrays in place of autovector in some places for efficiency. Simplified some logic by not trying to process some cache results before they're all ready.
Created new function `BlockBasedTable::GetCachePriority()` to reduce some pre-existing code duplication (and avoid making it worse).
Fixed at least one small bug from the prior confusing mixture of async and sync Lookups. In MaybeReadBlockAndLoadToCache(), called by RetrieveBlock(), called by MultiGet() with wait=false, is_cache_hit for the block_cache_tracer entry would not be set to true if the handle was pending after Lookup and before Wait.
## Intended follow-up work
* Figure out if there are any missing stats or block_cache_tracer work in refactored BlockBasedTable::MultiGet
* Stacked secondary caches (see above discussion)
* See if we can make up for the small MultiGet performance regression.
* Study more performance with SecondaryCache
* Items evicted from over-full LRUCache in Release were not being demoted to SecondaryCache, and still aren't to minimize unit test churn. Ideally they would be demoted, but it's an exceptional case so not a big deal.
* Use CreateStandalone for cache reservations (save unnecessary hash table operations). Not a big deal, but worthy cleanup.
* Somehow I got the contract for SecondaryCache::Insert wrong in #10945. (Doesn't take ownership!) That API comment needs to be fixed, but didn't want to mingle that in here.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11301
Test Plan:
## Unit tests
Generally updated to include HCC in SecondaryCache tests, though HyperClockCache has some different, less strict behaviors that leads to some tests not really being set up to work with it. Some of the tests remain disabled with it, but I think we have good coverage without them.
## Crash/stress test
Updated to use the new combination.
## Performance
First, let's check for regression on caches without secondary cache configured. Adding support for the eviction callback is likely to have a tiny effect, but it shouldn't be worrisome. LRUCache could benefit slightly from less logic around SecondaryCache handling. We can test with cache_bench default settings, built with DEBUG_LEVEL=0 and PORTABLE=0.
```
(while :; do base/cache_bench --cache_type=hyper_clock_cache | grep Rough; done) | awk '{ sum += $9; count++; print $0; print "Average: " int(sum / count) }'
```
**Before** this and #11299 (which could also have a small effect), running for about an hour, before & after running concurrently for each cache type:
HyperClockCache: 3168662 (average parallel ops/sec)
LRUCache: 2940127
**After** this and #11299, running for about an hour:
HyperClockCache: 3164862 (average parallel ops/sec) (0.12% slower)
LRUCache: 2940928 (0.03% faster)
This is an acceptable difference IMHO.
Next, let's consider essentially the worst case of new CPU overhead affecting overall performance. MultiGet uses the async lookup interface regardless of whether SecondaryCache or folly are used. We can configure a benchmark where all block cache queries are for data blocks, and all are hits.
Create DB and test (before and after tests running simultaneously):
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16
TEST_TMPDIR=/dev/shm base/db_bench -benchmarks=multireadrandom[-X30] -readonly -multiread_batched -batch_size=32 -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16
```
**Before**:
multireadrandom [AVG 30 runs] : 3444202 (± 57049) ops/sec; 240.9 (± 4.0) MB/sec
multireadrandom [MEDIAN 30 runs] : 3514443 ops/sec; 245.8 MB/sec
**After**:
multireadrandom [AVG 30 runs] : 3291022 (± 58851) ops/sec; 230.2 (± 4.1) MB/sec
multireadrandom [MEDIAN 30 runs] : 3366179 ops/sec; 235.4 MB/sec
So that's roughly a 3% regression, on kind of a *worst case* test of MultiGet CPU. Similar story with HyperClockCache:
**Before**:
multireadrandom [AVG 30 runs] : 3933777 (± 41840) ops/sec; 275.1 (± 2.9) MB/sec
multireadrandom [MEDIAN 30 runs] : 3970667 ops/sec; 277.7 MB/sec
**After**:
multireadrandom [AVG 30 runs] : 3755338 (± 30391) ops/sec; 262.6 (± 2.1) MB/sec
multireadrandom [MEDIAN 30 runs] : 3785696 ops/sec; 264.8 MB/sec
Roughly a 4-5% regression. Not ideal, but not the whole story, fortunately.
Let's also look at Get() in db_bench:
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X30] -readonly -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16
```
**Before**:
readrandom [AVG 30 runs] : 2198685 (± 13412) ops/sec; 153.8 (± 0.9) MB/sec
readrandom [MEDIAN 30 runs] : 2209498 ops/sec; 154.5 MB/sec
**After**:
readrandom [AVG 30 runs] : 2292814 (± 43508) ops/sec; 160.3 (± 3.0) MB/sec
readrandom [MEDIAN 30 runs] : 2365181 ops/sec; 165.4 MB/sec
That's showing roughly a 4% improvement, perhaps because of the secondary cache code that is no longer part of LRUCache. But weirdly, HyperClockCache is also showing 2-3% improvement:
**Before**:
readrandom [AVG 30 runs] : 2272333 (± 9992) ops/sec; 158.9 (± 0.7) MB/sec
readrandom [MEDIAN 30 runs] : 2273239 ops/sec; 159.0 MB/sec
**After**:
readrandom [AVG 30 runs] : 2332407 (± 11252) ops/sec; 163.1 (± 0.8) MB/sec
readrandom [MEDIAN 30 runs] : 2335329 ops/sec; 163.3 MB/sec
Reviewed By: ltamasi
Differential Revision: D44177044
Pulled By: pdillinger
fbshipit-source-id: e808e48ff3fe2f792a79841ba617be98e48689f5
2 years ago
|
|
|
bool use_cache, bool async_read) const {
|
Refactor to avoid confusing "raw block" (#10408)
Summary:
We have a lot of confusing code because of mixed, sometimes
completely opposite uses of of the term "raw block" or "raw contents",
sometimes within the same source file. For example, in `BlockBasedTableBuilder`,
`raw_block_contents` and `raw_size` generally referred to uncompressed block
contents and size, while `WriteRawBlock` referred to writing a block that
is already compressed if it is going to be. Meanwhile, in
`BlockBasedTable`, `raw_block_contents` either referred to a (maybe
compressed) block with trailer, or a maybe compressed block maybe
without trailer. (Note: left as follow-up work to use C++ typing to
better sort out the various kinds of BlockContents.)
This change primarily tries to apply some consistent terminology around
the kinds of block representations, avoiding the unclear "raw". (Any
meaning of "raw" assumes some bias toward the storage layer or toward
the logical data layer.) Preferred terminology:
* **Serialized block** - bytes that go into storage. For block-based table
(usually the case) this includes the block trailer. WART: block `size` may or
may not include the trailer; need to be clear about whether it does or not.
* **Maybe compressed block** - like a serialized block, but without the
trailer (or no promise of including a trailer). Must be accompanied by a
CompressionType.
* **Uncompressed block** - "payload" bytes that are either stored with no
compression, used as input to compression function, or result of
decompression function.
* **Parsed block** - an in-memory form of a block in block cache, as it is
used by the table reader. Different C++ types are used depending on the
block type (see block_like_traits.h).
Other refactorings:
* Misc corrections/improvements of internal API comments
* Remove a few misleading / unhelpful / redundant comments.
* Use move semantics in some places to simplify contracts
* Use better parameter names to indicate which parameters are used for
outputs
* Remove some extraneous `extern`
* Various clean-ups to `CacheDumperImpl` (mostly unnecessary code)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10408
Test Plan: existing tests
Reviewed By: akankshamahajan15
Differential Revision: D38172617
Pulled By: pdillinger
fbshipit-source-id: ccb99299f324ac5ca46996d34c5089621a4f260c
2 years ago
|
|
|
assert(out_parsed_block);
|
|
|
|
assert(out_parsed_block->IsEmpty());
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
|
|
|
|
Status s;
|
|
|
|
if (use_cache) {
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
s = MaybeReadBlockAndLoadToCache(
|
HyperClockCache support for SecondaryCache, with refactoring (#11301)
Summary:
Internally refactors SecondaryCache integration out of LRUCache specifically and into a wrapper/adapter class that works with various Cache implementations. Notably, this relies on separating the notion of async lookup handles from other cache handles, so that HyperClockCache doesn't have to deal with the problem of allocating handles from the hash table for lookups that might fail anyway, and might be on the same key without support for coalescing. (LRUCache's hash table can incorporate previously allocated handles thanks to its pointer indirection.) Specifically, I'm worried about the case in which hundreds of threads try to access the same block and probing in the hash table degrades to linear search on the pile of entries with the same key.
This change is a big step in the direction of supporting stacked SecondaryCaches, but there are obstacles to completing that. Especially, there is no SecondaryCache hook for evictions to pass from one to the next. It has been proposed that evictions be transmitted simply as the persisted data (as in SaveToCallback), but given the current structure provided by the CacheItemHelpers, that would require an extra copy of the block data, because there's intentionally no way to ask for a contiguous Slice of the data (to allow for flexibility in storage). `AsyncLookupHandle` and the re-worked `WaitAll()` should be essentially prepared for stacked SecondaryCaches, but several "TODO with stacked secondaries" issues remain in various places.
It could be argued that the stacking instead be done as a SecondaryCache adapter that wraps two (or more) SecondaryCaches, but at least with the current API that would require an extra heap allocation on SecondaryCache Lookup for a wrapper SecondaryCacheResultHandle that can transfer a Lookup between secondaries. We could also consider trying to unify the Cache and SecondaryCache APIs, though that might be difficult if `AsyncLookupHandle` is kept a fixed struct.
## cache.h (public API)
Moves `secondary_cache` option from LRUCacheOptions to ShardedCacheOptions so that it is applicable to HyperClockCache.
## advanced_cache.h (advanced public API)
* Add `Cache::CreateStandalone()` so that the SecondaryCache support wrapper can use it.
* Add `SetEvictionCallback()` / `eviction_callback_` so that the SecondaryCache support wrapper can use it. Only a single callback is supported for efficiency. If there is ever a need for more than one, hopefully that can be handled with a broadcast callback wrapper.
These are essentially the two "extra" pieces of `Cache` for pulling out specific SecondaryCache support from the `Cache` implementation. I think it's a good trade-off as these are reasonable, limited, and reusable "cut points" into the `Cache` implementations.
* Remove async capability from standard `Lookup()` (getting rid of awkward restrictions on pending Handles) and add `AsyncLookupHandle` and `StartAsyncLookup()`. As noted in the comments, the full struct of `AsyncLookupHandle` is exposed so that it can be stack allocated, for efficiency, though more data is being copied around than before, which could impact performance. (Lookup info -> AsyncLookupHandle -> Handle vs. Lookup info -> Handle)
I could foresee a future in which a Cache internally saves a pointer to the AsyncLookupHandle, which means it's dangerous to allow it to be copyable or even movable. It also means it's not compatible with std::vector (which I don't like requiring as an API parameter anyway), so `WaitAll()` expects any contiguous array of AsyncLookupHandles. I believe this is best for common case efficiency, while behaving well in other cases also. For example, `WaitAll()` has no effect on default-constructed AsyncLookupHandles, which look like a completed cache miss.
## cacheable_entry.h
A couple of functions are obsolete because Cache::Handle can no longer be pending.
## cache.cc
Provides default implementations for new or revamped Cache functions, especially appropriate for non-blocking caches.
## secondary_cache_adapter.{h,cc}
The full details of the Cache wrapper adding SecondaryCache support. Essentially replicates the SecondaryCache handling that was in LRUCache, but obviously refactored. There is a bit of logic duplication, where Lookup() is essentially a manually optimized version of StartAsyncLookup() and Wait(), but it's roughly a dozen lines of code.
## sharded_cache.h, typed_cache.h, charged_cache.{h,cc}, sim_cache.cc
Simply updated for Cache API changes.
## lru_cache.{h,cc}
Carefully remove SecondaryCache logic, implement `CreateStandalone` and eviction handler functionality.
## clock_cache.{h,cc}
Expose existing `CreateStandalone` functionality, add eviction handler functionality. Light refactoring.
## block_based_table_reader*
Mostly re-worked the only usage of async Lookup, which is in BlockBasedTable::MultiGet. Used arrays in place of autovector in some places for efficiency. Simplified some logic by not trying to process some cache results before they're all ready.
Created new function `BlockBasedTable::GetCachePriority()` to reduce some pre-existing code duplication (and avoid making it worse).
Fixed at least one small bug from the prior confusing mixture of async and sync Lookups. In MaybeReadBlockAndLoadToCache(), called by RetrieveBlock(), called by MultiGet() with wait=false, is_cache_hit for the block_cache_tracer entry would not be set to true if the handle was pending after Lookup and before Wait.
## Intended follow-up work
* Figure out if there are any missing stats or block_cache_tracer work in refactored BlockBasedTable::MultiGet
* Stacked secondary caches (see above discussion)
* See if we can make up for the small MultiGet performance regression.
* Study more performance with SecondaryCache
* Items evicted from over-full LRUCache in Release were not being demoted to SecondaryCache, and still aren't to minimize unit test churn. Ideally they would be demoted, but it's an exceptional case so not a big deal.
* Use CreateStandalone for cache reservations (save unnecessary hash table operations). Not a big deal, but worthy cleanup.
* Somehow I got the contract for SecondaryCache::Insert wrong in #10945. (Doesn't take ownership!) That API comment needs to be fixed, but didn't want to mingle that in here.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11301
Test Plan:
## Unit tests
Generally updated to include HCC in SecondaryCache tests, though HyperClockCache has some different, less strict behaviors that leads to some tests not really being set up to work with it. Some of the tests remain disabled with it, but I think we have good coverage without them.
## Crash/stress test
Updated to use the new combination.
## Performance
First, let's check for regression on caches without secondary cache configured. Adding support for the eviction callback is likely to have a tiny effect, but it shouldn't be worrisome. LRUCache could benefit slightly from less logic around SecondaryCache handling. We can test with cache_bench default settings, built with DEBUG_LEVEL=0 and PORTABLE=0.
```
(while :; do base/cache_bench --cache_type=hyper_clock_cache | grep Rough; done) | awk '{ sum += $9; count++; print $0; print "Average: " int(sum / count) }'
```
**Before** this and #11299 (which could also have a small effect), running for about an hour, before & after running concurrently for each cache type:
HyperClockCache: 3168662 (average parallel ops/sec)
LRUCache: 2940127
**After** this and #11299, running for about an hour:
HyperClockCache: 3164862 (average parallel ops/sec) (0.12% slower)
LRUCache: 2940928 (0.03% faster)
This is an acceptable difference IMHO.
Next, let's consider essentially the worst case of new CPU overhead affecting overall performance. MultiGet uses the async lookup interface regardless of whether SecondaryCache or folly are used. We can configure a benchmark where all block cache queries are for data blocks, and all are hits.
Create DB and test (before and after tests running simultaneously):
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16
TEST_TMPDIR=/dev/shm base/db_bench -benchmarks=multireadrandom[-X30] -readonly -multiread_batched -batch_size=32 -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16
```
**Before**:
multireadrandom [AVG 30 runs] : 3444202 (± 57049) ops/sec; 240.9 (± 4.0) MB/sec
multireadrandom [MEDIAN 30 runs] : 3514443 ops/sec; 245.8 MB/sec
**After**:
multireadrandom [AVG 30 runs] : 3291022 (± 58851) ops/sec; 230.2 (± 4.1) MB/sec
multireadrandom [MEDIAN 30 runs] : 3366179 ops/sec; 235.4 MB/sec
So that's roughly a 3% regression, on kind of a *worst case* test of MultiGet CPU. Similar story with HyperClockCache:
**Before**:
multireadrandom [AVG 30 runs] : 3933777 (± 41840) ops/sec; 275.1 (± 2.9) MB/sec
multireadrandom [MEDIAN 30 runs] : 3970667 ops/sec; 277.7 MB/sec
**After**:
multireadrandom [AVG 30 runs] : 3755338 (± 30391) ops/sec; 262.6 (± 2.1) MB/sec
multireadrandom [MEDIAN 30 runs] : 3785696 ops/sec; 264.8 MB/sec
Roughly a 4-5% regression. Not ideal, but not the whole story, fortunately.
Let's also look at Get() in db_bench:
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X30] -readonly -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16
```
**Before**:
readrandom [AVG 30 runs] : 2198685 (± 13412) ops/sec; 153.8 (± 0.9) MB/sec
readrandom [MEDIAN 30 runs] : 2209498 ops/sec; 154.5 MB/sec
**After**:
readrandom [AVG 30 runs] : 2292814 (± 43508) ops/sec; 160.3 (± 3.0) MB/sec
readrandom [MEDIAN 30 runs] : 2365181 ops/sec; 165.4 MB/sec
That's showing roughly a 4% improvement, perhaps because of the secondary cache code that is no longer part of LRUCache. But weirdly, HyperClockCache is also showing 2-3% improvement:
**Before**:
readrandom [AVG 30 runs] : 2272333 (± 9992) ops/sec; 158.9 (± 0.7) MB/sec
readrandom [MEDIAN 30 runs] : 2273239 ops/sec; 159.0 MB/sec
**After**:
readrandom [AVG 30 runs] : 2332407 (± 11252) ops/sec; 163.1 (± 0.8) MB/sec
readrandom [MEDIAN 30 runs] : 2335329 ops/sec; 163.3 MB/sec
Reviewed By: ltamasi
Differential Revision: D44177044
Pulled By: pdillinger
fbshipit-source-id: e808e48ff3fe2f792a79841ba617be98e48689f5
2 years ago
|
|
|
prefetch_buffer, ro, handle, uncompression_dict, for_compaction,
|
|
|
|
out_parsed_block, get_context, lookup_context,
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
/*contents=*/nullptr, async_read);
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
Refactor to avoid confusing "raw block" (#10408)
Summary:
We have a lot of confusing code because of mixed, sometimes
completely opposite uses of of the term "raw block" or "raw contents",
sometimes within the same source file. For example, in `BlockBasedTableBuilder`,
`raw_block_contents` and `raw_size` generally referred to uncompressed block
contents and size, while `WriteRawBlock` referred to writing a block that
is already compressed if it is going to be. Meanwhile, in
`BlockBasedTable`, `raw_block_contents` either referred to a (maybe
compressed) block with trailer, or a maybe compressed block maybe
without trailer. (Note: left as follow-up work to use C++ typing to
better sort out the various kinds of BlockContents.)
This change primarily tries to apply some consistent terminology around
the kinds of block representations, avoiding the unclear "raw". (Any
meaning of "raw" assumes some bias toward the storage layer or toward
the logical data layer.) Preferred terminology:
* **Serialized block** - bytes that go into storage. For block-based table
(usually the case) this includes the block trailer. WART: block `size` may or
may not include the trailer; need to be clear about whether it does or not.
* **Maybe compressed block** - like a serialized block, but without the
trailer (or no promise of including a trailer). Must be accompanied by a
CompressionType.
* **Uncompressed block** - "payload" bytes that are either stored with no
compression, used as input to compression function, or result of
decompression function.
* **Parsed block** - an in-memory form of a block in block cache, as it is
used by the table reader. Different C++ types are used depending on the
block type (see block_like_traits.h).
Other refactorings:
* Misc corrections/improvements of internal API comments
* Remove a few misleading / unhelpful / redundant comments.
* Use move semantics in some places to simplify contracts
* Use better parameter names to indicate which parameters are used for
outputs
* Remove some extraneous `extern`
* Various clean-ups to `CacheDumperImpl` (mostly unnecessary code)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10408
Test Plan: existing tests
Reviewed By: akankshamahajan15
Differential Revision: D38172617
Pulled By: pdillinger
fbshipit-source-id: ccb99299f324ac5ca46996d34c5089621a4f260c
2 years ago
|
|
|
if (out_parsed_block->GetValue() != nullptr ||
|
|
|
|
out_parsed_block->GetCacheHandle() != nullptr) {
|
|
|
|
assert(s.ok());
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
return s;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
Refactor to avoid confusing "raw block" (#10408)
Summary:
We have a lot of confusing code because of mixed, sometimes
completely opposite uses of of the term "raw block" or "raw contents",
sometimes within the same source file. For example, in `BlockBasedTableBuilder`,
`raw_block_contents` and `raw_size` generally referred to uncompressed block
contents and size, while `WriteRawBlock` referred to writing a block that
is already compressed if it is going to be. Meanwhile, in
`BlockBasedTable`, `raw_block_contents` either referred to a (maybe
compressed) block with trailer, or a maybe compressed block maybe
without trailer. (Note: left as follow-up work to use C++ typing to
better sort out the various kinds of BlockContents.)
This change primarily tries to apply some consistent terminology around
the kinds of block representations, avoiding the unclear "raw". (Any
meaning of "raw" assumes some bias toward the storage layer or toward
the logical data layer.) Preferred terminology:
* **Serialized block** - bytes that go into storage. For block-based table
(usually the case) this includes the block trailer. WART: block `size` may or
may not include the trailer; need to be clear about whether it does or not.
* **Maybe compressed block** - like a serialized block, but without the
trailer (or no promise of including a trailer). Must be accompanied by a
CompressionType.
* **Uncompressed block** - "payload" bytes that are either stored with no
compression, used as input to compression function, or result of
decompression function.
* **Parsed block** - an in-memory form of a block in block cache, as it is
used by the table reader. Different C++ types are used depending on the
block type (see block_like_traits.h).
Other refactorings:
* Misc corrections/improvements of internal API comments
* Remove a few misleading / unhelpful / redundant comments.
* Use move semantics in some places to simplify contracts
* Use better parameter names to indicate which parameters are used for
outputs
* Remove some extraneous `extern`
* Various clean-ups to `CacheDumperImpl` (mostly unnecessary code)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10408
Test Plan: existing tests
Reviewed By: akankshamahajan15
Differential Revision: D38172617
Pulled By: pdillinger
fbshipit-source-id: ccb99299f324ac5ca46996d34c5089621a4f260c
2 years ago
|
|
|
assert(out_parsed_block->IsEmpty());
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
|
|
|
|
const bool no_io = ro.read_tier == kBlockCacheTier;
|
|
|
|
if (no_io) {
|
|
|
|
return Status::Incomplete("no blocking io");
|
|
|
|
}
|
|
|
|
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
const bool maybe_compressed =
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
TBlocklike::kBlockType != BlockType::kFilter &&
|
|
|
|
TBlocklike::kBlockType != BlockType::kCompressionDictionary &&
|
|
|
|
rep_->blocks_maybe_compressed;
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
std::unique_ptr<TBlocklike> block;
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
|
|
|
|
{
|
|
|
|
Histograms histogram =
|
|
|
|
for_compaction ? READ_BLOCK_COMPACTION_MICROS : READ_BLOCK_GET_MICROS;
|
|
|
|
StopWatch sw(rep_->ioptions.clock, rep_->ioptions.stats, histogram);
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
s = ReadAndParseBlockFromFile(
|
|
|
|
rep_->file.get(), prefetch_buffer, rep_->footer, ro, handle, &block,
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
rep_->ioptions, rep_->create_context, maybe_compressed,
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
uncompression_dict, rep_->persistent_cache_options,
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
GetMemoryAllocator(rep_->table_options), for_compaction, async_read);
|
|
|
|
|
|
|
|
if (get_context) {
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
switch (TBlocklike::kBlockType) {
|
|
|
|
case BlockType::kIndex:
|
|
|
|
++(get_context->get_context_stats_.num_index_read);
|
|
|
|
break;
|
|
|
|
case BlockType::kFilter:
|
|
|
|
case BlockType::kFilterPartitionIndex:
|
|
|
|
++(get_context->get_context_stats_.num_filter_read);
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
}
|
|
|
|
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
Refactor to avoid confusing "raw block" (#10408)
Summary:
We have a lot of confusing code because of mixed, sometimes
completely opposite uses of of the term "raw block" or "raw contents",
sometimes within the same source file. For example, in `BlockBasedTableBuilder`,
`raw_block_contents` and `raw_size` generally referred to uncompressed block
contents and size, while `WriteRawBlock` referred to writing a block that
is already compressed if it is going to be. Meanwhile, in
`BlockBasedTable`, `raw_block_contents` either referred to a (maybe
compressed) block with trailer, or a maybe compressed block maybe
without trailer. (Note: left as follow-up work to use C++ typing to
better sort out the various kinds of BlockContents.)
This change primarily tries to apply some consistent terminology around
the kinds of block representations, avoiding the unclear "raw". (Any
meaning of "raw" assumes some bias toward the storage layer or toward
the logical data layer.) Preferred terminology:
* **Serialized block** - bytes that go into storage. For block-based table
(usually the case) this includes the block trailer. WART: block `size` may or
may not include the trailer; need to be clear about whether it does or not.
* **Maybe compressed block** - like a serialized block, but without the
trailer (or no promise of including a trailer). Must be accompanied by a
CompressionType.
* **Uncompressed block** - "payload" bytes that are either stored with no
compression, used as input to compression function, or result of
decompression function.
* **Parsed block** - an in-memory form of a block in block cache, as it is
used by the table reader. Different C++ types are used depending on the
block type (see block_like_traits.h).
Other refactorings:
* Misc corrections/improvements of internal API comments
* Remove a few misleading / unhelpful / redundant comments.
* Use move semantics in some places to simplify contracts
* Use better parameter names to indicate which parameters are used for
outputs
* Remove some extraneous `extern`
* Various clean-ups to `CacheDumperImpl` (mostly unnecessary code)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10408
Test Plan: existing tests
Reviewed By: akankshamahajan15
Differential Revision: D38172617
Pulled By: pdillinger
fbshipit-source-id: ccb99299f324ac5ca46996d34c5089621a4f260c
2 years ago
|
|
|
out_parsed_block->SetOwnedValue(std::move(block));
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
|
|
|
|
assert(s.ok());
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
BlockBasedTable::PartitionedIndexIteratorState::PartitionedIndexIteratorState(
|
|
|
|
const BlockBasedTable* table,
|
Meta-internal folly integration with F14FastMap (#9546)
Summary:
Especially after updating to C++17, I don't see a compelling case for
*requiring* any folly components in RocksDB. I was able to purge the existing
hard dependencies, and it can be quite difficult to strip out non-trivial components
from folly for use in RocksDB. (The prospect of doing that on F14 has changed
my mind on the best approach here.)
But this change creates an optional integration where we can plug in
components from folly at compile time, starting here with F14FastMap to replace
std::unordered_map when possible (probably no public APIs for example). I have
replaced the biggest CPU users of std::unordered_map with compile-time
pluggable UnorderedMap which will use F14FastMap when USE_FOLLY is set.
USE_FOLLY is always set in the Meta-internal buck build, and a simulation of
that is in the Makefile for public CI testing. A full folly build is not needed, but
checking out the full folly repo is much simpler for getting the dependency,
and anything else we might want to optionally integrate in the future.
Some picky details:
* I don't think the distributed mutex stuff is actually used, so it was easy to remove.
* I implemented an alternative to `folly::constexpr_log2` (which is much easier
in C++17 than C++11) so that I could pull out the hard dependencies on
`ConstexprMath.h`
* I had to add noexcept move constructors/operators to some types to make
F14's complainUnlessNothrowMoveAndDestroy check happy, and I added a
macro to make that easier in some common cases.
* Updated Meta-internal buck build to use folly F14Map (always)
No updates to HISTORY.md nor INSTALL.md as this is not (yet?) considered a
production integration for open source users.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9546
Test Plan:
CircleCI tests updated so that a couple of them use folly.
Most internal unit & stress/crash tests updated to use Meta-internal latest folly.
(Note: they should probably use buck but they currently use Makefile.)
