Summary:
Added some util function APIs to facilitate using the U64Ts.
The U64Ts format for encoding a timestamp is not entirely RocksDB internal. When users are using the user-defined timestamp feature from the transaction layer, its public APIs including `SetCommitTimestamp`, `GetCommitTimestamp`, `SetReadTimestampForValidation` are taking and returning timestamps as uint64_t. But if users want to use the APIs from the DB layer, including populating `ReadOptions.timestamp`, interpreting `Iterator::timestamp()`, these APIs are using and returning U64Ts timestamps as an encoded Slice. So these util functions are added to facilitate the usage.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11888
Reviewed By: ltamasi
Differential Revision: D49620709
Pulled By: jowlyzhang
fbshipit-source-id: ace8d782ee7c3372cf410abf761320d373e495e1
Summary:
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11870
Having a large number of merge operands applied at query time can have a significant effect on performance; therefore, applications might want limit the number of deltas for any given key. However, there is currently no way to establish the number of operands for certain types of queries. The ticker `READ_NUM_MERGE_OPERANDS` only provides aggregate (not per-read) information. The `PerfContext` counters `internal_merge_count` and `internal_merge_point_lookup_count` can be used to get this information on a per-query basis for iterators and single point lookups; however, there is no per-key breakdown for `MultiGet` type APIs. The patch addresses this issue by introducing a special kind of OK status which signals that an application-defined threshold on the number of merge operands has been exceeded for a given key. The threshold can be specified on a per-query basis using a new field in `ReadOptions`.
Reviewed By: jaykorean
Differential Revision: D49522786
fbshipit-source-id: 4265b3848d1be5ff313a3e8fb604ddf56411dd2c
Summary:
An internal user wants to be able to dynamically switch between Bloom and Ribbon filters, without a custom FilterPolicy. Making `filter_policy` mutable would actually make issue https://github.com/facebook/rocksdb/issues/10079 worse, because it would be a race on a pointer field, not just on scalars.
As a reasonable compromise until that is fixed, I am enabling dynamic control over Bloom vs. Ribbon choice by making
RibbonFilterPolicy::bloom_before_level mutable, and doing that safely by using an atomic.
I've also slightly tweaked the interpretation of that field so that setting it to INT_MAX really means "always Bloom."
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11838
Test Plan: unit tests added/extended. crash test updated for SetOptions call and tested under TSAN with amplified probability (lower set_options_one_in).
Reviewed By: ajkr
Differential Revision: D49296284
Pulled By: pdillinger
fbshipit-source-id: e4251c077510df9a9c719876f482448c0d15402a
Summary:
`GetEntity` API support for ReadOnly DB and Secondary DB.
- Introduced `GetImpl()` with `GetImplOptions` in `db_impl_readonly` and refactored current `Get()` logic into `GetImpl()` so that look up logic can be reused for `GetEntity()` (Following the same pattern as `DBImpl::Get()` and `DBImpl::GetEntity()`)
- Introduced `GetImpl()` with `GetImplOptions` in `db_impl_secondary` and refactored current `GetImpl()` logic. This is to make `DBImplSecondary::Get/GetEntity` consistent with `DBImpl::Get/GetEntity` and `DBImplReadOnly::Get/GetEntity`
- `GetImpl()` in `db_impl` is now virtual. both `db_impl_readonly` and `db_impl_secondary`'s `Get()` override are no longer needed since all three dbs now have the same `Get()` which calls `GetImpl()` internally.
- `GetImpl()` in `DBImplReadOnly` and `DBImplSecondary` now pass in `columns` instead of `nullptr` in lookup functions like `memtable->get()`
- Introduced `GetEntity()` API in `DBImplReadOnly` and `DBImplSecondary` which simply calls `GetImpl()` with `columns` set in `GetImplOptions`.
- Introduced `Env::IOActivity::kGetEntity` and set read_options.io_activity to `Env::IOActivity::kGetEntity` for `GetEntity()` operations (in db_impl)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11799
Test Plan:
**Unit Tests**
- Added verification in `DBWideBasicTest::PutEntity` by Reopening DB as ReadOnly with the same setup.
- Added verification in `DBSecondaryTest::ReopenAsSecondary` by calling `PutEntity()` and `GetEntity()` on top of existing `Put()` and `Get()`
- `make -j64 check`
**Crash Tests**
- `python3 tools/db_crashtest.py blackbox --max_key=25000000 --write_buffer_size=4194304 --max_bytes_for_level_base=2097152 --target_file_size_base=2097152 --periodic_compaction_seconds=0 --use_put_entity_one_in=10 --use_get_entity=1 --duration=60 --inter
val=10`
- `python3 tools/db_crashtest.py blackbox --simple --max_key=25000000 --write_buffer_size=4194304 --max_bytes_for_level_base=2097152 --target_file_size_base=2097152 --periodic_compaction_seconds=0 --use_put_entity_one_in=10 --use_get_entity=1 `
- `python3 tools/db_crashtest.py blackbox --cf_consistency --max_key=25000000 --write_buffer_size=4194304 --max_bytes_for_level_base=2097152 --target_file_size_base=2097152 --periodic_compaction_seconds=0 --use_put_entity_one_in=10 --use_get_entity=1 --duration=60 --inter
val=10`
Reviewed By: ltamasi
Differential Revision: D49037040
Pulled By: jaykorean
fbshipit-source-id: a0648253ded6e91af7953de364ed3c6bf163626b
Summary:
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11807
For now, RocksDB has limited support for using merge with wide columns: when a bunch of merge operands have to be applied to a wide-column base value, RocksDB currently passes only the value of the default column to the application's `MergeOperator`, which means there is no way to update any other columns during a merge. As a first step in making this more general, the patch adds a new API `FullMergeV3` to `MergeOperator`.
`FullMergeV3`'s interface enables applications to receive a plain, wide-column, or non-existent base value as merge input, and to produce a new plain value, a new wide-column value, or an existing operand as merge output. Note that there are no limitations on the column names and values if the merge result is a wide-column entity. Also, the interface is general in the sense that it makes it possible e.g. for a merge that takes a plain base value and some deltas to produce a wide-column entity as a result.
