mirror of https://github.com/facebook/rocksdb.git
38 Commits
| Author | SHA1 | Message | Date |
|---|---|---|---|
yuzhangyu@fb.com
|
1cfdece85d |
Run internal cpp modernizer on RocksDB repo (#12398)
Summary: When internal cpp modernizer attempts to format rocksdb code, it will replace macro `ROCKSDB_NAMESPACE` with its default definition `rocksdb` when collapsing nested namespace. We filed a feedback for the tool T180254030 and the team filed a bug for this: https://github.com/llvm/llvm-project/issues/83452. At the same time, they suggested us to run the modernizer tool ourselves so future auto codemod attempts will be smaller. This diff contains: Running `xplat/scripts/codemod_service/cpp_modernizer.sh` in fbcode/internal_repo_rocksdb/repo (excluding some directories in utilities/transactions/lock/range/range_tree/lib that has a non meta copyright comment) without swapping out the namespace macro `ROCKSDB_NAMESPACE` Followed by RocksDB's own `make format` Pull Request resolved: https://github.com/facebook/rocksdb/pull/12398 Test Plan: Auto tests Reviewed By: hx235 Differential Revision: D54382532 Pulled By: jowlyzhang fbshipit-source-id: e7d5b40f9b113b60e5a503558c181f080b9d02fa |
|
|
Peter Dillinger
|
88bc91f3cc |
Cap eviction effort (CPU under stress) in HyperClockCache (#12141)
Summary: HyperClockCache is intended to mitigate performance problems under stress conditions (as well as optimizing average-case parallel performance). In LRUCache, the biggest such problem is lock contention when one or a small number of cache entries becomes particularly hot. Regardless of cache sharding, accesses to any particular cache entry are linearized against a single mutex, which is held while each access updates the LRU list. All HCC variants are fully lock/wait-free for accessing blocks already in the cache, which fully mitigates this contention problem. However, HCC (and CLOCK in general) can exhibit extremely degraded performance under a different stress condition: when no (or almost no) entries in a cache shard are evictable (they are pinned). Unlike LRU which can find any evictable entries immediately (at the cost of more coordination / synchronization on each access), CLOCK has to search for evictable entries. Under the right conditions (almost exclusively MB-scale caches not GB-scale), the CPU cost of each cache miss could fall off a cliff and bog down the whole system. To effectively mitigate this problem (IMHO), I'm introducing a new default behavior and tuning parameter for HCC, `eviction_effort_cap`. See the comments on the new config parameter in the public API. Pull Request resolved: https://github.com/facebook/rocksdb/pull/12141 Test Plan: unit test included ## Performance test We can use cache_bench to validate no regression (CPU and memory) in normal operation, and to measure change in behavior when cache is almost entirely pinned. (TODO: I'm not sure why I had to get the pinned ratio parameter well over 1.0 to see truly bad performance, but the behavior is there.) Build with `make DEBUG_LEVEL=0 USE_CLANG=1 PORTABLE=0 cache_bench`. We also set MALLOC_CONF="narenas:1" for all these runs to essentially remove jemalloc variances from the results, so that the max RSS given by /usr/bin/time is essentially ideal (assuming the allocator minimizes fragmentation and other memory overheads well). Base command reproducing bad behavior: ``` ./cache_bench -cache_type=auto_hyper_clock_cache -threads=12 -histograms=0 -pinned_ratio=1.7 ``` ``` Before, LRU (alternate baseline not exhibiting bad behavior): Rough parallel ops/sec = 2290997 1088060 maxresident Before, AutoHCC (bad behavior): Rough parallel ops/sec = 141011 <- Yes, more than 10x slower 1083932 maxresident ``` Now let us sample a range of values in the solution space: ``` After, AutoHCC, eviction_effort_cap = 1: Rough parallel ops/sec = 3212586 2402216 maxresident After, AutoHCC, eviction_effort_cap = 10: Rough parallel ops/sec = 2371639 1248884 maxresident After, AutoHCC, eviction_effort_cap = 30: Rough parallel ops/sec = 1981092 1131596 maxresident After, AutoHCC, eviction_effort_cap = 100: Rough parallel ops/sec = 1446188 1090976 maxresident After, AutoHCC, eviction_effort_cap = 1000: Rough parallel ops/sec = 549568 1084064 maxresident ``` I looks like `cap=30` is a sweet spot balancing acceptable CPU and memory overheads, so is chosen as the default. ``` Change to -pinned_ratio=0.85 Before, LRU: Rough parallel ops/sec = 2108373 1078232 maxresident Before, AutoHCC, averaged over ~20 runs: Rough parallel ops/sec = 2164910 1077312 maxresident After, AutoHCC, eviction_effort_cap = 30, averaged over ~20 runs: Rough parallel ops/sec = 2145542 1077216 maxresident ``` The slight CPU improvement above is consistent with the cap, with no measurable memory overhead under moderate stress. ``` Change to -pinned_ratio=0.25 (low stress) Before, AutoHCC, averaged over ~20 runs: Rough parallel ops/sec = 2221149 1076540 maxresident After, AutoHCC, eviction_effort_cap = 30, averaged over ~20 runs: Rough parallel ops/sec = 2224521 1076664 maxresident ``` No measurable difference under normal circumstances. Some tests repeated with FixedHCC, with similar results. Reviewed By: anand1976 Differential Revision: D52174755 Pulled By: pdillinger fbshipit-source-id: d278108031b1220c1fa4c89c5a9d34b7cf4ef1b8 |
|
|
Peter Dillinger
|
92dc5f3e67 |
AutoHCC: fix a bug with "blind" Insert (#12046)
Summary: I have finally tracked down and fixed a bug affecting AutoHCC that was causing CI crash test assertion failures in AutoHCC when using secondary cache, but I was only able to reproduce locally a couple of times, after very long runs/repetitions. It turns out that the essential feature used by secondary cache to trigger the bug is Insert without keeping a handle, which is otherwise rarely used in RocksDB and not incorporated into cache_bench (also used for targeted correctness stress testing) until this change (new option `-blind_insert_percent`). The problem was in copying some logic from FixedHCC that makes the entry "sharable" but unreferenced once populated, if no reference is to be saved. The problem in AutoHCC is that we can only add the entry to a chain after it is in the sharable state, and must be removed from the chain while in the "under (de)construction" state and before it is back in the "empty" state. Also, it is possible for Lookup to find entries that are not connected to any chain, by design for efficiency, and for Release to erase_if_last_ref. Therefore, we could have * Thread 1 starts to Insert a cache entry without keeping ref, and pauses before adding to the chain. * Thread 2 finds it with Lookup optimizations, and then does Release with `erase_if_last_ref=true` causing it to trigger erasure on the entry. It successfully locks the home chain for the entry and purges any entries pending erasure. It is OK that this entry is not found on the chain, as another thread is allowed to remove it from the chain before we are able to (but after is it marked for (de)construction). And after the purge of the chain, the entry is marked empty. * Thread 1 resumes in adding the slot (presumed entry) to the home chain for what was being inserted, but that now violates invariants and sets up a race or double-chain-reference as another thread could insert a new entry in the slot and try to insert into a different chain. This is easily fixed by holding on to a reference until inserted onto the chain. Pull Request resolved: https://github.com/facebook/rocksdb/pull/12046 Test Plan: As I don't have a reliable local reproducer, I triggered 20 runs of internal CI on fbcode_blackbox_crash_test that were previously failing in AutoHCC with about 1/3 probability, and they all passed. Also re-enabling AutoHCC in the crash test with this change. (Revert https://github.com/facebook/rocksdb/issues/12000) Reviewed By: jowlyzhang Differential Revision: D51016979 Pulled By: pdillinger fbshipit-source-id: 3840fb829d65b97c779d8aed62a4a4a433aeff2b |