Example performance improvement: when filter partitions are pinned in cache,
they are tracked by PartitionedFilterBlockReader::filter_map_ and we can build
a test that exercises that heavily. Build DB with
```
TEST_TMPDIR=/dev/shm/rocksdb ./db_bench -benchmarks=fillrandom -num=10000000 -disable_wal=1 -write_buffer_size=30000000 -bloom_bits=16 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0 -partition_index_and_filters
```
and test with (simultaneous runs with & without folly, ~20 times each to see
convergence)
```
TEST_TMPDIR=/dev/shm/rocksdb ./db_bench_folly -readonly -use_existing_db -benchmarks=readrandom -num=10000000 -bloom_bits=16 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0 -partition_index_and_filters -duration=40 -pin_l0_filter_and_index_blocks_in_cache
```
Average ops/s no folly: 26229.2
Average ops/s with folly: 26853.3 (+2.4%)
Reviewed By: ajkr
Differential Revision: D34181736
Pulled By: pdillinger
fbshipit-source-id: ffa6ad5104c2880321d8a1aa7187e00ab0d02e94
3 years ago
|
|
|
UnorderedMap<uint64_t, CachableEntry<Block>>* block_map)
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
: table_(table), block_map_(block_map) {}
|
|
|
|
|
|
|
|
InternalIteratorBase<IndexValue>*
|
|
|
|
BlockBasedTable::PartitionedIndexIteratorState::NewSecondaryIterator(
|
|
|
|
const BlockHandle& handle) {
|
|
|
|
// Return a block iterator on the index partition
|
|
|
|
auto block = block_map_->find(handle.offset());
|
|
|
|
// block_map_ must be exhaustive
|
|
|
|
if (block == block_map_->end()) {
|
|
|
|
assert(false);
|
|
|
|
// Signal problem to caller
|
|
|
|
return nullptr;
|
|
|
|
}
|
|
|
|
const Rep* rep = table_->get_rep();
|
|
|
|
assert(rep);
|
|
|
|
|
|
|
|
Statistics* kNullStats = nullptr;
|
|
|
|
// We don't return pinned data from index blocks, so no need
|
|
|
|
// to set `block_contents_pinned`.
|
|
|
|
return block->second.GetValue()->NewIndexIterator(
|
|
|
|
rep->internal_comparator.user_comparator(),
|
|
|
|
rep->get_global_seqno(BlockType::kIndex), nullptr, kNullStats, true,
|
|
|
|
rep->index_has_first_key, rep->index_key_includes_seq,
|
|
|
|
rep->index_value_is_full);
|
|
|
|
}
|
|
|
|
|
|
|
|
// This will be broken if the user specifies an unusual implementation
|
|
|
|
// of Options.comparator, or if the user specifies an unusual
|
|
|
|
// definition of prefixes in BlockBasedTableOptions.filter_policy.
|
|
|
|
// In particular, we require the following three properties:
|
|
|
|
//
|
|
|
|
// 1) key.starts_with(prefix(key))
|
|
|
|
// 2) Compare(prefix(key), key) <= 0.
|
|
|
|
// 3) If Compare(key1, key2) <= 0, then Compare(prefix(key1), prefix(key2)) <= 0
|
|
|
|
//
|
Fix iterator reading filter block despite read_tier == kBlockCacheTier (#6562)
Summary:
We're seeing iterators with `ReadOptions::read_tier == kBlockCacheTier` sometimes doing file reads. Stack trace:
```
rocksdb::RandomAccessFileReader::Read(unsigned long, unsigned long, rocksdb::Slice*, char*, bool) const
rocksdb::BlockFetcher::ReadBlockContents()
rocksdb::Status rocksdb::BlockBasedTable::MaybeReadBlockAndLoadToCache<rocksdb::ParsedFullFilterBlock>(rocksdb::FilePrefetchBuffer*, rocksdb::ReadOptions const&, rocksdb::BlockHandle const&, rocksdb::UncompressionDict const&, rocksdb::CachableEntry<rocksdb::ParsedFullFilterBlock>*, rocksdb::BlockType, rocksdb::GetContext*, rocksdb::BlockCacheLookupContext*, rocksdb::BlockContents*) const
rocksdb::Status rocksdb::BlockBasedTable::RetrieveBlock<rocksdb::ParsedFullFilterBlock>(rocksdb::FilePrefetchBuffer*, rocksdb::ReadOptions const&, rocksdb::BlockHandle const&, rocksdb::UncompressionDict const&, rocksdb::CachableEntry<rocksdb::ParsedFullFilterBlock>*, rocksdb::BlockType, rocksdb::GetContext*, rocksdb::BlockCacheLookupContext*, bool, bool) const
rocksdb::FilterBlockReaderCommon<rocksdb::ParsedFullFilterBlock>::ReadFilterBlock(rocksdb::BlockBasedTable const*, rocksdb::FilePrefetchBuffer*, rocksdb::ReadOptions const&, bool, rocksdb::GetContext*, rocksdb::BlockCacheLookupContext*, rocksdb::CachableEntry<rocksdb::ParsedFullFilterBlock>*)
rocksdb::FilterBlockReaderCommon<rocksdb::ParsedFullFilterBlock>::GetOrReadFilterBlock(bool, rocksdb::GetContext*, rocksdb::BlockCacheLookupContext*, rocksdb::CachableEntry<rocksdb::ParsedFullFilterBlock>*) const
rocksdb::FullFilterBlockReader::MayMatch(rocksdb::Slice const&, bool, rocksdb::GetContext*, rocksdb::BlockCacheLookupContext*) const
rocksdb::FullFilterBlockReader::RangeMayExist(rocksdb::Slice const*, rocksdb::Slice const&, rocksdb::SliceTransform const*, rocksdb::Comparator const*, rocksdb::Slice const*, bool*, bool, rocksdb::BlockCacheLookupContext*)
rocksdb::BlockBasedTable::PrefixMayMatch(rocksdb::Slice const&, rocksdb::ReadOptions const&, rocksdb::SliceTransform const*, bool, rocksdb::BlockCacheLookupContext*) const
rocksdb::BlockBasedTableIterator<rocksdb::DataBlockIter, rocksdb::Slice>::SeekImpl(rocksdb::Slice const*)
rocksdb::ForwardIterator::SeekInternal(rocksdb::Slice const&, bool)
rocksdb::DBIter::Seek(rocksdb::Slice const&)
```
`BlockBasedTableIterator::CheckPrefixMayMatch` was missing a check for `kBlockCacheTier`. This PR adds it.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6562
Test Plan: deployed it to a logdevice test cluster and looked at logdevice's IO tracing.
Reviewed By: siying
Differential Revision: D20529368
Pulled By: al13n321
fbshipit-source-id: 65bf33964b1951464415c900336635fb20919611
5 years ago
|
|
|
// If read_options.read_tier == kBlockCacheTier, this method will do no I/O and
|
|
|
|
// will return true if the filter block is not in memory and not found in block
|
|
|
|
// cache.
|
|
|
|
//
|
|
|
|
// REQUIRES: this method shouldn't be called while the DB lock is held.
|
|
|
|
bool BlockBasedTable::PrefixRangeMayMatch(
|
|
|
|
const Slice& internal_key, const ReadOptions& read_options,
|
|
|
|
const SliceTransform* options_prefix_extractor,
|
Much better stats for seeks and prefix filtering (#11460)
Summary:
We want to know more about opportunities for better range filters, and the effectiveness of our own range filters. Currently the stats are very limited, essentially logging just hits and misses against prefix filters for range scans in BLOOM_FILTER_PREFIX_* without tracking the false positive rate. Perhaps confusingly, when prefix filters are used for point queries, the stats are currently going into the non-PREFIX tickers.
This change does several things:
* Introduce new stat tickers for seeks and related filtering, \*LEVEL_SEEK\*
* Most importantly, allows us to see opportunities for range filtering. Specifically, we can count how many times a seek in an SST file accesses at least one data block, and how many times at least one value() is then accessed. If a data block was accessed but no value(), we can generally assume that the key(s) seen was(were) not of interest so could have been filtered with the right kind of filter, avoiding the data block access.
* We can get the same level of detail when a filter (for now, prefix Bloom/ribbon) is used, or not. Specifically, we can infer a false positive rate for prefix filters (not available before) from the seek "false positive" rate: when a data block is accessed but no value() is called. (There can be other explanations for a seek false positive, but in typical iterator usage it would indicate a filter false positive.)
* For efficiency, I wanted to avoid making additional calls to the prefix extractor (or key comparisons, etc.), which would be required if we wanted to more precisely detect filter false positives. I believe that instrumenting value() is the best balance of efficiency vs. accurately measuring what we are often interested in.
* The stats are divided between last level and non-last levels, to help understand potential tiered storage use cases.
* The old BLOOM_FILTER_PREFIX_* stats have a different meaning: no longer referring to iterators but to point queries using prefix filters. BLOOM_FILTER_PREFIX_TRUE_POSITIVE is added for computing the prefix false positive rate on point queries, which can be due to filter false positives as well as different keys with the same prefix.
* Similarly, the non-PREFIX BLOOM_FILTER stats are now for whole key filtering only.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11460
Test Plan:
unit tests updated, including updating many to pop the stat value since last read to improve test
readability and maintainability.
Performance test shows a consistent small improvement with these changes, both with clang and with gcc. CPU profile indicates that RecordTick is using less CPU, and this makes sense at least for a high filter miss rate. Before, we were recording two ticks per filter miss in iterators (CHECKED & USEFUL) and now recording just one (FILTERED).
Create DB with
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=10000000 -disable_wal=1 -write_buffer_size=30000000 -bloom_bits=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0 -prefix_size=8
```
And run simultaneous before&after with
```
TEST_TMPDIR=/dev/shm ./db_bench -readonly -benchmarks=seekrandom[-X1000] -num=10000000 -bloom_bits=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0 -prefix_size=8 -seek_nexts=1 -duration=20 -seed=43 -threads=8 -cache_size=1000000000 -statistics
```
Before: seekrandom [AVG 275 runs] : 189680 (± 222) ops/sec; 18.4 (± 0.0) MB/sec
After: seekrandom [AVG 275 runs] : 197110 (± 208) ops/sec; 19.1 (± 0.0) MB/sec
Reviewed By: ajkr
Differential Revision: D46029177
Pulled By: pdillinger
fbshipit-source-id: cdace79a2ea548d46c5900b068c5b7c3a02e5822
2 years ago
|
|
|
const bool need_upper_bound_check, BlockCacheLookupContext* lookup_context,
|
|
|
|
bool* filter_checked) const {
|
|
|
|
if (!rep_->filter_policy) {
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
const SliceTransform* prefix_extractor;
|
|
|
|
|
|
|
|
if (rep_->table_prefix_extractor == nullptr) {
|
|
|
|
if (need_upper_bound_check) {
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
prefix_extractor = options_prefix_extractor;
|
|
|
|
} else {
|
|
|
|
prefix_extractor = rep_->table_prefix_extractor.get();
|
|
|
|
}
|
|
|
|
auto ts_sz = rep_->internal_comparator.user_comparator()->timestamp_size();
|
|
|
|
auto user_key_without_ts =
|
|
|
|
ExtractUserKeyAndStripTimestamp(internal_key, ts_sz);
|
|
|
|
if (!prefix_extractor->InDomain(user_key_without_ts)) {
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
bool may_match = true;
|
|
|
|
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
FilterBlockReader* const filter = rep_->filter.get();
|
Much better stats for seeks and prefix filtering (#11460)
Summary:
We want to know more about opportunities for better range filters, and the effectiveness of our own range filters. Currently the stats are very limited, essentially logging just hits and misses against prefix filters for range scans in BLOOM_FILTER_PREFIX_* without tracking the false positive rate. Perhaps confusingly, when prefix filters are used for point queries, the stats are currently going into the non-PREFIX tickers.
This change does several things:
* Introduce new stat tickers for seeks and related filtering, \*LEVEL_SEEK\*
* Most importantly, allows us to see opportunities for range filtering. Specifically, we can count how many times a seek in an SST file accesses at least one data block, and how many times at least one value() is then accessed. If a data block was accessed but no value(), we can generally assume that the key(s) seen was(were) not of interest so could have been filtered with the right kind of filter, avoiding the data block access.
* We can get the same level of detail when a filter (for now, prefix Bloom/ribbon) is used, or not. Specifically, we can infer a false positive rate for prefix filters (not available before) from the seek "false positive" rate: when a data block is accessed but no value() is called. (There can be other explanations for a seek false positive, but in typical iterator usage it would indicate a filter false positive.)
* For efficiency, I wanted to avoid making additional calls to the prefix extractor (or key comparisons, etc.), which would be required if we wanted to more precisely detect filter false positives. I believe that instrumenting value() is the best balance of efficiency vs. accurately measuring what we are often interested in.
* The stats are divided between last level and non-last levels, to help understand potential tiered storage use cases.
* The old BLOOM_FILTER_PREFIX_* stats have a different meaning: no longer referring to iterators but to point queries using prefix filters. BLOOM_FILTER_PREFIX_TRUE_POSITIVE is added for computing the prefix false positive rate on point queries, which can be due to filter false positives as well as different keys with the same prefix.
* Similarly, the non-PREFIX BLOOM_FILTER stats are now for whole key filtering only.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11460
Test Plan:
unit tests updated, including updating many to pop the stat value since last read to improve test
readability and maintainability.
Performance test shows a consistent small improvement with these changes, both with clang and with gcc. CPU profile indicates that RecordTick is using less CPU, and this makes sense at least for a high filter miss rate. Before, we were recording two ticks per filter miss in iterators (CHECKED & USEFUL) and now recording just one (FILTERED).
Create DB with
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=10000000 -disable_wal=1 -write_buffer_size=30000000 -bloom_bits=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0 -prefix_size=8
```
And run simultaneous before&after with
```
TEST_TMPDIR=/dev/shm ./db_bench -readonly -benchmarks=seekrandom[-X1000] -num=10000000 -bloom_bits=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0 -prefix_size=8 -seek_nexts=1 -duration=20 -seed=43 -threads=8 -cache_size=1000000000 -statistics
```
Before: seekrandom [AVG 275 runs] : 189680 (± 222) ops/sec; 18.4 (± 0.0) MB/sec
After: seekrandom [AVG 275 runs] : 197110 (± 208) ops/sec; 19.1 (± 0.0) MB/sec
Reviewed By: ajkr
Differential Revision: D46029177
Pulled By: pdillinger
fbshipit-source-id: cdace79a2ea548d46c5900b068c5b7c3a02e5822
2 years ago
|
|
|
*filter_checked = false;
|
|
|
|
if (filter != nullptr) {
|
Fix iterator reading filter block despite read_tier == kBlockCacheTier (#6562)
Summary:
We're seeing iterators with `ReadOptions::read_tier == kBlockCacheTier` sometimes doing file reads. Stack trace:
```
rocksdb::RandomAccessFileReader::Read(unsigned long, unsigned long, rocksdb::Slice*, char*, bool) const
rocksdb::BlockFetcher::ReadBlockContents()
rocksdb::Status rocksdb::BlockBasedTable::MaybeReadBlockAndLoadToCache<rocksdb::ParsedFullFilterBlock>(rocksdb::FilePrefetchBuffer*, rocksdb::ReadOptions const&, rocksdb::BlockHandle const&, rocksdb::UncompressionDict const&, rocksdb::CachableEntry<rocksdb::ParsedFullFilterBlock>*, rocksdb::BlockType, rocksdb::GetContext*, rocksdb::BlockCacheLookupContext*, rocksdb::BlockContents*) const
rocksdb::Status rocksdb::BlockBasedTable::RetrieveBlock<rocksdb::ParsedFullFilterBlock>(rocksdb::FilePrefetchBuffer*, rocksdb::ReadOptions const&, rocksdb::BlockHandle const&, rocksdb::UncompressionDict const&, rocksdb::CachableEntry<rocksdb::ParsedFullFilterBlock>*, rocksdb::BlockType, rocksdb::GetContext*, rocksdb::BlockCacheLookupContext*, bool, bool) const
rocksdb::FilterBlockReaderCommon<rocksdb::ParsedFullFilterBlock>::ReadFilterBlock(rocksdb::BlockBasedTable const*, rocksdb::FilePrefetchBuffer*, rocksdb::ReadOptions const&, bool, rocksdb::GetContext*, rocksdb::BlockCacheLookupContext*, rocksdb::CachableEntry<rocksdb::ParsedFullFilterBlock>*)
rocksdb::FilterBlockReaderCommon<rocksdb::ParsedFullFilterBlock>::GetOrReadFilterBlock(bool, rocksdb::GetContext*, rocksdb::BlockCacheLookupContext*, rocksdb::CachableEntry<rocksdb::ParsedFullFilterBlock>*) const
rocksdb::FullFilterBlockReader::MayMatch(rocksdb::Slice const&, bool, rocksdb::GetContext*, rocksdb::BlockCacheLookupContext*) const
rocksdb::FullFilterBlockReader::RangeMayExist(rocksdb::Slice const*, rocksdb::Slice const&, rocksdb::SliceTransform const*, rocksdb::Comparator const*, rocksdb::Slice const*, bool*, bool, rocksdb::BlockCacheLookupContext*)
rocksdb::BlockBasedTable::PrefixMayMatch(rocksdb::Slice const&, rocksdb::ReadOptions const&, rocksdb::SliceTransform const*, bool, rocksdb::BlockCacheLookupContext*) const
rocksdb::BlockBasedTableIterator<rocksdb::DataBlockIter, rocksdb::Slice>::SeekImpl(rocksdb::Slice const*)
rocksdb::ForwardIterator::SeekInternal(rocksdb::Slice const&, bool)
rocksdb::DBIter::Seek(rocksdb::Slice const&)
```
`BlockBasedTableIterator::CheckPrefixMayMatch` was missing a check for `kBlockCacheTier`. This PR adds it.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6562
Test Plan: deployed it to a logdevice test cluster and looked at logdevice's IO tracing.
Reviewed By: siying
Differential Revision: D20529368
Pulled By: al13n321
fbshipit-source-id: 65bf33964b1951464415c900336635fb20919611
5 years ago
|
|
|
const bool no_io = read_options.read_tier == kBlockCacheTier;
|
|
|
|
|
Remove deprecated block-based filter (#10184)
Summary:
In https://github.com/facebook/rocksdb/issues/9535, release 7.0, we hid the old block-based filter from being created using
the public API, because of its inefficiency. Although we normally maintain read compatibility
on old DBs forever, filters are not required for reading a DB, only for optimizing read
performance. Thus, it should be acceptable to remove this code and the substantial
maintenance burden it carries as useful features are developed and validated (such
as user timestamp).
This change completely removes the code for reading and writing the old block-based
filters, net removing about 1370 lines of code no longer needed. Options removed from
testing / benchmarking tools. The prior existence is only evident in a couple of places:
* `CacheEntryRole::kDeprecatedFilterBlock` - We can update this public API enum in
a major release to minimize source code incompatibilities.
* A warning is logged when an old table file is opened that used the old block-based
filter. This is provided as a courtesy, and would be a pain to unit test, so manual testing
should suffice. Unfortunately, sst_dump does not tell you whether a file uses
block-based filter, and the structure of the code makes it very difficult to fix.
* To detect that case, `kObsoleteFilterBlockPrefix` (renamed from `kFilterBlockPrefix`)
for metaindex is maintained (for now).
Other notes:
* In some cases where numbers are associated with filter configurations, we have had to
update the assigned numbers so that they all correspond to something that exists.
* Fixed potential stat counting bug by assuming `filter_checked = false` for cases
like `filter == nullptr` rather than assuming `filter_checked = true`
* Removed obsolete `block_offset` and `prefix_extractor` parameters from several
functions.
* Removed some unnecessary checks `if (!table_prefix_extractor() && !prefix_extractor)`
because the caller guarantees the prefix extractor exists and is compatible
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10184
Test Plan:
tests updated, manually test new warning in LOG using base version to
generate a DB
Reviewed By: riversand963
Differential Revision: D37212647
Pulled By: pdillinger
fbshipit-source-id: 06ee020d8de3b81260ffc36ad0c1202cbf463a80
2 years ago
|
|
|
const Slice* const const_ikey_ptr = &internal_key;
|
|
|
|
may_match = filter->RangeMayExist(
|
|
|
|
read_options.iterate_upper_bound, user_key_without_ts, prefix_extractor,
|
|
|
|
rep_->internal_comparator.user_comparator(), const_ikey_ptr,
|
Much better stats for seeks and prefix filtering (#11460)
Summary:
We want to know more about opportunities for better range filters, and the effectiveness of our own range filters. Currently the stats are very limited, essentially logging just hits and misses against prefix filters for range scans in BLOOM_FILTER_PREFIX_* without tracking the false positive rate. Perhaps confusingly, when prefix filters are used for point queries, the stats are currently going into the non-PREFIX tickers.
This change does several things:
* Introduce new stat tickers for seeks and related filtering, \*LEVEL_SEEK\*
* Most importantly, allows us to see opportunities for range filtering. Specifically, we can count how many times a seek in an SST file accesses at least one data block, and how many times at least one value() is then accessed. If a data block was accessed but no value(), we can generally assume that the key(s) seen was(were) not of interest so could have been filtered with the right kind of filter, avoiding the data block access.
* We can get the same level of detail when a filter (for now, prefix Bloom/ribbon) is used, or not. Specifically, we can infer a false positive rate for prefix filters (not available before) from the seek "false positive" rate: when a data block is accessed but no value() is called. (There can be other explanations for a seek false positive, but in typical iterator usage it would indicate a filter false positive.)
* For efficiency, I wanted to avoid making additional calls to the prefix extractor (or key comparisons, etc.), which would be required if we wanted to more precisely detect filter false positives. I believe that instrumenting value() is the best balance of efficiency vs. accurately measuring what we are often interested in.
* The stats are divided between last level and non-last levels, to help understand potential tiered storage use cases.
* The old BLOOM_FILTER_PREFIX_* stats have a different meaning: no longer referring to iterators but to point queries using prefix filters. BLOOM_FILTER_PREFIX_TRUE_POSITIVE is added for computing the prefix false positive rate on point queries, which can be due to filter false positives as well as different keys with the same prefix.
* Similarly, the non-PREFIX BLOOM_FILTER stats are now for whole key filtering only.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11460
Test Plan:
unit tests updated, including updating many to pop the stat value since last read to improve test
readability and maintainability.
Performance test shows a consistent small improvement with these changes, both with clang and with gcc. CPU profile indicates that RecordTick is using less CPU, and this makes sense at least for a high filter miss rate. Before, we were recording two ticks per filter miss in iterators (CHECKED & USEFUL) and now recording just one (FILTERED).
Create DB with
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=10000000 -disable_wal=1 -write_buffer_size=30000000 -bloom_bits=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0 -prefix_size=8
```
And run simultaneous before&after with
```
TEST_TMPDIR=/dev/shm ./db_bench -readonly -benchmarks=seekrandom[-X1000] -num=10000000 -bloom_bits=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0 -prefix_size=8 -seek_nexts=1 -duration=20 -seed=43 -threads=8 -cache_size=1000000000 -statistics
```
Before: seekrandom [AVG 275 runs] : 189680 (± 222) ops/sec; 18.4 (± 0.0) MB/sec
After: seekrandom [AVG 275 runs] : 197110 (± 208) ops/sec; 19.1 (± 0.0) MB/sec
Reviewed By: ajkr
Differential Revision: D46029177
Pulled By: pdillinger
fbshipit-source-id: cdace79a2ea548d46c5900b068c5b7c3a02e5822
2 years ago
|
|
|
filter_checked, need_upper_bound_check, no_io, lookup_context,
|
|
|
|
read_options);
|
|
|
|
}
|
|
|
|
|
|
|
|
return may_match;
|
|
|
|
}
|
|
|
|
|
|
|
|
bool BlockBasedTable::PrefixExtractorChanged(
|
|
|
|
const SliceTransform* prefix_extractor) const {
|
|
|
|
if (prefix_extractor == nullptr) {
|
|
|
|
return true;
|
|
|
|
} else if (prefix_extractor == rep_->table_prefix_extractor.get()) {
|
|
|
|
return false;
|
|
|
|
} else {
|
|
|
|
return PrefixExtractorChangedHelper(rep_->table_properties.get(),
|
|
|
|
prefix_extractor);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
Much better stats for seeks and prefix filtering (#11460)
Summary:
We want to know more about opportunities for better range filters, and the effectiveness of our own range filters. Currently the stats are very limited, essentially logging just hits and misses against prefix filters for range scans in BLOOM_FILTER_PREFIX_* without tracking the false positive rate. Perhaps confusingly, when prefix filters are used for point queries, the stats are currently going into the non-PREFIX tickers.
This change does several things:
* Introduce new stat tickers for seeks and related filtering, \*LEVEL_SEEK\*
* Most importantly, allows us to see opportunities for range filtering. Specifically, we can count how many times a seek in an SST file accesses at least one data block, and how many times at least one value() is then accessed. If a data block was accessed but no value(), we can generally assume that the key(s) seen was(were) not of interest so could have been filtered with the right kind of filter, avoiding the data block access.
* We can get the same level of detail when a filter (for now, prefix Bloom/ribbon) is used, or not. Specifically, we can infer a false positive rate for prefix filters (not available before) from the seek "false positive" rate: when a data block is accessed but no value() is called. (There can be other explanations for a seek false positive, but in typical iterator usage it would indicate a filter false positive.)
* For efficiency, I wanted to avoid making additional calls to the prefix extractor (or key comparisons, etc.), which would be required if we wanted to more precisely detect filter false positives. I believe that instrumenting value() is the best balance of efficiency vs. accurately measuring what we are often interested in.
* The stats are divided between last level and non-last levels, to help understand potential tiered storage use cases.
* The old BLOOM_FILTER_PREFIX_* stats have a different meaning: no longer referring to iterators but to point queries using prefix filters. BLOOM_FILTER_PREFIX_TRUE_POSITIVE is added for computing the prefix false positive rate on point queries, which can be due to filter false positives as well as different keys with the same prefix.
* Similarly, the non-PREFIX BLOOM_FILTER stats are now for whole key filtering only.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11460
Test Plan:
unit tests updated, including updating many to pop the stat value since last read to improve test
readability and maintainability.
Performance test shows a consistent small improvement with these changes, both with clang and with gcc. CPU profile indicates that RecordTick is using less CPU, and this makes sense at least for a high filter miss rate. Before, we were recording two ticks per filter miss in iterators (CHECKED & USEFUL) and now recording just one (FILTERED).
Create DB with
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=10000000 -disable_wal=1 -write_buffer_size=30000000 -bloom_bits=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0 -prefix_size=8
```
And run simultaneous before&after with
```
TEST_TMPDIR=/dev/shm ./db_bench -readonly -benchmarks=seekrandom[-X1000] -num=10000000 -bloom_bits=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0 -prefix_size=8 -seek_nexts=1 -duration=20 -seed=43 -threads=8 -cache_size=1000000000 -statistics
```
Before: seekrandom [AVG 275 runs] : 189680 (± 222) ops/sec; 18.4 (± 0.0) MB/sec
After: seekrandom [AVG 275 runs] : 197110 (± 208) ops/sec; 19.1 (± 0.0) MB/sec
Reviewed By: ajkr
Differential Revision: D46029177
Pulled By: pdillinger
fbshipit-source-id: cdace79a2ea548d46c5900b068c5b7c3a02e5822
2 years ago
|
|
|
Statistics* BlockBasedTable::GetStatistics() const {
|
|
|
|
return rep_->ioptions.stats;
|
|
|
|
}
|
|
|
|
bool BlockBasedTable::IsLastLevel() const {
|
|
|
|
return rep_->level == rep_->ioptions.num_levels - 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
InternalIterator* BlockBasedTable::NewIterator(
|
|
|
|
const ReadOptions& read_options, const SliceTransform* prefix_extractor,
|
|
|
|
Arena* arena, bool skip_filters, TableReaderCaller caller,
|
Properly report IO errors when IndexType::kBinarySearchWithFirstKey is used (#6621)
Summary:
Context: Index type `kBinarySearchWithFirstKey` added the ability for sst file iterator to sometimes report a key from index without reading the corresponding data block. This is useful when sst blocks are cut at some meaningful boundaries (e.g. one block per key prefix), and many seeks land between blocks (e.g. for each prefix, the ranges of keys in different sst files are nearly disjoint, so a typical seek needs to read a data block from only one file even if all files have the prefix). But this added a new error condition, which rocksdb code was really not equipped to deal with: `InternalIterator::value()` may fail with an IO error or Status::Incomplete, but it's just a method returning a Slice, with no way to report error instead. Before this PR, this type of error wasn't handled at all (an empty slice was returned), and kBinarySearchWithFirstKey implementation was considered a prototype.
Now that we (LogDevice) have experimented with kBinarySearchWithFirstKey for a while and confirmed that it's really useful, this PR is adding the missing error handling.
It's a pretty inconvenient situation implementation-wise. The error needs to be reported from InternalIterator when trying to access value. But there are ~700 call sites of `InternalIterator::value()`, most of which either can't hit the error condition (because the iterator is reading from memtable or from index or something) or wouldn't benefit from the deferred loading of the value (e.g. compaction iterator that reads all values anyway). Adding error handling to all these call sites would needlessly bloat the code. So instead I made the deferred value loading optional: only the call sites that may use deferred loading have to call the new method `PrepareValue()` before calling `value()`. The feature is enabled with a new bool argument `allow_unprepared_value` to a bunch of methods that create iterators (it wouldn't make sense to put it in ReadOptions because it's completely internal to iterators, with virtually no user-visible effect). Lmk if you have better ideas.
Note that the deferred value loading only happens for *internal* iterators. The user-visible iterator (DBIter) always prepares the value before returning from Seek/Next/etc. We could go further and add an API to defer that value loading too, but that's most likely not useful for LogDevice, so it doesn't seem worth the complexity for now.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6621
Test Plan: make -j5 check . Will also deploy to some logdevice test clusters and look at stats.
Reviewed By: siying
Differential Revision: D20786930
Pulled By: al13n321
fbshipit-source-id: 6da77d918bad3780522e918f17f4d5513d3e99ee
5 years ago
|
|
|
size_t compaction_readahead_size, bool allow_unprepared_value) {
|
|
|
|
BlockCacheLookupContext lookup_context{caller};
|
|
|
|
bool need_upper_bound_check =
|
|
|
|
read_options.auto_prefix_mode || PrefixExtractorChanged(prefix_extractor);
|
De-template block based table iterator (#6531)
Summary:
Right now block based table iterator is used as both of iterating data for block based table, and for the index iterator for partitioend index. This was initially convenient for introducing a new iterator and block type for new index format, while reducing code change. However, these two usage doesn't go with each other very well. For example, Prev() is never called for partitioned index iterator, and some other complexity is maintained in block based iterators, which is not needed for index iterator but maintainers will always need to reason about it. Furthermore, the template usage is not following Google C++ Style which we are following, and makes a large chunk of code tangled together. This commit separate the two iterators. Right now, here is what it is done:
1. Copy the block based iterator code into partitioned index iterator, and de-template them.
2. Remove some code not needed for partitioned index. The upper bound check and tricks are removed. We never tested performance for those tricks when partitioned index is enabled in the first place. It's unlikelyl to generate performance regression, as creating new partitioned index block is much rarer than data blocks.
3. Separate out the prefetch logic to a helper class and both classes call them.
This commit will enable future follow-ups. One direction is that we might separate index iterator interface for data blocks and index blocks, as they are quite different.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6531
Test Plan: build using make and cmake. And build release
Differential Revision: D20473108
fbshipit-source-id: e48011783b339a4257c204cc07507b171b834b0f
5 years ago
|
|
|
std::unique_ptr<InternalIteratorBase<IndexValue>> index_iter(NewIndexIterator(
|
|
|
|
read_options,
|
|
|
|
/*disable_prefix_seek=*/need_upper_bound_check &&
|
De-template block based table iterator (#6531)
Summary:
Right now block based table iterator is used as both of iterating data for block based table, and for the index iterator for partitioend index. This was initially convenient for introducing a new iterator and block type for new index format, while reducing code change. However, these two usage doesn't go with each other very well. For example, Prev() is never called for partitioned index iterator, and some other complexity is maintained in block based iterators, which is not needed for index iterator but maintainers will always need to reason about it. Furthermore, the template usage is not following Google C++ Style which we are following, and makes a large chunk of code tangled together. This commit separate the two iterators. Right now, here is what it is done:
1. Copy the block based iterator code into partitioned index iterator, and de-template them.
2. Remove some code not needed for partitioned index. The upper bound check and tricks are removed. We never tested performance for those tricks when partitioned index is enabled in the first place. It's unlikelyl to generate performance regression, as creating new partitioned index block is much rarer than data blocks.
