Summary:
With atomic flush, RocksDB background flush will flush memtables of a column family up to the largest memtable id in the immutable memtable list. This can introduce a bug in the following scenario. A user thread inserts into a column family until the memtable is full and triggers a flush. This will add the column family to flush_scheduler_. Then the user thread writes another record to the column family. In the PreprocessWrite function, the user thread picks the column family from flush_scheduler_ and schedules a flush request. The flush request gaurantees to flush all the memtables up to the current largest memtable ID of the immutable memtable list. Then the user thread writes new data to the newly-created active memtable. After the write returns, the user thread closes the db. This can cause assertion failure when the background flush thread tries to install superversion for the column family. The solution is to not install flush results if the db has already set `shutting_down_` to true.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5254
Differential Revision: D15124149
Pulled By: riversand963
fbshipit-source-id: 0a667a41339dedb5a18bcb01b0bf11c275c04df0
Summary:
Improve the iterators performance when the user explicitly sets the readahead size via `ReadOptions.readahead_size`.
1. Stop creating new table readers when the user explicitly sets readahead size.
2. Make use of an internal buffer based on `FilePrefetchBuffer` instead of using `ReadaheadRandomAccessFileReader`, to handle the user readahead requests (for both buffered and direct io cases).
3. Add `readahead_size` to db_bench.
**Benchmarks:**
https://gist.github.com/sagar0/53693edc320a18abeaeca94ca32f5737
For 1 MB readahead, Buffered IO performance improves by 28% and Direct IO performance improves by 50%.
For 512KB readahead, Buffered IO performance improves by 30% and Direct IO performance improves by 67%.
**Test Plan:**
Updated `DBIteratorTest.ReadAhead` test to make sure that:
- no new table readers are created for iterators on setting ReadOptions.readahead_size
- At least "readahead" number of bytes are actually getting read on each iterator read.
TODO later:
- Use similar logic for compactions as well.
- This ties in nicely with #4052 and paves the way for removing ReadaheadRandomAcessFile later.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5246
Differential Revision: D15107946
Pulled By: sagar0
fbshipit-source-id: 2c1149729ca7d779e4e8b7710ba6f4e8cbfd3bea
Summary:
The newly added test CompactionJobTest.SnapshotRefresh sets the snapshot refresh period to 0 to stress the feature. This results into large number of refresh events, which in turn results into an UBSAN failure when a bitwise shift operand goes beyond the uint64_t size.
The patch fixes that by simplifying the shift logic to be done only by 2 bits after each refresh. Furthermore it verifies that the shift operation does not result in decreasing the refresh period.
Testing:
COMPILE_WITH_UBSAN=1 make -j32 compaction_job_test
./compaction_job_test --gtest_filter=CompactionJobTest.SnapshotRefresh
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5257
Differential Revision: D15106463
Pulled By: maysamyabandeh
fbshipit-source-id: f2718898ea7ba4fa9f7e87b70cf98fe647c0de80
Summary:
Part of compaction cpu goes to processing snapshot list, the larger the list the bigger the overhead. Although the lifetime of most of the snapshots is much shorter than the lifetime of compactions, the compaction conservatively operates on the list of snapshots that it initially obtained. This patch allows the snapshot list to be updated via a callback if the compaction is taking long. This should let the compaction to continue more efficiently with much smaller snapshot list.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5099
Differential Revision: D15086710
Pulled By: maysamyabandeh
fbshipit-source-id: 7649f56c3b6b2fb334962048150142a3bf9c1a12
Summary:
- By providing the "env" field in any text-based options (i.e., string, map, or file), we can use `NewCustomObject` to deserialize the text value into an actual `Env` object.
- Currently factory functions for `Env` registered with object registry should only return pointer to static `Env` objects. That's because `DBOptions::env` is a raw pointer so we cannot easily delegate cleanup.