|
|
Peter Dillinger
|
a399bbc037 |
More fixes and enhancements for cache_bench (#12041)
Summary: Mostly things for using cache_bench for stress/correctness testing. * Make secondary_cache_uri option work with HCC (forgot to update when secondary support was added for HCC) * Add -pinned_ratio option to keep more than just one entry per thread pinned. This can be important for testing eviction stress. * Add -vary_capacity_ratio for testing dynamically changing capacity. Also added some overrides to CacheWrapper to help with diagnostic output. Pull Request resolved: https://github.com/facebook/rocksdb/pull/12041 Test Plan: manual, make check Reviewed By: jowlyzhang Differential Revision: D51013430 Pulled By: pdillinger fbshipit-source-id: 7914adc1218f0afacace05ccd77d3bfb91a878d0 |
|
|
anand76
|
269478ee46 |
Support compressed and local flash secondary cache stacking (#11812)
Summary: This PR implements support for a three tier cache - primary block cache, compressed secondary cache, and a nvm (local flash) secondary cache. This allows more effective utilization of the nvm cache, and minimizes the number of reads from local flash by caching compressed blocks in the compressed secondary cache. The basic design is as follows - 1. A new secondary cache implementation, ```TieredSecondaryCache```, is introduced. It keeps the compressed and nvm secondary caches and manages the movement of blocks between them and the primary block cache. To setup a three tier cache, we allocate a ```CacheWithSecondaryAdapter```, with a ```TieredSecondaryCache``` instance as the secondary cache. 2. The table reader passes both the uncompressed and compressed block to ```FullTypedCacheInterface::InsertFull```, allowing the block cache to optionally store the compressed block. 3. When there's a miss, the block object is constructed and inserted in the primary cache, and the compressed block is inserted into the nvm cache by calling ```InsertSaved```. This avoids the overhead of recompressing the block, as well as avoiding putting more memory pressure on the compressed secondary cache. 4. When there's a hit in the nvm cache, we attempt to insert the block in the compressed secondary cache and the primary cache, subject to the admission policy of those caches (i.e admit on second access). Blocks/items evicted from any tier are simply discarded. We can easily implement additional admission policies if desired. Todo (In a subsequent PR): 1. Add to db_bench and run benchmarks 2. Add to db_stress Pull Request resolved: https://github.com/facebook/rocksdb/pull/11812 Reviewed By: pdillinger Differential Revision: D49461842 Pulled By: anand1976 fbshipit-source-id: b40ac1330ef7cd8c12efa0a3ca75128e602e3a0b |
|
|
Peter Dillinger
|
fe3405e80f |
Automatic table sizing for HyperClockCache (AutoHCC) (#11738)
Summary: This change add an experimental next-generation HyperClockCache (HCC) with automatic sizing of the underlying hash table. Both the existing version (stable) and the new version (experimental for now) of HCC are available depending on whether an estimated average entry charge is provided in HyperClockCacheOptions. Internally, we call the two implementations AutoHyperClockCache (new) and FixedHyperClockCache (existing). The performance characteristics and much of the underlying logic are similar enough that AutoHCC is likely to make FixedHCC obsolete, and so it's best considered an evolution of the same technology or solution rather than an alternative. More specifically, both implementations share essentially the same logic for managing the state of individual entries in the cache, including metadata for reference counting and counting clocks for eviction. This metadata, which I like to call the "low-level HCC protocol," includes a read-write lock on entries, but relaxed consistency requirements on the cache (e.g. allowing rare duplication) means high-level cache operations never wait for these low-level per-entry locks. FixedHCC is fully wait-free. AutoHCC is different in how entries are indexed into an efficient hash table. AutoHCC is "essentially wait-free" as there is no pattern of typical high-level operations on a large cache that can lead to one thread waiting on another to complete some work, though it can happen in some unusual/unlucky cases, or atypical uses such as erasing specific cache keys. Table growth and entry reclamation is more complex in AutoHCC compared to FixedHCC, so uses some localized locking to manage that. AutoHCC uses linear hashing to grow the table as needed, with low latency and to a precise size. AutoHCC depends on anonymous mmap support from the OS (currently verified working on Linux, MacOS, and Windows) to allow the array underlying a hash table to grow in place without wasting resident memory on space reserved but unused. AutoHCC uses a form of chaining while FixedHCC uses open addressing and double hashing. More specifics: * In developing this PR, a rare availability bug (minor) was noticed in the existing HCC implementation of Release()+erase_if_last_ref, which is now inherited into AutoHCC. Fixing this without a performance regression will not be simple, so is left for follow-up work. * Some existing unit tests required adjustment of operational parameters or conditions to work with the new behaviors of AutoHCC. A number of bugs were found and fixed in the validation process, including getting unit tests in good working order. * Added an option to cache_bench, `-degenerate_hash_bits` for correctness stress testing described below. For this, the tool uses the reverse-engineered hash function for HCC to generate keys in which the specified number of hash bits, in critical positions, have a fixed value. Essentially each degenerate hash bit will half the number of chain heads utilized and double the average chain length. Pull Request resolved: https://github.com/facebook/rocksdb/pull/11738 Test Plan: unit tests updated, and already added to db crash test. Also ## Correctness The code includes generous assertions to check for unexpected states, especially at destruction time, so should be able to detect critical concurrency bugs. Less serious "availability bugs" in which cache data is hidden or cleanly lost are more difficult to detect, but also less scary for data correctness (as long as performance is good and the design is sound). In average operation, the structure is extremely low stress and low contention (see next section) so stressing the corner case logic requires artificially stressing the operating conditions. First, we