For backward compatibility, the default implementation of `FullMergeV3` falls back to `FullMergeV2` and implements the current logic where merge operands are applied to the default column of the base entity and any other columns are unchanged. (Note that with `FullMergeV3` in the `MergeOperator` interface, this behavior will become customizable.)
This patch just introduces the new API and the default backward compatible implementation. I plan to integrate `FullMergeV3` into the query and compaction logic in subsequent diffs.
Reviewed By: jaykorean
Differential Revision: D49117253
fbshipit-source-id: 109e016f25cd130fc504790818d927bae7fec6bd
Summary:
There was a `#include "port/lang.h"` situated inside an `extern "C" {` which just started causing the header to be unusuable in some contexts. This was a regression on the CircleCI job build-linux-unity-and-headers in https://github.com/facebook/rocksdb/issues/11792
The include, and another like it, now appears obsolete so removed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11797
Test Plan: local `make check-headers` and `make`, CI
Reviewed By: jaykorean
Differential Revision: D48976826
Pulled By: pdillinger
fbshipit-source-id: 131d66969e045f2ded0f8936924ee30c9ef2655a
Summary:
For a SST file that uses user-defined timestamp aware comparators, if a lower or upper bound is set, sst_dump tool doesn't handle it well. This PR adds support for that. While working on this `MaybeAddTimestampsToRange` is moved to the udt_util.h file to be shared.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11757
Test Plan:
make all check
for changes in db_impl.cc and db_impl_compaction_flush.cc
for changes in sst_file_dumper.cc, I manually tested this change handles specifying bounds for UDT use cases. It probably should have a unit test file eventually.
Reviewed By: ltamasi
Differential Revision: D48668048
Pulled By: jowlyzhang
fbshipit-source-id: 1560465f40e44668d6d82a7439fe9012be0e74a8
Summary:
This PR adds a missing piece for the UDT in memtable only feature, which is to automatically increase `full_history_ts_low` when flush happens during recovery.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11774
Test Plan:
Added unit test
make all check
Reviewed By: ltamasi
Differential Revision: D48799109
Pulled By: jowlyzhang
fbshipit-source-id: fd681ed66d9d40904ca2c919b2618eb692686035
Summary:
Having a synthetic implementation of `TimedWait()` in `SystemClock` will allow us to add `SyncPoint`s while mutex is released, which was previously impossible since the lock was released and reacquired all within `pthread_cond_timedwait()`. Additionally, integrating `TimedWait()` with `MockSystemClock` allows us to cleanup some workarounds in the test code. In this PR I only cleaned up the `GenericRateLimiter` test workaround.
This is related to the intended follow-up mentioned in https://github.com/facebook/rocksdb/issues/7101's description. There are a couple differences:
(1) This PR does not include removing the particular workaround that initially motivated it. Actually, the `Timer` class uses `InstrumentedCondVar`, so the interface introduced here is inadequate to remove that workaround. On the bright side, the interface introduced in this PR can be changed as needed since it can neither be used nor extended externally, due to using forward-declared `port::CondVar*` in the interface.
(2) This PR only makes the change in `SystemClock` not `Env`. Older revisions of this PR included `Env::TimedWait()` and `SpecialEnv::TimedWait()`; however, since they were unused it probably makes sense to defer adding them until when they are needed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11753
Reviewed By: pdillinger
Differential Revision: D48654995
Pulled By: ajkr
fbshipit-source-id: 15e19f2454b64d4ec7f50e328691c66ca9911122
Summary:
* JemallocNodumpAllocator was passing a size_t to FastRange32, which could cause compilation errors or warnings (seen with clang)
* Fixed the order of arguments to match what would be used with modulo operator (%), for clarity.
Fixes https://github.com/facebook/rocksdb/issues/11006
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11707
Test Plan: no functional change, existing tests
Reviewed By: ajkr
Differential Revision: D48435149
Pulled By: pdillinger
fbshipit-source-id: e6e8b107ded4eceda37db20df59985c846a2546b
Summary:
As titled. User-defined timestamp feature users sometimes directly call the user comparator to do validation on their side too. Having access to the root comparator can help make their code consistent for when UDT is enabled and disabled.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11704
Reviewed By: ltamasi
Differential Revision: D48355090
Pulled By: jowlyzhang
fbshipit-source-id: 26bc73543bfb379ef548d1361803d6f8c308cef6
Summary:
**Context/Summary:**
- Similar to https://github.com/facebook/rocksdb/pull/11288 but for user read such as `Get(), MultiGet(), DBIterator::XXX(), Verify(File)Checksum()`.
- For this, I refactored some user-facing `MultiGet` calls in `TransactionBase` and various types of `DB` so that it does not call a user-facing `Get()` but `GetImpl()` for passing the `ReadOptions::io_activity` check (see PR conversation)
- New user read stats breakdown are guarded by `kExceptDetailedTimers` since measurement shows they have 4-5% regression to the upstream/main.