- Note I did not add `env` to `db_option_type_info`. It wasn't needed for (de)serialization, and I believe we don't want to do verification on `env`, even by checking name. That's because the user should be able to copy their DB from Linux to Windows, change envs, and not see an option verification error.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5237
Differential Revision: D15056360
Pulled By: siying
fbshipit-source-id: 4b5f0b83297a5058f8949ec955dbf27d98d73d7e
Summary:
Currently one thread in RocksDB keeps a WAL file open while another thread
deletes it. Although the first thread never writes to the WAL again, it still
tries to close it in the end. This is fine on POSIX, but can be problematic on
other platforms, e.g. HDFS, etc.. It will either cause a lot of warning messages or
throw exceptions. The solution is to let the second thread close the WAL before deleting it.
RocksDB keeps the writers of the logs to delete in `logs_to_free_`, which is passed to `job_context` during `FindObsoleteFiles` (holding mutex). Then in `PurgeObsoleteFiles` (without mutex), these writers should close the logs.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5233
Differential Revision: D15032670
Pulled By: riversand963
fbshipit-source-id: c55e8a612db8cc2306644001a5e6d53842a8f754
Summary:
We have a DB with ~4k column families and ~70k files. On shutdown, destroying the 4k ColumnFamilyHandle-s takes over 2 minutes. Most of this time is spent in VersionSet::AddLiveFiles() called from FindObsoleteFiles() from ~ColumnFamilyHandleImpl(). It's just iterating over the list of files in memory. This seems completely unnecessary as no obsolete files are actually found since the CFs are not even dropped. This PR fixes that.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5238
Differential Revision: D15056342
Pulled By: siying
fbshipit-source-id: 2aa342ef3770b4aa384ce81f8768e485480e4f08
Summary: PR https://github.com/facebook/rocksdb/pull/4899 implemented the general framework for RocksDB secondary instances. This PR adds the support for WAL tailing in `OpenAsSecondary`, which means after the `OpenAsSecondary` call, the secondary is now able to see primary's writes that are yet to be flushed. The secondary can see primary's writes in the WAL up to the moment of `OpenAsSecondary` call starts.
Differential Revision: D15059905
Pulled By: miasantreble
fbshipit-source-id: 44f71f548a30b38179a7940165e138f622de1f10
Summary:
MultiGet batching was implemented in #5011 in order to reduce CPU utilization when looking up multiple keys at once. This PR implements corresponding ```MultiGet``` and ```MultiGetSingleCFForUpdate``` in ```rocksdb::Transaction``` that call the underlying batching implementation.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5210
Differential Revision: D15048164
Pulled By: anand1976
fbshipit-source-id: c52f6043102ab0cbc723f4cba2a7b7d1767f6f52
Summary:
In some cases, we want to known the smallest and largest sequence numbers of sstable files, to help us get more details.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5231
Differential Revision: D15038087
Pulled By: siying
fbshipit-source-id: c473c1ca07b53efe2f1884fa1ecdc8686f455ed8
Summary:
I needed this change to be able to build the v6.0.1 release on Windows.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5227
Differential Revision: D15033815
Pulled By: sagar0
fbshipit-source-id: 579f3b8e694c34c0d43527eb2fa37175e37f5911
Summary:
It's hard to get DBIter to directly use InternalIterator::NextAndGetResult() because the code change would be complicated. Instead, use IteratorWrapper, where Next() is already using NextAndGetResult(). Performance number is hard to measure because it is small and ther is variation. I run readseq many times, and there seems to be 1% gain.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5214
Differential Revision: D15003635
Pulled By: siying
fbshipit-source-id: 17af1965c409c2fe90cd85037fbd2c5a1364f82a
Summary:
The existing implementation does not guarantee bytes reach disk every `bytes_per_sync` when writing SST files, or every `wal_bytes_per_sync` when writing WALs. This can cause confusing behavior for users who enable this feature to avoid large syncs during flush and compaction, but then end up hitting them anyways.