keep the structure small to increase the number of threads hitting the same chain or entry, and just one cache shard. Second, we artificially degrade the hashing so that chains are much longer than typical, using the new `-degenerate_hash_bits` option to cache_bench. Third, we re-create the structure from scratch frequently in order to exercise the Grow logic repeatedly and to get the benefit of the consistency checks in the structure's destructor in debug builds. For cache_bench this also means disabling the single-threaded "populate cache" step (normally used for steady state performance testing). And of course use many more threads than cores to have many preemptions. An effective test for working out bugs was this (using debug build of course): ``` while ./cache_bench -cache_type=auto_hyper_clock_cache -histograms=0 -cache_size=8000000 -threads=100 -populate_cache=0 -ops_per_thread=10000 -degenerate_hash_bits=6 -num_shard_bits=0; do :; done ``` Or even smaller cases. This setup has around 27 utilized chains, with around 35 entries each, and yield-waits more than 1 million times per second (very high contention; see next section). I have let this run for hours searching for any lingering issues. I've also run cache_bench under ASAN, UBSAN, and TSAN. ## Essentially wait free There is a counter for number of yield() calls when one thread is waiting on another. When we pre-populate the structure in a single thread, ``` ./cache_bench -cache_type=auto_hyper_clock_cache -histograms=0 -populate_cache=1 -ops_per_thread=200000 2>&1 | grep Yield ``` We see something on the order of 1 yield call per second across 16 threads, even when we load the system other other jobs (parallel compilation). With -populate_cache=0, there are more yield opportunities with parallel table growth. On an otherwise unloaded system, we still see very small (single digit) yield counts, with a chance of getting into the thousands, and getting into 10s of thousands per second during table growth phase if the system is loaded with other jobs. However, I am not worried about this if performance is still good (see next section). ## Overall performance Although cache_bench initially suggested performance very close to FixedHCC, there was a very noticeable performance hit under a db_bench setup like used in validating https://github.com/facebook/rocksdb/issues/10626. Much of the difference has been reduced by optimizing Lookup with a "naive" pass that will almost always find entries quickly, and only falling back to the careful Lookup algorithm when not found in the first pass. Setups (chosen to be sensitive to block cache performance), and compiled with USE_CLANG=1 JEMALLOC=1 PORTABLE=0 DEBUG_LEVEL=0: ``` TEST_TMPDIR=/dev/shm base/db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16 ``` ### No regression on FixedHCC Running before & after builds at the same time on a 48 core machine. ``` TEST_TMPDIR=/dev/shm /usr/bin/time ./db_bench -benchmarks=readrandom[-X10],block_cache_entry_stats,cache_report_problems -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=610000000 -duration 20 -threads=24 -cache_type=fixed_hyper_clock_cache -seed=1234 ``` Before: readrandom [AVG 10 runs] : 847234 (± 8150) ops/sec; 59.2 (± 0.6) MB/sec 703MB max RSS After: readrandom [AVG 10 runs] : 851021 (± 7929) ops/sec; 59.5 (± 0.6) MB/sec 706MB max RSS Probably no material difference. ### Single-threaded performance Using `[-X2]` and `-threads=1` and `-duration=30`, running all three at the same time: lru_cache: 55100 ops/sec, then 55862 ops/sec (627MB max RSS) fixed_hyper_clock_cache: 60496 ops/sec, then 61231 ops/sec (626MB max RSS) auto_hyper_clock_cache: 47560 ops/sec, then 56081 ops/sec (626MB max RSS) So AutoHCC has more ramp-up cost in the first pass as the cache grows to the appropriate size. (In single-threaded operation, the parallelizability and per-op low latency of table growth is overall slower.) However, once up to size, its performance is comparable to LRUCache. FixedHCC's lean operations still win overall when a good estimate is available. If we look at HCC table stats, we can see that this configuration is not favorable to AutoHCC (and I have verified that other memory sizes do not yield substantially different results, until shards are under-sized for the full filters): FixedHCC: Slot occupancy stats: Overall 47% (124991/262144), Min/Max/Window = 28%/64%/500, MaxRun{Pos/Neg} = 17/22 AutoHCC: Slot occupancy stats: Overall 59% (125781/209682), Min/Max/Window = 43%/82%/500, MaxRun{Pos/Neg} = 76/16 Head occupancy stats: Overall 43% (92259/209682), Min/Max/Window = 24%/74%/500, MaxRun{Pos/Neg} = 19/26 Entries at home count: 53350 FixedHCC configuration is relatively good for speed, and not ideal for space utilization. As is typical, AutoHCC has tighter control on metadata usage (209682 x 64 bytes rather than 262144 x 64 bytes), and the higher load factor is slightly worse for speed. LRUCache also has more metadata usage, at 199680 x 96 bytes of tracked metadata (plus roughly another 10% of that untracked in the head pointers), and that metadata is subject to fragmentation. ### Parallel performance, high hit rate Now using `[-X10]` and `-threads=10`, all three at the same time lru_cache: [AVG 10 runs] : 263629 (± 1425) ops/sec; 18.4 (± 0.1) MB/sec 655MB max RSS, 97.1% cache hit rate fixed_hyper_clock_cache: [AVG 10 runs] : 479590 (± 8114) ops/sec; 33.5 (± 0.6) MB/sec 651MB max RSS, 97.1% cache hit rate auto_hyper_clock_cache: [AVG 10 runs] : 418687 (± 5915) ops/sec; 29.3 (± 0.4) MB/sec 657MB max RSS, 97.1% cache hit rate Even with just 10-way parallelism for each cache (though 30+/48 cores busy overall), LRUCache is already showing performance degradation, while AutoHCC is in the neighborhood of FixedHCC. And that brings us to the question of how AutoHCC holds up under extreme parallelism, so now independent runs with `-threads=100` (overloading 48 cores). lru_cache: 438613 ops/sec, 827MB max RSS fixed_hyper_clock_cache: 1651310 ops/sec, 812MB max RSS auto_hyper_clock_cache: 1505875 ops/sec, 821MB max RSS (Yield count: 1089 over 30s) Clearly, AutoHCC holds up extremely well under extreme parallelism, even closing some of the modest performance gap with FixedHCC. ### Parallel performance, low hit rate To get down to roughly 50% cache hit rate, we use `-cache_index_and_filter_blocks=0 -cache_size=1650000000` with `-threads=10`. Here the extra cost of running counting clock eviction, especially on the chains of AutoHCC, are evident, especially with the lower contention of cache_index_and_filter_blocks=0: lru_cache: 725231 ops/sec, 1770MB max RSS, 51.3% hit rate fixed_hyper_clock_cache: 638620 ops/sec, 1765MB max RSS, 50.2% hit rate auto_hyper_clock_cache: 541018 ops/sec, 1777MB max RSS, 50.8% hit rate Reviewed By: jowlyzhang Differential Revision: D48784755 Pulled By: pdillinger fbshipit-source-id: e79813dc087474ac427637dd282a14fa3011a6e4 |
|
|
Peter Dillinger
|
d3420464c3 |
cache_bench enhancements for jemalloc etc. (#11758)
Summary:
* Add some options to cache_bench to use JemallocNodumpAllocator
* Make num_shard_bits option use and report cache-specific defaults
* Add a usleep option to sleep between operations, for simulating a workload with more CPU idle/wait time.