3. Separate out the prefetch logic to a helper class and both classes call them.
This commit will enable future follow-ups. One direction is that we might separate index iterator interface for data blocks and index blocks, as they are quite different.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6531
Test Plan: build using make and cmake. And build release
Differential Revision: D20473108
fbshipit-source-id: e48011783b339a4257c204cc07507b171b834b0f
5 years ago
|
|
|
rep_->index_type == BlockBasedTableOptions::kHashSearch,
|
|
|
|
/*input_iter=*/nullptr, /*get_context=*/nullptr, &lookup_context));
|
|
|
|
if (arena == nullptr) {
|
De-template block based table iterator (#6531)
Summary:
Right now block based table iterator is used as both of iterating data for block based table, and for the index iterator for partitioend index. This was initially convenient for introducing a new iterator and block type for new index format, while reducing code change. However, these two usage doesn't go with each other very well. For example, Prev() is never called for partitioned index iterator, and some other complexity is maintained in block based iterators, which is not needed for index iterator but maintainers will always need to reason about it. Furthermore, the template usage is not following Google C++ Style which we are following, and makes a large chunk of code tangled together. This commit separate the two iterators. Right now, here is what it is done:
1. Copy the block based iterator code into partitioned index iterator, and de-template them.
2. Remove some code not needed for partitioned index. The upper bound check and tricks are removed. We never tested performance for those tricks when partitioned index is enabled in the first place. It's unlikelyl to generate performance regression, as creating new partitioned index block is much rarer than data blocks.
3. Separate out the prefetch logic to a helper class and both classes call them.
This commit will enable future follow-ups. One direction is that we might separate index iterator interface for data blocks and index blocks, as they are quite different.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6531
Test Plan: build using make and cmake. And build release
Differential Revision: D20473108
fbshipit-source-id: e48011783b339a4257c204cc07507b171b834b0f
5 years ago
|
|
|
return new BlockBasedTableIterator(
|
|
|
|
this, read_options, rep_->internal_comparator, std::move(index_iter),
|
|
|
|
!skip_filters && !read_options.total_order_seek &&
|
|
|
|
prefix_extractor != nullptr,
|
De-template block based table iterator (#6531)
Summary:
Right now block based table iterator is used as both of iterating data for block based table, and for the index iterator for partitioend index. This was initially convenient for introducing a new iterator and block type for new index format, while reducing code change. However, these two usage doesn't go with each other very well. For example, Prev() is never called for partitioned index iterator, and some other complexity is maintained in block based iterators, which is not needed for index iterator but maintainers will always need to reason about it. Furthermore, the template usage is not following Google C++ Style which we are following, and makes a large chunk of code tangled together. This commit separate the two iterators. Right now, here is what it is done:
1. Copy the block based iterator code into partitioned index iterator, and de-template them.
2. Remove some code not needed for partitioned index. The upper bound check and tricks are removed. We never tested performance for those tricks when partitioned index is enabled in the first place. It's unlikelyl to generate performance regression, as creating new partitioned index block is much rarer than data blocks.
3. Separate out the prefetch logic to a helper class and both classes call them.
This commit will enable future follow-ups. One direction is that we might separate index iterator interface for data blocks and index blocks, as they are quite different.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6531
Test Plan: build using make and cmake. And build release
Differential Revision: D20473108
fbshipit-source-id: e48011783b339a4257c204cc07507b171b834b0f
5 years ago
|
|
|
need_upper_bound_check, prefix_extractor, caller,
|
Properly report IO errors when IndexType::kBinarySearchWithFirstKey is used (#6621)
Summary:
Context: Index type `kBinarySearchWithFirstKey` added the ability for sst file iterator to sometimes report a key from index without reading the corresponding data block. This is useful when sst blocks are cut at some meaningful boundaries (e.g. one block per key prefix), and many seeks land between blocks (e.g. for each prefix, the ranges of keys in different sst files are nearly disjoint, so a typical seek needs to read a data block from only one file even if all files have the prefix). But this added a new error condition, which rocksdb code was really not equipped to deal with: `InternalIterator::value()` may fail with an IO error or Status::Incomplete, but it's just a method returning a Slice, with no way to report error instead. Before this PR, this type of error wasn't handled at all (an empty slice was returned), and kBinarySearchWithFirstKey implementation was considered a prototype.
Now that we (LogDevice) have experimented with kBinarySearchWithFirstKey for a while and confirmed that it's really useful, this PR is adding the missing error handling.
It's a pretty inconvenient situation implementation-wise. The error needs to be reported from InternalIterator when trying to access value. But there are ~700 call sites of `InternalIterator::value()`, most of which either can't hit the error condition (because the iterator is reading from memtable or from index or something) or wouldn't benefit from the deferred loading of the value (e.g. compaction iterator that reads all values anyway). Adding error handling to all these call sites would needlessly bloat the code. So instead I made the deferred value loading optional: only the call sites that may use deferred loading have to call the new method `PrepareValue()` before calling `value()`. The feature is enabled with a new bool argument `allow_unprepared_value` to a bunch of methods that create iterators (it wouldn't make sense to put it in ReadOptions because it's completely internal to iterators, with virtually no user-visible effect). Lmk if you have better ideas.
Note that the deferred value loading only happens for *internal* iterators. The user-visible iterator (DBIter) always prepares the value before returning from Seek/Next/etc. We could go further and add an API to defer that value loading too, but that's most likely not useful for LogDevice, so it doesn't seem worth the complexity for now.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6621
Test Plan: make -j5 check . Will also deploy to some logdevice test clusters and look at stats.
Reviewed By: siying
Differential Revision: D20786930
Pulled By: al13n321
fbshipit-source-id: 6da77d918bad3780522e918f17f4d5513d3e99ee
5 years ago
|
|
|
compaction_readahead_size, allow_unprepared_value);
|
|
|
|
} else {
|
De-template block based table iterator (#6531)
Summary:
Right now block based table iterator is used as both of iterating data for block based table, and for the index iterator for partitioend index. This was initially convenient for introducing a new iterator and block type for new index format, while reducing code change. However, these two usage doesn't go with each other very well. For example, Prev() is never called for partitioned index iterator, and some other complexity is maintained in block based iterators, which is not needed for index iterator but maintainers will always need to reason about it. Furthermore, the template usage is not following Google C++ Style which we are following, and makes a large chunk of code tangled together. This commit separate the two iterators. Right now, here is what it is done:
1. Copy the block based iterator code into partitioned index iterator, and de-template them.
2. Remove some code not needed for partitioned index. The upper bound check and tricks are removed. We never tested performance for those tricks when partitioned index is enabled in the first place. It's unlikelyl to generate performance regression, as creating new partitioned index block is much rarer than data blocks.
3. Separate out the prefetch logic to a helper class and both classes call them.
This commit will enable future follow-ups. One direction is that we might separate index iterator interface for data blocks and index blocks, as they are quite different.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6531
Test Plan: build using make and cmake. And build release
Differential Revision: D20473108
fbshipit-source-id: e48011783b339a4257c204cc07507b171b834b0f
5 years ago
|
|
|
auto* mem = arena->AllocateAligned(sizeof(BlockBasedTableIterator));
|
|
|
|
return new (mem) BlockBasedTableIterator(
|
|
|
|
this, read_options, rep_->internal_comparator, std::move(index_iter),
|
|
|
|
!skip_filters && !read_options.total_order_seek &&
|
|
|
|
prefix_extractor != nullptr,
|
De-template block based table iterator (#6531)
Summary:
Right now block based table iterator is used as both of iterating data for block based table, and for the index iterator for partitioend index. This was initially convenient for introducing a new iterator and block type for new index format, while reducing code change. However, these two usage doesn't go with each other very well. For example, Prev() is never called for partitioned index iterator, and some other complexity is maintained in block based iterators, which is not needed for index iterator but maintainers will always need to reason about it. Furthermore, the template usage is not following Google C++ Style which we are following, and makes a large chunk of code tangled together. This commit separate the two iterators. Right now, here is what it is done:
1. Copy the block based iterator code into partitioned index iterator, and de-template them.
2. Remove some code not needed for partitioned index. The upper bound check and tricks are removed. We never tested performance for those tricks when partitioned index is enabled in the first place. It's unlikelyl to generate performance regression, as creating new partitioned index block is much rarer than data blocks.
3. Separate out the prefetch logic to a helper class and both classes call them.
This commit will enable future follow-ups. One direction is that we might separate index iterator interface for data blocks and index blocks, as they are quite different.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6531
Test Plan: build using make and cmake. And build release
Differential Revision: D20473108
fbshipit-source-id: e48011783b339a4257c204cc07507b171b834b0f
5 years ago
|
|
|
need_upper_bound_check, prefix_extractor, caller,
|
Properly report IO errors when IndexType::kBinarySearchWithFirstKey is used (#6621)
Summary:
Context: Index type `kBinarySearchWithFirstKey` added the ability for sst file iterator to sometimes report a key from index without reading the corresponding data block. This is useful when sst blocks are cut at some meaningful boundaries (e.g. one block per key prefix), and many seeks land between blocks (e.g. for each prefix, the ranges of keys in different sst files are nearly disjoint, so a typical seek needs to read a data block from only one file even if all files have the prefix). But this added a new error condition, which rocksdb code was really not equipped to deal with: `InternalIterator::value()` may fail with an IO error or Status::Incomplete, but it's just a method returning a Slice, with no way to report error instead. Before this PR, this type of error wasn't handled at all (an empty slice was returned), and kBinarySearchWithFirstKey implementation was considered a prototype.
Now that we (LogDevice) have experimented with kBinarySearchWithFirstKey for a while and confirmed that it's really useful, this PR is adding the missing error handling.
It's a pretty inconvenient situation implementation-wise. The error needs to be reported from InternalIterator when trying to access value. But there are ~700 call sites of `InternalIterator::value()`, most of which either can't hit the error condition (because the iterator is reading from memtable or from index or something) or wouldn't benefit from the deferred loading of the value (e.g. compaction iterator that reads all values anyway). Adding error handling to all these call sites would needlessly bloat the code. So instead I made the deferred value loading optional: only the call sites that may use deferred loading have to call the new method `PrepareValue()` before calling `value()`. The feature is enabled with a new bool argument `allow_unprepared_value` to a bunch of methods that create iterators (it wouldn't make sense to put it in ReadOptions because it's completely internal to iterators, with virtually no user-visible effect). Lmk if you have better ideas.
Note that the deferred value loading only happens for *internal* iterators. The user-visible iterator (DBIter) always prepares the value before returning from Seek/Next/etc. We could go further and add an API to defer that value loading too, but that's most likely not useful for LogDevice, so it doesn't seem worth the complexity for now.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6621
Test Plan: make -j5 check . Will also deploy to some logdevice test clusters and look at stats.
Reviewed By: siying
Differential Revision: D20786930
Pulled By: al13n321
fbshipit-source-id: 6da77d918bad3780522e918f17f4d5513d3e99ee
5 years ago
|
|
|
compaction_readahead_size, allow_unprepared_value);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
FragmentedRangeTombstoneIterator* BlockBasedTable::NewRangeTombstoneIterator(
|
|
|
|
const ReadOptions& read_options) {
|
Cache fragmented range tombstones in BlockBasedTableReader (#4493)
Summary:
This allows tombstone fragmenting to only be performed when the table is opened, and cached for subsequent accesses.
On the same DB used in #4449, running `readrandom` results in the following:
```
readrandom : 0.983 micros/op 1017076 ops/sec; 78.3 MB/s (63103 of 100000 found)
```
Now that Get performance in the presence of range tombstones is reasonable, I also compared the performance between a DB with range tombstones, "expanded" range tombstones (several point tombstones that cover the same keys the equivalent range tombstone would cover, a common workaround for DeleteRange), and no range tombstones. The created DBs had 5 million keys each, and DeleteRange was called at regular intervals (depending on the total number of range tombstones being written) after 4.5 million Puts. The table below summarizes the results of a `readwhilewriting` benchmark (in order to provide somewhat more realistic results):
```
Tombstones? | avg micros/op | stddev micros/op | avg ops/s | stddev ops/s
----------------- | ------------- | ---------------- | ------------ | ------------
None | 0.6186 | 0.04637 | 1,625,252.90 | 124,679.41
500 Expanded | 0.6019 | 0.03628 | 1,666,670.40 | 101,142.65
500 Unexpanded | 0.6435 | 0.03994 | 1,559,979.40 | 104,090.52
1k Expanded | 0.6034 | 0.04349 | 1,665,128.10 | 125,144.57
1k Unexpanded | 0.6261 | 0.03093 | 1,600,457.50 | 79,024.94
5k Expanded | 0.6163 | 0.05926 | 1,636,668.80 | 154,888.85
5k Unexpanded | 0.6402 | 0.04002 | 1,567,804.70 | 100,965.55
10k Expanded | 0.6036 | 0.05105 | 1,667,237.70 | 142,830.36
10k Unexpanded | 0.6128 | 0.02598 | 1,634,633.40 | 72,161.82
25k Expanded | 0.6198 | 0.04542 | 1,620,980.50 | 116,662.93
25k Unexpanded | 0.5478 | 0.0362 | 1,833,059.10 | 121,233.81
50k Expanded | 0.5104 | 0.04347 | 1,973,107.90 | 184,073.49
50k Unexpanded | 0.4528 | 0.03387 | 2,219,034.50 | 170,984.32
```
After a large enough quantity of range tombstones are written, range tombstone Gets can become faster than reading from an equivalent DB with several point tombstones.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4493
Differential Revision: D10842844
Pulled By: abhimadan
fbshipit-source-id: a7d44534f8120e6aabb65779d26c6b9df954c509
6 years ago
|
|
|
if (rep_->fragmented_range_dels == nullptr) {
|
|
|
|
return nullptr;
|
|
|
|
}
|
|
|
|
SequenceNumber snapshot = kMaxSequenceNumber;
|
|
|
|
if (read_options.snapshot != nullptr) {
|
|
|
|
snapshot = read_options.snapshot->GetSequenceNumber();
|
|
|
|
}
|
|
|
|
return new FragmentedRangeTombstoneIterator(rep_->fragmented_range_dels,
|
|
|
|
rep_->internal_comparator,
|
|
|
|
snapshot, read_options.timestamp);
|
Cache fragmented range tombstones in BlockBasedTableReader (#4493)
Summary:
This allows tombstone fragmenting to only be performed when the table is opened, and cached for subsequent accesses.
On the same DB used in #4449, running `readrandom` results in the following:
```
readrandom : 0.983 micros/op 1017076 ops/sec; 78.3 MB/s (63103 of 100000 found)
```
Now that Get performance in the presence of range tombstones is reasonable, I also compared the performance between a DB with range tombstones, "expanded" range tombstones (several point tombstones that cover the same keys the equivalent range tombstone would cover, a common workaround for DeleteRange), and no range tombstones. The created DBs had 5 million keys each, and DeleteRange was called at regular intervals (depending on the total number of range tombstones being written) after 4.5 million Puts. The table below summarizes the results of a `readwhilewriting` benchmark (in order to provide somewhat more realistic results):
```
Tombstones? | avg micros/op | stddev micros/op | avg ops/s | stddev ops/s
----------------- | ------------- | ---------------- | ------------ | ------------
None | 0.6186 | 0.04637 | 1,625,252.90 | 124,679.41
500 Expanded | 0.6019 | 0.03628 | 1,666,670.40 | 101,142.65
500 Unexpanded | 0.6435 | 0.03994 | 1,559,979.40 | 104,090.52
1k Expanded | 0.6034 | 0.04349 | 1,665,128.10 | 125,144.57
1k Unexpanded | 0.6261 | 0.03093 | 1,600,457.50 | 79,024.94
5k Expanded | 0.6163 | 0.05926 | 1,636,668.80 | 154,888.85
5k Unexpanded | 0.6402 | 0.04002 | 1,567,804.70 | 100,965.55
10k Expanded | 0.6036 | 0.05105 | 1,667,237.70 | 142,830.36
10k Unexpanded | 0.6128 | 0.02598 | 1,634,633.40 | 72,161.82
25k Expanded | 0.6198 | 0.04542 | 1,620,980.50 | 116,662.93
25k Unexpanded | 0.5478 | 0.0362 | 1,833,059.10 | 121,233.81
50k Expanded | 0.5104 | 0.04347 | 1,973,107.90 | 184,073.49
50k Unexpanded | 0.4528 | 0.03387 | 2,219,034.50 | 170,984.32
```
After a large enough quantity of range tombstones are written, range tombstone Gets can become faster than reading from an equivalent DB with several point tombstones.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4493
Differential Revision: D10842844
Pulled By: abhimadan
fbshipit-source-id: a7d44534f8120e6aabb65779d26c6b9df954c509
6 years ago
|
|
|
}
|
|
|
|
|
|
|
|
bool BlockBasedTable::FullFilterKeyMayMatch(
|
|
|
|
FilterBlockReader* filter, const Slice& internal_key, const bool no_io,
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
const SliceTransform* prefix_extractor, GetContext* get_context,
|
|
|
|
BlockCacheLookupContext* lookup_context,
|
|
|
|
const ReadOptions& read_options) const {
|
Remove deprecated block-based filter (#10184)
Summary:
In https://github.com/facebook/rocksdb/issues/9535, release 7.0, we hid the old block-based filter from being created using
the public API, because of its inefficiency. Although we normally maintain read compatibility
on old DBs forever, filters are not required for reading a DB, only for optimizing read
performance. Thus, it should be acceptable to remove this code and the substantial
maintenance burden it carries as useful features are developed and validated (such
as user timestamp).
This change completely removes the code for reading and writing the old block-based
filters, net removing about 1370 lines of code no longer needed. Options removed from
testing / benchmarking tools. The prior existence is only evident in a couple of places:
* `CacheEntryRole::kDeprecatedFilterBlock` - We can update this public API enum in
a major release to minimize source code incompatibilities.
* A warning is logged when an old table file is opened that used the old block-based
filter. This is provided as a courtesy, and would be a pain to unit test, so manual testing
should suffice. Unfortunately, sst_dump does not tell you whether a file uses
block-based filter, and the structure of the code makes it very difficult to fix.
* To detect that case, `kObsoleteFilterBlockPrefix` (renamed from `kFilterBlockPrefix`)
for metaindex is maintained (for now).
Other notes:
* In some cases where numbers are associated with filter configurations, we have had to
update the assigned numbers so that they all correspond to something that exists.
* Fixed potential stat counting bug by assuming `filter_checked = false` for cases
like `filter == nullptr` rather than assuming `filter_checked = true`
* Removed obsolete `block_offset` and `prefix_extractor` parameters from several
functions.
* Removed some unnecessary checks `if (!table_prefix_extractor() && !prefix_extractor)`
because the caller guarantees the prefix extractor exists and is compatible
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10184
Test Plan:
tests updated, manually test new warning in LOG using base version to
generate a DB
Reviewed By: riversand963
Differential Revision: D37212647
Pulled By: pdillinger
fbshipit-source-id: 06ee020d8de3b81260ffc36ad0c1202cbf463a80
2 years ago
|
|
|
if (filter == nullptr) {
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
Slice user_key = ExtractUserKey(internal_key);
|
|
|
|
const Slice* const const_ikey_ptr = &internal_key;
|
|
|
|
bool may_match = true;
|
|
|
|
size_t ts_sz = rep_->internal_comparator.user_comparator()->timestamp_size();
|
|
|
|
Slice user_key_without_ts = StripTimestampFromUserKey(user_key, ts_sz);
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
if (rep_->whole_key_filtering) {
|
|
|
|
may_match = filter->KeyMayMatch(user_key_without_ts, no_io, const_ikey_ptr,
|
|
|
|
get_context, lookup_context, read_options);
|
Much better stats for seeks and prefix filtering (#11460)
Summary:
We want to know more about opportunities for better range filters, and the effectiveness of our own range filters. Currently the stats are very limited, essentially logging just hits and misses against prefix filters for range scans in BLOOM_FILTER_PREFIX_* without tracking the false positive rate. Perhaps confusingly, when prefix filters are used for point queries, the stats are currently going into the non-PREFIX tickers.
This change does several things:
* Introduce new stat tickers for seeks and related filtering, \*LEVEL_SEEK\*
* Most importantly, allows us to see opportunities for range filtering. Specifically, we can count how many times a seek in an SST file accesses at least one data block, and how many times at least one value() is then accessed. If a data block was accessed but no value(), we can generally assume that the key(s) seen was(were) not of interest so could have been filtered with the right kind of filter, avoiding the data block access.
* We can get the same level of detail when a filter (for now, prefix Bloom/ribbon) is used, or not. Specifically, we can infer a false positive rate for prefix filters (not available before) from the seek "false positive" rate: when a data block is accessed but no value() is called. (There can be other explanations for a seek false positive, but in typical iterator usage it would indicate a filter false positive.)
* For efficiency, I wanted to avoid making additional calls to the prefix extractor (or key comparisons, etc.), which would be required if we wanted to more precisely detect filter false positives. I believe that instrumenting value() is the best balance of efficiency vs. accurately measuring what we are often interested in.
* The stats are divided between last level and non-last levels, to help understand potential tiered storage use cases.
* The old BLOOM_FILTER_PREFIX_* stats have a different meaning: no longer referring to iterators but to point queries using prefix filters. BLOOM_FILTER_PREFIX_TRUE_POSITIVE is added for computing the prefix false positive rate on point queries, which can be due to filter false positives as well as different keys with the same prefix.
* Similarly, the non-PREFIX BLOOM_FILTER stats are now for whole key filtering only.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11460
Test Plan:
unit tests updated, including updating many to pop the stat value since last read to improve test
readability and maintainability.
Performance test shows a consistent small improvement with these changes, both with clang and with gcc. CPU profile indicates that RecordTick is using less CPU, and this makes sense at least for a high filter miss rate. Before, we were recording two ticks per filter miss in iterators (CHECKED & USEFUL) and now recording just one (FILTERED).
Create DB with
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=10000000 -disable_wal=1 -write_buffer_size=30000000 -bloom_bits=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0 -prefix_size=8
```
And run simultaneous before&after with
```
TEST_TMPDIR=/dev/shm ./db_bench -readonly -benchmarks=seekrandom[-X1000] -num=10000000 -bloom_bits=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0 -prefix_size=8 -seek_nexts=1 -duration=20 -seed=43 -threads=8 -cache_size=1000000000 -statistics
```
Before: seekrandom [AVG 275 runs] : 189680 (± 222) ops/sec; 18.4 (± 0.0) MB/sec
After: seekrandom [AVG 275 runs] : 197110 (± 208) ops/sec; 19.1 (± 0.0) MB/sec
Reviewed By: ajkr
Differential Revision: D46029177
Pulled By: pdillinger
fbshipit-source-id: cdace79a2ea548d46c5900b068c5b7c3a02e5822
2 years ago
|
|
|
if (may_match) {
|
|
|
|
RecordTick(rep_->ioptions.stats, BLOOM_FILTER_FULL_POSITIVE);
|
|
|
|
PERF_COUNTER_BY_LEVEL_ADD(bloom_filter_full_positive, 1, rep_->level);
|
|
|
|
} else {
|
|
|
|
RecordTick(rep_->ioptions.stats, BLOOM_FILTER_USEFUL);
|
|
|
|
PERF_COUNTER_BY_LEVEL_ADD(bloom_filter_useful, 1, rep_->level);
|
|
|
|
}
|
|
|
|
} else if (!PrefixExtractorChanged(prefix_extractor) &&
|
Much better stats for seeks and prefix filtering (#11460)
Summary:
We want to know more about opportunities for better range filters, and the effectiveness of our own range filters. Currently the stats are very limited, essentially logging just hits and misses against prefix filters for range scans in BLOOM_FILTER_PREFIX_* without tracking the false positive rate. Perhaps confusingly, when prefix filters are used for point queries, the stats are currently going into the non-PREFIX tickers.
This change does several things:
* Introduce new stat tickers for seeks and related filtering, \*LEVEL_SEEK\*
* Most importantly, allows us to see opportunities for range filtering. Specifically, we can count how many times a seek in an SST file accesses at least one data block, and how many times at least one value() is then accessed. If a data block was accessed but no value(), we can generally assume that the key(s) seen was(were) not of interest so could have been filtered with the right kind of filter, avoiding the data block access.
* We can get the same level of detail when a filter (for now, prefix Bloom/ribbon) is used, or not. Specifically, we can infer a false positive rate for prefix filters (not available before) from the seek "false positive" rate: when a data block is accessed but no value() is called. (There can be other explanations for a seek false positive, but in typical iterator usage it would indicate a filter false positive.)
* For efficiency, I wanted to avoid making additional calls to the prefix extractor (or key comparisons, etc.), which would be required if we wanted to more precisely detect filter false positives. I believe that instrumenting value() is the best balance of efficiency vs. accurately measuring what we are often interested in.
* The stats are divided between last level and non-last levels, to help understand potential tiered storage use cases.
* The old BLOOM_FILTER_PREFIX_* stats have a different meaning: no longer referring to iterators but to point queries using prefix filters. BLOOM_FILTER_PREFIX_TRUE_POSITIVE is added for computing the prefix false positive rate on point queries, which can be due to filter false positives as well as different keys with the same prefix.
* Similarly, the non-PREFIX BLOOM_FILTER stats are now for whole key filtering only.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11460
Test Plan:
unit tests updated, including updating many to pop the stat value since last read to improve test
readability and maintainability.
Performance test shows a consistent small improvement with these changes, both with clang and with gcc. CPU profile indicates that RecordTick is using less CPU, and this makes sense at least for a high filter miss rate. Before, we were recording two ticks per filter miss in iterators (CHECKED & USEFUL) and now recording just one (FILTERED).
Create DB with
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=10000000 -disable_wal=1 -write_buffer_size=30000000 -bloom_bits=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0 -prefix_size=8
```
And run simultaneous before&after with
```
TEST_TMPDIR=/dev/shm ./db_bench -readonly -benchmarks=seekrandom[-X1000] -num=10000000 -bloom_bits=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0 -prefix_size=8 -seek_nexts=1 -duration=20 -seed=43 -threads=8 -cache_size=1000000000 -statistics
```
Before: seekrandom [AVG 275 runs] : 189680 (± 222) ops/sec; 18.4 (± 0.0) MB/sec
After: seekrandom [AVG 275 runs] : 197110 (± 208) ops/sec; 19.1 (± 0.0) MB/sec
Reviewed By: ajkr
Differential Revision: D46029177
Pulled By: pdillinger
fbshipit-source-id: cdace79a2ea548d46c5900b068c5b7c3a02e5822
2 years ago
|
|
|
prefix_extractor->InDomain(user_key_without_ts)) {
|
|
|
|
// FIXME ^^^: there should be no reason for Get() to depend on current
|
|
|
|
// prefix_extractor at all. It should always use table_prefix_extractor.
|
Much better stats for seeks and prefix filtering (#11460)
Summary:
We want to know more about opportunities for better range filters, and the effectiveness of our own range filters. Currently the stats are very limited, essentially logging just hits and misses against prefix filters for range scans in BLOOM_FILTER_PREFIX_* without tracking the false positive rate. Perhaps confusingly, when prefix filters are used for point queries, the stats are currently going into the non-PREFIX tickers.
This change does several things:
* Introduce new stat tickers for seeks and related filtering, \*LEVEL_SEEK\*
* Most importantly, allows us to see opportunities for range filtering. Specifically, we can count how many times a seek in an SST file accesses at least one data block, and how many times at least one value() is then accessed. If a data block was accessed but no value(), we can generally assume that the key(s) seen was(were) not of interest so could have been filtered with the right kind of filter, avoiding the data block access.
* We can get the same level of detail when a filter (for now, prefix Bloom/ribbon) is used, or not. Specifically, we can infer a false positive rate for prefix filters (not available before) from the seek "false positive" rate: when a data block is accessed but no value() is called. (There can be other explanations for a seek false positive, but in typical iterator usage it would indicate a filter false positive.)
* For efficiency, I wanted to avoid making additional calls to the prefix extractor (or key comparisons, etc.), which would be required if we wanted to more precisely detect filter false positives. I believe that instrumenting value() is the best balance of efficiency vs. accurately measuring what we are often interested in.
* The stats are divided between last level and non-last levels, to help understand potential tiered storage use cases.
* The old BLOOM_FILTER_PREFIX_* stats have a different meaning: no longer referring to iterators but to point queries using prefix filters. BLOOM_FILTER_PREFIX_TRUE_POSITIVE is added for computing the prefix false positive rate on point queries, which can be due to filter false positives as well as different keys with the same prefix.
* Similarly, the non-PREFIX BLOOM_FILTER stats are now for whole key filtering only.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11460
Test Plan:
unit tests updated, including updating many to pop the stat value since last read to improve test
readability and maintainability.
Performance test shows a consistent small improvement with these changes, both with clang and with gcc. CPU profile indicates that RecordTick is using less CPU, and this makes sense at least for a high filter miss rate. Before, we were recording two ticks per filter miss in iterators (CHECKED & USEFUL) and now recording just one (FILTERED).