- Misc
- More refactoring: with https://github.com/facebook/rocksdb/pull/11288, we complete passing `ReadOptions/IOOptions` to FS level. So we can now replace the previously [added](https://github.com/facebook/rocksdb/pull/9424) `rate_limiter_priority` parameter in `RandomAccessFileReader`'s `Read/MultiRead/Prefetch()` with `IOOptions::rate_limiter_priority`
- Also, `ReadAsync()` call time is measured in `SST_READ_MICRO` now
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11444
Test Plan:
- CI fake db crash/stress test
- Microbenchmarking
**Build** `make clean && ROCKSDB_NO_FBCODE=1 DEBUG_LEVEL=0 make -jN db_basic_bench`
- google benchmark version: 604f6fd3f4
- db_basic_bench_base: upstream
- db_basic_bench_pr: db_basic_bench_base + this PR
- asyncread_db_basic_bench_base: upstream + [db basic bench patch for IteratorNext](https://github.com/facebook/rocksdb/compare/main...hx235:rocksdb:micro_bench_async_read)
- asyncread_db_basic_bench_pr: asyncread_db_basic_bench_base + this PR
**Test**
Get
```
TEST_TMPDIR=/dev/shm ./db_basic_bench_{null_stat|base|pr} --benchmark_filter=DBGet/comp_style:0/max_data:134217728/per_key_size:256/enable_statistics:1/negative_query:0/enable_filter:0/mmap:1/threads:1 --benchmark_repetitions=1000
```
Result
```
Coming soon
```
AsyncRead
```
TEST_TMPDIR=/dev/shm ./asyncread_db_basic_bench_{base|pr} --benchmark_filter=IteratorNext/comp_style:0/max_data:134217728/per_key_size:256/enable_statistics:1/async_io:1/include_detailed_timers:0 --benchmark_repetitions=1000 > syncread_db_basic_bench_{base|pr}.out
```
Result
```
Base:
1956,1956,1968,1977,1979,1986,1988,1988,1988,1990,1991,1991,1993,1993,1993,1993,1994,1996,1997,1997,1997,1998,1999,2001,2001,2002,2004,2007,2007,2008,
PR (2.3% regression, due to measuring `SST_READ_MICRO` that wasn't measured before):
1993,2014,2016,2022,2024,2027,2027,2028,2028,2030,2031,2031,2032,2032,2038,2039,2042,2044,2044,2047,2047,2047,2048,2049,2050,2052,2052,2052,2053,2053,
```
Reviewed By: ajkr
Differential Revision: D45918925
Pulled By: hx235
fbshipit-source-id: 58a54560d9ebeb3a59b6d807639692614dad058a
Summary:
BottomNBits() - there is a single fast instruction for this on x86 since BMI2, but testing with godbolt indicates you need at least GCC 10 for the compiler to choose that instruction from the obvious C++ code. https://godbolt.org/z/5a7Ysd41h
BitwiseAnd() - this is a convenience function that works around the language flaw that the type of the result of x & y is the larger of the two input types, when it should be the smaller. This can save some ugly static_cast.
I expect to use both of these in coming HyperClockCache developments, and have applied them in a couple of places in existing code.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11660
Test Plan: unit tests added
Reviewed By: jowlyzhang
Differential Revision: D47935531
Pulled By: pdillinger
fbshipit-source-id: d148c43a1e51df4a1c549b93aaf2725a3f8d3bd6
Summary:
... to improve data integrity validation during compaction.
A new option `compaction_verify_record_count` is introduced for this verification and is enabled by default. One exception when the verification is not done is when a compaction filter returns kRemoveAndSkipUntil which can cause CompactionIterator to seek until some key and hence not able to keep track of the number of keys processed.
For expected number of input keys, we sum over the number of total keys - number of range tombstones across compaction input files (`CompactionJob::UpdateCompactionStats()`). Table properties are consulted if `FileMetaData` is not initialized for some input file. Since table properties for all input files were also constructed during `DBImpl::NotifyOnCompactionBegin()`, `Compaction::GetTableProperties()` is introduced to reduce duplicated code.
For actual number of keys processed, each subcompaction will record its number of keys processed to `sub_compact->compaction_job_stats.num_input_records` and aggregated when all subcompactions finish (`CompactionJob::AggregateCompactionStats()`). In the case when some subcompaction encountered kRemoveAndSkipUntil from compaction filter and does not have accurate count, it propagates this information through `sub_compact->compaction_job_stats.has_num_input_records`.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11571
Test Plan:
* Add a new unit test `DBCompactionTest.VerifyRecordCount` for the corruption case.
* All other unit tests for non-corrupted case.
* Ran crash test for a few hours: `python3 ./tools/db_crashtest.py whitebox --simple`
Reviewed By: ajkr
Differential Revision: D47131965
Pulled By: cbi42
fbshipit-source-id: cc8e94565dd526c4347e9d3843ecf32f6727af92
Summary:
Add a built-in comparator that supports uint64_t style user-defined timestamps for ReverseBytewiseComparator.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11647
Test Plan:
Added a test wrapper for retrieving this comparator from registry and used it in this test:
`./udt_util_test`
Reviewed By: ltamasi
Differential Revision: D47848303
Pulled By: jowlyzhang
fbshipit-source-id: 5af5534a8c2d9195997d0308c8e194c1c797548c
Summary:
Fix use_after_free bug in async_io MultiReads when underlying FS enabled kFSBuffer. kFSBuffer is when underlying FS pass their own buffer instead of using RocksDB scratch in FSReadRequest
Since it's an experimental feature, added a hack for now to fix the bug.
Planning to make public API change to remove const from the callback as it doesn't make sense to use const.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11645
Test Plan: tested locally
Reviewed By: ltamasi
Differential Revision: D47819907
Pulled By: akankshamahajan15
fbshipit-source-id: 1faf5ef795bf27e2b3a60960374d91274931df8d
Summary:
Add support to allow enabling / disabling user-defined timestamps feature for an existing column family in combination with the in-Memtable only feature.
To do this, this PR includes:
1) Log the `persist_user_defined_timestamps` option per column family in Manifest to facilitate detecting an attempt to enable / disable UDT. This entry is enforced to be logged in the same VersionEdit as the user comparator name entry.
2) User-defined timestamps related options are validated when re-opening a column family, including user comparator name and the `persist_user_defined_timestamps` flag. These type of settings and settings change are considered valid:
a) no user comparator change and no effective `persist_user_defined_timestamp` flag change.
b) switch user comparator to enable UDT provided the immediately effective `persist_user_defined_timestamps` flag
is false.
c) switch user comparator to disable UDT provided that the before-change `persist_user_defined_timestamps` is
already false.
3) when an attempt to enable UDT is detected, we mark all its existing SST files as "having no UDT" by marking its `FileMetaData.user_defined_timestamps_persisted` flag to false and handle their file boundaries `FileMetaData.smallest`, `FileMetaData.largest` by padding a min timestamp.