* Use const& for JemallocAllocatorOptions, to improve API usability (e.g. can bind to temporary `{}`)
* InstallStackTraceHandler()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11758
Test Plan: manual
Reviewed By: jowlyzhang
Differential Revision: D48668479
Pulled By: pdillinger
fbshipit-source-id: b6032fbe09444cdb8f1443a5e017d2eea4f6205a
|
|
|
Peter Dillinger
|
ef6f025563 |
Placeholder for AutoHyperClockCache, more (#11692)
Summary:
* The plan is for AutoHyperClockCache to be selected when HyperClockCacheOptions::estimated_entry_charge == 0, and in that case to use a new configuration option min_avg_entry_charge for determining an extreme case maximum size for the hash table. For the placeholder, a hack is in place in HyperClockCacheOptions::MakeSharedCache() to make the unit tests happy despite the new options not really making sense with the current implementation.
* Mostly updating and refactoring tests to test both the current HCC (internal name FixedHyperClockCache) and a placeholder for the new version (internal name AutoHyperClockCache).
* Simplify some existing tests not to depend directly on cache type.
* Type-parameterize the shard-level unit tests, which unfortunately requires more syntax like `this->` in places for disambiguation.
* Added means of choosing auto_hyper_clock_cache to cache_bench, db_bench, and db_stress, including add to crash test.
* Add another templated class BaseHyperClockCache to reduce future copy-paste
* Added ReportProblems support to cache_bench
* Added a DEBUG-level diagnostic to ReportProblems for the variance in load factor throughout the table, which will become more of a concern with linear hashing to be used in the Auto implementation. Example with current Fixed HCC:
```
2023/08/10-13:41:41.602450 6ac36 [DEBUG] [che/clock_cache.cc:1507] Slot occupancy stats: Overall 49% (129008/262144), Min/Max/Window = 39%/60%/500, MaxRun{Pos/Neg} = 18/17
```
In other words, with overall occupancy of 49%, the lowest across any 500 contiguous cells is 39% and highest 60%. Longest run of occupied is 18 and longest run of unoccupied is 17. This seems consistent with random samples from a uniform distribution.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11692
Test Plan: Shouldn't be any meaningful changes yet to production code or to what is tested, but there is temporary redundancy in testing until the new implementation is plugged in.
Reviewed By: jowlyzhang
Differential Revision: D48247413
Pulled By: pdillinger
fbshipit-source-id: 11541f996d97af403c2e43c92fb67ff22dd0b5da
|
|
|
Peter Dillinger
|
99daea3481 |
Prepare tests for new HCC naming (#11676)
Summary: I'm anticipating using the public name HyperClockCache for both the current version with a fixed-size table and the upcoming version with an automatically growing table. However, for simplicity of testing them as substantially distinct implementations, I want to give them distinct internal names, like FixedHyperClockCache and AutoHyperClockCache. This change anticipates that by renaming to FixedHyperClockCache and assuming for now that all the unit tests run on HCC will run and behave similarly for the automatic HCC. Obviously updates will need to be made, but I'm trying to avoid uninteresting find & replace updates in what will be a large and engineering-heavy PR for AutoHCC Pull Request resolved: https://github.com/facebook/rocksdb/pull/11676 Test Plan: no behavior change intended, except logging will now use the name FixedHyperClockCache Reviewed By: ajkr Differential Revision: D48103165 Pulled By: pdillinger fbshipit-source-id: a33f1901488fea102164c2318e2f2b156aaba736 |
|
|
Peter Dillinger
|
cdb11f5ce6 |
More minor HCC refactoring + typed mmap (#11670)
Summary: More code leading up to dynamic HCC. * Small enhancements to cache_bench * Extra assertion in Unref * Improve a CAS loop in ChargeUsageMaybeEvictStrict * Put load factor constants in appropriate class * Move `standalone` field to HyperClockTable::HandleImpl because it can be encoded differently in the upcoming dynamic HCC. * Add a typed version of MemMapping to simplify some future code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/11670 Test Plan: existing tests, unit test added for TypedMemMapping Reviewed By: jowlyzhang Differential Revision: D48056464 Pulled By: pdillinger fbshipit-source-id: 186b7d3105c5d6d2eb6a592369bc10a97ee14a15 |
|
|
Peter Dillinger
|
f9de217353 |
Some cache_bench enhancements (#11661)
Summary: ... used in validating some HyperClockCache development in progress. * Revamp the "populate cache" step to avoid redundant insertions (very rare in practice) and more consistently approach the desired resident_ratio while maintaining appropriate skew (still not perfect). * Track and print hit ratio on lookups, to ensure a fair comparison is happening between implementations etc. * Add an option to disable tracking and printing histograms (lots of output) * Add an option to specify a random seed (for more reproducibility) * Remove confusing/redundant "-skewed" option Uses BitwiseAnd from https://github.com/facebook/rocksdb/issues/11660 (tested there) Pull Request resolved: https://github.com/facebook/rocksdb/pull/11661 Test Plan: manual Reviewed By: akankshamahajan15, jowlyzhang Differential Revision: D47937671 Pulled By: pdillinger fbshipit-source-id: 85a2bb881b1bca4f63e015bac684105fd91c9f35 |
|
|
Peter Dillinger
|
b1b6f87fbe |
Some small improvements to HyperClockCache (#11601)
Summary: Stacked on https://github.com/facebook/rocksdb/issues/11572 * Minimize use of std::function and lambdas to minimize chances of compiler heap-allocating closures (unnecessary stress on allocator). It appears that converting FindSlot to a template enables inlining the lambda parameters, avoiding heap allocations. * Clean up some logic with FindSlot (FIXMEs from https://github.com/facebook/rocksdb/issues/11572) * Fix handling of rare case of probing all slots, with new unit test. (Previously Insert would not roll back displacements in that case, which would kill performance if it were to happen.) * Add an -early_exit option to cache_bench for gathering memory stats before deallocation. Pull Request resolved: https://github.com/facebook/rocksdb/pull/11601 Test Plan: unit test added for probing all slots ## Seeing heap allocations Run `MALLOC_CONF="stats_print:true" ./cache_bench -cache_type=hyper_clock_cache` before https://github.com/facebook/rocksdb/issues/11572 vs. after this change. Before, we see this in the interesting bin statistics: ``` size nrequests ---- --------- 32 578460 64 24340 8192 578460 ``` And after: ``` size nrequests ---- --------- 32 (insignificant) 64 24370 8192 579130 ``` ## Performance test Build with `make USE_CLANG=1 PORTABLE=0 DEBUG_LEVEL=0 -j32 cache_bench` Run `./cache_bench -cache_type=hyper_clock_cache -ops_per_thread=5000000` in before and after configurations, simultaneously: ``` Before: Complete in 33.244 s; Rough parallel ops/sec = 2406442 After: Complete in 32.773 s; Rough parallel ops/sec = 2441019 ``` Reviewed By: jowlyzhang Differential Revision: D47375092 Pulled By: pdillinger fbshipit-source-id: 46f0f57257ddb374290a0a38c651764ea60ba410 |