Create DB with
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=10000000 -disable_wal=1 -write_buffer_size=30000000 -bloom_bits=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0 -prefix_size=8
```
And run simultaneous before&after with
```
TEST_TMPDIR=/dev/shm ./db_bench -readonly -benchmarks=seekrandom[-X1000] -num=10000000 -bloom_bits=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0 -prefix_size=8 -seek_nexts=1 -duration=20 -seed=43 -threads=8 -cache_size=1000000000 -statistics
```
Before: seekrandom [AVG 275 runs] : 189680 (± 222) ops/sec; 18.4 (± 0.0) MB/sec
After: seekrandom [AVG 275 runs] : 197110 (± 208) ops/sec; 19.1 (± 0.0) MB/sec
Reviewed By: ajkr
Differential Revision: D46029177
Pulled By: pdillinger
fbshipit-source-id: cdace79a2ea548d46c5900b068c5b7c3a02e5822
2 years ago
|
|
|
may_match = filter->PrefixMayMatch(
|
|
|
|
prefix_extractor->Transform(user_key_without_ts), no_io, const_ikey_ptr,
|
|
|
|
get_context, lookup_context, read_options);
|
|
|
|
RecordTick(rep_->ioptions.stats, BLOOM_FILTER_PREFIX_CHECKED);
|
|
|
|
if (may_match) {
|
|
|
|
// Includes prefix stats
|
|
|
|
PERF_COUNTER_BY_LEVEL_ADD(bloom_filter_full_positive, 1, rep_->level);
|
|
|
|
} else {
|
|
|
|
RecordTick(rep_->ioptions.stats, BLOOM_FILTER_PREFIX_USEFUL);
|
|
|
|
// Includes prefix stats
|
|
|
|
PERF_COUNTER_BY_LEVEL_ADD(bloom_filter_useful, 1, rep_->level);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return may_match;
|
|
|
|
}
|
|
|
|
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
6 years ago
|
|
|
void BlockBasedTable::FullFilterKeysMayMatch(
|
|
|
|
FilterBlockReader* filter, MultiGetRange* range, const bool no_io,
|
|
|
|
const SliceTransform* prefix_extractor,
|
|
|
|
BlockCacheLookupContext* lookup_context,
|
|
|
|
const ReadOptions& read_options) const {
|
Remove deprecated block-based filter (#10184)
Summary:
In https://github.com/facebook/rocksdb/issues/9535, release 7.0, we hid the old block-based filter from being created using
the public API, because of its inefficiency. Although we normally maintain read compatibility
on old DBs forever, filters are not required for reading a DB, only for optimizing read
performance. Thus, it should be acceptable to remove this code and the substantial
maintenance burden it carries as useful features are developed and validated (such
as user timestamp).
This change completely removes the code for reading and writing the old block-based
filters, net removing about 1370 lines of code no longer needed. Options removed from
testing / benchmarking tools. The prior existence is only evident in a couple of places:
* `CacheEntryRole::kDeprecatedFilterBlock` - We can update this public API enum in
a major release to minimize source code incompatibilities.
* A warning is logged when an old table file is opened that used the old block-based
filter. This is provided as a courtesy, and would be a pain to unit test, so manual testing
should suffice. Unfortunately, sst_dump does not tell you whether a file uses
block-based filter, and the structure of the code makes it very difficult to fix.
* To detect that case, `kObsoleteFilterBlockPrefix` (renamed from `kFilterBlockPrefix`)
for metaindex is maintained (for now).
Other notes:
* In some cases where numbers are associated with filter configurations, we have had to
update the assigned numbers so that they all correspond to something that exists.
* Fixed potential stat counting bug by assuming `filter_checked = false` for cases
like `filter == nullptr` rather than assuming `filter_checked = true`
* Removed obsolete `block_offset` and `prefix_extractor` parameters from several
functions.
* Removed some unnecessary checks `if (!table_prefix_extractor() && !prefix_extractor)`
because the caller guarantees the prefix extractor exists and is compatible
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10184
Test Plan:
tests updated, manually test new warning in LOG using base version to
generate a DB
Reviewed By: riversand963
Differential Revision: D37212647
Pulled By: pdillinger
fbshipit-source-id: 06ee020d8de3b81260ffc36ad0c1202cbf463a80
2 years ago
|
|
|
if (filter == nullptr) {
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
6 years ago
|
|
|
return;
|
|
|
|
}
|
|
|
|
uint64_t before_keys = range->KeysLeft();
|
|
|
|
assert(before_keys > 0); // Caller should ensure
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
if (rep_->whole_key_filtering) {
|
|
|
|
filter->KeysMayMatch(range, no_io, lookup_context, read_options);
|
|
|
|
uint64_t after_keys = range->KeysLeft();
|
|
|
|
if (after_keys) {
|
|
|
|
RecordTick(rep_->ioptions.stats, BLOOM_FILTER_FULL_POSITIVE, after_keys);
|
|
|
|
PERF_COUNTER_BY_LEVEL_ADD(bloom_filter_full_positive, after_keys,
|
|
|
|
rep_->level);
|
|
|
|
}
|
|
|
|
uint64_t filtered_keys = before_keys - after_keys;
|
|
|
|
if (filtered_keys) {
|
|
|
|
RecordTick(rep_->ioptions.stats, BLOOM_FILTER_USEFUL, filtered_keys);
|
|
|
|
PERF_COUNTER_BY_LEVEL_ADD(bloom_filter_useful, filtered_keys,
|
|
|
|
rep_->level);
|
|
|
|
}
|
|
|
|
} else if (!PrefixExtractorChanged(prefix_extractor)) {
|
|
|
|
// FIXME ^^^: there should be no reason for MultiGet() to depend on current
|
|
|
|
// prefix_extractor at all. It should always use table_prefix_extractor.
|
|
|
|
filter->PrefixesMayMatch(range, prefix_extractor, false, lookup_context,
|
|
|
|
read_options);
|
|
|
|
RecordTick(rep_->ioptions.stats, BLOOM_FILTER_PREFIX_CHECKED, before_keys);
|
|
|
|
uint64_t after_keys = range->KeysLeft();
|
Much better stats for seeks and prefix filtering (#11460)
Summary:
We want to know more about opportunities for better range filters, and the effectiveness of our own range filters. Currently the stats are very limited, essentially logging just hits and misses against prefix filters for range scans in BLOOM_FILTER_PREFIX_* without tracking the false positive rate. Perhaps confusingly, when prefix filters are used for point queries, the stats are currently going into the non-PREFIX tickers.
This change does several things:
* Introduce new stat tickers for seeks and related filtering, \*LEVEL_SEEK\*
* Most importantly, allows us to see opportunities for range filtering. Specifically, we can count how many times a seek in an SST file accesses at least one data block, and how many times at least one value() is then accessed. If a data block was accessed but no value(), we can generally assume that the key(s) seen was(were) not of interest so could have been filtered with the right kind of filter, avoiding the data block access.
* We can get the same level of detail when a filter (for now, prefix Bloom/ribbon) is used, or not. Specifically, we can infer a false positive rate for prefix filters (not available before) from the seek "false positive" rate: when a data block is accessed but no value() is called. (There can be other explanations for a seek false positive, but in typical iterator usage it would indicate a filter false positive.)
* For efficiency, I wanted to avoid making additional calls to the prefix extractor (or key comparisons, etc.), which would be required if we wanted to more precisely detect filter false positives. I believe that instrumenting value() is the best balance of efficiency vs. accurately measuring what we are often interested in.
* The stats are divided between last level and non-last levels, to help understand potential tiered storage use cases.
* The old BLOOM_FILTER_PREFIX_* stats have a different meaning: no longer referring to iterators but to point queries using prefix filters. BLOOM_FILTER_PREFIX_TRUE_POSITIVE is added for computing the prefix false positive rate on point queries, which can be due to filter false positives as well as different keys with the same prefix.
* Similarly, the non-PREFIX BLOOM_FILTER stats are now for whole key filtering only.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11460
Test Plan:
unit tests updated, including updating many to pop the stat value since last read to improve test
readability and maintainability.
Performance test shows a consistent small improvement with these changes, both with clang and with gcc. CPU profile indicates that RecordTick is using less CPU, and this makes sense at least for a high filter miss rate. Before, we were recording two ticks per filter miss in iterators (CHECKED & USEFUL) and now recording just one (FILTERED).
Create DB with
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=10000000 -disable_wal=1 -write_buffer_size=30000000 -bloom_bits=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0 -prefix_size=8
```
And run simultaneous before&after with
```
TEST_TMPDIR=/dev/shm ./db_bench -readonly -benchmarks=seekrandom[-X1000] -num=10000000 -bloom_bits=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0 -prefix_size=8 -seek_nexts=1 -duration=20 -seed=43 -threads=8 -cache_size=1000000000 -statistics
```
Before: seekrandom [AVG 275 runs] : 189680 (± 222) ops/sec; 18.4 (± 0.0) MB/sec
After: seekrandom [AVG 275 runs] : 197110 (± 208) ops/sec; 19.1 (± 0.0) MB/sec
Reviewed By: ajkr
Differential Revision: D46029177
Pulled By: pdillinger
fbshipit-source-id: cdace79a2ea548d46c5900b068c5b7c3a02e5822
2 years ago
|
|
|
if (after_keys) {
|
|
|
|
// Includes prefix stats
|
|
|
|
PERF_COUNTER_BY_LEVEL_ADD(bloom_filter_full_positive, after_keys,
|
|
|
|
rep_->level);
|
|
|
|
}
|
|
|
|
uint64_t filtered_keys = before_keys - after_keys;
|
|
|
|
if (filtered_keys) {
|
|
|
|
RecordTick(rep_->ioptions.stats, BLOOM_FILTER_PREFIX_USEFUL,
|
|
|
|
filtered_keys);
|
Much better stats for seeks and prefix filtering (#11460)
Summary:
We want to know more about opportunities for better range filters, and the effectiveness of our own range filters. Currently the stats are very limited, essentially logging just hits and misses against prefix filters for range scans in BLOOM_FILTER_PREFIX_* without tracking the false positive rate. Perhaps confusingly, when prefix filters are used for point queries, the stats are currently going into the non-PREFIX tickers.
This change does several things:
* Introduce new stat tickers for seeks and related filtering, \*LEVEL_SEEK\*
* Most importantly, allows us to see opportunities for range filtering. Specifically, we can count how many times a seek in an SST file accesses at least one data block, and how many times at least one value() is then accessed. If a data block was accessed but no value(), we can generally assume that the key(s) seen was(were) not of interest so could have been filtered with the right kind of filter, avoiding the data block access.
* We can get the same level of detail when a filter (for now, prefix Bloom/ribbon) is used, or not. Specifically, we can infer a false positive rate for prefix filters (not available before) from the seek "false positive" rate: when a data block is accessed but no value() is called. (There can be other explanations for a seek false positive, but in typical iterator usage it would indicate a filter false positive.)
* For efficiency, I wanted to avoid making additional calls to the prefix extractor (or key comparisons, etc.), which would be required if we wanted to more precisely detect filter false positives. I believe that instrumenting value() is the best balance of efficiency vs. accurately measuring what we are often interested in.
* The stats are divided between last level and non-last levels, to help understand potential tiered storage use cases.
* The old BLOOM_FILTER_PREFIX_* stats have a different meaning: no longer referring to iterators but to point queries using prefix filters. BLOOM_FILTER_PREFIX_TRUE_POSITIVE is added for computing the prefix false positive rate on point queries, which can be due to filter false positives as well as different keys with the same prefix.
* Similarly, the non-PREFIX BLOOM_FILTER stats are now for whole key filtering only.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11460
Test Plan:
unit tests updated, including updating many to pop the stat value since last read to improve test
readability and maintainability.
Performance test shows a consistent small improvement with these changes, both with clang and with gcc. CPU profile indicates that RecordTick is using less CPU, and this makes sense at least for a high filter miss rate. Before, we were recording two ticks per filter miss in iterators (CHECKED & USEFUL) and now recording just one (FILTERED).
Create DB with
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=10000000 -disable_wal=1 -write_buffer_size=30000000 -bloom_bits=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0 -prefix_size=8
```
And run simultaneous before&after with
```
TEST_TMPDIR=/dev/shm ./db_bench -readonly -benchmarks=seekrandom[-X1000] -num=10000000 -bloom_bits=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0 -prefix_size=8 -seek_nexts=1 -duration=20 -seed=43 -threads=8 -cache_size=1000000000 -statistics
```
Before: seekrandom [AVG 275 runs] : 189680 (± 222) ops/sec; 18.4 (± 0.0) MB/sec
After: seekrandom [AVG 275 runs] : 197110 (± 208) ops/sec; 19.1 (± 0.0) MB/sec
Reviewed By: ajkr
Differential Revision: D46029177
Pulled By: pdillinger
fbshipit-source-id: cdace79a2ea548d46c5900b068c5b7c3a02e5822
2 years ago
|
|
|
// Includes prefix stats
|
|
|
|
PERF_COUNTER_BY_LEVEL_ADD(bloom_filter_useful, filtered_keys,
|
|
|
|
rep_->level);
|
|
|
|
}
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
6 years ago
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
Status BlockBasedTable::ApproximateKeyAnchors(const ReadOptions& read_options,
|
|
|
|
std::vector<Anchor>& anchors) {
|
|
|
|
// We iterator the whole index block here. More efficient implementation
|
|
|
|
// is possible if we push this operation into IndexReader. For example, we
|
|
|
|
// can directly sample from restart block entries in the index block and
|
|
|
|
// only read keys needed. Here we take a simple solution. Performance is
|
|
|
|
// likely not to be a problem. We are compacting the whole file, so all
|
|
|
|
// keys will be read out anyway. An extra read to index block might be
|
|
|
|
// a small share of the overhead. We can try to optimize if needed.
|
Use CacheDependencies() at start of ApproximateKeyAnchors() (#11230)
Summary: Pull Request resolved: https://github.com/facebook/rocksdb/pull/11230
Test Plan:
- setup command: `$ ./db_bench -benchmarks=fillrandom,compact -compression_type=none -num=1000000 -write_buffer_size=4194304 -target_file_size_base=4194304 -use_direct_io_for_flush_and_compaction=true -partition_index_and_filters=true -bloom_bits=10 -metadata_block_size=1024`
- measure small read count bucketed by size: `$ strace -fye pread64 ./db_bench.ctrl -use_existing_db=true -benchmarks=compact -compaction_readahead_size=4194304 -compression_type=none -num=1000000 -write_buffer_size=4194304 -target_file_size_base=4194304 -use_direct_io_for_flush_and_compaction=true -partition_index_and_filters=true -bloom_bits=10 -metadata_block_size=1024 -subcompactions=4 -cache_size=1048576000 2>&1 >/dev/null | awk '/= [0-9]+$/{print "[", int($NF / 1024), "KB,", int(1 + $NF / 1024), "KB)"}' | sort -n -k 2 | uniq -c | head -3`
- before:
```
1119 [ 0 KB, 1 KB)
1 [ 6 KB, 7 KB)
2 [ 7 KB, 8 KB)
```
- after:
```
242 [ 0 KB, 1 KB)
1 [ 6 KB, 7 KB)
2 [ 7 KB, 8 KB)
```
Reviewed By: pdillinger
Differential Revision: D43388507
Pulled By: ajkr
fbshipit-source-id: a02413c9f615b00784700646825a9870ee10f3a7
2 years ago
|
|
|
//
|
|
|
|
// `CacheDependencies()` brings all the blocks into cache using one I/O. That
|
|
|
|
// way the full index scan usually finds the index data it is looking for in
|
|
|
|
// cache rather than doing an I/O for each "dependency" (partition).
|
Record and use the tail size to prefetch table tail (#11406)
Summary:
**Context:**
We prefetch the tail part of a SST file (i.e, the blocks after data blocks till the end of the file) during each SST file open in hope to prefetch all the stuff at once ahead of time for later read e.g, footer, meta index, filter/index etc. The existing approach to estimate the tail size to prefetch is through `TailPrefetchStats` heuristics introduced in https://github.com/facebook/rocksdb/pull/4156, which has caused small reads in unlucky case (e.g, small read into the tail buffer during table open in thread 1 under the same BlockBasedTableFactory object can make thread 2's tail prefetching use a small size that it shouldn't) and is hard to debug. Therefore we decide to record the exact tail size and use it directly to prefetch tail of the SST instead of relying heuristics.
**Summary:**
- Obtain and record in manifest the tail size in `BlockBasedTableBuilder::Finish()`
- For backward compatibility, we fall back to TailPrefetchStats and last to simple heuristics that the tail size is a linear portion of the file size - see PR conversation for more.
- Make`tail_start_offset` part of the table properties and deduct tail size to record in manifest for external files (e.g, file ingestion, import CF) and db repair (with no access to manifest).
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11406
Test Plan:
1. New UT
2. db bench
Note: db bench on /tmp/ where direct read is supported is too slow to finish and the default pinning setting in db bench is not helpful to profile # sst read of Get. Therefore I hacked the following to obtain the following comparison.
```
diff --git a/table/block_based/block_based_table_reader.cc b/table/block_based/block_based_table_reader.cc
index bd5669f0f..791484c1f 100644
--- a/table/block_based/block_based_table_reader.cc
+++ b/table/block_based/block_based_table_reader.cc
@@ -838,7 +838,7 @@ Status BlockBasedTable::PrefetchTail(
&tail_prefetch_size);
// Try file system prefetch
- if (!file->use_direct_io() && !force_direct_prefetch) {
+ if (false && !file->use_direct_io() && !force_direct_prefetch) {
if (!file->Prefetch(prefetch_off, prefetch_len, ro.rate_limiter_priority)
.IsNotSupported()) {
prefetch_buffer->reset(new FilePrefetchBuffer(
diff --git a/tools/db_bench_tool.cc b/tools/db_bench_tool.cc
index ea40f5fa0..39a0ac385 100644
--- a/tools/db_bench_tool.cc
+++ b/tools/db_bench_tool.cc
@@ -4191,6 +4191,8 @@ class Benchmark {
std::shared_ptr<TableFactory>(NewCuckooTableFactory(table_options));
} else {
BlockBasedTableOptions block_based_options;
+ block_based_options.metadata_cache_options.partition_pinning =
+ PinningTier::kAll;
block_based_options.checksum =
static_cast<ChecksumType>(FLAGS_checksum_type);
if (FLAGS_use_hash_search) {
```
Create DB
```
./db_bench --bloom_bits=3 --use_existing_db=1 --seed=1682546046158958 --partition_index_and_filters=1 --statistics=1 -db=/dev/shm/testdb/ -benchmarks=readrandom -key_size=3200 -value_size=512 -num=1000000 -write_buffer_size=6550000 -disable_auto_compactions=false -target_file_size_base=6550000 -compression_type=none
```
ReadRandom
```
./db_bench --bloom_bits=3 --use_existing_db=1 --seed=1682546046158958 --partition_index_and_filters=1 --statistics=1 -db=/dev/shm/testdb/ -benchmarks=readrandom -key_size=3200 -value_size=512 -num=1000000 -write_buffer_size=6550000 -disable_auto_compactions=false -target_file_size_base=6550000 -compression_type=none
```
(a) Existing (Use TailPrefetchStats for tail size + use seperate prefetch buffer in PartitionedFilter/IndexReader::CacheDependencies())
```
rocksdb.table.open.prefetch.tail.hit COUNT : 3395
rocksdb.sst.read.micros P50 : 5.655570 P95 : 9.931396 P99 : 14.845454 P100 : 585.000000 COUNT : 999905 SUM : 6590614
```
(b) This PR (Record tail size + use the same tail buffer in PartitionedFilter/IndexReader::CacheDependencies())
```
rocksdb.table.open.prefetch.tail.hit COUNT : 14257
rocksdb.sst.read.micros P50 : 5.173347 P95 : 9.015017 P99 : 12.912610 P100 : 228.000000 COUNT : 998547 SUM : 5976540
```
As we can see, we increase the prefetch tail hit count and decrease SST read count with this PR
3. Test backward compatibility by stepping through reading with post-PR code on a db generated pre-PR.
Reviewed By: pdillinger
Differential Revision: D45413346
Pulled By: hx235
fbshipit-source-id: 7d5e36a60a72477218f79905168d688452a4c064
2 years ago
|
|
|
Status s = rep_->index_reader->CacheDependencies(
|
|
|
|
read_options, false /* pin */, nullptr /* prefetch_buffer */);
|
Use CacheDependencies() at start of ApproximateKeyAnchors() (#11230)
Summary: Pull Request resolved: https://github.com/facebook/rocksdb/pull/11230
Test Plan:
- setup command: `$ ./db_bench -benchmarks=fillrandom,compact -compression_type=none -num=1000000 -write_buffer_size=4194304 -target_file_size_base=4194304 -use_direct_io_for_flush_and_compaction=true -partition_index_and_filters=true -bloom_bits=10 -metadata_block_size=1024`
- measure small read count bucketed by size: `$ strace -fye pread64 ./db_bench.ctrl -use_existing_db=true -benchmarks=compact -compaction_readahead_size=4194304 -compression_type=none -num=1000000 -write_buffer_size=4194304 -target_file_size_base=4194304 -use_direct_io_for_flush_and_compaction=true -partition_index_and_filters=true -bloom_bits=10 -metadata_block_size=1024 -subcompactions=4 -cache_size=1048576000 2>&1 >/dev/null | awk '/= [0-9]+$/{print "[", int($NF / 1024), "KB,", int(1 + $NF / 1024), "KB)"}' | sort -n -k 2 | uniq -c | head -3`
- before:
```
1119 [ 0 KB, 1 KB)
1 [ 6 KB, 7 KB)
2 [ 7 KB, 8 KB)
```
- after:
```
242 [ 0 KB, 1 KB)
1 [ 6 KB, 7 KB)
2 [ 7 KB, 8 KB)
```
Reviewed By: pdillinger
Differential Revision: D43388507
Pulled By: ajkr
fbshipit-source-id: a02413c9f615b00784700646825a9870ee10f3a7
2 years ago
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
IndexBlockIter iiter_on_stack;
|
|
|
|
auto iiter = NewIndexIterator(
|
|
|
|
read_options, /*disable_prefix_seek=*/false, &iiter_on_stack,
|
|
|
|
/*get_context=*/nullptr, /*lookup_context=*/nullptr);
|
|
|
|
std::unique_ptr<InternalIteratorBase<IndexValue>> iiter_unique_ptr;
|
|
|
|
if (iiter != &iiter_on_stack) {
|
|
|
|
iiter_unique_ptr.reset(iiter);
|
|
|
|
}
|
|
|
|
|
|
|
|
// If needed the threshold could be more adaptive. For example, it can be
|
|
|
|
// based on size, so that a larger will be sampled to more partitions than a
|
|
|
|
// smaller file. The size might also need to be passed in by the caller based
|
|
|
|
// on total compaction size.
|
|
|
|
const uint64_t kMaxNumAnchors = uint64_t{128};
|
|
|
|
uint64_t num_blocks = this->GetTableProperties()->num_data_blocks;
|
|
|
|
uint64_t num_blocks_per_anchor = num_blocks / kMaxNumAnchors;
|
|
|
|
if (num_blocks_per_anchor == 0) {
|
|
|
|
num_blocks_per_anchor = 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
uint64_t count = 0;
|
|
|
|
std::string last_key;
|
|
|
|
uint64_t range_size = 0;
|
|
|
|
uint64_t prev_offset = 0;
|
|
|
|
for (iiter->SeekToFirst(); iiter->Valid(); iiter->Next()) {
|
|
|
|
const BlockHandle& bh = iiter->value().handle;
|
|
|
|
range_size += bh.offset() + bh.size() - prev_offset;
|
|
|
|
prev_offset = bh.offset() + bh.size();
|
|
|
|
if (++count % num_blocks_per_anchor == 0) {
|
|
|
|
count = 0;
|
|
|
|
anchors.emplace_back(iiter->user_key(), range_size);
|
|
|
|
range_size = 0;
|
|
|
|
} else {
|
|
|
|
last_key = iiter->user_key().ToString();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if (count != 0) {
|
|
|
|
anchors.emplace_back(last_key, range_size);
|
|
|
|
}
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
|
|
|
bool BlockBasedTable::TimestampMayMatch(const ReadOptions& read_options) const {
|
|
|
|
if (read_options.timestamp != nullptr && !rep_->min_timestamp.empty()) {
|
|
|
|
RecordTick(rep_->ioptions.stats, TIMESTAMP_FILTER_TABLE_CHECKED);
|
|
|
|
auto read_ts = read_options.timestamp;
|
|
|
|
auto comparator = rep_->internal_comparator.user_comparator();
|
|
|
|
if (comparator->CompareTimestamp(*read_ts, rep_->min_timestamp) < 0) {
|
|
|
|
RecordTick(rep_->ioptions.stats, TIMESTAMP_FILTER_TABLE_FILTERED);
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
Skip bottom-level filter block caching when hit-optimized
Summary:
When Get() or NewIterator() trigger file loads, skip caching the filter block if
(1) optimize_filters_for_hits is set and (2) the file is on the bottommost
level. Also skip checking filters under the same conditions, which means that
for a preloaded file or a file that was trivially-moved to the bottom level, its
filter block will eventually expire from the cache.
- added parameters/instance variables in various places in order to propagate the config ("skip_filters") from version_set to block_based_table_reader
- in BlockBasedTable::Rep, this optimization prevents filter from being loaded when the file is opened simply by setting filter_policy = nullptr
- in BlockBasedTable::Get/BlockBasedTable::NewIterator, this optimization prevents filter from being used (even if it was loaded already) by setting filter = nullptr
Test Plan:
updated unit test:
$ ./db_test --gtest_filter=DBTest.OptimizeFiltersForHits
will also run 'make check'
Reviewers: sdong, igor, paultuckfield, anthony, rven, kradhakrishnan, IslamAbdelRahman, yhchiang
Reviewed By: yhchiang
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D51633
9 years ago
|
|
|
Status BlockBasedTable::Get(const ReadOptions& read_options, const Slice& key,
|
|
|
|
GetContext* get_context,
|
|
|
|
const SliceTransform* prefix_extractor,
|
|
|
|
bool skip_filters) {
|
|
|
|
// Similar to Bloom filter !may_match
|
|
|
|
// If timestamp is beyond the range of the table, skip
|
|
|
|
if (!TimestampMayMatch(read_options)) {
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
assert(key.size() >= 8); // key must be internal key
|
|
|
|
assert(get_context != nullptr);
|
|
|
|
Status s;
|
|
|
|
const bool no_io = read_options.read_tier == kBlockCacheTier;
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
|
|
|
|
FilterBlockReader* const filter =
|
|
|
|
!skip_filters ? rep_->filter.get() : nullptr;
|
|
|
|
|
|
|
|
// First check the full filter
|
|
|
|
// If full filter not useful, Then go into each block
|
|
|
|
uint64_t tracing_get_id = get_context->get_tracing_get_id();
|
|
|
|
BlockCacheLookupContext lookup_context{
|
|
|
|
TableReaderCaller::kUserGet, tracing_get_id,
|
|
|
|
/*get_from_user_specified_snapshot=*/read_options.snapshot != nullptr};
|
|
|
|
if (block_cache_tracer_ && block_cache_tracer_->is_tracing_enabled()) {
|
|
|
|
// Trace the key since it contains both user key and sequence number.
|
|
|
|
lookup_context.referenced_key = key.ToString();
|
|
|
|
lookup_context.get_from_user_specified_snapshot =
|
|
|
|
read_options.snapshot != nullptr;
|
|
|
|
}
|
|
|
|
TEST_SYNC_POINT("BlockBasedTable::Get:BeforeFilterMatch");
|
|
|
|
const bool may_match =
|
|
|
|
FullFilterKeyMayMatch(filter, key, no_io, prefix_extractor, get_context,
|
|
|
|
&lookup_context, read_options);
|
|
|
|
TEST_SYNC_POINT("BlockBasedTable::Get:AfterFilterMatch");
|
Much better stats for seeks and prefix filtering (#11460)
Summary:
We want to know more about opportunities for better range filters, and the effectiveness of our own range filters. Currently the stats are very limited, essentially logging just hits and misses against prefix filters for range scans in BLOOM_FILTER_PREFIX_* without tracking the false positive rate. Perhaps confusingly, when prefix filters are used for point queries, the stats are currently going into the non-PREFIX tickers.
This change does several things:
* Introduce new stat tickers for seeks and related filtering, \*LEVEL_SEEK\*
* Most importantly, allows us to see opportunities for range filtering. Specifically, we can count how many times a seek in an SST file accesses at least one data block, and how many times at least one value() is then accessed. If a data block was accessed but no value(), we can generally assume that the key(s) seen was(were) not of interest so could have been filtered with the right kind of filter, avoiding the data block access.
* We can get the same level of detail when a filter (for now, prefix Bloom/ribbon) is used, or not. Specifically, we can infer a false positive rate for prefix filters (not available before) from the seek "false positive" rate: when a data block is accessed but no value() is called. (There can be other explanations for a seek false positive, but in typical iterator usage it would indicate a filter false positive.)
* For efficiency, I wanted to avoid making additional calls to the prefix extractor (or key comparisons, etc.), which would be required if we wanted to more precisely detect filter false positives. I believe that instrumenting value() is the best balance of efficiency vs. accurately measuring what we are often interested in.
* The stats are divided between last level and non-last levels, to help understand potential tiered storage use cases.
* The old BLOOM_FILTER_PREFIX_* stats have a different meaning: no longer referring to iterators but to point queries using prefix filters. BLOOM_FILTER_PREFIX_TRUE_POSITIVE is added for computing the prefix false positive rate on point queries, which can be due to filter false positives as well as different keys with the same prefix.
* Similarly, the non-PREFIX BLOOM_FILTER stats are now for whole key filtering only.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11460
Test Plan:
unit tests updated, including updating many to pop the stat value since last read to improve test
readability and maintainability.
Performance test shows a consistent small improvement with these changes, both with clang and with gcc. CPU profile indicates that RecordTick is using less CPU, and this makes sense at least for a high filter miss rate. Before, we were recording two ticks per filter miss in iterators (CHECKED & USEFUL) and now recording just one (FILTERED).
