Summary:
This diff adds support for concurrent adds to the skiplist memtable
implementations. Memory allocation is made thread-safe by the addition of
a spinlock, with small per-core buffers to avoid contention. Concurrent
memtable writes are made via an additional method and don't impose a
performance overhead on the non-concurrent case, so parallelism can be
selected on a per-batch basis.
Write thread synchronization is an increasing bottleneck for higher levels
of concurrency, so this diff adds --enable_write_thread_adaptive_yield
(default off). This feature causes threads joining a write batch
group to spin for a short time (default 100 usec) using sched_yield,
rather than going to sleep on a mutex. If the timing of the yield calls
indicates that another thread has actually run during the yield then
spinning is avoided. This option improves performance for concurrent
situations even without parallel adds, although it has the potential to
increase CPU usage (and the heuristic adaptation is not yet mature).
Parallel writes are not currently compatible with
inplace updates, update callbacks, or delete filtering.
Enable it with --allow_concurrent_memtable_write (and
--enable_write_thread_adaptive_yield). Parallel memtable writes
are performance neutral when there is no actual parallelism, and in
my experiments (SSD server-class Linux and varying contention and key
sizes for fillrandom) they are always a performance win when there is
more than one thread.
Statistics are updated earlier in the write path, dropping the number
of DB mutex acquisitions from 2 to 1 for almost all cases.
This diff was motivated and inspired by Yahoo's cLSM work. It is more
conservative than cLSM: RocksDB's write batch group leader role is
preserved (along with all of the existing flush and write throttling
logic) and concurrent writers are blocked until all memtable insertions
have completed and the sequence number has been advanced, to preserve
linearizability.
My test config is "db_bench -benchmarks=fillrandom -threads=$T
-batch_size=1 -memtablerep=skip_list -value_size=100 --num=1000000/$T
-level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999
-disable_auto_compactions --max_write_buffer_number=8
-max_background_flushes=8 --disable_wal --write_buffer_size=160000000
--block_size=16384 --allow_concurrent_memtable_write" on a two-socket
Xeon E5-2660 @ 2.2Ghz with lots of memory and an SSD hard drive. With 1
thread I get ~440Kops/sec. Peak performance for 1 socket (numactl
-N1) is slightly more than 1Mops/sec, at 16 threads. Peak performance
across both sockets happens at 30 threads, and is ~900Kops/sec, although
with fewer threads there is less performance loss when the system has
background work.
Test Plan:
1. concurrent stress tests for InlineSkipList and DynamicBloom
2. make clean; make check
3. make clean; DISABLE_JEMALLOC=1 make valgrind_check; valgrind db_bench
4. make clean; COMPILE_WITH_TSAN=1 make all check; db_bench
5. make clean; COMPILE_WITH_ASAN=1 make all check; db_bench
6. make clean; OPT=-DROCKSDB_LITE make check
7. verify no perf regressions when disabled
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: MarkCallaghan, IslamAbdelRahman, anthony, yhchiang, rven, sdong, guyg8, kradhakrishnan, dhruba
Differential Revision: https://reviews.facebook.net/D50589
Summary: DBTest.HardLimit fails in appveyor build. Use special mem table to make the test behavior depends less on platform
Test Plan: Run the test with JEMALLOC both on and off.
Reviewers: yhchiang, kradhakrishnan, rven, anthony, IslamAbdelRahman
Reviewed By: IslamAbdelRahman
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D52317
Summary: DBTest.DelayedWriteRate has sign and unsign comparisons that break Windows build. Fix it.
Test Plan: Build and run the test modified.
Reviewers: IslamAbdelRahman, rven, anthony, yhchiang, kradhakrishnan
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D52311
Summary: We now have a mechanism to further slowdown writes. Double default options.delayed_write_rate to try to keep the default behavior closer to it used to be.
Test Plan: Run all tests.
Reviewers: IslamAbdelRahman
Reviewed By: IslamAbdelRahman
Subscribers: yhchiang, kradhakrishnan, rven, leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D52281
Summary: It's usually hard for users to set a value of options.delayed_write_rate. With this diff, after slowdown condition triggers, we greedily reduce write rate if estimated pending compaction bytes increase. If estimated compaction pending bytes drop, we increase the write rate.