4) while enabling / disabling UDT feature, timestamp size inconsistency in existing WAL logs are handled to make it compatible with the running user comparator.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11623
Test Plan:
```
make all check
./db_with_timestamp_basic_test --gtest-filter="*EnableDisableUDT*"
./db_wal_test --gtest_filter="*EnableDisableUDT*"
```
Reviewed By: ltamasi
Differential Revision: D47636862
Pulled By: jowlyzhang
fbshipit-source-id: dcd19f67292da3c3cc9584c09ad00331c9ab9322
Summary:
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11624
<queue> must be included to use std::queue.
Reviewed By: pdillinger
Differential Revision: D47562433
fbshipit-source-id: 7c5b19fd9e411694c782dfc0dff0231d4f92ef24
Summary:
I made some changes to add OpenBSD support.
Second time doing something like this, so I apologize in advance if I'm doing something wrong (had some minor hiccups with how github worked).
Fixes https://github.com/facebook/rocksdb/issues/11220
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11255
Reviewed By: akankshamahajan15
Differential Revision: D46361706
Pulled By: ajkr
fbshipit-source-id: 90922fa30197fe6d6f3c0e3ecca2dbb92c337277
Summary:
Switch from std::unordered_map to RocksDB UnorderedMap for all the places that logging user-defined timestamp size in WAL used.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11507
Test Plan:
```
make all check
```
Reviewed By: ltamasi
Differential Revision: D46448975
Pulled By: jowlyzhang
fbshipit-source-id: bdb4d56a723b697a33daaf0f856a61d49a367a99
Summary:
Start logging the timestamp size record in WAL and use the record during recovery. Currently, user comparator cannot be different from what was used to create a column family, so the timestamp size record is just used to confirm it's consistent with the timestamp size the running user comparator indicates.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11471
Test Plan:
```
make all check
./db_secondary_test
./db_wal_test --gtest_filter="*WithTimestamp*"
./repair_test --gtest_filter="*WithTimestamp*"
```
Reviewed By: ltamasi
Differential Revision: D46236769
Pulled By: jowlyzhang
fbshipit-source-id: f6c60b5c8defdb05021c63df302ccc0be1275ad0
Summary:
Currently it's easy to use a ton of memory with many small OptimisticTransactionDB instances, because each one by default allocates a million mutexes (40 bytes each on my compiler) for validating transactions. It even puts a lot of pressure on the allocator by allocating each one individually!
In this change:
* Create a new object and option that enables sharing these buckets of mutexes between instances. This is generally good for load balancing potential contention as various DBs become hotter or colder with txn writes. About the only cases where this sharing wouldn't make sense (e.g. each DB usually written by one thread) are cases that would be better off with OccValidationPolicy::kValidateSerial which doesn't use the buckets anyway.
* Allocate the mutexes in a contiguous array, for efficiency
* Add an option to ensure the mutexes are cache-aligned. In several other places we use cache-aligned mutexes but OptimisticTransactionDB historically does not. It should be a space-time trade-off the user can choose.
* Provide some visibility into the memory used by the mutex buckets with an ApproximateMemoryUsage() function (also used in unit testing)
* Share code with other users of "striped" mutexes, appropriate refactoring for customization & efficiency (e.g. using FastRange instead of modulus)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11439
Test Plan: unit tests added. Ran sized-up versions of stress test in unit test, including a before-and-after performance test showing no consistent difference. (NOTE: OptimisticTransactionDB not currently covered by db_stress!)
Reviewed By: ltamasi
Differential Revision: D45796393
Pulled By: pdillinger
fbshipit-source-id: ae2b3a26ad91ceeec15debcdc63ff48df6736a54
Summary:
Add a util method `HandleWriteBatchTimestampSizeDifference` to handle a `WriteBatch` read from WAL log when user-defined timestamp size record is written and read. Two check modes are added: `kVerifyConsistency` that just verifies the recorded timestamp size are consistent with the running ones. This mode is to be used by `db_impl_secondary` for opening a DB as secondary instance. It will also be used by `db_impl_open` before the user comparator switch support is added to make a column switch between enabling/disable UDT feature. The other mode `kReconcileInconsistency` will be used by `db_impl_open` later when user comparator can be changed.
Another change is to extract a method `CollectColumnFamilyIdsFromWriteBatch` in db_secondary_impl.h into its standalone util file so it can be shared.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11451
Test Plan:
```
make check
./udt_util_test
```
Reviewed By: ltamasi
Differential Revision: D45894386
Pulled By: jowlyzhang
fbshipit-source-id: b96790777f154cddab6d45d9ba2e5d20ebc6fe9d
Summary:
We want to know more about opportunities for better range filters, and the effectiveness of our own range filters. Currently the stats are very limited, essentially logging just hits and misses against prefix filters for range scans in BLOOM_FILTER_PREFIX_* without tracking the false positive rate. Perhaps confusingly, when prefix filters are used for point queries, the stats are currently going into the non-PREFIX tickers.
This change does several things:
* Introduce new stat tickers for seeks and related filtering, \*LEVEL_SEEK\*
* Most importantly, allows us to see opportunities for range filtering. Specifically, we can count how many times a seek in an SST file accesses at least one data block, and how many times at least one value() is then accessed. If a data block was accessed but no value(), we can generally assume that the key(s) seen was(were) not of interest so could have been filtered with the right kind of filter, avoiding the data block access.
* We can get the same level of detail when a filter (for now, prefix Bloom/ribbon) is used, or not. Specifically, we can infer a false positive rate for prefix filters (not available before) from the seek "false positive" rate: when a data block is accessed but no value() is called. (There can be other explanations for a seek false positive, but in typical iterator usage it would indicate a filter false positive.)
* For efficiency, I wanted to avoid making additional calls to the prefix extractor (or key comparisons, etc.), which would be required if we wanted to more precisely detect filter false positives. I believe that instrumenting value() is the best balance of efficiency vs. accurately measuring what we are often interested in.