Create DB with
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=10000000 -disable_wal=1 -write_buffer_size=30000000 -bloom_bits=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0 -prefix_size=8
```
And run simultaneous before&after with
```
TEST_TMPDIR=/dev/shm ./db_bench -readonly -benchmarks=seekrandom[-X1000] -num=10000000 -bloom_bits=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0 -prefix_size=8 -seek_nexts=1 -duration=20 -seed=43 -threads=8 -cache_size=1000000000 -statistics
```
Before: seekrandom [AVG 275 runs] : 189680 (± 222) ops/sec; 18.4 (± 0.0) MB/sec
After: seekrandom [AVG 275 runs] : 197110 (± 208) ops/sec; 19.1 (± 0.0) MB/sec
Reviewed By: ajkr
Differential Revision: D46029177
Pulled By: pdillinger
fbshipit-source-id: cdace79a2ea548d46c5900b068c5b7c3a02e5822
2 years ago
|
|
|
if (may_match) {
|
|
|
|
IndexBlockIter iiter_on_stack;
|
|
|
|
// if prefix_extractor found in block differs from options, disable
|
|
|
|
// BlockPrefixIndex. Only do this check when index_type is kHashSearch.
|
|
|
|
bool need_upper_bound_check = false;
|
|
|
|
if (rep_->index_type == BlockBasedTableOptions::kHashSearch) {
|
|
|
|
need_upper_bound_check = PrefixExtractorChanged(prefix_extractor);
|
|
|
|
}
|
|
|
|
auto iiter =
|
|
|
|
NewIndexIterator(read_options, need_upper_bound_check, &iiter_on_stack,
|
|
|
|
get_context, &lookup_context);
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
std::unique_ptr<InternalIteratorBase<IndexValue>> iiter_unique_ptr;
|
|
|
|
if (iiter != &iiter_on_stack) {
|
|
|
|
iiter_unique_ptr.reset(iiter);
|
|
|
|
}
|
|
|
|
|
|
|
|
size_t ts_sz =
|
|
|
|
rep_->internal_comparator.user_comparator()->timestamp_size();
|
|
|
|
bool matched = false; // if such user key matched a key in SST
|
|
|
|
bool done = false;
|
|
|
|
for (iiter->Seek(key); iiter->Valid() && !done; iiter->Next()) {
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
IndexValue v = iiter->value();
|
|
|
|
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
if (!v.first_internal_key.empty() && !skip_filters &&
|
|
|
|
UserComparatorWrapper(rep_->internal_comparator.user_comparator())
|
Fix a bug in key comparison when index type is kBinarySearchWithFirstKey (#8062)
Summary:
When timestamp is enabled, key comparison should take this into account.
In `BlockBasedTableReader::Get()`, `BlockBasedTableReader::MultiGet()`,
assume the target key is `key`, and the timestamp upper bound is `ts`.
The highest key in current block is (key, ts1), while the lowest key in next
block is (key, ts2).
If
```
ts1 > ts > ts2
```
then
```
(key, ts1) < (key, ts) < (key, ts2)
```
It can be shown that if `Compare()` is used, then we will mistakenly skip the next
block. Instead, we should use `CompareWithoutTimestamp()`.
The majority of this PR makes some existing tests in `db_with_timestamp_basic_test.cc`
parameterized so that different index types can be tested. A new unit test is
also added for more coverage.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8062
Test Plan: make check
Reviewed By: ltamasi
Differential Revision: D27057557
Pulled By: riversand963
fbshipit-source-id: c1062fa7c159ed600a1ad7e461531d52265021f1
4 years ago
|
|
|
.CompareWithoutTimestamp(
|
|
|
|
ExtractUserKey(key),
|
|
|
|
ExtractUserKey(v.first_internal_key)) < 0) {
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
// The requested key falls between highest key in previous block and
|
|
|
|
// lowest key in current block.
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
BlockCacheLookupContext lookup_data_block_context{
|
|
|
|
TableReaderCaller::kUserGet, tracing_get_id,
|
|
|
|
/*get_from_user_specified_snapshot=*/read_options.snapshot !=
|
|
|
|
nullptr};
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
bool does_referenced_key_exist = false;
|
|
|
|
DataBlockIter biter;
|
|
|
|
uint64_t referenced_data_size = 0;
|
|
|
|
Status tmp_status;
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
NewDataBlockIterator<DataBlockIter>(
|
|
|
|
read_options, v.handle, &biter, BlockType::kData, get_context,
|
|
|
|
&lookup_data_block_context, /*prefetch_buffer=*/nullptr,
|
|
|
|
/*for_compaction=*/false, /*async_read=*/false, tmp_status);
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
|
|
|
|
if (no_io && biter.status().IsIncomplete()) {
|
|
|
|
// couldn't get block from block_cache
|
|
|
|
// Update Saver.state to Found because we are only looking for
|
|
|
|
// whether we can guarantee the key is not there when "no_io" is set
|
|
|
|
get_context->MarkKeyMayExist();
|
|
|
|
s = biter.status();
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
break;
|
|
|
|
}
|
|
|
|
if (!biter.status().ok()) {
|
|
|
|
s = biter.status();
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
bool may_exist = biter.SeekForGet(key);
|
|
|
|
// If user-specified timestamp is supported, we cannot end the search
|
|
|
|
// just because hash index lookup indicates the key+ts does not exist.
|
|
|
|
if (!may_exist && ts_sz == 0) {
|
|
|
|
// HashSeek cannot find the key this block and the the iter is not
|
|
|
|
// the end of the block, i.e. cannot be in the following blocks
|
|
|
|
// either. In this case, the seek_key cannot be found, so we break
|
|
|
|
// from the top level for-loop.
|
|
|
|
done = true;
|
|
|
|
} else {
|
|
|
|
// Call the *saver function on each entry/block until it returns false
|
|
|
|
for (; biter.Valid(); biter.Next()) {
|
|
|
|
ParsedInternalKey parsed_key;
|
|
|
|
Status pik_status = ParseInternalKey(
|
|
|
|
biter.key(), &parsed_key, false /* log_err_key */); // TODO
|
|
|
|
if (!pik_status.ok()) {
|
|
|
|
s = pik_status;
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
}
|
|
|
|
|
|
|
|
if (!get_context->SaveValue(
|
|
|
|
parsed_key, biter.value(), &matched,
|
|
|
|
biter.IsValuePinned() ? &biter : nullptr)) {
|
|
|
|
if (get_context->State() == GetContext::GetState::kFound) {
|
|
|
|
does_referenced_key_exist = true;
|
|
|
|
referenced_data_size = biter.key().size() + biter.value().size();
|
|
|
|
}
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
done = true;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
s = biter.status();
|
|
|
|
if (!s.ok()) {
|
|
|
|
break;
|
|
|
|
}
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
}
|
|
|
|
// Write the block cache access record.
|
|
|
|
if (block_cache_tracer_ && block_cache_tracer_->is_tracing_enabled()) {
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
// Avoid making copy of block_key, cf_name, and referenced_key when
|
|
|
|
// constructing the access record.
|
|
|
|
Slice referenced_key;
|
|
|
|
if (does_referenced_key_exist) {
|
|
|
|
referenced_key = biter.key();
|
|
|
|
} else {
|
|
|
|
referenced_key = key;
|
|
|
|
}
|
Refactor block cache tracing w/improved MultiGet (#11339)
Summary:
After https://github.com/facebook/rocksdb/issues/11301, I wasn't sure whether I had regressed block cache tracing with MultiGet. Demo PR https://github.com/facebook/rocksdb/issues/11330 shows the flawed state of tracing MultiGet before my change, and based on the unit test, there was essentially no change in tracing behavior with https://github.com/facebook/rocksdb/issues/11301. This change is to leave that code and behavior better than I found it.
This change is not intended to change any production behaviors except when block cache tracing is active, though might improve general read path efficiency by disabling some related tracking when such tracing is disabled.
More detail on production code:
* Refactoring to consolidate the construction of BlockCacheTraceRecord, and other related functionality, in block-based table reader, though it's somewhat awkward to preserve an optimization to avoid copying Slices into temporary strings in BlockCacheLookupContext.
* Accurately track cache hits and misses (etc.) for each data block accessed by a MultiGet(). (Previously reported hits as misses.)
* Reduced repeated checking of `block_cache_tracer_` state (by creating lookup_context only when active) for efficiency and to reduce the risk of corner case bugs where tracing is enabled or disabled for different parts of a read op. (See a TODO below)
* Improved estimate calculation for num_keys_in_block (see code comment)
Possible follow-up:
* `XXX:` use_cache=true means double cache query? (possible double-query of block cache when allow_mmap_reads=true)
* `TODO:` need more than one lookup_context here to track individual filter and index partition hits and misses
* `TODO:` optimize more state checks of `block_cache_tracer_` down to `lookup_context != nullptr`
* Pre-existing `XXX:` There appear to be 'break' statements above that bypass this writing of the block cache trace record
* Expand test coverage (see below)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11339
Test Plan:
* Added a basic unit test for block cache tracing MultiGet, for now just covering one data block with two keys.
* Added HitMissCountingCache to independently verify that the actual block cache trace and expected block cache trace also agree with the actual number of cache hits / misses (nothing missing or mislabeled). For now only used with MultiGet test.
* Better testing of num_keys_in_block, for now just with MultiGet
* Misc improvements to table_test to improve clarity, such as making it clear that certain keys are auto-inserted at the start of every test.
Performance test:
Testing multireadrandom as in https://github.com/facebook/rocksdb/issues/11301, except averaging over distinct runs rather than [-X30] which doesn't seem to sufficiently reset after each run to work as an independent test run.
Base with revert of 11301: 3148926 ops/sec
Base: 3019146 ops/sec
New: 2999529 ops/sec
Possibly a tiny MultiGet CPU regression with this change. We are now always allocating an additional vector for the LookupContexts. I'm still contemplating options to try to correct the regression in https://github.com/facebook/rocksdb/issues/11301.
Testing readrandom:
Base with revert of 11301: 2311988
Base: 2281726
New: 2299722
Possibly a tiny Get CPU improvement with this change. We are now avoiding some unnecessary LookupContext population.
Reviewed By: akankshamahajan15
Differential Revision: D44557845
Pulled By: pdillinger
fbshipit-source-id: b841691799d2a48fb59cc8880dc7cbb1e107ae3d
2 years ago
|
|
|
FinishTraceRecord(lookup_data_block_context,
|
|
|
|
lookup_data_block_context.block_key, referenced_key,
|
|
|
|
does_referenced_key_exist, referenced_data_size);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (done) {
|
|
|
|
// Avoid the extra Next which is expensive in two-level indexes
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
Remove deprecated block-based filter (#10184)
Summary:
In https://github.com/facebook/rocksdb/issues/9535, release 7.0, we hid the old block-based filter from being created using
the public API, because of its inefficiency. Although we normally maintain read compatibility
on old DBs forever, filters are not required for reading a DB, only for optimizing read
performance. Thus, it should be acceptable to remove this code and the substantial
maintenance burden it carries as useful features are developed and validated (such
as user timestamp).
This change completely removes the code for reading and writing the old block-based
filters, net removing about 1370 lines of code no longer needed. Options removed from
testing / benchmarking tools. The prior existence is only evident in a couple of places:
* `CacheEntryRole::kDeprecatedFilterBlock` - We can update this public API enum in
a major release to minimize source code incompatibilities.
* A warning is logged when an old table file is opened that used the old block-based
filter. This is provided as a courtesy, and would be a pain to unit test, so manual testing
should suffice. Unfortunately, sst_dump does not tell you whether a file uses
block-based filter, and the structure of the code makes it very difficult to fix.
* To detect that case, `kObsoleteFilterBlockPrefix` (renamed from `kFilterBlockPrefix`)
for metaindex is maintained (for now).
Other notes:
* In some cases where numbers are associated with filter configurations, we have had to
update the assigned numbers so that they all correspond to something that exists.
* Fixed potential stat counting bug by assuming `filter_checked = false` for cases
like `filter == nullptr` rather than assuming `filter_checked = true`
* Removed obsolete `block_offset` and `prefix_extractor` parameters from several
functions.
* Removed some unnecessary checks `if (!table_prefix_extractor() && !prefix_extractor)`
because the caller guarantees the prefix extractor exists and is compatible
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10184
Test Plan:
tests updated, manually test new warning in LOG using base version to
generate a DB
Reviewed By: riversand963
Differential Revision: D37212647
Pulled By: pdillinger
fbshipit-source-id: 06ee020d8de3b81260ffc36ad0c1202cbf463a80
2 years ago
|
|
|
if (matched && filter != nullptr) {
|
Much better stats for seeks and prefix filtering (#11460)
Summary:
We want to know more about opportunities for better range filters, and the effectiveness of our own range filters. Currently the stats are very limited, essentially logging just hits and misses against prefix filters for range scans in BLOOM_FILTER_PREFIX_* without tracking the false positive rate. Perhaps confusingly, when prefix filters are used for point queries, the stats are currently going into the non-PREFIX tickers.
This change does several things:
* Introduce new stat tickers for seeks and related filtering, \*LEVEL_SEEK\*
* Most importantly, allows us to see opportunities for range filtering. Specifically, we can count how many times a seek in an SST file accesses at least one data block, and how many times at least one value() is then accessed. If a data block was accessed but no value(), we can generally assume that the key(s) seen was(were) not of interest so could have been filtered with the right kind of filter, avoiding the data block access.
* We can get the same level of detail when a filter (for now, prefix Bloom/ribbon) is used, or not. Specifically, we can infer a false positive rate for prefix filters (not available before) from the seek "false positive" rate: when a data block is accessed but no value() is called. (There can be other explanations for a seek false positive, but in typical iterator usage it would indicate a filter false positive.)
* For efficiency, I wanted to avoid making additional calls to the prefix extractor (or key comparisons, etc.), which would be required if we wanted to more precisely detect filter false positives. I believe that instrumenting value() is the best balance of efficiency vs. accurately measuring what we are often interested in.
* The stats are divided between last level and non-last levels, to help understand potential tiered storage use cases.
* The old BLOOM_FILTER_PREFIX_* stats have a different meaning: no longer referring to iterators but to point queries using prefix filters. BLOOM_FILTER_PREFIX_TRUE_POSITIVE is added for computing the prefix false positive rate on point queries, which can be due to filter false positives as well as different keys with the same prefix.
* Similarly, the non-PREFIX BLOOM_FILTER stats are now for whole key filtering only.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11460
Test Plan:
unit tests updated, including updating many to pop the stat value since last read to improve test
readability and maintainability.
Performance test shows a consistent small improvement with these changes, both with clang and with gcc. CPU profile indicates that RecordTick is using less CPU, and this makes sense at least for a high filter miss rate. Before, we were recording two ticks per filter miss in iterators (CHECKED & USEFUL) and now recording just one (FILTERED).
Create DB with
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=10000000 -disable_wal=1 -write_buffer_size=30000000 -bloom_bits=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0 -prefix_size=8
```
And run simultaneous before&after with
```
TEST_TMPDIR=/dev/shm ./db_bench -readonly -benchmarks=seekrandom[-X1000] -num=10000000 -bloom_bits=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0 -prefix_size=8 -seek_nexts=1 -duration=20 -seed=43 -threads=8 -cache_size=1000000000 -statistics
```
Before: seekrandom [AVG 275 runs] : 189680 (± 222) ops/sec; 18.4 (± 0.0) MB/sec
After: seekrandom [AVG 275 runs] : 197110 (± 208) ops/sec; 19.1 (± 0.0) MB/sec
Reviewed By: ajkr
Differential Revision: D46029177
Pulled By: pdillinger
fbshipit-source-id: cdace79a2ea548d46c5900b068c5b7c3a02e5822
2 years ago
|
|
|
if (rep_->whole_key_filtering) {
|
|
|
|
RecordTick(rep_->ioptions.stats, BLOOM_FILTER_FULL_TRUE_POSITIVE);
|
|
|
|
} else {
|
|
|
|
RecordTick(rep_->ioptions.stats, BLOOM_FILTER_PREFIX_TRUE_POSITIVE);
|
|
|
|
}
|
|
|
|
// Includes prefix stats
|
|
|
|
PERF_COUNTER_BY_LEVEL_ADD(bloom_filter_full_true_positive, 1,
|
|
|
|
rep_->level);
|
|
|
|
}
|
Much better stats for seeks and prefix filtering (#11460)
Summary:
We want to know more about opportunities for better range filters, and the effectiveness of our own range filters. Currently the stats are very limited, essentially logging just hits and misses against prefix filters for range scans in BLOOM_FILTER_PREFIX_* without tracking the false positive rate. Perhaps confusingly, when prefix filters are used for point queries, the stats are currently going into the non-PREFIX tickers.
This change does several things:
* Introduce new stat tickers for seeks and related filtering, \*LEVEL_SEEK\*
* Most importantly, allows us to see opportunities for range filtering. Specifically, we can count how many times a seek in an SST file accesses at least one data block, and how many times at least one value() is then accessed. If a data block was accessed but no value(), we can generally assume that the key(s) seen was(were) not of interest so could have been filtered with the right kind of filter, avoiding the data block access.
* We can get the same level of detail when a filter (for now, prefix Bloom/ribbon) is used, or not. Specifically, we can infer a false positive rate for prefix filters (not available before) from the seek "false positive" rate: when a data block is accessed but no value() is called. (There can be other explanations for a seek false positive, but in typical iterator usage it would indicate a filter false positive.)
* For efficiency, I wanted to avoid making additional calls to the prefix extractor (or key comparisons, etc.), which would be required if we wanted to more precisely detect filter false positives. I believe that instrumenting value() is the best balance of efficiency vs. accurately measuring what we are often interested in.
* The stats are divided between last level and non-last levels, to help understand potential tiered storage use cases.
* The old BLOOM_FILTER_PREFIX_* stats have a different meaning: no longer referring to iterators but to point queries using prefix filters. BLOOM_FILTER_PREFIX_TRUE_POSITIVE is added for computing the prefix false positive rate on point queries, which can be due to filter false positives as well as different keys with the same prefix.
* Similarly, the non-PREFIX BLOOM_FILTER stats are now for whole key filtering only.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11460
Test Plan:
unit tests updated, including updating many to pop the stat value since last read to improve test
readability and maintainability.
Performance test shows a consistent small improvement with these changes, both with clang and with gcc. CPU profile indicates that RecordTick is using less CPU, and this makes sense at least for a high filter miss rate. Before, we were recording two ticks per filter miss in iterators (CHECKED & USEFUL) and now recording just one (FILTERED).
Create DB with
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=10000000 -disable_wal=1 -write_buffer_size=30000000 -bloom_bits=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0 -prefix_size=8
```
And run simultaneous before&after with
```
TEST_TMPDIR=/dev/shm ./db_bench -readonly -benchmarks=seekrandom[-X1000] -num=10000000 -bloom_bits=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0 -prefix_size=8 -seek_nexts=1 -duration=20 -seed=43 -threads=8 -cache_size=1000000000 -statistics
```
Before: seekrandom [AVG 275 runs] : 189680 (± 222) ops/sec; 18.4 (± 0.0) MB/sec
After: seekrandom [AVG 275 runs] : 197110 (± 208) ops/sec; 19.1 (± 0.0) MB/sec
Reviewed By: ajkr
Differential Revision: D46029177
Pulled By: pdillinger
fbshipit-source-id: cdace79a2ea548d46c5900b068c5b7c3a02e5822
2 years ago
|
|
|
|
|
|
|
if (s.ok() && !iiter->status().IsNotFound()) {
|
|
|
|
s = iiter->status();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
Status BlockBasedTable::MultiGetFilter(const ReadOptions& read_options,
|
|
|
|
const SliceTransform* prefix_extractor,
|
|
|
|
MultiGetRange* mget_range) {
|
|
|
|
if (mget_range->empty()) {
|
|
|
|
// Caller should ensure non-empty (performance bug)
|
|
|
|
assert(false);
|
|
|
|
return Status::OK(); // Nothing to do
|
|
|
|
}
|
|
|
|
|
|
|
|
FilterBlockReader* const filter = rep_->filter.get();
|
|
|
|
if (!filter) {
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
|
|
|
// First check the full filter
|
|
|
|
// If full filter not useful, Then go into each block
|
|
|
|
const bool no_io = read_options.read_tier == kBlockCacheTier;
|
|
|
|
uint64_t tracing_mget_id = BlockCacheTraceHelper::kReservedGetId;
|
|
|
|
if (mget_range->begin()->get_context) {
|
|
|
|
tracing_mget_id = mget_range->begin()->get_context->get_tracing_get_id();
|
|
|
|
}
|
|
|
|
BlockCacheLookupContext lookup_context{
|
|
|
|
TableReaderCaller::kUserMultiGet, tracing_mget_id,
|
|
|
|
/*_get_from_user_specified_snapshot=*/read_options.snapshot != nullptr};
|
|
|
|
FullFilterKeysMayMatch(filter, mget_range, no_io, prefix_extractor,
|
|
|
|
&lookup_context, read_options);
|
|
|
|
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
|
|
|
Status BlockBasedTable::Prefetch(const ReadOptions& read_options,
|
|
|
|
const Slice* const begin,
|
|
|
|
const Slice* const end) {
|
|
|
|
auto& comparator = rep_->internal_comparator;
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
UserComparatorWrapper user_comparator(comparator.user_comparator());
|
|
|
|
// pre-condition
|
|
|
|
if (begin && end && comparator.Compare(*begin, *end) > 0) {
|
|
|
|
return Status::InvalidArgument(*begin, *end);
|
|
|
|
}
|
|
|
|
BlockCacheLookupContext lookup_context{TableReaderCaller::kPrefetch};
|
|
|
|
IndexBlockIter iiter_on_stack;
|
|
|
|
auto iiter = NewIndexIterator(read_options, /*need_upper_bound_check=*/false,
|
|
|
|
&iiter_on_stack, /*get_context=*/nullptr,
|
|
|
|
&lookup_context);
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
std::unique_ptr<InternalIteratorBase<IndexValue>> iiter_unique_ptr;
|
|
|
|
if (iiter != &iiter_on_stack) {
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
iiter_unique_ptr = std::unique_ptr<InternalIteratorBase<IndexValue>>(iiter);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!iiter->status().ok()) {
|
|
|
|
// error opening index iterator
|
|
|
|
return iiter->status();
|
|
|
|
}
|
|
|
|
|
|
|
|
// indicates if we are on the last page that need to be pre-fetched
|
|
|
|
bool prefetching_boundary_page = false;
|
|
|
|
|
|
|
|
for (begin ? iiter->Seek(*begin) : iiter->SeekToFirst(); iiter->Valid();
|
|
|
|
iiter->Next()) {
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
BlockHandle block_handle = iiter->value().handle;
|
|
|
|
const bool is_user_key = !rep_->index_key_includes_seq;
|
|
|
|
if (end &&
|
|
|
|
((!is_user_key && comparator.Compare(iiter->key(), *end) >= 0) ||
|
|
|
|
(is_user_key &&
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
user_comparator.Compare(iiter->key(), ExtractUserKey(*end)) >= 0))) {
|
|
|
|
if (prefetching_boundary_page) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
// The index entry represents the last key in the data block.
|
|
|
|
// We should load this page into memory as well, but no more
|
|
|
|
prefetching_boundary_page = true;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Load the block specified by the block_handle into the block cache
|
|
|
|
DataBlockIter biter;
|
|
|
|
Status tmp_status;
|
|
|
|
NewDataBlockIterator<DataBlockIter>(
|
|
|
|
read_options, block_handle, &biter, /*type=*/BlockType::kData,
|
|
|
|
/*get_context=*/nullptr, &lookup_context,
|
|
|
|
/*prefetch_buffer=*/nullptr, /*for_compaction=*/false,
|
|
|
|
/*async_read=*/false, tmp_status);
|
|
|
|
|
|
|
|
if (!biter.status().ok()) {
|
|
|
|
// there was an unexpected error while pre-fetching
|
|
|
|
return biter.status();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
|
|
|
Status BlockBasedTable::VerifyChecksum(const ReadOptions& read_options,
|
|
|
|
TableReaderCaller caller) {
|
|
|
|
Status s;
|
|
|
|
// Check Meta blocks
|
|
|
|
std::unique_ptr<Block> metaindex;
|
|
|
|
std::unique_ptr<InternalIterator> metaindex_iter;
|
|
|
|
s = ReadMetaIndexBlock(read_options, nullptr /* prefetch buffer */,
|
|
|
|
&metaindex, &metaindex_iter);
|
|
|
|
if (s.ok()) {
|
|
|
|
s = VerifyChecksumInMetaBlocks(read_options, metaindex_iter.get());
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
// Check Data blocks
|
|
|
|
IndexBlockIter iiter_on_stack;
|
|
|
|
BlockCacheLookupContext context{caller};
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
InternalIteratorBase<IndexValue>* iiter = NewIndexIterator(
|
|
|
|
read_options, /*disable_prefix_seek=*/false, &iiter_on_stack,
|
|
|
|
/*get_context=*/nullptr, &context);
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
std::unique_ptr<InternalIteratorBase<IndexValue>> iiter_unique_ptr;
|
|
|
|
if (iiter != &iiter_on_stack) {
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
iiter_unique_ptr = std::unique_ptr<InternalIteratorBase<IndexValue>>(iiter);
|
|
|
|
}
|
|
|
|
if (!iiter->status().ok()) {
|
|
|
|
// error opening index iterator
|
|
|
|
return iiter->status();
|
|
|
|
}
|
|
|
|
s = VerifyChecksumInBlocks(read_options, iiter);
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
Status BlockBasedTable::VerifyChecksumInBlocks(
|
|
|
|
const ReadOptions& read_options,
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
InternalIteratorBase<IndexValue>* index_iter) {
|
|
|
|
Status s;
|
|
|
|
// We are scanning the whole file, so no need to do exponential
|
|
|
|
// increasing of the buffer size.
|
|
|
|
size_t readahead_size = (read_options.readahead_size != 0)
|
|
|
|
? read_options.readahead_size
|
|
|
|
: rep_->table_options.max_auto_readahead_size;
|
|
|
|
// FilePrefetchBuffer doesn't work in mmap mode and readahead is not
|
|
|
|
// needed there.
|
|
|
|
FilePrefetchBuffer prefetch_buffer(
|
|
|
|
readahead_size /* readahead_size */,
|
|
|
|
readahead_size /* max_readahead_size */,
|
|
|
|
!rep_->ioptions.allow_mmap_reads /* enable */);
|
|
|
|
|
|
|
|
for (index_iter->SeekToFirst(); index_iter->Valid(); index_iter->Next()) {
|
|
|
|
s = index_iter->status();
|
|
|
|
if (!s.ok()) {
|
|
|
|
break;
|
|
|
|
}
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
BlockHandle handle = index_iter->value().handle;
|
|
|
|
BlockContents contents;
|
|
|
|
BlockFetcher block_fetcher(
|
|
|
|
rep_->file.get(), &prefetch_buffer, rep_->footer, read_options, handle,
|
|
|
|
&contents, rep_->ioptions, false /* decompress */,
|
|
|
|
false /*maybe_compressed*/, BlockType::kData,
|
|
|
|
UncompressionDict::GetEmptyDict(), rep_->persistent_cache_options);
|
|
|
|
s = block_fetcher.ReadBlockContents();
|
|
|
|
if (!s.ok()) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if (s.ok()) {
|
|
|
|
// In the case of two level indexes, we would have exited the above loop
|
|
|
|
// by checking index_iter->Valid(), but Valid() might have returned false
|
|
|
|
// due to an IO error. So check the index_iter status
|
|
|
|
s = index_iter->status();
|
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
BlockType BlockBasedTable::GetBlockTypeForMetaBlockByName(
|
|
|
|
const Slice& meta_block_name) {
|
|
|
|
if (meta_block_name.starts_with(kFullFilterBlockPrefix)) {
|
|
|
|
return BlockType::kFilter;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (meta_block_name.starts_with(kPartitionedFilterBlockPrefix)) {
|
|
|
|
return BlockType::kFilterPartitionIndex;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (meta_block_name == kPropertiesBlockName) {
|
|
|
|
return BlockType::kProperties;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (meta_block_name == kCompressionDictBlockName) {
|
|
|
|
return BlockType::kCompressionDictionary;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (meta_block_name == kRangeDelBlockName) {
|
|
|
|
return BlockType::kRangeDeletion;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (meta_block_name == kHashIndexPrefixesBlock) {
|
|
|
|
return BlockType::kHashIndexPrefixes;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (meta_block_name == kHashIndexPrefixesMetadataBlock) {
|
|
|
|
return BlockType::kHashIndexMetadata;
|
|
|
|
}
|
|
|
|
|
Remove deprecated block-based filter (#10184)
Summary:
In https://github.com/facebook/rocksdb/issues/9535, release 7.0, we hid the old block-based filter from being created using
the public API, because of its inefficiency. Although we normally maintain read compatibility
on old DBs forever, filters are not required for reading a DB, only for optimizing read
performance. Thus, it should be acceptable to remove this code and the substantial
maintenance burden it carries as useful features are developed and validated (such
as user timestamp).
This change completely removes the code for reading and writing the old block-based
filters, net removing about 1370 lines of code no longer needed. Options removed from
testing / benchmarking tools. The prior existence is only evident in a couple of places:
* `CacheEntryRole::kDeprecatedFilterBlock` - We can update this public API enum in
a major release to minimize source code incompatibilities.
* A warning is logged when an old table file is opened that used the old block-based
filter. This is provided as a courtesy, and would be a pain to unit test, so manual testing
should suffice. Unfortunately, sst_dump does not tell you whether a file uses
block-based filter, and the structure of the code makes it very difficult to fix.