|
|
Peter Dillinger
|
204fcff751 |
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
|
|
|
Peter Dillinger
|
ccaa3225b0 |
Simplify tracking entries already in SecondaryCache (#11299)
Summary: In preparation for factoring secondary cache support out of individual Cache implementations, we can get rid of the "in secondary cache" flag on entries through a workable hack: when an entry is promoted from secondary, it is inserted in primary using a helper that lacks secondary cache support, thus preventing re-insertion into secondary cache through existing logic. This adds to the complexity of building CacheItemHelpers, because you always have to be able to get to an equivalent helper without secondary cache support, but that complexity is reasonably isolated within RocksDB typed_cache.h and test code. gcc-7 seems to have problems with constexpr constructor referencing `this` so removed constexpr support on CacheItemHelper. Also refactored some related test code to share common code / functionality. Pull Request resolved: https://github.com/facebook/rocksdb/pull/11299 Test Plan: existing tests Reviewed By: anand1976 Differential Revision: D44101453 Pulled By: pdillinger fbshipit-source-id: 7a59d0a3938ee40159c90c3e65d7004f6a272345 |
|
|
Peter Dillinger
|
3cacd4b4ec |
Put Cache and CacheWrapper in new public header (#11192)
Summary: The definition of the Cache class should not be needed by the vast majority of RocksDB users, so I think it is just distracting to include it in cache.h, which is primarily needed for configuring and creating caches. This change moves the class to a new header advanced_cache.h. It is just cut-and-paste except for modifying the class API comment. In general, operations on shared_ptr<Cache> should continue to work when only a forward declaration of Cache is available, as long as all the Cache instances provided are already shared_ptr. See https://stackoverflow.com/a/17650101/454544 Also, the most common way to customize a Cache is by wrapping an existing implementation, so it makes sense to provide CacheWrapper in the public API. This was a cut-and-paste job except removing the implementation of Name() so that derived classes must provide it. Intended follow-up: consolidate Release() into one function to reduce customization bugs / confusion Pull Request resolved: https://github.com/facebook/rocksdb/pull/11192 Test Plan: `make check` Reviewed By: anand1976 Differential Revision: D43055487 Pulled By: pdillinger fbshipit-source-id: 7b05492df35e0f30b581b4c24c579bc275b6d110 |
|
|
sdong
|
4720ba4391 |
Remove RocksDB LITE (#11147)
Summary: We haven't been actively mantaining RocksDB LITE recently and the size must have been gone up significantly. We are removing the support. Most of changes were done through following comments: unifdef -m -UROCKSDB_LITE `git grep -l ROCKSDB_LITE | egrep '[.](cc|h)'` by Peter Dillinger. Others changes were manually applied to build scripts, CircleCI manifests, ROCKSDB_LITE is used in an expression and file db_stress_test_base.cc. Pull Request resolved: https://github.com/facebook/rocksdb/pull/11147 Test Plan: See CI Reviewed By: pdillinger Differential Revision: D42796341 fbshipit-source-id: 4920e15fc2060c2cd2221330a6d0e5e65d4b7fe2 |
|
|
Peter Dillinger
|
9f7801c5f1 |
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 |
|
|
Peter Dillinger
|
32520df1d9 |
Remove prototype FastLRUCache (#10954)
Summary: This was just a stepping stone to what eventually became HyperClockCache, and is now just more code to maintain. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10954 Test Plan: tests updated Reviewed By: akankshamahajan15 Differential Revision: D41310123 Pulled By: pdillinger fbshipit-source-id: 618ee148a1a0a29ee756ba8fe28359617b7cd67c |
|
|
Peter Dillinger
|
0f91c72adc |
Call experimental new clock cache HyperClockCache (#10684)
Summary: This change establishes a distinctive name for the experimental new lock-free clock cache (originally developed by guidotag and revamped in PR https://github.com/facebook/rocksdb/issues/10626). A few reasons: * We want to make it clear that this is a fundamentally different implementation vs. the old clock cache, to avoid people saying "I already tried clock cache." * We want to highlight the key feature: it's fast (especially under parallel load) * Because it requires an estimated charge per entry, it is not drop-in API compatible with old clock cache. This estimate might always be required for highest performance, and giving it a distinct name should reduce confusion about the distinct API requirements. * We might develop a variant requiring the same estimate parameter but with LRU eviction. In that case, using the name HyperLRUCache should make things more clear. (FastLRUCache is just a prototype that might soon be removed.) Some API detail: * To reduce copy-pasting parameter lists, etc. as in LRUCache construction, I have a `MakeSharedCache()` function on `HyperClockCacheOptions` instead of `NewHyperClockCache()`. * Changes -cache_type=clock_cache to -cache_type=hyper_clock_cache for applicable tools. I think this is more consistent / sustainable for reasons already stated. For performance tests see https://github.com/facebook/rocksdb/pull/10626 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10684 Test Plan: no interesting functional changes; tests updated Reviewed By: anand1976 Differential Revision: D39547800 Pulled By: pdillinger fbshipit-source-id: 5c0fe1b5cf3cb680ab369b928c8569682b9795bf |
|
|
Peter Dillinger
|
5724348689 |
Revamp, optimize new experimental clock cache (#10626)
Summary:
* Consolidates most metadata into a single word per slot so that more
can be accomplished with a single atomic update. In the common case,
Lookup was previously about 4 atomic updates, now just 1 atomic update.
Common case Release was previously 1 atomic read + 1 atomic update,
now just 1 atomic update.
* Eliminate spins / waits / yields, which likely threaten some "lock free"
benefits. Compare-exchange loops are only used in explicit Erase, and
strict_capacity_limit=true Insert. Eviction uses opportunistic compare-
exchange.
* Relaxes some aggressiveness and guarantees. For example,
* Duplicate Inserts will sometimes go undetected and the shadow duplicate
will age out with eviction.
* In many cases, the older Inserted value for a given cache key will be kept
(i.e. Insert does not support overwrite).
* Entries explicitly erased (rather than evicted) might not be freed
immediately in some rare cases.
* With strict_capacity_limit=false, capacity limit is not tracked/enforced as
precisely as LRUCache, but is self-correcting and should only deviate by a
very small number of extra or fewer entries.
* Use smaller "computed default" number of cache shards in many cases,
because benefits to larger usage tracking / eviction pools outweigh the small
cost of more lock-free atomic contention. The improvement in CPU and I/O
is dramatic in some limit-memory cases.