Test Plan:
Add a unit test
Test with db_bench setting:
TEST_TMPDIR=/dev/shm/ ./db_bench --benchmarks=fillrandom -num=10000000 --soft_pending_compaction_bytes_limit=1000000000 --hard_pending_compaction_bytes_limit=3000000000 --delayed_write_rate=100000000
and make sure without the commit, write stop will happen, but with the commit, it will not happen.
Reviewers: igor, anthony, rven, yhchiang, kradhakrishnan, IslamAbdelRahman
Reviewed By: IslamAbdelRahman
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D52131
Summary:
Missed this in https://reviews.facebook.net/D51633 because I didn't
wait for 'make commit-prereq' to finish
Test Plan: make clean && USE_CLANG=1 make -j32 all
Reviewers: IslamAbdelRahman, sdong
Reviewed By: sdong
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D52275
Summary:
When Get() or NewIterator() trigger file loads, skip caching the filter block if
(1) optimize_filters_for_hits is set and (2) the file is on the bottommost
level. Also skip checking filters under the same conditions, which means that
for a preloaded file or a file that was trivially-moved to the bottom level, its
filter block will eventually expire from the cache.
- added parameters/instance variables in various places in order to propagate the config ("skip_filters") from version_set to block_based_table_reader
- in BlockBasedTable::Rep, this optimization prevents filter from being loaded when the file is opened simply by setting filter_policy = nullptr
- in BlockBasedTable::Get/BlockBasedTable::NewIterator, this optimization prevents filter from being used (even if it was loaded already) by setting filter = nullptr
Test Plan:
updated unit test:
$ ./db_test --gtest_filter=DBTest.OptimizeFiltersForHits
will also run 'make check'
Reviewers: sdong, igor, paultuckfield, anthony, rven, kradhakrishnan, IslamAbdelRahman, yhchiang
Reviewed By: yhchiang
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D51633
Summary:
Add CompactionReason to CompactionJobInfo
This will allow users to understand why compaction started which will help options tuning
Test Plan:
added new tests
make check -j64
Reviewers: yhchiang, anthony, kradhakrishnan, sdong, rven
Reviewed By: rven
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D51975
Summary:
This patch fixes https://github.com/facebook/mysql-5.6/issues/121
There is a recent change in rocksdb to disable auto compactions on startup: https://reviews.facebook.net/D51147. However, there is a small timing window where a column family needs to be compacted and schedules a compaction, but the scheduled compaction fails when it checks the disable_auto_compactions setting. The expectation is once the application is ready, it will call EnableAutoCompactions() to allow new compactions to go through. However, if the Column family is stalled because L0 is full, and no writes can go through, it is possible the column family may never have a new compaction request get scheduled. EnableAutoCompaction() should probably schedule an new flush and compaction event when it resets disable_auto_compaction.
Using InstallSuperVersionAndScheduleWork, we call SchedulePendingFlush,
SchedulePendingCompaction, as well as MaybeScheduleFlushOrcompaction on all the
column families to avoid the situation above.
This is still a first pass for feedback.
Could also just call SchedePendingFlush and SchedulePendingCompaction directly.
Test Plan:
Run on Asan build
cd _build-5.6-ASan/ && ./mysql-test/mtr --mem --big --testcase-timeout=36000 --suite-timeout=12000 --parallel=16 --suite=rocksdb,rocksdb_rpl,rocksdb_sys_vars --mysqld=--default-storage-engine=rocksdb --mysqld=--skip-innodb --mysqld=--default-tmp-storage-engine=MyISAM --mysqld=--rocksdb rocksdb_rpl.rpl_rocksdb_stress_crash --repeat=1000
Ensure that it no longer hangs during the test.
Reviewers: hermanlee4, yhchiang, anthony
Reviewed By: anthony
Subscribers: leveldb, yhchiang, dhruba
Differential Revision: https://reviews.facebook.net/D51747
Summary: Fix a bug that options.soft_pending_compaction_bytes_limit is not actually set with --soft_pending_compaction_bytes_limit
Test Plan: Run db_bench with this parameter and make sure the parameter is set correctly.