* To detect that case, `kObsoleteFilterBlockPrefix` (renamed from `kFilterBlockPrefix`)
for metaindex is maintained (for now).
Other notes:
* In some cases where numbers are associated with filter configurations, we have had to
update the assigned numbers so that they all correspond to something that exists.
* Fixed potential stat counting bug by assuming `filter_checked = false` for cases
like `filter == nullptr` rather than assuming `filter_checked = true`
* Removed obsolete `block_offset` and `prefix_extractor` parameters from several
functions.
* Removed some unnecessary checks `if (!table_prefix_extractor() && !prefix_extractor)`
because the caller guarantees the prefix extractor exists and is compatible
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10184
Test Plan:
tests updated, manually test new warning in LOG using base version to
generate a DB
Reviewed By: riversand963
Differential Revision: D37212647
Pulled By: pdillinger
fbshipit-source-id: 06ee020d8de3b81260ffc36ad0c1202cbf463a80
2 years ago
|
|
|
if (meta_block_name.starts_with(kObsoleteFilterBlockPrefix)) {
|
|
|
|
// Obsolete but possible in old files
|
|
|
|
return BlockType::kInvalid;
|
|
|
|
}
|
|
|
|
|
|
|
|
assert(false);
|
|
|
|
return BlockType::kInvalid;
|
|
|
|
}
|
|
|
|
|
|
|
|
Status BlockBasedTable::VerifyChecksumInMetaBlocks(
|
|
|
|
const ReadOptions& read_options, InternalIteratorBase<Slice>* index_iter) {
|
|
|
|
Status s;
|
|
|
|
for (index_iter->SeekToFirst(); index_iter->Valid(); index_iter->Next()) {
|
|
|
|
s = index_iter->status();
|
|
|
|
if (!s.ok()) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
BlockHandle handle;
|
|
|
|
Slice input = index_iter->value();
|
|
|
|
s = handle.DecodeFrom(&input);
|
|
|
|
BlockContents contents;
|
|
|
|
const Slice meta_block_name = index_iter->key();
|
|
|
|
if (meta_block_name == kPropertiesBlockName) {
|
Improve / clean up meta block code & integrity (#9163)
Summary:
* Checksums are now checked on meta blocks unless specifically
suppressed or not applicable (e.g. plain table). (Was other way around.)
This means a number of cases that were not checking checksums now are,
including direct read TableProperties in Version::GetTableProperties
(fixed in meta_blocks ReadTableProperties), reading any block from
PersistentCache (fixed in BlockFetcher), read TableProperties in
SstFileDumper (ldb/sst_dump/BackupEngine) before table reader open,
maybe more.
* For that to work, I moved the global_seqno+TableProperties checksum
logic to the shared table/ code, because that is used by many utilies
such as SstFileDumper.
* Also for that to work, we have to know when we're dealing with a block
that has a checksum (trailer), so added that capability to Footer based
on magic number, and from there BlockFetcher.
* Knowledge of trailer presence has also fixed a problem where other
table formats were reading blocks including bytes for a non-existant
trailer--and awkwardly kind-of not using them, e.g. no shared code
checking checksums. (BlockFetcher compression type was populated
incorrectly.) Now we only read what is needed.
* Minimized code duplication and differing/incompatible/awkward
abstractions in meta_blocks.{cc,h} (e.g. SeekTo in metaindex block
without parsing block handle)
* Moved some meta block handling code from table_properties*.*
* Moved some code specific to block-based table from shared table/ code
to BlockBasedTable class. The checksum stuff means we can't completely
separate it, but things that don't need to be in shared table/ code
should not be.
* Use unique_ptr rather than raw ptr in more places. (Note: you can
std::move from unique_ptr to shared_ptr.)
Without enhancements to GetPropertiesOfAllTablesTest (see below),
net reduction of roughly 100 lines of code.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9163
Test Plan:
existing tests and
* Enhanced DBTablePropertiesTest.GetPropertiesOfAllTablesTest to verify that
checksums are now checked on direct read of table properties by TableCache
(new test would fail before this change)
* Also enhanced DBTablePropertiesTest.GetPropertiesOfAllTablesTest to test
putting table properties under old meta name
* Also generally enhanced that same test to actually test what it was
supposed to be testing already, by kicking things out of table cache when
we don't want them there.
Reviewed By: ajkr, mrambacher
Differential Revision: D32514757
Pulled By: pdillinger
fbshipit-source-id: 507964b9311d186ae8d1131182290cbd97a99fa9
3 years ago
|
|
|
// Unfortunate special handling for properties block checksum w/
|
|
|
|
// global seqno
|
|
|
|
std::unique_ptr<TableProperties> table_properties;
|
|
|
|
s = ReadTablePropertiesHelper(read_options, handle, rep_->file.get(),
|
Improve / clean up meta block code & integrity (#9163)
Summary:
* Checksums are now checked on meta blocks unless specifically
suppressed or not applicable (e.g. plain table). (Was other way around.)
This means a number of cases that were not checking checksums now are,
including direct read TableProperties in Version::GetTableProperties
(fixed in meta_blocks ReadTableProperties), reading any block from
PersistentCache (fixed in BlockFetcher), read TableProperties in
SstFileDumper (ldb/sst_dump/BackupEngine) before table reader open,
maybe more.
* For that to work, I moved the global_seqno+TableProperties checksum
logic to the shared table/ code, because that is used by many utilies
such as SstFileDumper.
* Also for that to work, we have to know when we're dealing with a block
that has a checksum (trailer), so added that capability to Footer based
on magic number, and from there BlockFetcher.
* Knowledge of trailer presence has also fixed a problem where other
table formats were reading blocks including bytes for a non-existant
trailer--and awkwardly kind-of not using them, e.g. no shared code
checking checksums. (BlockFetcher compression type was populated
incorrectly.) Now we only read what is needed.
* Minimized code duplication and differing/incompatible/awkward
abstractions in meta_blocks.{cc,h} (e.g. SeekTo in metaindex block
without parsing block handle)
* Moved some meta block handling code from table_properties*.*
* Moved some code specific to block-based table from shared table/ code
to BlockBasedTable class. The checksum stuff means we can't completely
separate it, but things that don't need to be in shared table/ code
should not be.
* Use unique_ptr rather than raw ptr in more places. (Note: you can
std::move from unique_ptr to shared_ptr.)
Without enhancements to GetPropertiesOfAllTablesTest (see below),
net reduction of roughly 100 lines of code.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9163
Test Plan:
existing tests and
* Enhanced DBTablePropertiesTest.GetPropertiesOfAllTablesTest to verify that
checksums are now checked on direct read of table properties by TableCache
(new test would fail before this change)
* Also enhanced DBTablePropertiesTest.GetPropertiesOfAllTablesTest to test
putting table properties under old meta name
* Also generally enhanced that same test to actually test what it was
supposed to be testing already, by kicking things out of table cache when
we don't want them there.
Reviewed By: ajkr, mrambacher
Differential Revision: D32514757
Pulled By: pdillinger
fbshipit-source-id: 507964b9311d186ae8d1131182290cbd97a99fa9
3 years ago
|
|
|
nullptr /* prefetch_buffer */, rep_->footer,
|
|
|
|
rep_->ioptions, &table_properties,
|
|
|
|
nullptr /* memory_allocator */);
|
|
|
|
} else {
|
|
|
|
s = BlockFetcher(
|
|
|
|
rep_->file.get(), nullptr /* prefetch buffer */, rep_->footer,
|
|
|
|
read_options, handle, &contents, rep_->ioptions,
|
Improve / clean up meta block code & integrity (#9163)
Summary:
* Checksums are now checked on meta blocks unless specifically
suppressed or not applicable (e.g. plain table). (Was other way around.)
This means a number of cases that were not checking checksums now are,
including direct read TableProperties in Version::GetTableProperties
(fixed in meta_blocks ReadTableProperties), reading any block from
PersistentCache (fixed in BlockFetcher), read TableProperties in
SstFileDumper (ldb/sst_dump/BackupEngine) before table reader open,
maybe more.
* For that to work, I moved the global_seqno+TableProperties checksum
logic to the shared table/ code, because that is used by many utilies
such as SstFileDumper.
* Also for that to work, we have to know when we're dealing with a block
that has a checksum (trailer), so added that capability to Footer based
on magic number, and from there BlockFetcher.
* Knowledge of trailer presence has also fixed a problem where other
table formats were reading blocks including bytes for a non-existant
trailer--and awkwardly kind-of not using them, e.g. no shared code
checking checksums. (BlockFetcher compression type was populated
incorrectly.) Now we only read what is needed.
* Minimized code duplication and differing/incompatible/awkward
abstractions in meta_blocks.{cc,h} (e.g. SeekTo in metaindex block
without parsing block handle)
* Moved some meta block handling code from table_properties*.*
* Moved some code specific to block-based table from shared table/ code
to BlockBasedTable class. The checksum stuff means we can't completely
separate it, but things that don't need to be in shared table/ code
should not be.
* Use unique_ptr rather than raw ptr in more places. (Note: you can
std::move from unique_ptr to shared_ptr.)
Without enhancements to GetPropertiesOfAllTablesTest (see below),
net reduction of roughly 100 lines of code.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9163
Test Plan:
existing tests and
* Enhanced DBTablePropertiesTest.GetPropertiesOfAllTablesTest to verify that
checksums are now checked on direct read of table properties by TableCache
(new test would fail before this change)
* Also enhanced DBTablePropertiesTest.GetPropertiesOfAllTablesTest to test
putting table properties under old meta name
* Also generally enhanced that same test to actually test what it was
supposed to be testing already, by kicking things out of table cache when
we don't want them there.
Reviewed By: ajkr, mrambacher
Differential Revision: D32514757
Pulled By: pdillinger
fbshipit-source-id: 507964b9311d186ae8d1131182290cbd97a99fa9
3 years ago
|
|
|
false /* decompress */, false /*maybe_compressed*/,
|
|
|
|
GetBlockTypeForMetaBlockByName(meta_block_name),
|
|
|
|
UncompressionDict::GetEmptyDict(), rep_->persistent_cache_options)
|
|
|
|
.ReadBlockContents();
|
|
|
|
}
|
|
|
|
if (!s.ok()) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
bool BlockBasedTable::TEST_BlockInCache(const BlockHandle& handle) const {
|
|
|
|
assert(rep_ != nullptr);
|
|
|
|
|
|
|
|
Cache* const cache = rep_->table_options.block_cache.get();
|
|
|
|
if (cache == nullptr) {
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
New stable, fixed-length cache keys (#9126)
Summary:
This change standardizes on a new 16-byte cache key format for
block cache (incl compressed and secondary) and persistent cache (but
not table cache and row cache).
The goal is a really fast cache key with practically ideal stability and
uniqueness properties without external dependencies (e.g. from FileSystem).
A fixed key size of 16 bytes should enable future optimizations to the
concurrent hash table for block cache, which is a heavy CPU user /
bottleneck, but there appears to be measurable performance improvement
even with no changes to LRUCache.
This change replaces a lot of disjointed and ugly code handling cache
keys with calls to a simple, clean new internal API (cache_key.h).
(Preserving the old cache key logic under an option would be very ugly
and likely negate the performance gain of the new approach. Complete
replacement carries some inherent risk, but I think that's acceptable
with sufficient analysis and testing.)
The scheme for encoding new cache keys is complicated but explained
in cache_key.cc.
Also: EndianSwapValue is moved to math.h to be next to other bit
operations. (Explains some new include "math.h".) ReverseBits operation
added and unit tests added to hash_test for both.
Fixes https://github.com/facebook/rocksdb/issues/7405 (presuming a root cause)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9126
Test Plan:
### Basic correctness
Several tests needed updates to work with the new functionality, mostly
because we are no longer relying on filesystem for stable cache keys
so table builders & readers need more context info to agree on cache
keys. This functionality is so core, a huge number of existing tests
exercise the cache key functionality.
### Performance
Create db with
`TEST_TMPDIR=/dev/shm ./db_bench -bloom_bits=10 -benchmarks=fillrandom -num=3000000 -partition_index_and_filters`
And test performance with
`TEST_TMPDIR=/dev/shm ./db_bench -readonly -use_existing_db -bloom_bits=10 -benchmarks=readrandom -num=3000000 -duration=30 -cache_index_and_filter_blocks -cache_size=250000 -threads=4`
using DEBUG_LEVEL=0 and simultaneous before & after runs.
Before ops/sec, avg over 100 runs: 121924
After ops/sec, avg over 100 runs: 125385 (+2.8%)
### Collision probability
I have built a tool, ./cache_bench -stress_cache_key to broadly simulate host-wide cache activity
over many months, by making some pessimistic simplifying assumptions:
* Every generated file has a cache entry for every byte offset in the file (contiguous range of cache keys)
* All of every file is cached for its entire lifetime
We use a simple table with skewed address assignment and replacement on address collision
to simulate files coming & going, with quite a variance (super-Poisson) in ages. Some output
with `./cache_bench -stress_cache_key -sck_keep_bits=40`:
```
Total cache or DBs size: 32TiB Writing 925.926 MiB/s or 76.2939TiB/day
Multiply by 9.22337e+18 to correct for simulation losses (but still assume whole file cached)
```
These come from default settings of 2.5M files per day of 32 MB each, and
`-sck_keep_bits=40` means that to represent a single file, we are only keeping 40 bits of
the 128-bit cache key. With file size of 2\*\*25 contiguous keys (pessimistic), our simulation
is about 2\*\*(128-40-25) or about 9 billion billion times more prone to collision than reality.
More default assumptions, relatively pessimistic:
* 100 DBs in same process (doesn't matter much)
* Re-open DB in same process (new session ID related to old session ID) on average
every 100 files generated
* Restart process (all new session IDs unrelated to old) 24 times per day
After enough data, we get a result at the end:
```
(keep 40 bits) 17 collisions after 2 x 90 days, est 10.5882 days between (9.76592e+19 corrected)
```
If we believe the (pessimistic) simulation and the mathematical generalization, we would need to run a billion machines all for 97 billion days to expect a cache key collision. To help verify that our generalization ("corrected") is robust, we can make our simulation more precise with `-sck_keep_bits=41` and `42`, which takes more running time to get enough data:
```
(keep 41 bits) 16 collisions after 4 x 90 days, est 22.5 days between (1.03763e+20 corrected)
(keep 42 bits) 19 collisions after 10 x 90 days, est 47.3684 days between (1.09224e+20 corrected)
```
The generalized prediction still holds. With the `-sck_randomize` option, we can see that we are beating "random" cache keys (except offsets still non-randomized) by a modest amount (roughly 20x less collision prone than random), which should make us reasonably comfortable even in "degenerate" cases:
```
197 collisions after 1 x 90 days, est 0.456853 days between (4.21372e+18 corrected)
```
I've run other tests to validate other conditions behave as expected, never behaving "worse than random" unless we start chopping off structured data.
Reviewed By: zhichao-cao
Differential Revision: D33171746
Pulled By: pdillinger
fbshipit-source-id: f16a57e369ed37be5e7e33525ace848d0537c88f
3 years ago
|
|
|
CacheKey key = GetCacheKey(rep_->base_cache_key, handle);
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
|
New stable, fixed-length cache keys (#9126)
Summary:
This change standardizes on a new 16-byte cache key format for
block cache (incl compressed and secondary) and persistent cache (but
not table cache and row cache).
The goal is a really fast cache key with practically ideal stability and
uniqueness properties without external dependencies (e.g. from FileSystem).
A fixed key size of 16 bytes should enable future optimizations to the
concurrent hash table for block cache, which is a heavy CPU user /
bottleneck, but there appears to be measurable performance improvement
even with no changes to LRUCache.
This change replaces a lot of disjointed and ugly code handling cache
keys with calls to a simple, clean new internal API (cache_key.h).
(Preserving the old cache key logic under an option would be very ugly
and likely negate the performance gain of the new approach. Complete
replacement carries some inherent risk, but I think that's acceptable
with sufficient analysis and testing.)
The scheme for encoding new cache keys is complicated but explained
in cache_key.cc.
Also: EndianSwapValue is moved to math.h to be next to other bit
operations. (Explains some new include "math.h".) ReverseBits operation
added and unit tests added to hash_test for both.
Fixes https://github.com/facebook/rocksdb/issues/7405 (presuming a root cause)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9126
Test Plan:
### Basic correctness
Several tests needed updates to work with the new functionality, mostly
because we are no longer relying on filesystem for stable cache keys
so table builders & readers need more context info to agree on cache
keys. This functionality is so core, a huge number of existing tests
exercise the cache key functionality.
### Performance
Create db with
`TEST_TMPDIR=/dev/shm ./db_bench -bloom_bits=10 -benchmarks=fillrandom -num=3000000 -partition_index_and_filters`
And test performance with
`TEST_TMPDIR=/dev/shm ./db_bench -readonly -use_existing_db -bloom_bits=10 -benchmarks=readrandom -num=3000000 -duration=30 -cache_index_and_filter_blocks -cache_size=250000 -threads=4`
using DEBUG_LEVEL=0 and simultaneous before & after runs.
Before ops/sec, avg over 100 runs: 121924
After ops/sec, avg over 100 runs: 125385 (+2.8%)
### Collision probability
I have built a tool, ./cache_bench -stress_cache_key to broadly simulate host-wide cache activity
over many months, by making some pessimistic simplifying assumptions:
* Every generated file has a cache entry for every byte offset in the file (contiguous range of cache keys)
* All of every file is cached for its entire lifetime
We use a simple table with skewed address assignment and replacement on address collision
to simulate files coming & going, with quite a variance (super-Poisson) in ages. Some output
with `./cache_bench -stress_cache_key -sck_keep_bits=40`:
```
Total cache or DBs size: 32TiB Writing 925.926 MiB/s or 76.2939TiB/day
Multiply by 9.22337e+18 to correct for simulation losses (but still assume whole file cached)
```
These come from default settings of 2.5M files per day of 32 MB each, and
`-sck_keep_bits=40` means that to represent a single file, we are only keeping 40 bits of
the 128-bit cache key. With file size of 2\*\*25 contiguous keys (pessimistic), our simulation
is about 2\*\*(128-40-25) or about 9 billion billion times more prone to collision than reality.
More default assumptions, relatively pessimistic:
* 100 DBs in same process (doesn't matter much)
* Re-open DB in same process (new session ID related to old session ID) on average
every 100 files generated
* Restart process (all new session IDs unrelated to old) 24 times per day
After enough data, we get a result at the end:
```
(keep 40 bits) 17 collisions after 2 x 90 days, est 10.5882 days between (9.76592e+19 corrected)
```
If we believe the (pessimistic) simulation and the mathematical generalization, we would need to run a billion machines all for 97 billion days to expect a cache key collision. To help verify that our generalization ("corrected") is robust, we can make our simulation more precise with `-sck_keep_bits=41` and `42`, which takes more running time to get enough data:
```
(keep 41 bits) 16 collisions after 4 x 90 days, est 22.5 days between (1.03763e+20 corrected)
(keep 42 bits) 19 collisions after 10 x 90 days, est 47.3684 days between (1.09224e+20 corrected)
```
The generalized prediction still holds. With the `-sck_randomize` option, we can see that we are beating "random" cache keys (except offsets still non-randomized) by a modest amount (roughly 20x less collision prone than random), which should make us reasonably comfortable even in "degenerate" cases:
```
197 collisions after 1 x 90 days, est 0.456853 days between (4.21372e+18 corrected)
```
I've run other tests to validate other conditions behave as expected, never behaving "worse than random" unless we start chopping off structured data.
Reviewed By: zhichao-cao
Differential Revision: D33171746
Pulled By: pdillinger
fbshipit-source-id: f16a57e369ed37be5e7e33525ace848d0537c88f
3 years ago
|
|
|
Cache::Handle* const cache_handle = cache->Lookup(key.AsSlice());
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
if (cache_handle == nullptr) {
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
cache->Release(cache_handle);
|
|
|
|
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
bool BlockBasedTable::TEST_KeyInCache(const ReadOptions& options,
|
|
|
|
const Slice& key) {
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
std::unique_ptr<InternalIteratorBase<IndexValue>> iiter(NewIndexIterator(
|
|
|
|
options, /*need_upper_bound_check=*/false, /*input_iter=*/nullptr,
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
/*get_context=*/nullptr, /*lookup_context=*/nullptr));
|
|
|
|
iiter->Seek(key);
|
|
|
|
assert(iiter->Valid());
|
|
|
|
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
return TEST_BlockInCache(iiter->value().handle);
|
|
|
|
}
|
|
|
|
|
|
|
|
// REQUIRES: The following fields of rep_ should have already been populated:
|
|
|
|
// 1. file
|
|
|
|
// 2. index_handle,
|
|
|
|
// 3. options
|
|
|
|
// 4. internal_comparator
|
|
|
|
// 5. index_type
|
|
|
|
Status BlockBasedTable::CreateIndexReader(
|
|
|
|
const ReadOptions& ro, FilePrefetchBuffer* prefetch_buffer,
|
|
|
|
InternalIterator* meta_iter, bool use_cache, bool prefetch, bool pin,
|
|
|
|
BlockCacheLookupContext* lookup_context,
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
std::unique_ptr<IndexReader>* index_reader) {
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
switch (rep_->index_type) {
|
|
|
|
case BlockBasedTableOptions::kTwoLevelIndexSearch: {
|
|
|
|
return PartitionIndexReader::Create(this, ro, prefetch_buffer, use_cache,
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
prefetch, pin, lookup_context,
|
|
|
|
index_reader);
|
|
|
|
}
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
case BlockBasedTableOptions::kBinarySearch:
|
|
|
|
FALLTHROUGH_INTENDED;
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
case BlockBasedTableOptions::kBinarySearchWithFirstKey: {
|
|
|
|
return BinarySearchIndexReader::Create(this, ro, prefetch_buffer,
|
|
|
|
use_cache, prefetch, pin,
|
|
|
|
lookup_context, index_reader);
|
|
|
|
}
|
|
|
|
case BlockBasedTableOptions::kHashSearch: {
|
|
|
|
if (!rep_->table_prefix_extractor) {
|
|
|
|
ROCKS_LOG_WARN(rep_->ioptions.logger,
|
|
|
|
"Missing prefix extractor for hash index. Fall back to"
|
|
|
|
" binary search index.");
|
|
|
|
return BinarySearchIndexReader::Create(this, ro, prefetch_buffer,
|
|
|
|
use_cache, prefetch, pin,
|
|
|
|
lookup_context, index_reader);
|
|
|
|
} else {
|
|
|
|
return HashIndexReader::Create(this, ro, prefetch_buffer, meta_iter,
|
|
|
|
use_cache, prefetch, pin, lookup_context,
|
|
|
|
index_reader);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
default: {
|
|
|
|
std::string error_message =
|
|
|
|
"Unrecognized index type: " + std::to_string(rep_->index_type);
|
|
|
|
return Status::InvalidArgument(error_message.c_str());
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
For ApproximateSizes, pro-rate table metadata size over data blocks (#6784)
Summary:
The implementation of GetApproximateSizes was inconsistent in
its treatment of the size of non-data blocks of SST files, sometimes
including and sometimes now. This was at its worst with large portion
of table file used by filters and querying a small range that crossed
a table boundary: the size estimate would include large filter size.
It's conceivable that someone might want only to know the size in terms
of data blocks, but I believe that's unlikely enough to ignore for now.
Similarly, there's no evidence the internal function AppoximateOffsetOf
is used for anything other than a one-sided ApproximateSize, so I intend
to refactor to remove redundancy in a follow-up commit.
So to fix this, GetApproximateSizes (and implementation details
ApproximateSize and ApproximateOffsetOf) now consistently include in
their returned sizes a portion of table file metadata (incl filters
and indexes) based on the size portion of the data blocks in range. In
other words, if a key range covers data blocks that are X% by size of all
the table's data blocks, returned approximate size is X% of the total
file size. It would technically be more accurate to attribute metadata
based on number of keys, but that's not computationally efficient with
data available and rarely a meaningful difference.
Also includes miscellaneous comment improvements / clarifications.
Also included is a new approximatesizerandom benchmark for db_bench.
No significant performance difference seen with this change, whether ~700 ops/sec with cache_index_and_filter_blocks and small cache or ~150k ops/sec without cache_index_and_filter_blocks.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6784
Test Plan:
Test added to DBTest.ApproximateSizesFilesWithErrorMargin.
Old code running new test...
[ RUN ] DBTest.ApproximateSizesFilesWithErrorMargin
db/db_test.cc:1562: Failure
Expected: (size) <= (11 * 100), actual: 9478 vs 1100
Other tests updated to reflect consistent accounting of metadata.
Reviewed By: siying
Differential Revision: D21334706
Pulled By: pdillinger
fbshipit-source-id: 6f86870e45213334fedbe9c73b4ebb1d8d611185
5 years ago
|
|
|
uint64_t BlockBasedTable::ApproximateDataOffsetOf(
|
|
|
|
const InternalIteratorBase<IndexValue>& index_iter,
|
|
|
|
uint64_t data_size) const {
|
Handle failures in block-based table size/offset approximation (#9615)
Summary:
In crash test with fault injection, we were seeing stack traces like the following:
```
https://github.com/facebook/rocksdb/issues/3 0x00007f75f763c533 in __GI___assert_fail (assertion=assertion@entry=0x1c5b2a0 "end_offset >= start_offset", file=file@entry=0x1c580a0 "table/block_based/block_based_table_reader.cc", line=line@entry=3245,
function=function@entry=0x1c60e60 "virtual uint64_t rocksdb::BlockBasedTable::ApproximateSize(const rocksdb::Slice&, const rocksdb::Slice&, rocksdb::TableReaderCaller)") at assert.c:101
https://github.com/facebook/rocksdb/issues/4 0x00000000010ea9b4 in rocksdb::BlockBasedTable::ApproximateSize (this=<optimized out>, start=..., end=..., caller=<optimized out>) at table/block_based/block_based_table_reader.cc:3224
https://github.com/facebook/rocksdb/issues/5 0x0000000000be61fb in rocksdb::TableCache::ApproximateSize (this=0x60f0000161b0, start=..., end=..., fd=..., caller=caller@entry=rocksdb::kCompaction, internal_comparator=..., prefix_extractor=...) at db/table_cache.cc:719
https://github.com/facebook/rocksdb/issues/6 0x0000000000c3eaec in rocksdb::VersionSet::ApproximateSize (this=<optimized out>, v=<optimized out>, f=..., start=..., end=..., caller=<optimized out>) at ./db/version_set.h:850
https://github.com/facebook/rocksdb/issues/7 0x0000000000c6ebc3 in rocksdb::VersionSet::ApproximateSize (this=<optimized out>, options=..., v=v@entry=0x621000047500, start=..., end=..., start_level=start_level@entry=0, end_level=<optimized out>, caller=<optimized out>)
at db/version_set.cc:5657
https://github.com/facebook/rocksdb/issues/8 0x000000000166e894 in rocksdb::CompactionJob::GenSubcompactionBoundaries (this=<optimized out>) at ./include/rocksdb/options.h:1869
https://github.com/facebook/rocksdb/issues/9 0x000000000168c526 in rocksdb::CompactionJob::Prepare (this=this@entry=0x7f75f3ffcf00) at db/compaction/compaction_job.cc:546
```
The problem occurred in `ApproximateSize()` when the index `Seek()` for the first `ApproximateDataOffsetOf()` encountered an I/O error, while the second `Seek()` did not. In the old code that scenario caused `start_offset == data_size` , thus it was easy to trip the assertion that `end_offset >= start_offset`.
The fix is to set `start_offset == 0` when the first index `Seek()` fails, and `end_offset == data_size` when the second index `Seek()` fails. I doubt these give an "on average correct" answer for how this function is used, but I/O errors in index seeks are hopefully rare, it looked consistent with what was already there, and it was easier to calculate.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9615
Test Plan:
run the repro command for a while and stopped seeing coredumps -
```
$ while ! ./db_stress --block_size=128 --cache_size=32768 --clear_column_family_one_in=0 --column_families=1 --continuous_verification_interval=0 --db=/dev/shm/rocksdb_crashtest --delpercent=4 --delrangepercent=1 --destroy_db_initially=0 --expected_values_dir=/dev/shm/rocksdb_crashtest_expected --index_type=2 --iterpercent=10 --kill_random_test=18887 --max_key=1000000 --max_bytes_for_level_base=2048576 --nooverwritepercent=1 --open_files=-1 --open_read_fault_one_in=32 --ops_per_thread=1000000 --prefixpercent=5 --read_fault_one_in=0 --readpercent=45 --reopen=0 --skip_verifydb=1 --subcompactions=2 --target_file_size_base=524288 --test_batches_snapshots=0 --value_size_mult=32 --write_buffer_size=524288 --writepercent=35 ; do : ; done
```
Reviewed By: pdillinger
Differential Revision: D34383069
Pulled By: ajkr
fbshipit-source-id: fac26c3b20ea962e75387515ba5f2724dc48719f
3 years ago
|
|
|
assert(index_iter.status().ok());
|
|
|
|
if (index_iter.Valid()) {
|
|
|
|
BlockHandle handle = index_iter.value().handle;
|
For ApproximateSizes, pro-rate table metadata size over data blocks (#6784)
Summary:
The implementation of GetApproximateSizes was inconsistent in
its treatment of the size of non-data blocks of SST files, sometimes
including and sometimes now. This was at its worst with large portion
of table file used by filters and querying a small range that crossed
a table boundary: the size estimate would include large filter size.
It's conceivable that someone might want only to know the size in terms
of data blocks, but I believe that's unlikely enough to ignore for now.
Similarly, there's no evidence the internal function AppoximateOffsetOf
is used for anything other than a one-sided ApproximateSize, so I intend
to refactor to remove redundancy in a follow-up commit.