* Even without the sharding change, the eviction algorithm is likely more
effective than LRU overall because it's more stateful, even though the
"hot path" state tracking for it is essentially free with ref counting. It
is like a generalized CLOCK with aging (see code comments). I don't have
performance numbers showing a specific improvement, but in theory, for a
Poisson access pattern to each block, keeping some state allows better
estimation of time to next access (Poisson interval) than strict LRU. The
bounded randomness in CLOCK can also reduce "cliff" effect for repeated
range scans approaching and exceeding cache size.
## Hot path algorithm comparison
Rough descriptions, focusing on number and kind of atomic operations:
* Old `Lookup()` (2-5 atomic updates per probe):
```
Loop:
Increment internal ref count at slot
If possible hit:
Check flags atomic (and non-atomic fields)
If cache hit:
Three distinct updates to 'flags' atomic
Increment refs for internal-to-external
Return
Decrement internal ref count
while atomic read 'displacements' > 0
```
* New `Lookup()` (1-2 atomic updates per probe):
```
Loop:
Increment acquire counter in meta word (optimistic)
If visible entry (already read meta word):
If match (read non-atomic fields):
Return
Else:
Decrement acquire counter in meta word
Else if invisible entry (rare, already read meta word):
Decrement acquire counter in meta word
while atomic read 'displacements' > 0
```
* Old `Release()` (1 atomic update, conditional on atomic read, rarely more):
```
Read atomic ref count
If last reference and invisible (rare):
Use CAS etc. to remove
Return
Else:
Decrement ref count
```
* New `Release()` (1 unconditional atomic update, rarely more):
```
Increment release counter in meta word
If last reference and invisible (rare):
Use CAS etc. to remove
Return
```
## Performance test setup
Build DB with
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16
```
Test with
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics
```
Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations:
base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6)
folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry)
gt_clock: experimental ClockCache before this change
new_clock: experimental ClockCache with this change
## Performance test results
First test "hot path" read performance, with block cache large enough for whole DB:
4181MB 1thread base -> kops/s: 47.761
4181MB 1thread folly -> kops/s: 45.877
4181MB 1thread gt_clock -> kops/s: 51.092
4181MB 1thread new_clock -> kops/s: 53.944
4181MB 16thread base -> kops/s: 284.567
4181MB 16thread folly -> kops/s: 249.015
4181MB 16thread gt_clock -> kops/s: 743.762
4181MB 16thread new_clock -> kops/s: 861.821
4181MB 24thread base -> kops/s: 303.415
4181MB 24thread folly -> kops/s: 266.548
4181MB 24thread gt_clock -> kops/s: 975.706
4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944)
4181MB 32thread base -> kops/s: 311.251
4181MB 32thread folly -> kops/s: 274.952
4181MB 32thread gt_clock -> kops/s: 1045.98
4181MB 32thread new_clock -> kops/s: 1370.38
4181MB 48thread base -> kops/s: 310.504
4181MB 48thread folly -> kops/s: 268.322
4181MB 48thread gt_clock -> kops/s: 1195.65
4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944)
4181MB 64thread base -> kops/s: 307.839
4181MB 64thread folly -> kops/s: 272.172
4181MB 64thread gt_clock -> kops/s: 1204.47
4181MB 64thread new_clock -> kops/s: 1615.37
4181MB 128thread base -> kops/s: 310.934
4181MB 128thread folly -> kops/s: 267.468
4181MB 128thread gt_clock -> kops/s: 1188.75
4181MB 128thread new_clock -> kops/s: 1595.46
Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x.
Now test a large block cache with low miss ratio, but some eviction is required:
1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23
1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43
1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4
1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56
1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59
1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8
1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89
1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45
1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98
1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91
1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26
1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63
610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137
610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996
610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934
610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5
610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402
610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742
610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062
610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453
610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457
610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426
610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273
610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812
The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.)
Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc.
233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371
233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293
233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844
233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461
233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227
233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738
233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688
233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402
233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84
233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785
233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94
233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016
89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086
89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984
89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441
89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754
89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812
89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418
89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422
89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293
89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43
89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824
89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32
89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223
^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.)
Even smaller cache size:
34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914
34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281
34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523
34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125
34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48
34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531
34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465
34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793
34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484
34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457
34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41
34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52
As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn:
13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328
13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633
13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684
13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383
13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492
13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863
13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121
13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758
13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539
13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098
13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77
13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27
gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention:
13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852
13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516
13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688