* The stats are divided between last level and non-last levels, to help understand potential tiered storage use cases.
* The old BLOOM_FILTER_PREFIX_* stats have a different meaning: no longer referring to iterators but to point queries using prefix filters. BLOOM_FILTER_PREFIX_TRUE_POSITIVE is added for computing the prefix false positive rate on point queries, which can be due to filter false positives as well as different keys with the same prefix.
* Similarly, the non-PREFIX BLOOM_FILTER stats are now for whole key filtering only.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11460
Test Plan:
unit tests updated, including updating many to pop the stat value since last read to improve test
readability and maintainability.
Performance test shows a consistent small improvement with these changes, both with clang and with gcc. CPU profile indicates that RecordTick is using less CPU, and this makes sense at least for a high filter miss rate. Before, we were recording two ticks per filter miss in iterators (CHECKED & USEFUL) and now recording just one (FILTERED).
Create DB with
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=10000000 -disable_wal=1 -write_buffer_size=30000000 -bloom_bits=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0 -prefix_size=8
```
And run simultaneous before&after with
```
TEST_TMPDIR=/dev/shm ./db_bench -readonly -benchmarks=seekrandom[-X1000] -num=10000000 -bloom_bits=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0 -prefix_size=8 -seek_nexts=1 -duration=20 -seed=43 -threads=8 -cache_size=1000000000 -statistics
```
Before: seekrandom [AVG 275 runs] : 189680 (± 222) ops/sec; 18.4 (± 0.0) MB/sec
After: seekrandom [AVG 275 runs] : 197110 (± 208) ops/sec; 19.1 (± 0.0) MB/sec
Reviewed By: ajkr
Differential Revision: D46029177
Pulled By: pdillinger
fbshipit-source-id: cdace79a2ea548d46c5900b068c5b7c3a02e5822
Summary:
Apply a small (and automatic) set of IntelliJ Java inspections/repairs to the Java interface to RocksDB Java and its tests.
Partly enabled by the fact that we now (from RocksDB7) require java 8.
Explicit <p> in empty lines in javadoc comments.
Parameters and variables made final where possible.
Anonymous subclasses converted lambdas.
Some tests which previously used other assertion models were converted to assertj, e.g. (assertThat(actual).isEqualTo(expected)
In a very few cases tests were found to be inoperative or broken, and were repaired. No problems with actual RocksDB behaviour were observed.
This PR is intended to replace https://github.com/facebook/rocksdb/pull/9618 - that PR was not merged, and attempts to rebase it have yielded a questionable looking diff, so we choose to go back to square 1 here, and implement a conservative set of changes.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10951
Reviewed By: anand1976
Differential Revision: D45057849
Pulled By: ajkr
fbshipit-source-id: e4ea46bfc80518ae86f37702b03ca9352bc11c3d
Summary:
In IDE navigation I find it annoying that there are two statistics.h files (etc.) and often land on the wrong one. Here I migrate several headers to use the blah.h <- blah_impl.h <- blah.cc idiom. Although clang-format wants "blah.h" to be the top include for "blah.cc", I think overall this is an improvement.
No public API changes.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11408
Test Plan: existing tests
Reviewed By: ltamasi
Differential Revision: D45456696
Pulled By: pdillinger
fbshipit-source-id: 809d931253f3272c908cf5facf7e1d32fc507373
Summary:
This patch adds support to write and read a user-defined timestamp size record in log writer and log reader. It will be used by WAL logs to persist the user-defined timestamp format for subsequent WriteBatch records. Reading and writing UDT sizes for WAL logs are not included in this patch. It will be in a follow up.
The syntax for the record is: at write time, one such record is added when log writer encountered any non-zero UDT size it hasn't recorded so far. At read time, all such records read up to a point are accumulated and applicable to all subsequent WriteBatch records.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11433
Test Plan:
```
make clean && make -j32 all
./log_test --gtest_filter="*WithTimestampSize*"
```
Reviewed By: ltamasi
Differential Revision: D45678708
Pulled By: jowlyzhang
fbshipit-source-id: b770c8f45bb7b9383b14aac9f22af781304fb41d
Summary:
Context:
This pull request update is in response to a comment made on https://github.com/facebook/rocksdb/pull/8596#discussion_r680264932. The current implementation of RefillBytesAndGrantRequestsLocked() drains all available_bytes, but the first request after the last wave of requesting/bytes granting is done is not being handled in the same way.
This creates a scenario where if a request for a large amount of bytes is enqueued first, but there are not enough available_bytes to fulfill it, the request is put to sleep until the next refill time. Meanwhile, a later request for a smaller number of bytes comes in and is granted immediately. This behavior is not fair as the earlier request was made first.
To address this issue, we have made changes to the code to exhaust the remaining available bytes from the request and queue the remaining. With this change, requests are granted in the order they are received, ensuring that earlier requests are not unfairly delayed by later, smaller requests. The specific scenario described above will no longer occur with this change. Also consolidated `granted` and `request_bytes` as part of the change since `granted` is equivalent to `request_bytes == 0`
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11425
Test Plan: Added `AvailableByteSizeExhaustTest`
Reviewed By: hx235
Differential Revision: D45570711
Pulled By: jaykorean
fbshipit-source-id: a7117ed17bf4b8a7ae0f76124cb41902db1a2592
Summary:
**Background** - runtime detection of certain x86 CPU features was added for optimizing CRC32c checksums, where performance is dramatically affected by the availability of certain CPU instructions and code using intrinsics for those instructions. And Java builds with native library try to be broadly compatible but performant.