So to fix this, GetApproximateSizes (and implementation details
ApproximateSize and ApproximateOffsetOf) now consistently include in
their returned sizes a portion of table file metadata (incl filters
and indexes) based on the size portion of the data blocks in range. In
other words, if a key range covers data blocks that are X% by size of all
the table's data blocks, returned approximate size is X% of the total
file size. It would technically be more accurate to attribute metadata
based on number of keys, but that's not computationally efficient with
data available and rarely a meaningful difference.
Also includes miscellaneous comment improvements / clarifications.
Also included is a new approximatesizerandom benchmark for db_bench.
No significant performance difference seen with this change, whether ~700 ops/sec with cache_index_and_filter_blocks and small cache or ~150k ops/sec without cache_index_and_filter_blocks.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6784
Test Plan:
Test added to DBTest.ApproximateSizesFilesWithErrorMargin.
Old code running new test...
[ RUN ] DBTest.ApproximateSizesFilesWithErrorMargin
db/db_test.cc:1562: Failure
Expected: (size) <= (11 * 100), actual: 9478 vs 1100
Other tests updated to reflect consistent accounting of metadata.
Reviewed By: siying
Differential Revision: D21334706
Pulled By: pdillinger
fbshipit-source-id: 6f86870e45213334fedbe9c73b4ebb1d8d611185
5 years ago
|
|
|
return handle.offset();
|
|
|
|
} else {
|
For ApproximateSizes, pro-rate table metadata size over data blocks (#6784)
Summary:
The implementation of GetApproximateSizes was inconsistent in
its treatment of the size of non-data blocks of SST files, sometimes
including and sometimes now. This was at its worst with large portion
of table file used by filters and querying a small range that crossed
a table boundary: the size estimate would include large filter size.
It's conceivable that someone might want only to know the size in terms
of data blocks, but I believe that's unlikely enough to ignore for now.
Similarly, there's no evidence the internal function AppoximateOffsetOf
is used for anything other than a one-sided ApproximateSize, so I intend
to refactor to remove redundancy in a follow-up commit.
So to fix this, GetApproximateSizes (and implementation details
ApproximateSize and ApproximateOffsetOf) now consistently include in
their returned sizes a portion of table file metadata (incl filters
and indexes) based on the size portion of the data blocks in range. In
other words, if a key range covers data blocks that are X% by size of all
the table's data blocks, returned approximate size is X% of the total
file size. It would technically be more accurate to attribute metadata
based on number of keys, but that's not computationally efficient with
data available and rarely a meaningful difference.
Also includes miscellaneous comment improvements / clarifications.
Also included is a new approximatesizerandom benchmark for db_bench.
No significant performance difference seen with this change, whether ~700 ops/sec with cache_index_and_filter_blocks and small cache or ~150k ops/sec without cache_index_and_filter_blocks.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6784
Test Plan:
Test added to DBTest.ApproximateSizesFilesWithErrorMargin.
Old code running new test...
[ RUN ] DBTest.ApproximateSizesFilesWithErrorMargin
db/db_test.cc:1562: Failure
Expected: (size) <= (11 * 100), actual: 9478 vs 1100
Other tests updated to reflect consistent accounting of metadata.
Reviewed By: siying
Differential Revision: D21334706
Pulled By: pdillinger
fbshipit-source-id: 6f86870e45213334fedbe9c73b4ebb1d8d611185
5 years ago
|
|
|
// The iterator is past the last key in the file.
|
|
|
|
return data_size;
|
|
|
|
}
|
For ApproximateSizes, pro-rate table metadata size over data blocks (#6784)
Summary:
The implementation of GetApproximateSizes was inconsistent in
its treatment of the size of non-data blocks of SST files, sometimes
including and sometimes now. This was at its worst with large portion
of table file used by filters and querying a small range that crossed
a table boundary: the size estimate would include large filter size.
It's conceivable that someone might want only to know the size in terms
of data blocks, but I believe that's unlikely enough to ignore for now.
Similarly, there's no evidence the internal function AppoximateOffsetOf
is used for anything other than a one-sided ApproximateSize, so I intend
to refactor to remove redundancy in a follow-up commit.
So to fix this, GetApproximateSizes (and implementation details
ApproximateSize and ApproximateOffsetOf) now consistently include in
their returned sizes a portion of table file metadata (incl filters
and indexes) based on the size portion of the data blocks in range. In
other words, if a key range covers data blocks that are X% by size of all
the table's data blocks, returned approximate size is X% of the total
file size. It would technically be more accurate to attribute metadata
based on number of keys, but that's not computationally efficient with
data available and rarely a meaningful difference.
Also includes miscellaneous comment improvements / clarifications.
Also included is a new approximatesizerandom benchmark for db_bench.
No significant performance difference seen with this change, whether ~700 ops/sec with cache_index_and_filter_blocks and small cache or ~150k ops/sec without cache_index_and_filter_blocks.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6784
Test Plan:
Test added to DBTest.ApproximateSizesFilesWithErrorMargin.
Old code running new test...
[ RUN ] DBTest.ApproximateSizesFilesWithErrorMargin
db/db_test.cc:1562: Failure
Expected: (size) <= (11 * 100), actual: 9478 vs 1100
Other tests updated to reflect consistent accounting of metadata.
Reviewed By: siying
Differential Revision: D21334706
Pulled By: pdillinger
fbshipit-source-id: 6f86870e45213334fedbe9c73b4ebb1d8d611185
5 years ago
|
|
|
}
|
|
|
|
|
For ApproximateSizes, pro-rate table metadata size over data blocks (#6784)
Summary:
The implementation of GetApproximateSizes was inconsistent in
its treatment of the size of non-data blocks of SST files, sometimes
including and sometimes now. This was at its worst with large portion
of table file used by filters and querying a small range that crossed
a table boundary: the size estimate would include large filter size.
It's conceivable that someone might want only to know the size in terms
of data blocks, but I believe that's unlikely enough to ignore for now.
Similarly, there's no evidence the internal function AppoximateOffsetOf
is used for anything other than a one-sided ApproximateSize, so I intend
to refactor to remove redundancy in a follow-up commit.
So to fix this, GetApproximateSizes (and implementation details
ApproximateSize and ApproximateOffsetOf) now consistently include in
their returned sizes a portion of table file metadata (incl filters
and indexes) based on the size portion of the data blocks in range. In
other words, if a key range covers data blocks that are X% by size of all
the table's data blocks, returned approximate size is X% of the total
file size. It would technically be more accurate to attribute metadata
based on number of keys, but that's not computationally efficient with
data available and rarely a meaningful difference.
Also includes miscellaneous comment improvements / clarifications.
Also included is a new approximatesizerandom benchmark for db_bench.
No significant performance difference seen with this change, whether ~700 ops/sec with cache_index_and_filter_blocks and small cache or ~150k ops/sec without cache_index_and_filter_blocks.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6784
Test Plan:
Test added to DBTest.ApproximateSizesFilesWithErrorMargin.
Old code running new test...
[ RUN ] DBTest.ApproximateSizesFilesWithErrorMargin
db/db_test.cc:1562: Failure
Expected: (size) <= (11 * 100), actual: 9478 vs 1100
Other tests updated to reflect consistent accounting of metadata.
Reviewed By: siying
Differential Revision: D21334706
Pulled By: pdillinger
fbshipit-source-id: 6f86870e45213334fedbe9c73b4ebb1d8d611185
5 years ago
|
|
|
uint64_t BlockBasedTable::GetApproximateDataSize() {
|
|
|
|
// Should be in table properties unless super old version
|
|
|
|
if (rep_->table_properties) {
|
|
|
|
return rep_->table_properties->data_size;
|
|
|
|
}
|
|
|
|
// Fall back to rough estimate from footer
|
|
|
|
return rep_->footer.metaindex_handle().offset();
|
|
|
|
}
|
|
|
|
|
|
|
|
uint64_t BlockBasedTable::ApproximateOffsetOf(const ReadOptions& read_options,
|
|
|
|
const Slice& key,
|
|
|
|
TableReaderCaller caller) {
|
For ApproximateSizes, pro-rate table metadata size over data blocks (#6784)
Summary:
The implementation of GetApproximateSizes was inconsistent in
its treatment of the size of non-data blocks of SST files, sometimes
including and sometimes now. This was at its worst with large portion
of table file used by filters and querying a small range that crossed
a table boundary: the size estimate would include large filter size.
It's conceivable that someone might want only to know the size in terms
of data blocks, but I believe that's unlikely enough to ignore for now.
Similarly, there's no evidence the internal function AppoximateOffsetOf
is used for anything other than a one-sided ApproximateSize, so I intend
to refactor to remove redundancy in a follow-up commit.
So to fix this, GetApproximateSizes (and implementation details
ApproximateSize and ApproximateOffsetOf) now consistently include in
their returned sizes a portion of table file metadata (incl filters
and indexes) based on the size portion of the data blocks in range. In
other words, if a key range covers data blocks that are X% by size of all
the table's data blocks, returned approximate size is X% of the total
file size. It would technically be more accurate to attribute metadata
based on number of keys, but that's not computationally efficient with
data available and rarely a meaningful difference.
Also includes miscellaneous comment improvements / clarifications.
Also included is a new approximatesizerandom benchmark for db_bench.
No significant performance difference seen with this change, whether ~700 ops/sec with cache_index_and_filter_blocks and small cache or ~150k ops/sec without cache_index_and_filter_blocks.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6784
Test Plan:
Test added to DBTest.ApproximateSizesFilesWithErrorMargin.
Old code running new test...
[ RUN ] DBTest.ApproximateSizesFilesWithErrorMargin
db/db_test.cc:1562: Failure
Expected: (size) <= (11 * 100), actual: 9478 vs 1100
Other tests updated to reflect consistent accounting of metadata.
Reviewed By: siying
Differential Revision: D21334706
Pulled By: pdillinger
fbshipit-source-id: 6f86870e45213334fedbe9c73b4ebb1d8d611185
5 years ago
|
|
|
uint64_t data_size = GetApproximateDataSize();
|
|
|
|
if (UNLIKELY(data_size == 0)) {
|
|
|
|
// Hmm. Let's just split in half to avoid skewing one way or another,
|
|
|
|
// since we don't know whether we're operating on lower bound or
|
|
|
|
// upper bound.
|
|
|
|
return rep_->file_size / 2;
|
|
|
|
}
|
|
|
|
|
|
|
|
BlockCacheLookupContext context(caller);
|
|
|
|
IndexBlockIter iiter_on_stack;
|
|
|
|
ReadOptions ro;
|
|
|
|
ro.total_order_seek = true;
|
|
|
|
ro.io_activity = read_options.io_activity;
|
|
|
|
auto index_iter =
|
|
|
|
NewIndexIterator(ro, /*disable_prefix_seek=*/true,
|
|
|
|
/*input_iter=*/&iiter_on_stack, /*get_context=*/nullptr,
|
|
|
|
/*lookup_context=*/&context);
|
|
|
|
std::unique_ptr<InternalIteratorBase<IndexValue>> iiter_unique_ptr;
|
|
|
|
if (index_iter != &iiter_on_stack) {
|
|
|
|
iiter_unique_ptr.reset(index_iter);
|
|
|
|
}
|
|
|
|
|
|
|
|
index_iter->Seek(key);
|
Handle failures in block-based table size/offset approximation (#9615)
Summary:
In crash test with fault injection, we were seeing stack traces like the following:
```
https://github.com/facebook/rocksdb/issues/3 0x00007f75f763c533 in __GI___assert_fail (assertion=assertion@entry=0x1c5b2a0 "end_offset >= start_offset", file=file@entry=0x1c580a0 "table/block_based/block_based_table_reader.cc", line=line@entry=3245,
function=function@entry=0x1c60e60 "virtual uint64_t rocksdb::BlockBasedTable::ApproximateSize(const rocksdb::Slice&, const rocksdb::Slice&, rocksdb::TableReaderCaller)") at assert.c:101
https://github.com/facebook/rocksdb/issues/4 0x00000000010ea9b4 in rocksdb::BlockBasedTable::ApproximateSize (this=<optimized out>, start=..., end=..., caller=<optimized out>) at table/block_based/block_based_table_reader.cc:3224
https://github.com/facebook/rocksdb/issues/5 0x0000000000be61fb in rocksdb::TableCache::ApproximateSize (this=0x60f0000161b0, start=..., end=..., fd=..., caller=caller@entry=rocksdb::kCompaction, internal_comparator=..., prefix_extractor=...) at db/table_cache.cc:719
https://github.com/facebook/rocksdb/issues/6 0x0000000000c3eaec in rocksdb::VersionSet::ApproximateSize (this=<optimized out>, v=<optimized out>, f=..., start=..., end=..., caller=<optimized out>) at ./db/version_set.h:850
https://github.com/facebook/rocksdb/issues/7 0x0000000000c6ebc3 in rocksdb::VersionSet::ApproximateSize (this=<optimized out>, options=..., v=v@entry=0x621000047500, start=..., end=..., start_level=start_level@entry=0, end_level=<optimized out>, caller=<optimized out>)
at db/version_set.cc:5657
https://github.com/facebook/rocksdb/issues/8 0x000000000166e894 in rocksdb::CompactionJob::GenSubcompactionBoundaries (this=<optimized out>) at ./include/rocksdb/options.h:1869
https://github.com/facebook/rocksdb/issues/9 0x000000000168c526 in rocksdb::CompactionJob::Prepare (this=this@entry=0x7f75f3ffcf00) at db/compaction/compaction_job.cc:546
```
The problem occurred in `ApproximateSize()` when the index `Seek()` for the first `ApproximateDataOffsetOf()` encountered an I/O error, while the second `Seek()` did not. In the old code that scenario caused `start_offset == data_size` , thus it was easy to trip the assertion that `end_offset >= start_offset`.
The fix is to set `start_offset == 0` when the first index `Seek()` fails, and `end_offset == data_size` when the second index `Seek()` fails. I doubt these give an "on average correct" answer for how this function is used, but I/O errors in index seeks are hopefully rare, it looked consistent with what was already there, and it was easier to calculate.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9615
Test Plan:
run the repro command for a while and stopped seeing coredumps -
```
$ while ! ./db_stress --block_size=128 --cache_size=32768 --clear_column_family_one_in=0 --column_families=1 --continuous_verification_interval=0 --db=/dev/shm/rocksdb_crashtest --delpercent=4 --delrangepercent=1 --destroy_db_initially=0 --expected_values_dir=/dev/shm/rocksdb_crashtest_expected --index_type=2 --iterpercent=10 --kill_random_test=18887 --max_key=1000000 --max_bytes_for_level_base=2048576 --nooverwritepercent=1 --open_files=-1 --open_read_fault_one_in=32 --ops_per_thread=1000000 --prefixpercent=5 --read_fault_one_in=0 --readpercent=45 --reopen=0 --skip_verifydb=1 --subcompactions=2 --target_file_size_base=524288 --test_batches_snapshots=0 --value_size_mult=32 --write_buffer_size=524288 --writepercent=35 ; do : ; done
```
Reviewed By: pdillinger
Differential Revision: D34383069
Pulled By: ajkr
fbshipit-source-id: fac26c3b20ea962e75387515ba5f2724dc48719f
3 years ago
|
|
|
uint64_t offset;
|
|
|
|
if (index_iter->status().ok()) {
|
|
|
|
offset = ApproximateDataOffsetOf(*index_iter, data_size);
|
|
|
|
} else {
|
|
|
|
// Split in half to avoid skewing one way or another,
|
|
|
|
// since we don't know whether we're operating on lower bound or
|
|
|
|
// upper bound.
|
|
|
|
return rep_->file_size / 2;
|
|
|
|
}
|
For ApproximateSizes, pro-rate table metadata size over data blocks (#6784)
Summary:
The implementation of GetApproximateSizes was inconsistent in
its treatment of the size of non-data blocks of SST files, sometimes
including and sometimes now. This was at its worst with large portion
of table file used by filters and querying a small range that crossed
a table boundary: the size estimate would include large filter size.
It's conceivable that someone might want only to know the size in terms
of data blocks, but I believe that's unlikely enough to ignore for now.
Similarly, there's no evidence the internal function AppoximateOffsetOf
is used for anything other than a one-sided ApproximateSize, so I intend
to refactor to remove redundancy in a follow-up commit.
So to fix this, GetApproximateSizes (and implementation details
ApproximateSize and ApproximateOffsetOf) now consistently include in
their returned sizes a portion of table file metadata (incl filters
and indexes) based on the size portion of the data blocks in range. In
other words, if a key range covers data blocks that are X% by size of all
the table's data blocks, returned approximate size is X% of the total
file size. It would technically be more accurate to attribute metadata
based on number of keys, but that's not computationally efficient with
data available and rarely a meaningful difference.
Also includes miscellaneous comment improvements / clarifications.
Also included is a new approximatesizerandom benchmark for db_bench.
No significant performance difference seen with this change, whether ~700 ops/sec with cache_index_and_filter_blocks and small cache or ~150k ops/sec without cache_index_and_filter_blocks.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6784
Test Plan:
Test added to DBTest.ApproximateSizesFilesWithErrorMargin.
Old code running new test...
[ RUN ] DBTest.ApproximateSizesFilesWithErrorMargin
db/db_test.cc:1562: Failure
Expected: (size) <= (11 * 100), actual: 9478 vs 1100
Other tests updated to reflect consistent accounting of metadata.
Reviewed By: siying
Differential Revision: D21334706
Pulled By: pdillinger
fbshipit-source-id: 6f86870e45213334fedbe9c73b4ebb1d8d611185
5 years ago
|
|
|
|
|
|
|
// Pro-rate file metadata (incl filters) size-proportionally across data
|
|
|
|
// blocks.
|
|
|
|
double size_ratio =
|
|
|
|
static_cast<double>(offset) / static_cast<double>(data_size);
|
|
|
|
return static_cast<uint64_t>(size_ratio *
|
|
|
|
static_cast<double>(rep_->file_size));
|
|
|
|
}
|
|
|
|
|
|
|
|
uint64_t BlockBasedTable::ApproximateSize(const ReadOptions& read_options,
|
|
|
|
const Slice& start, const Slice& end,
|
|
|
|
TableReaderCaller caller) {
|
|
|
|
assert(rep_->internal_comparator.Compare(start, end) <= 0);
|
|
|
|
|
For ApproximateSizes, pro-rate table metadata size over data blocks (#6784)
Summary:
The implementation of GetApproximateSizes was inconsistent in
its treatment of the size of non-data blocks of SST files, sometimes
including and sometimes now. This was at its worst with large portion
of table file used by filters and querying a small range that crossed
a table boundary: the size estimate would include large filter size.
It's conceivable that someone might want only to know the size in terms
of data blocks, but I believe that's unlikely enough to ignore for now.
Similarly, there's no evidence the internal function AppoximateOffsetOf
is used for anything other than a one-sided ApproximateSize, so I intend
to refactor to remove redundancy in a follow-up commit.
So to fix this, GetApproximateSizes (and implementation details
ApproximateSize and ApproximateOffsetOf) now consistently include in
their returned sizes a portion of table file metadata (incl filters
and indexes) based on the size portion of the data blocks in range. In
other words, if a key range covers data blocks that are X% by size of all
the table's data blocks, returned approximate size is X% of the total
file size. It would technically be more accurate to attribute metadata
based on number of keys, but that's not computationally efficient with
data available and rarely a meaningful difference.
Also includes miscellaneous comment improvements / clarifications.
Also included is a new approximatesizerandom benchmark for db_bench.
No significant performance difference seen with this change, whether ~700 ops/sec with cache_index_and_filter_blocks and small cache or ~150k ops/sec without cache_index_and_filter_blocks.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6784
Test Plan:
Test added to DBTest.ApproximateSizesFilesWithErrorMargin.
Old code running new test...
[ RUN ] DBTest.ApproximateSizesFilesWithErrorMargin
db/db_test.cc:1562: Failure
Expected: (size) <= (11 * 100), actual: 9478 vs 1100
Other tests updated to reflect consistent accounting of metadata.
Reviewed By: siying
Differential Revision: D21334706
Pulled By: pdillinger
fbshipit-source-id: 6f86870e45213334fedbe9c73b4ebb1d8d611185
5 years ago
|
|
|
uint64_t data_size = GetApproximateDataSize();
|
|
|
|
if (UNLIKELY(data_size == 0)) {
|
|
|
|
// Hmm. Assume whole file is involved, since we have lower and upper
|
Handle failures in block-based table size/offset approximation (#9615)
Summary:
In crash test with fault injection, we were seeing stack traces like the following:
```
https://github.com/facebook/rocksdb/issues/3 0x00007f75f763c533 in __GI___assert_fail (assertion=assertion@entry=0x1c5b2a0 "end_offset >= start_offset", file=file@entry=0x1c580a0 "table/block_based/block_based_table_reader.cc", line=line@entry=3245,
function=function@entry=0x1c60e60 "virtual uint64_t rocksdb::BlockBasedTable::ApproximateSize(const rocksdb::Slice&, const rocksdb::Slice&, rocksdb::TableReaderCaller)") at assert.c:101
https://github.com/facebook/rocksdb/issues/4 0x00000000010ea9b4 in rocksdb::BlockBasedTable::ApproximateSize (this=<optimized out>, start=..., end=..., caller=<optimized out>) at table/block_based/block_based_table_reader.cc:3224
https://github.com/facebook/rocksdb/issues/5 0x0000000000be61fb in rocksdb::TableCache::ApproximateSize (this=0x60f0000161b0, start=..., end=..., fd=..., caller=caller@entry=rocksdb::kCompaction, internal_comparator=..., prefix_extractor=...) at db/table_cache.cc:719
https://github.com/facebook/rocksdb/issues/6 0x0000000000c3eaec in rocksdb::VersionSet::ApproximateSize (this=<optimized out>, v=<optimized out>, f=..., start=..., end=..., caller=<optimized out>) at ./db/version_set.h:850
https://github.com/facebook/rocksdb/issues/7 0x0000000000c6ebc3 in rocksdb::VersionSet::ApproximateSize (this=<optimized out>, options=..., v=v@entry=0x621000047500, start=..., end=..., start_level=start_level@entry=0, end_level=<optimized out>, caller=<optimized out>)
at db/version_set.cc:5657
https://github.com/facebook/rocksdb/issues/8 0x000000000166e894 in rocksdb::CompactionJob::GenSubcompactionBoundaries (this=<optimized out>) at ./include/rocksdb/options.h:1869
https://github.com/facebook/rocksdb/issues/9 0x000000000168c526 in rocksdb::CompactionJob::Prepare (this=this@entry=0x7f75f3ffcf00) at db/compaction/compaction_job.cc:546
```
The problem occurred in `ApproximateSize()` when the index `Seek()` for the first `ApproximateDataOffsetOf()` encountered an I/O error, while the second `Seek()` did not. In the old code that scenario caused `start_offset == data_size` , thus it was easy to trip the assertion that `end_offset >= start_offset`.
The fix is to set `start_offset == 0` when the first index `Seek()` fails, and `end_offset == data_size` when the second index `Seek()` fails. I doubt these give an "on average correct" answer for how this function is used, but I/O errors in index seeks are hopefully rare, it looked consistent with what was already there, and it was easier to calculate.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9615
Test Plan:
run the repro command for a while and stopped seeing coredumps -
```
$ while ! ./db_stress --block_size=128 --cache_size=32768 --clear_column_family_one_in=0 --column_families=1 --continuous_verification_interval=0 --db=/dev/shm/rocksdb_crashtest --delpercent=4 --delrangepercent=1 --destroy_db_initially=0 --expected_values_dir=/dev/shm/rocksdb_crashtest_expected --index_type=2 --iterpercent=10 --kill_random_test=18887 --max_key=1000000 --max_bytes_for_level_base=2048576 --nooverwritepercent=1 --open_files=-1 --open_read_fault_one_in=32 --ops_per_thread=1000000 --prefixpercent=5 --read_fault_one_in=0 --readpercent=45 --reopen=0 --skip_verifydb=1 --subcompactions=2 --target_file_size_base=524288 --test_batches_snapshots=0 --value_size_mult=32 --write_buffer_size=524288 --writepercent=35 ; do : ; done
```
Reviewed By: pdillinger
Differential Revision: D34383069
Pulled By: ajkr
fbshipit-source-id: fac26c3b20ea962e75387515ba5f2724dc48719f
3 years ago
|
|
|
// bound. This likely skews the estimate if we consider that this function
|
|
|
|
// is typically called with `[start, end]` fully contained in the file's
|
|
|
|
// key-range.
|
For ApproximateSizes, pro-rate table metadata size over data blocks (#6784)
Summary:
The implementation of GetApproximateSizes was inconsistent in
its treatment of the size of non-data blocks of SST files, sometimes
including and sometimes now. This was at its worst with large portion
of table file used by filters and querying a small range that crossed
a table boundary: the size estimate would include large filter size.
It's conceivable that someone might want only to know the size in terms
of data blocks, but I believe that's unlikely enough to ignore for now.
Similarly, there's no evidence the internal function AppoximateOffsetOf
is used for anything other than a one-sided ApproximateSize, so I intend
to refactor to remove redundancy in a follow-up commit.
So to fix this, GetApproximateSizes (and implementation details
ApproximateSize and ApproximateOffsetOf) now consistently include in
their returned sizes a portion of table file metadata (incl filters
and indexes) based on the size portion of the data blocks in range. In
other words, if a key range covers data blocks that are X% by size of all
the table's data blocks, returned approximate size is X% of the total
file size. It would technically be more accurate to attribute metadata
based on number of keys, but that's not computationally efficient with
data available and rarely a meaningful difference.
Also includes miscellaneous comment improvements / clarifications.
Also included is a new approximatesizerandom benchmark for db_bench.
No significant performance difference seen with this change, whether ~700 ops/sec with cache_index_and_filter_blocks and small cache or ~150k ops/sec without cache_index_and_filter_blocks.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6784
Test Plan:
Test added to DBTest.ApproximateSizesFilesWithErrorMargin.
Old code running new test...
[ RUN ] DBTest.ApproximateSizesFilesWithErrorMargin
db/db_test.cc:1562: Failure
Expected: (size) <= (11 * 100), actual: 9478 vs 1100
Other tests updated to reflect consistent accounting of metadata.
Reviewed By: siying
Differential Revision: D21334706
Pulled By: pdillinger
fbshipit-source-id: 6f86870e45213334fedbe9c73b4ebb1d8d611185
5 years ago
|
|
|
return rep_->file_size;
|
|
|
|
}
|
|
|
|
|
|
|
|
BlockCacheLookupContext context(caller);
|
|
|
|
IndexBlockIter iiter_on_stack;
|
|
|
|
ReadOptions ro;
|
|
|
|
ro.total_order_seek = true;
|
|
|
|
ro.io_activity = read_options.io_activity;
|
|
|
|
auto index_iter =
|
|
|
|
NewIndexIterator(ro, /*disable_prefix_seek=*/true,
|
|
|
|
/*input_iter=*/&iiter_on_stack, /*get_context=*/nullptr,
|
|
|
|
/*lookup_context=*/&context);
|
|
|
|
std::unique_ptr<InternalIteratorBase<IndexValue>> iiter_unique_ptr;
|
|
|
|
if (index_iter != &iiter_on_stack) {
|
|
|
|
iiter_unique_ptr.reset(index_iter);
|
|
|
|
}
|
|
|
|
|
|
|
|
index_iter->Seek(start);
|
Handle failures in block-based table size/offset approximation (#9615)
Summary:
In crash test with fault injection, we were seeing stack traces like the following:
```
https://github.com/facebook/rocksdb/issues/3 0x00007f75f763c533 in __GI___assert_fail (assertion=assertion@entry=0x1c5b2a0 "end_offset >= start_offset", file=file@entry=0x1c580a0 "table/block_based/block_based_table_reader.cc", line=line@entry=3245,
function=function@entry=0x1c60e60 "virtual uint64_t rocksdb::BlockBasedTable::ApproximateSize(const rocksdb::Slice&, const rocksdb::Slice&, rocksdb::TableReaderCaller)") at assert.c:101
https://github.com/facebook/rocksdb/issues/4 0x00000000010ea9b4 in rocksdb::BlockBasedTable::ApproximateSize (this=<optimized out>, start=..., end=..., caller=<optimized out>) at table/block_based/block_based_table_reader.cc:3224
https://github.com/facebook/rocksdb/issues/5 0x0000000000be61fb in rocksdb::TableCache::ApproximateSize (this=0x60f0000161b0, start=..., end=..., fd=..., caller=caller@entry=rocksdb::kCompaction, internal_comparator=..., prefix_extractor=...) at db/table_cache.cc:719
https://github.com/facebook/rocksdb/issues/6 0x0000000000c3eaec in rocksdb::VersionSet::ApproximateSize (this=<optimized out>, v=<optimized out>, f=..., start=..., end=..., caller=<optimized out>) at ./db/version_set.h:850
https://github.com/facebook/rocksdb/issues/7 0x0000000000c6ebc3 in rocksdb::VersionSet::ApproximateSize (this=<optimized out>, options=..., v=v@entry=0x621000047500, start=..., end=..., start_level=start_level@entry=0, end_level=<optimized out>, caller=<optimized out>)
at db/version_set.cc:5657
https://github.com/facebook/rocksdb/issues/8 0x000000000166e894 in rocksdb::CompactionJob::GenSubcompactionBoundaries (this=<optimized out>) at ./include/rocksdb/options.h:1869
https://github.com/facebook/rocksdb/issues/9 0x000000000168c526 in rocksdb::CompactionJob::Prepare (this=this@entry=0x7f75f3ffcf00) at db/compaction/compaction_job.cc:546
```
The problem occurred in `ApproximateSize()` when the index `Seek()` for the first `ApproximateDataOffsetOf()` encountered an I/O error, while the second `Seek()` did not. In the old code that scenario caused `start_offset == data_size` , thus it was easy to trip the assertion that `end_offset >= start_offset`.