My understanding of the existing behavior is we used `sync_file_range` with `SYNC_FILE_RANGE_WRITE` to submit ranges for async writeback, such that we could continue processing the next range of bytes while that I/O is happening. I believe we can preserve that benefit while also limiting how far the processing can get ahead of the I/O, which prevents huge syncs from happening when the file finishes.
Consider this `sync_file_range` usage: `sync_file_range(fd_, 0, static_cast<off_t>(offset + nbytes), SYNC_FILE_RANGE_WAIT_BEFORE | SYNC_FILE_RANGE_WRITE)`. Expanding the range to start at 0 and adding the `SYNC_FILE_RANGE_WAIT_BEFORE` flag causes any pending writeback (like from a previous call to `sync_file_range`) to finish before it proceeds to submit the latest `nbytes` for writeback. The latest `nbytes` are still written back asynchronously, unless processing exceeds I/O speed, in which case the following `sync_file_range` will need to wait on it.
There is a second change in this PR to use `fdatasync` when `sync_file_range` is unavailable (determined statically) or has some known problem with the underlying filesystem (determined dynamically).
The above two changes only apply when the user enables a new option, `strict_bytes_per_sync`.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5183
Differential Revision: D14953553
Pulled By: siying
fbshipit-source-id: 445c3862e019fb7b470f9c7f314fc231b62706e9
Summary:
Introduce BlockBasedTableOptions::index_shortening to give users control on which key shortening techniques to be used in building index blocks. Before this patch, both separators and successor keys where shortened in indexes. With this patch, the default is set to kShortenSeparators to only shorten the separators. Since each index block has many separators and only one successor (last key), the change should not have negative impact on index block size. However it should prevent many unnecessary block loads where due to approximation introduced by shorted successor, seek would land us to the previous block and then fix it by moving to the next one.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5174
Differential Revision: D14884185
Pulled By: al13n321
fbshipit-source-id: 1b08bc8c03edcf09b6b8c16e9a7eea08ad4dd534
Summary:
Savepoints are assumed to be used in a stack-wise fashion (only
the top element should be used), so they were stored by `WriteBatch`
in a member variable `save_points` using an std::stack.
Conceptually this is fine, but the implementation had a few issues:
- the `save_points_` instance variable was a plain pointer to a heap-
allocated `SavePoints` struct. The destructor of `WriteBatch` simply
deletes this pointer. However, the copy constructor of WriteBatch
just copied that pointer, meaning that copying a WriteBatch with
active savepoints will very likely have crashed before. Now a proper
copy of the savepoints is made in the copy constructor, and not just
a copy of the pointer
- `save_points_` was an std::stack, which defaults to `std::deque` for
the underlying container. A deque is a bit over the top here, as we
only need access to the most recent savepoint (i.e. stack.top()) but
never any elements at the front. std::deque is rather expensive to
initialize in common environments. For example, the STL implementation
shipped with GNU g++ will perform a heap allocation of more than 500
bytes to create an empty deque object. Although the `save_points_`
container is created lazily by RocksDB, moving from a deque to a plain
`std::vector` is much more memory-efficient. So `save_points_` is now
a vector.
- `save_points_` was changed from a plain pointer to an `std::unique_ptr`,
making ownership more explicit.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5192
Differential Revision: D15024074
Pulled By: maysamyabandeh
fbshipit-source-id: 5b128786d3789cde94e46465c9e91badd07a25d7
Summary:
Fix HISTORY.md by removing a few items from 6.1.1 history as they did not make into the 6.1.fb branch.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5224
Differential Revision: D15017030
Pulled By: sagar0
fbshipit-source-id: 090724d326d29168952e06dc1a5090c03fdd739e
Summary:
Cleanup a couple of stray includes left by #5011.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5219
Differential Revision: D15007244
Pulled By: anand1976
fbshipit-source-id: 15ca1d4f977b5b60e99df3bfb8fc3db217d19bdd
Summary:
My compiler doesn't inline DBIter::Next() to arena wrapped iterator, even if it is a direct forward. Adding this annotation makes it inlined. It might not always work but inlinging this function to arena wrapped iterator always feels like the right decision.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5217
Differential Revision: D15004086
Pulled By: siying
fbshipit-source-id: a4cffd79c6fb092669a3a90633c9aa5e494f8a66
Summary:
We found an issue in Periodic Compactions (introduced in #5166) where files were not being picked up for compactions as all the SST files created with older versions of RocksDB have `file_creation_time` as 0. (Note that `file_creation_time` is a new table property introduced in #5166).