What has changed is that CRC32c is no longer the most efficient cheecksum on contemporary x86_64 hardware, nor the default checksum. XXH3 is generally faster and not as dramatically impacted by the availability of certain CPU instructions. For example, on my Skylake system using db_bench (similar on an older Skylake system without AVX512):
PORTABLE=1 empty USE_SSE : xxh3->8 GB/s crc32c->0.8 GB/s (no SSE4.2 nor AVX2 instructions)
PORTABLE=1 USE_SSE=1 : xxh3->19 GB/s crc32c->16 GB/s (with SSE4.2 and AVX2)
PORTABLE=0 USE_SSE ignored: xxh3->28 GB/s crc32c->16 GB/s (also some AVX512)
Testing a ~10 year old system, with SSE4.2 but without AVX2, crc32c is a similar speed to the new systems but xxh3 is only about half that speed, also 8GB/s like the non-AVX2 compile above. Given that xxh3 has specific optimization for AVX2, I think we can infer that that crc32c is only fastest for that ~2008-2013 period when SSE4.2 was included but not AVX2. And given that xxh3 is only about 2x slower on these systems (not like >10x slower for unoptimized crc32c), I don't think we need to invest too much in optimally adapting to these old cases.
x86 hardware that doesn't support fast CRC32c is now extremely rare, so requiring a custom build to support such hardware is fine IMHO.
**This change** does two related things:
* Remove runtime CPU detection for optimizing CRC32c on x86. Maintaining this code is non-zero work, and compiling special code that doesn't work on the configured target instruction set for code generation is always dubious. (On the one hand we have to ensure the CRC32c code uses SSE4.2 but on the other hand we have to ensure nothing else does.)
* Detect CPU features in source code, not in build scripts. Although there are some hypothetical advantages to detectiong in build scripts (compiler generality), RocksDB supports at least three build systems: make, cmake, and buck. It's not practical to support feature detection on all three, and we have suffered from missed optimization opportunities by relying on missing or incomplete detection in cmake and buck. We also depend on some components like xxhash that do source code detection anyway.
**In more detail:**
* `HAVE_SSE42`, `HAVE_AVX2`, and `HAVE_PCLMUL` replaced by standard macros `__SSE4_2__`, `__AVX2__`, and `__PCLMUL__`.
* MSVC does not provide high fidelity defines for SSE, PCLMUL, or POPCNT, but we can infer those from `__AVX__` or `__AVX2__` in a compatibility header. In rare cases of false negative or false positive feature detection, a build engineer should be able to set defines to work around the issue.
* `__POPCNT__` is another standard define, but we happen to only need it on MSVC, where it is set by that compatibility header, or can be set by the build engineer.
* `PORTABLE` can be set to a CPU type, e.g. "haswell", to compile for that CPU type.
* `USE_SSE` is deprecated, now equivalent to PORTABLE=haswell, which roughly approximates its old behavior.
Notably, this change should enable more builds to use the AVX2-optimized Bloom filter implementation.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11419
Test Plan:
existing tests, CI
Manual performance tests after the change match the before above (none expected with make build).
We also see AVX2 optimized Bloom filter code enabled when expected, by injecting a compiler error. (Performance difference is not big on my current CPU.)
Reviewed By: ajkr
Differential Revision: D45489041
Pulled By: pdillinger
fbshipit-source-id: 60ceb0dd2aa3b365c99ed08a8b2a087a9abb6a70
Summary:
## Option API updates
* Add new CompressionOptions::max_compressed_bytes_per_kb, which corresponds to 1024.0 / min allowable compression ratio. This avoids the hard-coded minimum ratio of 8/7.
* Remove unnecessary constructor for CompressionOptions.
* Document undocumented CompressionOptions. Use idiom for default values shown clearly in one place (not precariously repeated).
## Stat API updates
* Deprecate the BYTES_COMPRESSED, BYTES_DECOMPRESSED histograms. Histograms incur substantial extra space & time costs compared to tickers, and the distribution of uncompressed data block sizes tends to be uninteresting. If we're interested in that distribution, I don't see why it should be limited to blocks stored as compressed.
* Deprecate the NUMBER_BLOCK_NOT_COMPRESSED ticker, because the name is very confusing.
* New or existing tickers relevant to compression:
* BYTES_COMPRESSED_FROM
* BYTES_COMPRESSED_TO
* BYTES_COMPRESSION_BYPASSED
* BYTES_COMPRESSION_REJECTED
* COMPACT_WRITE_BYTES + FLUSH_WRITE_BYTES (both existing)
* NUMBER_BLOCK_COMPRESSED (existing)
* NUMBER_BLOCK_COMPRESSION_BYPASSED
* NUMBER_BLOCK_COMPRESSION_REJECTED
* BYTES_DECOMPRESSED_FROM
* BYTES_DECOMPRESSED_TO
We can compute a number of things with these stats:
* "Successful" compression ratio: BYTES_COMPRESSED_FROM / BYTES_COMPRESSED_TO
* Compression ratio of data on which compression was attempted: (BYTES_COMPRESSED_FROM + BYTES_COMPRESSION_REJECTED) / (BYTES_COMPRESSED_TO + BYTES_COMPRESSION_REJECTED)
* Compression ratio of data that could be eligible for compression: (BYTES_COMPRESSED_FROM + X) / (BYTES_COMPRESSED_TO + X) where X = BYTES_COMPRESSION_REJECTED + NUMBER_BLOCK_COMPRESSION_REJECTED
* Overall SST compression ratio (compression disabled vs. actual): (Y - BYTES_COMPRESSED_TO + BYTES_COMPRESSED_FROM) / Y where Y = COMPACT_WRITE_BYTES + FLUSH_WRITE_BYTES
Keeping _REJECTED separate from _BYPASSED helps us to understand "wasted" CPU time in compression.
## BlockBasedTableBuilder
Various small refactorings, optimizations, and name clean-ups.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11388
Test Plan:
unit tests added
* `options_settable_test.cc`: use non-deprecated idiom for configuring CompressionOptions from string. The old idiom is tested elsewhere and does not need to be updated to support the new field.
Reviewed By: ajkr
Differential Revision: D45128202
Pulled By: pdillinger
fbshipit-source-id: 5a652bf5c022b7ec340cf79018cccf0686962803
Summary:
**Context:**
The existing stat rocksdb.sst.read.micros does not reflect each of compaction and flush cases but aggregate them, which is not so helpful for us to understand IO read behavior of each of them.