The fix is to set `start_offset == 0` when the first index `Seek()` fails, and `end_offset == data_size` when the second index `Seek()` fails. I doubt these give an "on average correct" answer for how this function is used, but I/O errors in index seeks are hopefully rare, it looked consistent with what was already there, and it was easier to calculate.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9615
Test Plan:
run the repro command for a while and stopped seeing coredumps -
```
$ while ! ./db_stress --block_size=128 --cache_size=32768 --clear_column_family_one_in=0 --column_families=1 --continuous_verification_interval=0 --db=/dev/shm/rocksdb_crashtest --delpercent=4 --delrangepercent=1 --destroy_db_initially=0 --expected_values_dir=/dev/shm/rocksdb_crashtest_expected --index_type=2 --iterpercent=10 --kill_random_test=18887 --max_key=1000000 --max_bytes_for_level_base=2048576 --nooverwritepercent=1 --open_files=-1 --open_read_fault_one_in=32 --ops_per_thread=1000000 --prefixpercent=5 --read_fault_one_in=0 --readpercent=45 --reopen=0 --skip_verifydb=1 --subcompactions=2 --target_file_size_base=524288 --test_batches_snapshots=0 --value_size_mult=32 --write_buffer_size=524288 --writepercent=35 ; do : ; done
```
Reviewed By: pdillinger
Differential Revision: D34383069
Pulled By: ajkr
fbshipit-source-id: fac26c3b20ea962e75387515ba5f2724dc48719f
3 years ago
|
|
|
uint64_t start_offset;
|
|
|
|
if (index_iter->status().ok()) {
|
|
|
|
start_offset = ApproximateDataOffsetOf(*index_iter, data_size);
|
|
|
|
} else {
|
|
|
|
// Assume file is involved from the start. This likely skews the estimate
|
|
|
|
// but is consistent with the above error handling.
|
|
|
|
start_offset = 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
index_iter->Seek(end);
|
Handle failures in block-based table size/offset approximation (#9615)
Summary:
In crash test with fault injection, we were seeing stack traces like the following:
```
https://github.com/facebook/rocksdb/issues/3 0x00007f75f763c533 in __GI___assert_fail (assertion=assertion@entry=0x1c5b2a0 "end_offset >= start_offset", file=file@entry=0x1c580a0 "table/block_based/block_based_table_reader.cc", line=line@entry=3245,
function=function@entry=0x1c60e60 "virtual uint64_t rocksdb::BlockBasedTable::ApproximateSize(const rocksdb::Slice&, const rocksdb::Slice&, rocksdb::TableReaderCaller)") at assert.c:101
https://github.com/facebook/rocksdb/issues/4 0x00000000010ea9b4 in rocksdb::BlockBasedTable::ApproximateSize (this=<optimized out>, start=..., end=..., caller=<optimized out>) at table/block_based/block_based_table_reader.cc:3224
https://github.com/facebook/rocksdb/issues/5 0x0000000000be61fb in rocksdb::TableCache::ApproximateSize (this=0x60f0000161b0, start=..., end=..., fd=..., caller=caller@entry=rocksdb::kCompaction, internal_comparator=..., prefix_extractor=...) at db/table_cache.cc:719
https://github.com/facebook/rocksdb/issues/6 0x0000000000c3eaec in rocksdb::VersionSet::ApproximateSize (this=<optimized out>, v=<optimized out>, f=..., start=..., end=..., caller=<optimized out>) at ./db/version_set.h:850
https://github.com/facebook/rocksdb/issues/7 0x0000000000c6ebc3 in rocksdb::VersionSet::ApproximateSize (this=<optimized out>, options=..., v=v@entry=0x621000047500, start=..., end=..., start_level=start_level@entry=0, end_level=<optimized out>, caller=<optimized out>)
at db/version_set.cc:5657
https://github.com/facebook/rocksdb/issues/8 0x000000000166e894 in rocksdb::CompactionJob::GenSubcompactionBoundaries (this=<optimized out>) at ./include/rocksdb/options.h:1869
https://github.com/facebook/rocksdb/issues/9 0x000000000168c526 in rocksdb::CompactionJob::Prepare (this=this@entry=0x7f75f3ffcf00) at db/compaction/compaction_job.cc:546
```
The problem occurred in `ApproximateSize()` when the index `Seek()` for the first `ApproximateDataOffsetOf()` encountered an I/O error, while the second `Seek()` did not. In the old code that scenario caused `start_offset == data_size` , thus it was easy to trip the assertion that `end_offset >= start_offset`.
The fix is to set `start_offset == 0` when the first index `Seek()` fails, and `end_offset == data_size` when the second index `Seek()` fails. I doubt these give an "on average correct" answer for how this function is used, but I/O errors in index seeks are hopefully rare, it looked consistent with what was already there, and it was easier to calculate.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9615
Test Plan:
run the repro command for a while and stopped seeing coredumps -
```
$ while ! ./db_stress --block_size=128 --cache_size=32768 --clear_column_family_one_in=0 --column_families=1 --continuous_verification_interval=0 --db=/dev/shm/rocksdb_crashtest --delpercent=4 --delrangepercent=1 --destroy_db_initially=0 --expected_values_dir=/dev/shm/rocksdb_crashtest_expected --index_type=2 --iterpercent=10 --kill_random_test=18887 --max_key=1000000 --max_bytes_for_level_base=2048576 --nooverwritepercent=1 --open_files=-1 --open_read_fault_one_in=32 --ops_per_thread=1000000 --prefixpercent=5 --read_fault_one_in=0 --readpercent=45 --reopen=0 --skip_verifydb=1 --subcompactions=2 --target_file_size_base=524288 --test_batches_snapshots=0 --value_size_mult=32 --write_buffer_size=524288 --writepercent=35 ; do : ; done
```
Reviewed By: pdillinger
Differential Revision: D34383069
Pulled By: ajkr
fbshipit-source-id: fac26c3b20ea962e75387515ba5f2724dc48719f
3 years ago
|
|
|
uint64_t end_offset;
|
|
|
|
if (index_iter->status().ok()) {
|
|
|
|
end_offset = ApproximateDataOffsetOf(*index_iter, data_size);
|
|
|
|
} else {
|
|
|
|
// Assume file is involved until the end. This likely skews the estimate
|
|
|
|
// but is consistent with the above error handling.
|
|
|
|
end_offset = data_size;
|
|
|
|
}
|
|
|
|
|
|
|
|
assert(end_offset >= start_offset);
|
For ApproximateSizes, pro-rate table metadata size over data blocks (#6784)
Summary:
The implementation of GetApproximateSizes was inconsistent in
its treatment of the size of non-data blocks of SST files, sometimes
including and sometimes now. This was at its worst with large portion
of table file used by filters and querying a small range that crossed
a table boundary: the size estimate would include large filter size.
It's conceivable that someone might want only to know the size in terms
of data blocks, but I believe that's unlikely enough to ignore for now.
Similarly, there's no evidence the internal function AppoximateOffsetOf
is used for anything other than a one-sided ApproximateSize, so I intend
to refactor to remove redundancy in a follow-up commit.
So to fix this, GetApproximateSizes (and implementation details
ApproximateSize and ApproximateOffsetOf) now consistently include in
their returned sizes a portion of table file metadata (incl filters
and indexes) based on the size portion of the data blocks in range. In
other words, if a key range covers data blocks that are X% by size of all
the table's data blocks, returned approximate size is X% of the total
file size. It would technically be more accurate to attribute metadata
based on number of keys, but that's not computationally efficient with
data available and rarely a meaningful difference.
Also includes miscellaneous comment improvements / clarifications.
Also included is a new approximatesizerandom benchmark for db_bench.
No significant performance difference seen with this change, whether ~700 ops/sec with cache_index_and_filter_blocks and small cache or ~150k ops/sec without cache_index_and_filter_blocks.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6784
Test Plan:
Test added to DBTest.ApproximateSizesFilesWithErrorMargin.
Old code running new test...
[ RUN ] DBTest.ApproximateSizesFilesWithErrorMargin
db/db_test.cc:1562: Failure
Expected: (size) <= (11 * 100), actual: 9478 vs 1100
Other tests updated to reflect consistent accounting of metadata.
Reviewed By: siying
Differential Revision: D21334706
Pulled By: pdillinger
fbshipit-source-id: 6f86870e45213334fedbe9c73b4ebb1d8d611185
5 years ago
|
|
|
// Pro-rate file metadata (incl filters) size-proportionally across data
|
|
|
|
// blocks.
|
|
|
|
double size_ratio = static_cast<double>(end_offset - start_offset) /
|
|
|
|
static_cast<double>(data_size);
|
|
|
|
return static_cast<uint64_t>(size_ratio *
|
|
|
|
static_cast<double>(rep_->file_size));
|
|
|
|
}
|
|
|
|
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
bool BlockBasedTable::TEST_FilterBlockInCache() const {
|
|
|
|
assert(rep_ != nullptr);
|
New stable, fixed-length cache keys (#9126)
Summary:
This change standardizes on a new 16-byte cache key format for
block cache (incl compressed and secondary) and persistent cache (but
not table cache and row cache).
The goal is a really fast cache key with practically ideal stability and
uniqueness properties without external dependencies (e.g. from FileSystem).
A fixed key size of 16 bytes should enable future optimizations to the
concurrent hash table for block cache, which is a heavy CPU user /
bottleneck, but there appears to be measurable performance improvement
even with no changes to LRUCache.
This change replaces a lot of disjointed and ugly code handling cache
keys with calls to a simple, clean new internal API (cache_key.h).
(Preserving the old cache key logic under an option would be very ugly
and likely negate the performance gain of the new approach. Complete
replacement carries some inherent risk, but I think that's acceptable
with sufficient analysis and testing.)
The scheme for encoding new cache keys is complicated but explained
in cache_key.cc.
Also: EndianSwapValue is moved to math.h to be next to other bit
operations. (Explains some new include "math.h".) ReverseBits operation
added and unit tests added to hash_test for both.
Fixes https://github.com/facebook/rocksdb/issues/7405 (presuming a root cause)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9126
Test Plan:
### Basic correctness
Several tests needed updates to work with the new functionality, mostly
because we are no longer relying on filesystem for stable cache keys
so table builders & readers need more context info to agree on cache
keys. This functionality is so core, a huge number of existing tests
exercise the cache key functionality.
### Performance
Create db with
`TEST_TMPDIR=/dev/shm ./db_bench -bloom_bits=10 -benchmarks=fillrandom -num=3000000 -partition_index_and_filters`
And test performance with
`TEST_TMPDIR=/dev/shm ./db_bench -readonly -use_existing_db -bloom_bits=10 -benchmarks=readrandom -num=3000000 -duration=30 -cache_index_and_filter_blocks -cache_size=250000 -threads=4`
using DEBUG_LEVEL=0 and simultaneous before & after runs.
Before ops/sec, avg over 100 runs: 121924
After ops/sec, avg over 100 runs: 125385 (+2.8%)
### Collision probability
I have built a tool, ./cache_bench -stress_cache_key to broadly simulate host-wide cache activity
over many months, by making some pessimistic simplifying assumptions:
* Every generated file has a cache entry for every byte offset in the file (contiguous range of cache keys)
* All of every file is cached for its entire lifetime
We use a simple table with skewed address assignment and replacement on address collision
to simulate files coming & going, with quite a variance (super-Poisson) in ages. Some output
with `./cache_bench -stress_cache_key -sck_keep_bits=40`:
```
Total cache or DBs size: 32TiB Writing 925.926 MiB/s or 76.2939TiB/day
Multiply by 9.22337e+18 to correct for simulation losses (but still assume whole file cached)
```
These come from default settings of 2.5M files per day of 32 MB each, and
`-sck_keep_bits=40` means that to represent a single file, we are only keeping 40 bits of
the 128-bit cache key. With file size of 2\*\*25 contiguous keys (pessimistic), our simulation
is about 2\*\*(128-40-25) or about 9 billion billion times more prone to collision than reality.
More default assumptions, relatively pessimistic:
* 100 DBs in same process (doesn't matter much)
* Re-open DB in same process (new session ID related to old session ID) on average
every 100 files generated
* Restart process (all new session IDs unrelated to old) 24 times per day
After enough data, we get a result at the end:
```
(keep 40 bits) 17 collisions after 2 x 90 days, est 10.5882 days between (9.76592e+19 corrected)
```
If we believe the (pessimistic) simulation and the mathematical generalization, we would need to run a billion machines all for 97 billion days to expect a cache key collision. To help verify that our generalization ("corrected") is robust, we can make our simulation more precise with `-sck_keep_bits=41` and `42`, which takes more running time to get enough data:
```
(keep 41 bits) 16 collisions after 4 x 90 days, est 22.5 days between (1.03763e+20 corrected)
(keep 42 bits) 19 collisions after 10 x 90 days, est 47.3684 days between (1.09224e+20 corrected)
```
The generalized prediction still holds. With the `-sck_randomize` option, we can see that we are beating "random" cache keys (except offsets still non-randomized) by a modest amount (roughly 20x less collision prone than random), which should make us reasonably comfortable even in "degenerate" cases:
```
197 collisions after 1 x 90 days, est 0.456853 days between (4.21372e+18 corrected)
```
I've run other tests to validate other conditions behave as expected, never behaving "worse than random" unless we start chopping off structured data.
Reviewed By: zhichao-cao
Differential Revision: D33171746
Pulled By: pdillinger
fbshipit-source-id: f16a57e369ed37be5e7e33525ace848d0537c88f
3 years ago
|
|
|
return rep_->filter_type != Rep::FilterType::kNoFilter &&
|
|
|
|
TEST_BlockInCache(rep_->filter_handle);
|
|
|
|
}
|
|
|
|
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
bool BlockBasedTable::TEST_IndexBlockInCache() const {
|
|
|
|
assert(rep_ != nullptr);
|
|
|
|
|
|
|
|
return TEST_BlockInCache(rep_->footer.index_handle());
|
|
|
|
}
|
|
|
|
|
|
|
|
Status BlockBasedTable::GetKVPairsFromDataBlocks(
|
|
|
|
const ReadOptions& read_options, std::vector<KVPairBlock>* kv_pair_blocks) {
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
std::unique_ptr<InternalIteratorBase<IndexValue>> blockhandles_iter(
|
|
|
|
NewIndexIterator(read_options, /*need_upper_bound_check=*/false,
|
|
|
|
/*input_iter=*/nullptr, /*get_context=*/nullptr,
|
|
|
|
/*lookup_contex=*/nullptr));
|
|
|
|
|
|
|
|
Status s = blockhandles_iter->status();
|
|
|
|
if (!s.ok()) {
|
|
|
|
// Cannot read Index Block
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
for (blockhandles_iter->SeekToFirst(); blockhandles_iter->Valid();
|
|
|
|
blockhandles_iter->Next()) {
|
|
|
|
s = blockhandles_iter->status();
|
|
|
|
|
|
|
|
if (!s.ok()) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
std::unique_ptr<InternalIterator> datablock_iter;
|
|
|
|
Status tmp_status;
|
|
|
|
datablock_iter.reset(NewDataBlockIterator<DataBlockIter>(
|
|
|
|
read_options, blockhandles_iter->value().handle,
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
/*input_iter=*/nullptr, /*type=*/BlockType::kData,
|
|
|
|
/*get_context=*/nullptr, /*lookup_context=*/nullptr,
|
|
|
|
/*prefetch_buffer=*/nullptr, /*for_compaction=*/false,
|
|
|
|
/*async_read=*/false, tmp_status));
|
|
|
|
s = datablock_iter->status();
|
|
|
|
|
|
|
|
if (!s.ok()) {
|
|
|
|
// Error reading the block - Skipped
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
KVPairBlock kv_pair_block;
|
|
|
|
for (datablock_iter->SeekToFirst(); datablock_iter->Valid();
|
|
|
|
datablock_iter->Next()) {
|
|
|
|
s = datablock_iter->status();
|
|
|
|
if (!s.ok()) {
|
|
|
|
// Error reading the block - Skipped
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
const Slice& key = datablock_iter->key();
|
|
|
|
const Slice& value = datablock_iter->value();
|
|
|
|
std::string key_copy = std::string(key.data(), key.size());
|
|
|
|
std::string value_copy = std::string(value.data(), value.size());
|
|
|
|
|
|
|
|
kv_pair_block.push_back(
|
|
|
|
std::make_pair(std::move(key_copy), std::move(value_copy)));
|
|
|
|
}
|
|
|
|
kv_pair_blocks->push_back(std::move(kv_pair_block));
|
|
|
|
}
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
Status BlockBasedTable::DumpTable(WritableFile* out_file) {
|
|
|
|
WritableFileStringStreamAdapter out_file_wrapper(out_file);
|
|
|
|
std::ostream out_stream(&out_file_wrapper);
|
|
|
|
// Output Footer
|
|
|
|
out_stream << "Footer Details:\n"
|
|
|
|
"--------------------------------------\n";
|
|
|
|
out_stream << " " << rep_->footer.ToString() << "\n";
|
|
|
|
|
|
|
|
// Output MetaIndex
|
|
|
|
out_stream << "Metaindex Details:\n"
|
|
|
|
"--------------------------------------\n";
|
|
|
|
std::unique_ptr<Block> metaindex;
|
|
|
|
std::unique_ptr<InternalIterator> metaindex_iter;
|
|
|
|
// TODO: plumb Env::IOActivity
|
|
|
|
const ReadOptions ro;
|
|
|
|
Status s = ReadMetaIndexBlock(ro, nullptr /* prefetch_buffer */, &metaindex,
|
|
|
|
&metaindex_iter);
|
|
|
|
if (s.ok()) {
|
|
|
|
for (metaindex_iter->SeekToFirst(); metaindex_iter->Valid();
|
|
|
|
metaindex_iter->Next()) {
|
|
|
|
s = metaindex_iter->status();
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
if (metaindex_iter->key() == kPropertiesBlockName) {
|
|
|
|
out_stream << " Properties block handle: "
|
|
|
|
<< metaindex_iter->value().ToString(true) << "\n";
|
|
|
|
} else if (metaindex_iter->key() == kCompressionDictBlockName) {
|
|
|
|
out_stream << " Compression dictionary block handle: "
|
|
|
|
<< metaindex_iter->value().ToString(true) << "\n";
|
|
|
|
} else if (strstr(metaindex_iter->key().ToString().c_str(),
|
|
|
|
"filter.rocksdb.") != nullptr) {
|
|
|
|
out_stream << " Filter block handle: "
|
|
|
|
<< metaindex_iter->value().ToString(true) << "\n";
|
|
|
|
} else if (metaindex_iter->key() == kRangeDelBlockName) {
|
|
|
|
out_stream << " Range deletion block handle: "
|
|
|
|
<< metaindex_iter->value().ToString(true) << "\n";
|
|
|
|
}
|
|
|
|
}
|
|
|
|
out_stream << "\n";
|
|
|
|
} else {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Output TableProperties
|
|
|
|
const ROCKSDB_NAMESPACE::TableProperties* table_properties;
|
|
|
|
table_properties = rep_->table_properties.get();
|
|
|
|
|
|
|
|
if (table_properties != nullptr) {
|
|
|
|
out_stream << "Table Properties:\n"
|
|
|
|
"--------------------------------------\n";
|
|
|
|
out_stream << " " << table_properties->ToString("\n ", ": ") << "\n";
|
|
|
|
}
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
5 years ago
|
|
|
|
|
|
|
if (rep_->filter) {
|
|
|
|
out_stream << "Filter Details:\n"
|
|
|
|
"--------------------------------------\n";
|
|
|
|
out_stream << " " << rep_->filter->ToString() << "\n";
|
|
|
|
}
|
|
|
|
|
|
|
|
// Output Index block
|
|
|
|
s = DumpIndexBlock(out_stream);
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Output compression dictionary
|
|
|
|
if (rep_->uncompression_dict_reader) {
|
|
|
|
CachableEntry<UncompressionDict> uncompression_dict;
|
|
|
|
s = rep_->uncompression_dict_reader->GetOrReadUncompressionDictionary(
|
|
|
|
nullptr /* prefetch_buffer */, ro, false /* no_io */,
|
|
|
|
false, /* verify_checksums */
|
|
|
|
nullptr /* get_context */, nullptr /* lookup_context */,
|
|
|
|
&uncompression_dict);
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
assert(uncompression_dict.GetValue());
|
|
|
|
|
|
|
|
const Slice& raw_dict = uncompression_dict.GetValue()->GetRawDict();
|
|
|
|
out_stream << "Compression Dictionary:\n"
|
|
|
|
"--------------------------------------\n";
|
|
|
|
out_stream << " size (bytes): " << raw_dict.size() << "\n\n";
|
|
|
|
out_stream << " HEX " << raw_dict.ToString(true) << "\n\n";
|
|
|
|
}
|
|
|
|
|
|
|
|
// Output range deletions block
|
|
|
|
auto* range_del_iter = NewRangeTombstoneIterator(ro);
|
|
|
|
if (range_del_iter != nullptr) {
|
|
|
|
range_del_iter->SeekToFirst();
|
|
|
|
if (range_del_iter->Valid()) {
|
|
|
|
out_stream << "Range deletions:\n"
|
|
|
|
"--------------------------------------\n";
|
|
|
|
for (; range_del_iter->Valid(); range_del_iter->Next()) {
|
|
|
|
DumpKeyValue(range_del_iter->key(), range_del_iter->value(),
|
|
|
|
out_stream);
|
|
|
|
}
|
|
|
|
out_stream << "\n";
|
|
|
|
}
|
|
|
|
delete range_del_iter;
|
|
|
|
}
|
|
|
|
// Output Data blocks
|
|
|
|
s = DumpDataBlocks(out_stream);
|
|
|
|
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!out_stream.good()) {
|
|
|
|
return Status::IOError("Failed to write to output file");
|
|
|
|
}
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
|
|
|
Status BlockBasedTable::DumpIndexBlock(std::ostream& out_stream) {
|
|
|
|
out_stream << "Index Details:\n"
|
|
|
|
"--------------------------------------\n";
|
|
|
|
// TODO: plumb Env::IOActivity
|
|
|
|
const ReadOptions read_options;
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
std::unique_ptr<InternalIteratorBase<IndexValue>> blockhandles_iter(
|
|
|
|
NewIndexIterator(read_options, /*need_upper_bound_check=*/false,
|
|
|
|
/*input_iter=*/nullptr, /*get_context=*/nullptr,
|
|
|
|
/*lookup_contex=*/nullptr));
|
|
|
|
Status s = blockhandles_iter->status();
|
|
|
|
if (!s.ok()) {
|
|
|
|
out_stream << "Can not read Index Block \n\n";
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
out_stream << " Block key hex dump: Data block handle\n";
|
|
|
|
out_stream << " Block key ascii\n\n";
|
|
|
|
for (blockhandles_iter->SeekToFirst(); blockhandles_iter->Valid();
|
|
|
|
blockhandles_iter->Next()) {
|
|
|
|
s = blockhandles_iter->status();
|
|
|
|
if (!s.ok()) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
Slice key = blockhandles_iter->key();
|
|
|
|
Slice user_key;
|
|
|
|
InternalKey ikey;
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
if (!rep_->index_key_includes_seq) {
|
|
|
|
user_key = key;
|
|
|
|
} else {
|
|
|
|
ikey.DecodeFrom(key);
|
|
|
|
user_key = ikey.user_key();
|
|
|
|
}
|
|
|
|
|
|
|
|
out_stream << " HEX " << user_key.ToString(true) << ": "
|
|
|
|
<< blockhandles_iter->value().ToString(true,
|
|
|
|
rep_->index_has_first_key)
|
|
|
|
<< " offset " << blockhandles_iter->value().handle.offset()
|
|
|
|
<< " size " << blockhandles_iter->value().handle.size() << "\n";
|
|
|
|
|
|
|
|
std::string str_key = user_key.ToString();
|
|
|
|
std::string res_key("");
|
|
|
|
char cspace = ' ';
|
|
|
|
for (size_t i = 0; i < str_key.size(); i++) {
|
|
|
|
res_key.append(&str_key[i], 1);
|
|
|
|
res_key.append(1, cspace);
|
|
|
|
}
|
|
|
|
out_stream << " ASCII " << res_key << "\n";
|
|
|
|
out_stream << " ------\n";
|
|
|
|
}
|
|
|
|
out_stream << "\n";
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
|
|
|
Status BlockBasedTable::DumpDataBlocks(std::ostream& out_stream) {
|
|
|
|
// TODO: plumb Env::IOActivity
|
|
|
|
const ReadOptions read_options;
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
std::unique_ptr<InternalIteratorBase<IndexValue>> blockhandles_iter(
|
|
|
|
NewIndexIterator(read_options, /*need_upper_bound_check=*/false,
|
|
|
|
/*input_iter=*/nullptr, /*get_context=*/nullptr,
|
|
|
|
/*lookup_contex=*/nullptr));
|
|
|
|
Status s = blockhandles_iter->status();
|
|
|
|
if (!s.ok()) {
|
|
|
|
out_stream << "Can not read Index Block \n\n";
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
uint64_t datablock_size_min = std::numeric_limits<uint64_t>::max();
|
|
|
|
uint64_t datablock_size_max = 0;
|
|
|
|
uint64_t datablock_size_sum = 0;
|
|
|
|
|
|
|
|
size_t block_id = 1;
|
|
|
|
for (blockhandles_iter->SeekToFirst(); blockhandles_iter->Valid();
|
|
|
|
block_id++, blockhandles_iter->Next()) {
|
|
|
|
s = blockhandles_iter->status();
|
|
|
|
if (!s.ok()) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
BlockHandle bh = blockhandles_iter->value().handle;
|
|
|
|
uint64_t datablock_size = bh.size();
|
|
|
|
datablock_size_min = std::min(datablock_size_min, datablock_size);
|
|
|
|
datablock_size_max = std::max(datablock_size_max, datablock_size);
|
|
|
|
datablock_size_sum += datablock_size;
|
|
|
|
|
|
|
|
out_stream << "Data Block # " << block_id << " @ "
|
|
|
|
<< blockhandles_iter->value().handle.ToString(true) << "\n";
|
|
|
|
out_stream << "--------------------------------------\n";
|
|
|
|
|
|
|
|
std::unique_ptr<InternalIterator> datablock_iter;
|
|
|
|
Status tmp_status;
|
|
|
|
datablock_iter.reset(NewDataBlockIterator<DataBlockIter>(
|
|
|
|
read_options, blockhandles_iter->value().handle,
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
|
|
|
/*input_iter=*/nullptr, /*type=*/BlockType::kData,
|
|
|
|
/*get_context=*/nullptr, /*lookup_context=*/nullptr,
|
|
|
|
/*prefetch_buffer=*/nullptr, /*for_compaction=*/false,
|
|
|
|
/*async_read=*/false, tmp_status));
|
|
|
|
s = datablock_iter->status();
|
|
|
|
|
|
|
|
if (!s.ok()) {
|
|
|
|
out_stream << "Error reading the block - Skipped \n\n";
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
for (datablock_iter->SeekToFirst(); datablock_iter->Valid();
|
|
|
|
datablock_iter->Next()) {
|
|
|
|
s = datablock_iter->status();
|
|
|
|
if (!s.ok()) {
|
|
|
|
out_stream << "Error reading the block - Skipped \n";
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
DumpKeyValue(datablock_iter->key(), datablock_iter->value(), out_stream);
|
|
|
|
}
|
|
|
|
out_stream << "\n";
|
|
|
|
}
|
|
|
|
|
|
|
|
uint64_t num_datablocks = block_id - 1;
|
|
|
|
if (num_datablocks) {
|
|
|
|
double datablock_size_avg =
|
|
|
|
static_cast<double>(datablock_size_sum) / num_datablocks;
|
|
|
|
out_stream << "Data Block Summary:\n";
|
|
|
|
out_stream << "--------------------------------------\n";
|
|
|
|
out_stream << " # data blocks: " << num_datablocks << "\n";
|
|
|
|
out_stream << " min data block size: " << datablock_size_min << "\n";
|
|
|
|
out_stream << " max data block size: " << datablock_size_max << "\n";
|
|
|
|
out_stream << " avg data block size: "
|
|
|
|
<< std::to_string(datablock_size_avg) << "\n";
|
|
|
|
}
|
|
|
|
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
|
|
|
void BlockBasedTable::DumpKeyValue(const Slice& key, const Slice& value,
|
|
|
|
std::ostream& out_stream) {
|
|
|
|
InternalKey ikey;
|
|
|
|
ikey.DecodeFrom(key);
|
|
|
|
|
|
|
|
out_stream << " HEX " << ikey.user_key().ToString(true) << ": "
|
|
|
|
<< value.ToString(true) << "\n";
|
|
|
|
|
|
|
|
std::string str_key = ikey.user_key().ToString();
|
|
|
|
std::string str_value = value.ToString();
|
|
|
|
std::string res_key(""), res_value("");
|
|
|
|
char cspace = ' ';
|
|
|
|
for (size_t i = 0; i < str_key.size(); i++) {
|
|
|
|
if (str_key[i] == '\0') {
|
|
|
|
res_key.append("\\0", 2);
|
|
|
|
} else {
|
|
|
|
res_key.append(&str_key[i], 1);
|
|
|
|
}
|
|
|
|
res_key.append(1, cspace);
|
|
|
|
}
|
|
|
|
for (size_t i = 0; i < str_value.size(); i++) {
|
|
|
|
if (str_value[i] == '\0') {
|
|
|
|
res_value.append("\\0", 2);
|
|
|
|
} else {
|
|
|
|
res_value.append(&str_value[i], 1);
|
|
|
|
}
|
|
|
|
res_value.append(1, cspace);
|
|
|
|
}
|
|
|
|
|
|
|
|
out_stream << " ASCII " << res_key << ": " << res_value << "\n";
|
|
|
|
out_stream << " ------\n";
|
|
|
|
}
|
|
|
|
|
|
|
|
} // namespace ROCKSDB_NAMESPACE
|