Reviewers: anthony, kradhakrishnan, yhchiang, IslamAbdelRahman, igor, rven
Reviewed By: rven
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D52125
Summary:
When there are waiting manual compactions, we need to signal
them after removing the current manual compaction from the deque.
Test Plan: ColumnFamilytTest.SameCFManualManualCommaction
Reviewers: anthony, IslamAbdelRahman, kradhakrishnan, sdong
Reviewed By: sdong
Subscribers: dhruba, yoshinorim
Differential Revision: https://reviews.facebook.net/D52119
Summary: Now if inserting to mem table is much faster than writing to files, there is no mechanism users can rely on to avoid stopping for reaching options.max_write_buffer_number. With the commit, if there are more than four maximum write buffers configured, we slow down to the rate of options.delayed_write_rate while we reach the last one.
Test Plan:
1. Add a new unit test.
2. Run db_bench with
./db_bench --benchmarks=fillrandom --num=10000000 --max_background_flushes=6 --batch_size=32 -max_write_buffer_number=4 --delayed_write_rate=500000 --statistics
based on hard drive and see stopping is avoided with the commit.
Reviewers: yhchiang, IslamAbdelRahman, anthony, rven, kradhakrishnan, igor
Reviewed By: igor
Subscribers: MarkCallaghan, leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D52047
Summary:
This patch update the Iterator API to introduce new functions that allow users to keep the Slices returned by key() valid as long as the Iterator is not deleted
ReadOptions::pin_data : If true keep loaded blocks in memory as long as the iterator is not deleted
Iterator::IsKeyPinned() : If true, this mean that the Slice returned by key() is valid as long as the iterator is not deleted
Also add a new option BlockBasedTableOptions::use_delta_encoding to allow users to disable delta_encoding if needed.
Benchmark results (using https://phabricator.fb.com/P20083553)
```
// $ du -h /home/tec/local/normal.4K.Snappy/db10077
// 6.1G /home/tec/local/normal.4K.Snappy/db10077
// $ du -h /home/tec/local/zero.8K.LZ4/db10077
// 6.4G /home/tec/local/zero.8K.LZ4/db10077
// Benchmarks for shard db10077
// _build/opt/rocks/benchmark/rocks_copy_benchmark \
// --normal_db_path="/home/tec/local/normal.4K.Snappy/db10077" \
// --zero_db_path="/home/tec/local/zero.8K.LZ4/db10077"
// First run
// ============================================================================
// rocks/benchmark/RocksCopyBenchmark.cpp relative time/iter iters/s
// ============================================================================
// BM_StringCopy 1.73s 576.97m
// BM_StringPiece 103.74% 1.67s 598.55m
// ============================================================================
// Match rate : 1000000 / 1000000
// Second run
// ============================================================================
// rocks/benchmark/RocksCopyBenchmark.cpp relative time/iter iters/s
// ============================================================================
// BM_StringCopy 611.99ms 1.63
// BM_StringPiece 203.76% 300.35ms 3.33
// ============================================================================
// Match rate : 1000000 / 1000000
```
Test Plan: Unit tests
Reviewers: sdong, igor, anthony, yhchiang, rven
Reviewed By: rven
Subscribers: dhruba, lovro, adsharma
Differential Revision: https://reviews.facebook.net/D48999
Summary:
List of changes:
1) Fix the snprintf() usage in cases where wrong variable was used to determine the output buffer size.
2) Remove unnecessary checks before calling delete operator.
3) Increase code correctness by using size_t type when getting vector's size.
4) Unify the coding style by removing namespace::std usage at the top of the file to confirm to the majority usage.
5) Fix various lint errors pointed out by 'arc lint'.
Test Plan:
Code review and build:
git diff
make clean
make -j 32 commit-prereq
arc lint
Reviewers: kradhakrishnan, sdong, rven, anthony, yhchiang, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D51849
Summary:
This diff provides a framework for doing manual
compactions in parallel with other compactions. We now have a deque of manual compactions. We also pass manual compactions as an argument from RunManualCompactions down to
BackgroundCompactions, so that RunManualCompactions can be reentrant.