To address this, Periodic compactions now fall back to looking at the `creation_time` table property or the file's modification time (as given by the Env) when `file_creation_time` table property is found to be 0.
Here how the file's modification time (and, in turn, the file age) is computed now:
1. Use `file_creation_time` table property if it is > 0.
1. If not, then use `creation_time` table property if it is > 0.
1. If not, then use file's mtime stat metadata given by the underlying Env.
Don't consider the file at all for compaction if the modification time cannot be correctly determined based on the above conditions.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5184
Differential Revision: D14907795
Pulled By: sagar0
fbshipit-source-id: 4bb2f3631f9a3e04470c674a1d13544584e1e56c
Summary:
Reorganize the code so that no function call into ReadRangeDelAggregator is needed if there is no tomb range stone.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5202
Differential Revision: D14968155
Pulled By: siying
fbshipit-source-id: 0bd61911293c7a27b4e1b8d57c66d0c4ad6a6a5f
Summary:
Several small changes for Next():
1. Reducing branching by always update local_stats_.next_count_++ even if statistics is null. This should be faster than a branching.
2. Replacing ResetInternalKeysSkippedCounter() in Next() because the valid_ check is not needed in this case.
3. iter_->Valid() should always be true for non merge case. Remove this check.
4. Adding an inline annotation. It ends up with not picked up by my compiler, but it shouldn't hurt.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5200
Differential Revision: D15000391
Pulled By: siying
fbshipit-source-id: be97f61c708968234fb8e5cf272b5c2ac07dc4dd
Summary:
In long scans, virtual function calls of Next(), Valid(), key() and value() are not trivial. By introducing NextAndGetResult(), Some of the Next(), Valid() and key() calls are consolidated into one virtual function call to reduce CPU.
Also did some inline tricks and add some "final" randomly in some functions. Even without the "final" annotation, most Next() calls are inlined with -O3, but sometimes with a final it is inlined by O2 too. It doesn't hurt to add those final annotations.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5197
Differential Revision: D14945977
Pulled By: siying
fbshipit-source-id: 7003969f9a5f1d5717f0bda503b91d19ba75ed88
Summary:
`GetOverlappingInputsRangeBinarySearch` firstly use binary search
to find a index in the given range `[begin, end]`. But after find
the index, then use linear search to find the `start_index` and
`end_index`. So the search process degraded to linear time.
Here optmize the search process with below changes:
- use `std::lower_bound` and `std::upper_bound` to get
`lg(n)` search complexity.
- use uniformed lambda for search process.
- simplify process for `within_interval` true or false.
- remove function `ExtendFileRangeWithinInterval`
and `ExtendFileRangeOverlappingInterval`.
Signed-off-by: JiYou <jiyou09@gmail.com>
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4987
Differential Revision: D14984192
Pulled By: riversand963
fbshipit-source-id: fae4b8e59a21b7e350718d60cdc94dd55ac81e89
Summary:
this PR fixes the following compile warning:
```
db/memtable.cc: In member function ‘virtual void rocksdb::MemTableIterator::Seek(const rocksdb::Slice&)’:
db/memtable.cc:321:22: error: declaration of ‘user_key’ shadows a member of 'this' [-Werror=shadow]
Slice user_key(ExtractUserKey(k));
^
db/memtable.cc: In member function ‘virtual void rocksdb::MemTableIterator::SeekForPrev(const rocksdb::Slice&)’:
db/memtable.cc:338:22: error: declaration of ‘user_key’ shadows a member of 'this' [-Werror=shadow]
Slice user_key(ExtractUserKey(k));
^
```
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5204
Differential Revision: D14970160
Pulled By: miasantreble
fbshipit-source-id: 388eb089f90c4528cc6d615dd4607fb53ceac705
Summary:
Depending on the config, manual compaction (leveled compaction style) does following compactions:
L0->L1
L1->L2
...