**Summary**
- Update `StopWatch` and `RandomAccessFileReader` to record `rocksdb.sst.read.micros` and `rocksdb.file.{flush/compaction}.read.micros`
- Fixed the default histogram in `RandomAccessFileReader`
- New field `ReadOptions/IOOptions::io_activity`; Pass `ReadOptions` through paths under db open, flush and compaction to where we can prepare `IOOptions` and pass it to `RandomAccessFileReader`
- Use `thread_status_util` for assertion in `DbStressFSWrapper` for continuous testing on we are passing correct `io_activity` under db open, flush and compaction
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11288
Test Plan:
- **Stress test**
- **Db bench 1: rocksdb.sst.read.micros COUNT ≈ sum of rocksdb.file.read.flush.micros's and rocksdb.file.read.compaction.micros's.** (without blob)
- May not be exactly the same due to `HistogramStat::Add` only guarantees atomic not accuracy across threads.
```
./db_bench -db=/dev/shm/testdb/ -statistics=true -benchmarks="fillseq" -key_size=32 -value_size=512 -num=50000 -write_buffer_size=655 -target_file_size_base=655 -disable_auto_compactions=false -compression_type=none -bloom_bits=3 (-use_plain_table=1 -prefix_size=10)
```
```
// BlockBasedTable
rocksdb.sst.read.micros P50 : 2.009374 P95 : 4.968548 P99 : 8.110362 P100 : 43.000000 COUNT : 40456 SUM : 114805
rocksdb.file.read.flush.micros P50 : 1.871841 P95 : 3.872407 P99 : 5.540541 P100 : 43.000000 COUNT : 2250 SUM : 6116
rocksdb.file.read.compaction.micros P50 : 2.023109 P95 : 5.029149 P99 : 8.196910 P100 : 26.000000 COUNT : 38206 SUM : 108689
// PlainTable
Does not apply
```
- **Db bench 2: performance**
**Read**
SETUP: db with 900 files
```
./db_bench -db=/dev/shm/testdb/ -benchmarks="fillseq" -key_size=32 -value_size=512 -num=50000 -write_buffer_size=655 -disable_auto_compactions=true -target_file_size_base=655 -compression_type=none
```run till convergence
```
./db_bench -seed=1678564177044286 -use_existing_db=true -db=/dev/shm/testdb -benchmarks=readrandom[-X60] -statistics=true -num=1000000 -disable_auto_compactions=true -compression_type=none -bloom_bits=3
```
Pre-change
`readrandom [AVG 60 runs] : 21568 (± 248) ops/sec`
Post-change (no regression, -0.3%)
`readrandom [AVG 60 runs] : 21486 (± 236) ops/sec`
**Compaction/Flush**run till convergence
```
./db_bench -db=/dev/shm/testdb2/ -seed=1678564177044286 -benchmarks="fillseq[-X60]" -key_size=32 -value_size=512 -num=50000 -write_buffer_size=655 -disable_auto_compactions=false -target_file_size_base=655 -compression_type=none
rocksdb.sst.read.micros COUNT : 33820
rocksdb.sst.read.flush.micros COUNT : 1800
rocksdb.sst.read.compaction.micros COUNT : 32020
```
Pre-change
`fillseq [AVG 46 runs] : 1391 (± 214) ops/sec; 0.7 (± 0.1) MB/sec`
Post-change (no regression, ~-0.4%)
`fillseq [AVG 46 runs] : 1385 (± 216) ops/sec; 0.7 (± 0.1) MB/sec`
Reviewed By: ajkr
Differential Revision: D44007011
Pulled By: hx235
fbshipit-source-id: a54c89e4846dfc9a135389edf3f3eedfea257132
Summary:
util/ribbon_test.cc: avoid ambiguous reversed operator error in c++20 (and enable checking for the error)
Code would produce errors like this, when compiled with -Wambiguous-reversed-operator under c++20.
```
util/ribbon_test.cc:695:20: error: ISO C++20 considers use of overloaded operator '!=' (with operand types 'KeyGen' (aka '(anonymous namespace)::StandardKeyGen') and 'KeyGen') to be ambiguou
s despite there being a unique best viable function with non-reversed arguments [-Werror,-Wambiguous-reversed-operator]
while (cur != batch_end) {
~~~ ^ ~~~~~~~~~
util/ribbon_test.cc:111:8: note: candidate function with non-reversed arguments
bool operator!=(const StandardKeyGen& other) {
^
util/ribbon_test.cc:107:8: note: ambiguous candidate function with reversed arguments
bool operator==(const StandardKeyGen& other) {
^
```
This will become a hard error in future standards.
Confirmed that no errors were generated when building using clang and c++20:
```
USE_CLANG=1 USE_COROUTINES=1 make
```
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11371
Reviewed By: meyering
Differential Revision: D44921027
Pulled By: cbi42
fbshipit-source-id: ef25b78260920a4d75a718310688d3a2487ffa87
Summary:
The primary purpose of the FactoryFunc was to support LITE mode where the ObjectRegistry was not available. With the removal of LITE mode, the function was no longer required.
Note that the MergeOperator had some private classes defined in header files. To gain access to their constructors (and name methods), the class definitions were moved into header files.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11203
Reviewed By: cbi42
Differential Revision: D43160255
Pulled By: pdillinger
fbshipit-source-id: f3a465fd5d1a7049b73ecf31e4b8c3762f6dae6c
Summary:
From HISTORY.md: Added a subcode of `Status::Corruption`, `Status::SubCode::kMergeOperatorFailed`, for users to identify corruption failures originating in the merge operator, as opposed to RocksDB's internally identified data corruptions.
This is a followup to https://github.com/facebook/rocksdb/issues/11092, where we gave users the ability to keep running a DB despite merge operator failing. Now that the DB keeps running despite such failures, they want to be able to distinguish such failures from real corruptions.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11231
Test Plan: updated unit test
Reviewed By: akankshamahajan15
Differential Revision: D43396607
Pulled By: ajkr
fbshipit-source-id: 17fbcc779ad724dafada8abd73efd38e1c5208b9
Summary:
The files in `port/`, such as `port_posix.h`, are layering over the system libraries, so shouldn't include the DB-specific files like `options.h`. This PR remove this dependency.