Parallelism is controlled by the two routines
ConflictingManualCompaction to allow/disallow new parallel/manual
compactions based on already existing ManualCompactions. In this diff, by default manual compactions still have to run exclusive of other compactions. However, by setting the compaction option, exclusive_manual_compaction to false, it is possible to run other compactions in parallel with a manual compaction. However, we are still restricted to one manual compaction per column family at a time. All of these restrictions will be relaxed in future diffs.
I will be adding more tests later.
Test Plan: Rocksdb regression + new tests + valgrind
Reviewers: igor, anthony, IslamAbdelRahman, kradhakrishnan, yhchiang, sdong
Reviewed By: sdong
Subscribers: yoshinorim, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D47973
Summary:
Currently, transactions can fail even if there is no actual write conflict. This is due to relying on only the memtables to check for write-conflicts. Users have to tune memtable settings to try to avoid this, but it's hard to figure out exactly how to tune these settings.
With this diff, TransactionDB will use both memtables and SST files to determine if there are any write conflicts. This relies on the fact that BlockBasedTable stores sequence numbers for all writes that happen after any open snapshot. Also, D50295 is needed to prevent SingleDelete from disappearing writes (the TODOs in this test code will be fixed once the other diff is approved and merged).
Note that Optimistic transactions will still rely on tuning memtable settings as we do not want to read from SST while on the write thread. Also, memtable settings can still be used to reduce how often TransactionDB needs to read SST files.
Test Plan: unit tests, db bench
Reviewers: rven, yhchiang, kradhakrishnan, IslamAbdelRahman, sdong
Reviewed By: sdong
Subscribers: dhruba, leveldb, yoshinorim
Differential Revision: https://reviews.facebook.net/D50475
Summary: For Transactions, we want to start using the SST files to do write conflict checking. To do this, we need to make sure that compaction never removes all writes if an earlier snapshot exists. So I had to change the way we process SingleDeletes to sometimes leave a SingleDelete behind when we encounter a Put followed by a SingleDelete. See the comments in this diff for a more detailed explanation.
Test Plan: added more unit tests
Reviewers: rven, igor, kradhakrishnan, IslamAbdelRahman, yhchiang, sdong
Reviewed By: sdong
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D50295
Summary: Deprecate options.soft_rate_limit, which is hard to tune, with options.soft_pending_compaction_bytes_limit, which would trigger the slowdown if estimated pending compaction bytes exceeds the threshold. The hope is to make it more striaght-forward to tune.
Test Plan: Modify DBTest.SoftLimit to cover options.soft_pending_compaction_bytes_limit instead; run all unit tests.
Reviewers: IslamAbdelRahman, yhchiang, rven, kradhakrishnan, igor, anthony
Reviewed By: anthony
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D51117
Summary: Introduce a compaction picking priority that picks files who contains the oldest rows to compact. This is a mode that slightly improves write amplification for random update cases.
Test Plan: Add a unit test and run it in valgrind too.
Reviewers: yhchiang, anthony, IslamAbdelRahman, rven, kradhakrishnan, MarkCallaghan, igor
Reviewed By: igor
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D51459
Summary: Now in benchmark "uncompress" in db_bench, we get size from compressed stream for all other compression types except Snappy, where we allocate memory based on parameter. Change it to match to behavior of other compression types.
Test Plan: Run ./db_bench --benchmarks=uncompress with snappy and other compression types.
Reviewers: yhchiang, kradhakrishnan, anthony, IslamAbdelRahman, igor
Reviewed By: igor
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D51681
Summary:
This patch fix a race condition in persisting options which will cause a crash when:
* Thread A obtain cf options and start to persist options based on that cf options.
* Thread B kicks in and finish DropColumnFamily and delete cf_handle.
* Thread A wakes up and tries to finish the persisting options and crashes.