13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707
13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57
13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219
13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871
13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626
Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN
Reviewed By: anand1976
Differential Revision: D39368406
Pulled By: pdillinger
fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
|
|
|
Gang Liao
|
275cd80cdb |
Add a blob-specific cache priority (#10461)
Summary: RocksDB's `Cache` abstraction currently supports two priority levels for items: high (used for frequently accessed/highly valuable SST metablocks like index/filter blocks) and low (used for SST data blocks). Blobs are typically lower-value targets for caching than data blocks, since 1) with BlobDB, data blocks containing blob references conceptually form an index structure which has to be consulted before we can read the blob value, and 2) cached blobs represent only a single key-value, while cached data blocks generally contain multiple KVs. Since we would like to make it possible to use the same backing cache for the block cache and the blob cache, it would make sense to add a new, lower-than-low cache priority level (bottom level) for blobs so data blocks are prioritized over them. This task is a part of https://github.com/facebook/rocksdb/issues/10156 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10461 Reviewed By: siying Differential Revision: D38672823 Pulled By: ltamasi fbshipit-source-id: 90cf7362036563d79891f47be2cc24b827482743 |
|
|
Peter Dillinger
|
86a1e3e0e7 |
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 |
|
|
Peter Dillinger
|
65036e4217 |
Revert "Add a blob-specific cache priority (#10309)" (#10434)
Summary:
This reverts commit
|
|
|
Gang Liao
|
8d178090be |
Add a blob-specific cache priority (#10309)
Summary: RocksDB's `Cache` abstraction currently supports two priority levels for items: high (used for frequently accessed/highly valuable SST metablocks like index/filter blocks) and low (used for SST data blocks). Blobs are typically lower-value targets for caching than data blocks, since 1) with BlobDB, data blocks containing blob references conceptually form an index structure which has to be consulted before we can read the blob value, and 2) cached blobs represent only a single key-value, while cached data blocks generally contain multiple KVs. Since we would like to make it possible to use the same backing cache for the block cache and the blob cache, it would make sense to add a new, lower-than-low cache priority level (bottom level) for blobs so data blocks are prioritized over them. This task is a part of https://github.com/facebook/rocksdb/issues/10156 Pull Request resolved: https://github.com/facebook/rocksdb/pull/10309 Reviewed By: ltamasi Differential Revision: D38211655 Pulled By: gangliao fbshipit-source-id: 65ef33337db4d85277cc6f9782d67c421ad71dd5 |
|
|
Peter Dillinger
|
01a2e20299 |
Account for DB ID in stress testing block cache keys (#10388)
Summary: I recently discovered that block cache keys are slightly lower quality than previously thought, because my stress testing tool failed to simulate the effect of DB ID differences. This change updates the tool and gives us data to guide future developments. (No changes to production code here and now.) Nevertheless, the following promise still holds ``` // In fact, if our SST files are all < 4TB (see // BlockBasedTable::kMaxFileSizeStandardEncoding), then SST files generated // in a single process are guaranteed to have unique cache keys, unless/until // number session ids * max file number = 2**86 ... ``` because although different DB IDs could cause collision in file number and offset data, that would have to be using the same DB session (lower) to cause a block cache key collision, which is not possible in the same process. (A session is associated with only one DB ID.) This change fixes cache_bench -stress_cache_key to set and reset DB IDs in a parameterized way to evaluate the effect. Previous results assumed to be representative (using -sck_keep_bits=43): ``` 15 collisions after 15 x 90 days, est 90 days between (1.03763e+20 corrected) ``` or expected collision on a single machine every 104 billion billion days (see "corrected" value). After accounting for DB IDs, test never really changing, intermediate, and very frequently changing (using default -sck_db_count=100): ``` -sck_newdb_nreopen=1000000000: 15 collisions after 2 x 90 days, est 12 days between (1.38351e+19 corrected) -sck_newdb_nreopen=10000: 17 collisions after 2 x 90 days, est 10.5882 days between (1.22074e+19 corrected) -sck_newdb_nreopen=100: 19 collisions after 2 x 90 days, est 9.47368 days between (1.09224e+19 corrected) ``` or roughly 10x more often than previously thought (still extremely if not impossibly rare), and better than random base cache keys (with -sck_randomize), though < 10x better than random: ``` 31 collisions after 1 x 90 days, est 2.90323 days between (3.34719e+18 corrected) ``` If we simply fixed this by ignoring DB ID for cache keys, we would potentially have a shortage of entropy for some cases, such as small file numbers and offsets (e.g. many short-lived processes each using SstFileWriter to create a small file), because existing DB session IDs only provide ~103 bits of entropy. We could upgrade the entropy in DB session IDs to accommodate, but it's not known what all would be affected by changing from 20 digit session IDs to something larger. Instead, my plan is to 1) Move to block cache keys derived from SST unique IDs (so that we can derive block cache keys from manifest data without reading file on storage), and show no significant regression in expected collision rate. 2) Generate better SST unique IDs in format_version=6 (https://github.com/facebook/rocksdb/issues/9058), which should have ~100x lower expected/predicted collision rate based on simulations with this stress test: ``` ./cache_bench -stress_cache_key -sck_keep_bits=39 -sck_newdb_nreopen=100 -sck_footer_unique_id ... 15 collisions after 19 x 90 days, est 114 days between (2.10293e+21 corrected) ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/10388 Test Plan: no production changes Reviewed By: jay-zhuang Differential Revision: D37986714 Pulled By: pdillinger fbshipit-source-id: e759b2469e3365cb01c6661a69e0ab849ef4c3df |
|
|
Guido Tagliavini Ponce
|
a543773bbc |
Add lean option to cache_bench (#10363)
Summary: Sometimes we may not want to include extra computation in our cache_bench experiments. Here we add a flag to avoid any extra work. We also moved the timer start after the key generation. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10363 Test Plan: Run cache_bench with and without the new flag and check that the appropriate code is being executed. Reviewed By: pdillinger Differential Revision: D37870416 Pulled By: guidotag fbshipit-source-id: f853207b6643b9328e774251c3f679b1fd78a11a |
|
|
Guido Tagliavini Ponce
|
9645e66fc9 |
Temporarily return a LRUCache from NewClockCache (#10351)
Summary: ClockCache is still in experimental stage, and currently fails some pre-release fbcode tests. See https://www.internalfb.com/diff/D37772011. API calls to construct ClockCache are done via the function NewClockCache. For now, NewClockCache calls will return an LRUCache (with appropriate arguments), which is stable. The idea that NewClockCache returns nullptr was also floated, but this would be interpreted as unsupported cache, and a default LRUCache would be constructed instead, potentially causing a performance regression that is harder to identify. A new version of the NewClockCache function was created for our internal tests. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10351 Test Plan: ``make -j24 check`` and re-run the pre-release tests. Reviewed By: pdillinger Differential Revision: D37802685 Pulled By: guidotag fbshipit-source-id: 0a8d10612ff21e576f7360cb13e20bc36e244972 |
|
|
Guido Tagliavini Ponce
|
57a0e2f304 |
Clock cache (#10273)