# How
The reason that `port_posix.h` (or `port_win.h`) include `options.h` is to use `CpuPriority`, as there is a method `SetCpuPriority()` in `port_posix.h` that uses `CpuPriority.`
- I think `SetCpuPriority()` make sense to exist in `port_posix.h` as it provides has platform-dependent implementation
- `CpuPriority` enum is defined in `env.h`, but used in `rocksdb/include` and `port/`.
Hence, let us define `CpuPriority` enum in a common file, say `port_defs.h`, such that both directories `rocksdb/include` and `port/` can include.
When we remove this dependency, some other files have compile errors because they can't find definitions, so add header files to resolve
# Test
make all check -j
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11214
Reviewed By: pdillinger
Differential Revision: D43196910
Pulled By: guowentian
fbshipit-source-id: 70deccb72844cfb08fcc994f76c6ef6df5d55ab9
Summary:
In anticipation of using this to represent sets of CacheEntryRole for including or excluding kinds of blocks in block cache tiers, add significant new features to SmallEnumSet, including at least:
* List initialization
* Applicative constexpr operations
* copy/move/equality ops
* begin/end/const_iterator for iteration
* Better comments
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11178
Test Plan: unit tests added/expanded
Reviewed By: ltamasi
Differential Revision: D42973723
Pulled By: pdillinger
fbshipit-source-id: 40783486feda931c3f7c6fcc9a300acd6a4b0a0a
Summary:
Fix a bug in the calculation of the input buffer address/offset in log_reader.cc. The bug is when consecutive fragments of a compressed record are located at the same offset in the log reader buffer, the second fragment input buffer is treated as a leftover from the previous input buffer. As a result, the offset in the `ZSTD_inBuffer` is not reset.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11198
Test Plan: Add a unit test in log_test.cc that fails without the fix and passes with it.
Reviewed By: ajkr, cbi42
Differential Revision: D43102692
Pulled By: anand1976
fbshipit-source-id: aa2648f4802c33991b76a3233c5a58d4cc9e77fd
Summary:
Currently, we incorrectly return a Status::Corruption to the MultiGet caller if the file system ReadAsync cannot issue a read and returns an error for some reason, such as IOStatus::NotSupported(). In this PR, we copy the ReadAsync error to the request status so it can be returned to the user.
Tests:
Update existing unit tests and add a new one for this scenario
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11171
Reviewed By: akankshamahajan15
Differential Revision: D42950057
Pulled By: anand1976
fbshipit-source-id: 85ffcb015fa6c064c311f8a28488fec78c487869
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
Summary:
Like other versions before, gcc 13 moved some includes around and as a result <cstdint> is no longer transitively included [1]. Explicitly include it for uint{32,64}_t.
[1] https://gcc.gnu.org/gcc-13/porting_to.html#header-dep-changes
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11118
Reviewed By: cbi42
Differential Revision: D42711356
Pulled By: ajkr
fbshipit-source-id: 5ea257b85b7017f40fd8fdbce965336da95c55b2
Summary:
Upgrading xxhash.h to latest dev version as of 1/17/2023, which is d7197ddea81364a539051f116ca77926100fc77f This should improve performance on some ARM machines.
I allowed some of our RocksDB-specific changes to be made obsolete where it seemed appropriate, for example
* xxhash.h has its own fallthrough marker (which I hope works for us)
* As in https://github.com/Cyan4973/xxHash/pull/549
Merging and resolving conflicts one way or the other was all that went into this diff. Except I had to mix the two sides around `defined(__loongarch64)`
How I did the upgrade (for future reference), so that I could use usual merge conflict resolution:
```
# New branch to help with merging
git checkout -b xxh_merge_base
# Check out RocksDB revision before last xxhash.h upgrade
git reset --hard 22161b7547652af82a5dc67458de9ca8946ac83d^
# Create a commit with the raw base version from xxHash repo (from xxHash repo)
git show 2c611a76f914828bed675f0f342d6c4199ffee1e:xxhash.h > ../rocksdb/util/xxhash.h
# In RocksDB repo
git commit -a
# Merge in the last xxhash.h upgrade
git merge 22161b7547
# Resolve conflict using committed version
git show 22161b7547652af82a5dc67458de9ca8946ac83d:util/xxhash.h > util/xxhash.h
git commit -a
# Catch up to upstream
git merge upstream/main
# Create a different branch for applying raw upgrade
git checkout -b xxh_upgrade_2023
# Find the RocksDB commit we made for the raw base version from xxHash
git log main..HEAD
# Rewind to it
git reset --hard 2428b727a9
# Copy in latest raw version (from xxHash repo)
cat xxhash.h > ../rocksdb/util/xxhash.h
# Merge in RocksDB changes, use typical tools for conflict resolution
git merge xxh_merge_base
```
Branch https://github.com/facebook/rocksdb/tree/xxhash_merge_base can be used as a base for future xxhash merges.
Fixes https://github.com/facebook/rocksdb/issues/11073
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11098
Test Plan:
existing tests (e.g. Bloom filter schema stability tests)
Also seems to include a small performance boost on my Intel dev machine, using `./db_bench --benchmarks=xxh3[-X50] 2>&1 | egrep -o 'operations;.*' | sort`
Fastest out of 50 runs, before: 15477.3 MB/s
Fastest out of 50 runs, after: 15850.7 MB/s, and 11 more runs faster than the "before" number
Slowest out of 50 runs, before: 12267.5 MB/s
Slowest out of 50 runs, after: 13897.1 MB/s
More repetitions show the distinction is repeatable
Reviewed By: hx235
Differential Revision: D42560010
Pulled By: pdillinger
fbshipit-source-id: c43ee52f1c5fe0ba3d6d6e4eebb22ded5f5492ea
Yu Zhang