Test Plan: Add a test in column_family_test that can reproduce the crash
Reviewers: anthony, IslamAbdelRahman, rven, kradhakrishnan, sdong
Reviewed By: sdong
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D51717
Summary:
D51183 was reverted due to breaking the LITE build.
This diff is the same as D51183 but with a fix for the LITE BUILD(D51693)
Test Plan: run all unit tests
Reviewers: sdong
Reviewed By: sdong
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D51711
Summary:
Fixing a valgrind failure in DBTestUniversalCompaction
in the IncreaseUniversalCompactionNumLevels test. Using
SpecialSkipList with 10 rows per file.
Test Plan: Run valgrind and functional tests.
Reviewers: anthony, yhchiang, kradhakrishnan, sdong
Reviewed By: sdong
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D51705
Summary:
D50475 enables using SST files for transaction write-conflict checking. In order for this to work, we need to make sure not to compact out SingleDeletes when there is an earlier transaction snapshot(D50295). If there is a long-held snapshot, this could reduce the benefit of the SingleDelete optimization.
This diff allows Transactions to mark snapshots as being used for write-conflict checking. Then, during compaction, we will be able to optimize SingleDeletes better in the future.
This diff adds a flag to SnapshotImpl which is used by Transactions. This diff also passes the earliest write-conflict snapshot's sequence number to CompactionIterator. This diff does not actually change Compaction (after this diff is pushed, D50295 will be able to use this information).
Test Plan: no behavior change, ran existing tests
Reviewers: rven, kradhakrishnan, yhchiang, IslamAbdelRahman, sdong
Reviewed By: sdong
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D51183
Summary: DBTest.DynamicCompactionOptions ocasionally fails during valgrind run. We sent a sleeping task to block compaction thread pool but we don't wait it to run.
Test Plan: Run the test multiple times in an environment which can cause failure.
Reviewers: rven, kradhakrishnan, igor, IslamAbdelRahman, anthony, yhchiang
Reviewed By: yhchiang
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D51687
Summary:
This patch fix a race condition in persisting options which will cause a crash when:
* Thread A obtain cf options and start to persist options based on that cf options.
* Thread B kicks in and finish DropColumnFamily and delete cf_handle.
* Thread A wakes up and tries to finish the persisting options and crashes.
Test Plan: Add a test in column_family_test that can reproduce the crash
Reviewers: anthony, IslamAbdelRahman, rven, kradhakrishnan, sdong
Reviewed By: sdong
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D51609
Summary:
Several tests in db_compaction_test are failing with aborts in
valgrind. These are LevelCompactionThirdPath, LevelCompactionPathUse and
CompressLevelCompaction. We now use the SpecialSkipListFactory to make
them more deterministic
Test Plan: valgrind
Reviewers: anthony, yhchiang, kradhakrishnan, IslamAbdelRahman, sdong
Reviewed By: sdong
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D51663
Summary: After the skip list optimization, ColumnFamilyTest.DifferentWriteBufferSizes can occasionally fail with flush triggering of column family 3. Insert more data to it to make sure flush will trigger.
Test Plan: Run it multiple times with both of jemaloc on and off and see it always passes. (Without thd commit the run with jemalloc fails with chance of about one in two)
Reviewers: rven, yhchiang, IslamAbdelRahman, anthony, kradhakrishnan, igor
Reviewed By: igor
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D51645
Summary:
db_universal_compaction_test is still failing because of
UniversalCompactionNumLevels/DBTestUniversalCompaction.UniversalCompactionSecondPathRatio/0
https://travis-ci.org/facebook/rocksdb/jobs/94949919
Use same approach to fix other tests to fix this test
Test Plan: Run ./db_universal_compaction_test on mac and make sure all the tests pass
Reviewers: kradhakrishnan, yhchiang, rven, anthony, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D51591
Summary: Skip list now cannot estimate memory across allocators
consistently and hence triggers flush at different time. This breaks certain
unit tests.
The fix is to adopt key count instead of size for flush.
Test Plan: Ran test on dev box and mac (where it used to fail)
Reviewers: sdong
CC: leveldb@
Task ID: #9273334
Blame Rev:
Nathan Bronson