Ln-1 -> Ln
Ln -> Ln
The final Ln -> Ln compaction is partly unnecessary as it recompacts all the files that were just generated by the Ln-1 -> Ln. We should avoid recompacting such files. This rule should be applied to Lmax only.
Resolves issue https://github.com/facebook/rocksdb/issues/4995
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5138
Differential Revision: D14940106
Pulled By: miasantreble
fbshipit-source-id: 8d3cf5507a17e76f3333cfd4bac5256d005636e5
Summary:
#4905 removed the implementation of `NewEmptyIterator` but kept its
declaration in the public header. This breaks some systems that depend on
RocksDB if the systems use `NewEmptyIterator`. Therefore, add it back to fix. cc maysamyabandeh please remind me if I miss anything here. Thanks
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5203
Differential Revision: D14968382
Pulled By: riversand963
fbshipit-source-id: 5fb86e99c8cfaf9f7a9473cdb1355d7558ff6e01
Summary:
Dummy cache size of 1MB is too large for small block sizes. Our GetDefaultCacheShardBits() use min_shard_size = 512L * 1024L to determine number of shards, so 1MB will excceeds the size of the whole shard and make the cache excceeds the budget.
Change it to 256KB accordingly.
There shouldn't be obvious performance impact, since inserting a cache entry every 256KB of memtable inserts is still infrequently enough.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5175
Differential Revision: D14954289
Pulled By: siying
fbshipit-source-id: 2c275255c1ac3992174e06529e44c55538325c94
Summary:
This is second attempt for #5101. Original commit message:
`BlockBasedTableIterator` avoid reading next block on `Next()` if it detects the iterator will be out of bound, by checking against index key. The optimization was added in #2239, and by the time it only check the bound per block. It seems later change make it a per-key check, which introduce unnecessary key comparisons.
This patch come with two fixes:
Fix 1: To optimize checking for bounds, we need comparing the bounds with index key as well. However BlockBasedTableIterator doesn't know whether its index iterator is internally using user keys or internal keys. The patch fixes that by extending InternalIterator with a user_key() function that is overridden by In IndexBlockIter.
Fix 2: In #5101 we return `IsOutOfBound()=true` when block index key is out of bound. But the index key can be larger than smallest key of the next file on the level. That file can be within upper bound and should not be filtered out.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5142
Differential Revision: D14907113
Pulled By: siying
fbshipit-source-id: ac95775c5b4e7b700f76ab43e39f45402c98fbfb
Summary:
Right now, two separate pieces of code are used to create WAL files in DBImpl::Open function of db_impl_open.cc and DBImpl::SwitchMemtable function of db_impl_write.cc. This code change simply creates 1 function called DBImpl::CreateWAL in db_impl_open.cc which is used to replace existing WAL creation logic in DBImpl::Open and DBImpl::SwitchMemtable.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5188
Differential Revision: D14942832
Pulled By: vjnadimpalli
fbshipit-source-id: d49230e04c36176015c8c1b422575872f92157fb
Summary:
Found this when test driving the new MultiGet. If you pass unsorted result with sorted_result = false you'll trigger the ASSERT incorrect even though we'll sort down below.