Summary: This is the initial step in the development of a lock-free clock cache. This PR includes the base hash table design (which we mostly ported over from FastLRUCache) and the clock eviction algorithm. Importantly, it's still _not_ lock-free---all operations use a shard lock. Besides the locking, there are other features left as future work: - Remove keys from the handles. Instead, use 128-bit bijective hashes of them for handle comparisons, probing (we need two 32-bit hashes of the key for double hashing) and sharding (we need one 6-bit hash). - Remove the clock_usage_ field, which is updated on every lookup. Even if it were atomically updated, it could cause memory invalidations across cores. - Middle insertions into the clock list. - A test that exercises the clock eviction policy. - Update the Java API of ClockCache and Java calls to C++. Along the way, we improved the code and comments quality of FastLRUCache. These changes are relatively minor. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10273 Test Plan: ``make -j24 check`` Reviewed By: pdillinger Differential Revision: D37522461 Pulled By: guidotag fbshipit-source-id: 3d70b737dbb70dcf662f00cef8c609750f083943 |
|
|
Guido Tagliavini Ponce
|
b52620ab0e |
Fix key size in cache_bench (#10234)
Summary: cache_bench wasn't generating 16B keys, which are necessary for FastLRUCache. Also: - Added asserts in cache_bench, which is assuming that inserts never fail. When they fail (for example, if we used keys of the wrong size), memory allocated to the values will becomes leaked, and eventually the program crashes. - Move kCacheKeySize to the right spot. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10234 Test Plan: ``make -j24 check``. Also, run cache_bench with FastLRUCache and check that memory usage doesn't blow up: ``./cache_bench -cache_type=fast_lru_cache -num_shard_bits=6 -skewed=true \ -lookup_insert_percent=100 -lookup_percent=0 -insert_percent=0 -erase_percent=0 \ -populate_cache=true -cache_size=1073741824 -ops_per_thread=10000000 \ -value_bytes=8192 -resident_ratio=1 -threads=16`` Reviewed By: pdillinger Differential Revision: D37382949 Pulled By: guidotag fbshipit-source-id: b697a942ebb215de5d341f98dc8566763436ba9b |
|
|
Peter Dillinger
|
1aac814578 |
Use optimized folly DistributedMutex in LRUCache when available (#10179)
Summary: folly DistributedMutex is faster than standard mutexes though imposes some static obligations on usage. See https://github.com/facebook/folly/blob/main/folly/synchronization/DistributedMutex.h for details. Here we use this alternative for our Cache implementations (especially LRUCache) for better locking performance, when RocksDB is compiled with folly. Also added information about which distributed mutex implementation is being used to cache_bench output and to DB LOG. Intended follow-up: * Use DMutex in more places, perhaps improving API to support non-scoped locking * Fix linking with fbcode compiler (needs ROCKSDB_NO_FBCODE=1 currently) Credit: Thanks Siying for reminding me about this line of work that was previously left unfinished. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10179 Test Plan: for correctness, existing tests. CircleCI config updated. Also Meta-internal buck build updated. For performance, ran simultaneous before & after cache_bench. Out of three comparison runs, the middle improvement to ops/sec was +21%: Baseline: USE_CLANG=1 DEBUG_LEVEL=0 make -j24 cache_bench (fbcode compiler) ``` Complete in 20.201 s; Rough parallel ops/sec = 1584062 Thread ops/sec = 107176 Operation latency (ns): Count: 32000000 Average: 9257.9421 StdDev: 122412.04 Min: 134 Median: 3623.0493 Max: 56918500 Percentiles: P50: 3623.05 P75: 10288.02 P99: 30219.35 P99.9: 683522.04 P99.99: 7302791.63 ``` New: (add USE_FOLLY=1) ``` Complete in 16.674 s; Rough parallel ops/sec = 1919135 (+21%) Thread ops/sec = 135487 Operation latency (ns): Count: 32000000 Average: 7304.9294 StdDev: 108530.28 Min: 132 Median: 3777.6012 Max: 91030902 Percentiles: P50: 3777.60 P75: 10169.89 P99: 24504.51 P99.9: 59721.59 P99.99: 1861151.83 ``` Reviewed By: anand1976 Differential Revision: D37182983 Pulled By: pdillinger fbshipit-source-id: a17eb05f25b832b6a2c1356f5c657e831a5af8d1 |
|
|
Guido Tagliavini Ponce
|
f105e1a501 |
Make the per-shard hash table fixed-size. (#10154)
Summary: We make the size of the per-shard hash table fixed. The base level of the hash table is now preallocated with the required capacity. The user must provide an estimate of the size of the values. Notice that even though the base level becomes fixed, the chains are still dynamic. Overall, the shard capacity mechanisms haven't changed, so we don't need to test this. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10154 Test Plan: `make -j24 check` Reviewed By: pdillinger Differential Revision: D37124451 Pulled By: guidotag fbshipit-source-id: cba6ac76052fe0ec60b8ff4211b3de7650e80d0c |
|
|
Guido Tagliavini Ponce
|
eb99e08076 |
Add support for FastLRUCache in cache_bench (#10095)
Summary: cache_bench can now run with FastLRUCache. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10095 Test Plan: - Temporarily add an ``assert(false)`` in the execution path that sets up the FastLRUCache. Run ``make -j24 cache_bench``. Then test the appropriate code is used by running ``./cache_bench -cache_type=fast_lru_cache`` and checking that the assert is called. Repeat for LRUCache. - Verify that FastLRUCache (currently a clone of LRUCache) has similar latency distribution than LRUCache, by comparing the outputs of ``./cache_bench -cache_type=fast_lru_cache`` and ``./cache_bench -cache_type=lru_cache``. Reviewed By: pdillinger Differential Revision: D36875834 Pulled By: guidotag fbshipit-source-id: eb2ad0bb32c2717a258a6ac66ed736e06f826cd8 |
|
|
Andrew Kryczka
|
54fb2a8975 |
Change type of cache buffer passed to `Cache::CreateCallback()` to `const void*` (#9595)
Summary: Pull Request resolved: https://github.com/facebook/rocksdb/pull/9595 Reviewed By: riversand963 Differential Revision: D34329906 Pulled By: ajkr fbshipit-source-id: 508601856fa9bee4d40f4a68d14d333ef2143d40 |
|
|
Peter Dillinger
|
afc280fdfd |
Enhance new cache key testing & comments (#9329)
Summary: Follow-up to https://github.com/facebook/rocksdb/issues/9126 Added new unit tests to validate some of the claims of guaranteed uniqueness within certain large bounds. Also cleaned up the cache_bench -stress-cache-key tool with better comments and description. Pull Request resolved: https://github.com/facebook/rocksdb/pull/9329 Test Plan: no changes to production code Reviewed By: mrambacher Differential Revision: D33269328 Pulled By: pdillinger fbshipit-source-id: 3a2b684a6b2b15f79dc872e563e3d16563be26de |
|
|
Peter Dillinger
|
0050a73a4f |
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 |
|
|
Yanqin Jin
|
42fef0224f |
Fix build for msvc (#9230)
Summary: Test plan With Visual Studio 2017. ``` cd rocksdb mkdir build && cd build cmake -G "Visual Studio 15 Win64" -DWITH_GFLAGS=1 .. MSBuild rocksdb.sln /m /TARGET:cache_bench /TARGET:db_bench /TARGET:db_stress ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/9230 Reviewed By: akankshamahajan15 Differential Revision: D32705095 Pulled By: riversand963 fbshipit-source-id: 101e3533f5178b24c0535ddc47a39347ccfcf92c |
|
|
mrambacher
|
570248aeff |
Make SecondaryCache Customizable (#8480)
Summary: Pull Request resolved: https://github.com/facebook/rocksdb/pull/8480 Reviewed By: zhichao-cao Differential Revision: D29528740 Pulled By: mrambacher fbshipit-source-id: fd0f70d15f66611c8498257a9973f7e98ca13839 |
|
|
anand76
|
13232e11d4 |
Allow cache_bench/db_bench to use a custom secondary cache (#8312)
Summary: This PR adds a ```-secondary_cache_uri``` option to the cache_bench and db_bench tools to allow the user to specify a custom secondary cache URI. The object registry is used to create an instance of the ```SecondaryCache``` object of the type specified in the URI. The main cache_bench code is packaged into a separate library, similar to db_bench. An example invocation of db_bench with a secondary cache URI - ```db_bench --env_uri=ws://ws.flash_sandbox.vll1_2/ -db=anand/nvm_cache_2 -use_existing_db=true -benchmarks=readrandom -num=30000000 -key_size=32 -value_size=256 -use_direct_reads=true -cache_size=67108864 -cache_index_and_filter_blocks=true -secondary_cache_uri='cachelibwrapper://filename=/home/anand76/nvm_cache/cache_file;size=2147483648;regionSize=16777216;admPolicy=random;admProbability=1.0;volatileSize=8388608;bktPower=20;lockPower=12' -partition_index_and_filters=true -duration=1800``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/8312 Reviewed By: zhichao-cao Differential Revision: D28544325 Pulled By: anand1976 fbshipit-source-id: 8f209b9af900c459dc42daa7a610d5f00176eeed |