I've also added simple test cover sorted_result=true/false scenario copied from MultiGetSimple.
anand1976
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5195
Differential Revision: D14935475
Pulled By: yizhang82
fbshipit-source-id: 1d2af5e3a003847d965066a16e3b19da68acf170
Summary:
This branch contains two small improvements:
* Create `LockMap` entries using `std::make_shared`. This saves one heap allocation per LockMap entry but also locates the control block and the LockMap object closely together in memory, which can help with caching
* Reorder the members of `TrackedTrxInfo`, so that the resulting struct uses less memory (at least on 64bit systems)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5193
Differential Revision: D14934536
Pulled By: maysamyabandeh
fbshipit-source-id: f7b49812bb4b6029eef9d131e7cd56260df5b28e
Summary:
Before using prefix extractor `InDomain()` should be check. All uses in memtable.cc didn't check `InDomain()`.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5190
Differential Revision: D14923773
Pulled By: miasantreble
fbshipit-source-id: b3ad60bcca5f3a1a2b929a6eb34b0b7ba6326f04
Summary:
In `PessimisticTransaction::TryLock`, we were calling `TrackKey` even when assume_tracked=true, which defeats the purpose of assume_tracked. Remove this.
For keys that are already tracked, TrackKey will actually bump some counters (num_reads/num_writes) which are consumed in `TransactionBaseImpl::GetTrackedKeysSinceSavePoint`, and this is used to determine which keys were tracked since the last savepoint. I believe this functionality should still work, since I think the user should not call GetForUpdate/Put(assume_tracked=true) across savepoints, and if they do, they should not expect the Put(assume_tracked=true) to show up as a tracked key in the second savepoint.
This is another 2-3% cpu improvement.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5173
Differential Revision: D14883809
Pulled By: lth
fbshipit-source-id: 7d09f0772da422384af0519773e310c22b0cbca3
Summary:
When ReadOption doesn't specify a snapshot, WritePrepared::Get used kMaxSequenceNumber to avoid the cost of creating a new snapshot object (that requires sync over db_mutex). This creates a race condition if it is reading from the writes of a transaction that had duplicate keys: each instance of duplicate key is inserted with a different sequence number and depending on the ordering the ::Get might skip the newer one and read the older one that is obsolete.
The patch fixes that by using last published seq as the snapshot sequence number. It also adds a check after the read is done to ensure that the max_evicted_seq has not advanced the aforementioned seq, which is a very unlikely event. If it did, then the read is not valid since the seq is not backed by an actually snapshot to let IsInSnapshot handle that properly when an overlapping commit is evicted from commit cache.
A unit test is added to reproduce the race condition with duplicate keys.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5147
Differential Revision: D14758815
Pulled By: maysamyabandeh
fbshipit-source-id: a56915657132cf6ba5e3f5ea1b5d78c803407719
Summary:
I would like to be able to read out the current Filter that has been set (or not) for a BlockBasedTableConfig. Added one public method to BlockBasedTableConfig:
public Filter filterPolicy() {
return filterPolicy;
}
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5186
Differential Revision: D14921415
Pulled By: siying
fbshipit-source-id: 2a63c8685480197862b49fc48916c757cd6daf95
Summary:
Since Statistics::measureTime() is deprecated, StatisticsImpl::measureTime() is not implemented. We realized that users might have a wrapped Statistics implementation in which measureTime() is implemented as forwarded to StatisticsImpl, and causes assert failure. In order to make the change less intrusive, we implement StatisticsImpl::measureTime(). We will revisit whether we need to remove it after several releases.
Also, add a test to make sure that a Statistics implementation using the old interface still works.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5181
Differential Revision: D14907089
Pulled By: siying
fbshipit-source-id: 29b6202fd04e30ed6f6adcaeb1000e87f10d1e1a
Summary:
When a new SST file is created via flush or compaction, we dump out the table properties, however only a few table properties are logged. The change here is to log all the table properties
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5168
Differential Revision: D14876928
Pulled By: vjnadimpalli
fbshipit-source-id: 1aca42ad00f9f650761d39e187f8beeb8700149b
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
