Summary:
db/version_builder.cc:
117 base_vstorage_->InternalComparator();
CID 1351713 (#1 of 1): Uninitialized pointer field (UNINIT_CTOR)
2. uninit_member: Non-static class member field level_zero_cmp_.internal_comparator is not initialized in this constructor nor in any functions that it calls.
db/version_edit.h:
145 FdWithKeyRange()
146 : fd(),
147 smallest_key(),
148 largest_key() {
CID 1418254 (#1 of 1): Uninitialized pointer field (UNINIT_CTOR)
2. uninit_member: Non-static class member file_metadata is not initialized in this constructor nor in any functions that it calls.
149 }
db/version_set.cc:
120 }
CID 1322789 (#1 of 1): Uninitialized pointer field (UNINIT_CTOR)
4. uninit_member: Non-static class member curr_file_level_ is not initialized in this constructor nor in any functions that it calls.
121 }
db/write_batch.cc:
939 assert(cf_mems_);
CID 1419862 (#1 of 1): Uninitialized scalar field (UNINIT_CTOR)
3. uninit_member: Non-static class member rebuilding_trx_seq_ is not initialized in this constructor nor in any functions that it calls.
940 }
Closes https://github.com/facebook/rocksdb/pull/3092
Differential Revision: D6505666
Pulled By: yiwu-arbug
fbshipit-source-id: fd2c68948a0280772691a419d72ac7e190951d86
Summary:
Add PreReleaseCallback to be called at the end of WriteImpl but before publishing the sequence number. The callback is used in WritePrepareTxn to i) update the commit map, ii) update the last published sequence number in the 2nd write queue. It also ensures that all the commits will go to the 2nd queue.
These changes will ensure that the commit map is updated before the sequence number is published and used by reading snapshots. If we use two write queues, the snapshots will use the seq number published by the 2nd queue. If we use one write queue (the default, the snapshots will use the last seq number in the memtable, which also indicates the last published seq number.
Closes https://github.com/facebook/rocksdb/pull/3205
Differential Revision: D6438959
Pulled By: maysamyabandeh
fbshipit-source-id: f8b6c434e94bc5f5ab9cb696879d4c23e2577ab9
Summary:
Refactor the logic around WriteCallback in the write path to clarify when and how exactly we advance the sequence number and making sure it is consistent across the code.
Closes https://github.com/facebook/rocksdb/pull/3168
Differential Revision: D6324312
Pulled By: maysamyabandeh
fbshipit-source-id: 9a34f479561fdb2a5d01ef6d37a28908d03bbe33
Summary:
The sequence number was not properly advanced after a rollback marker. The patch extends the existing unit tests to detect the bug and also fixes it.
Closes https://github.com/facebook/rocksdb/pull/3157
Differential Revision: D6304291
Pulled By: maysamyabandeh
fbshipit-source-id: 1b519c44a5371b802da49c9e32bd00087a8da401
Summary:
When testing rebuilding_trx_ in MemTableInserter might still be set before the tests finishes which would cause ASAN alarms for leaks. This patch deletes the pointers in MemTableInserter destructor.
Closes https://github.com/facebook/rocksdb/pull/3162
Differential Revision: D6317113
Pulled By: maysamyabandeh
fbshipit-source-id: a68be70709a4fff7ac2b768660119311968f9c21
Summary:
Add tests to ensure that WritePrepared and WriteCommitted policies are cross compatible when the db WAL is empty. This is important when the admin want to switch between the policies. In such case, before the switch the admin needs to empty the WAL by i) committing/rollbacking all the pending transactions, ii) FlushMemTables
Closes https://github.com/facebook/rocksdb/pull/3118
Differential Revision: D6227247
Pulled By: maysamyabandeh
fbshipit-source-id: bcde3d92c1e89cda3b9cfa69f6a20af5d8993db7
Summary:
Previously setting `write_buffer_size` with `SetOptions` would only apply to new memtables. An internal user wanted it to take effect immediately, instead of at an arbitrary future point, to prevent OOM.
This PR makes the memtable's size mutable, and makes `SetOptions()` mutate it. There is one case when we preserve the old behavior, which is when memtable prefix bloom filter is enabled and the user is increasing the memtable's capacity. That's because the prefix bloom filter's size is fixed and wouldn't work as well on a larger memtable.
Closes https://github.com/facebook/rocksdb/pull/3119
Differential Revision: D6228304
Pulled By: ajkr
fbshipit-source-id: e44bd9d10a5f8c9d8c464bf7436070bb3eafdfc9
Summary:
GetCommitTimeWriteBatch is currently used to store some state as part of commit in 2PC. In MyRocks it is specifically used to store some data that would be needed only during recovery. So it is not need to be stored in memtable right after each commit.
This patch enables an optimization to write the GetCommitTimeWriteBatch only to the WAL. The batch will be written to memtable during recovery when the WAL is replayed. To cover the case when WAL is deleted after memtable flush, the batch is also buffered and written to memtable right before each memtable flush.
Closes https://github.com/facebook/rocksdb/pull/3071
Differential Revision: D6148023
Pulled By: maysamyabandeh
fbshipit-source-id: 2d09bae5565abe2017c0327421010d5c0d55eaa7
Summary:
Right now in `PutCFImpl` we always increment NUMBER_KEYS_UPDATED counter for both in-place update or insertion. This PR fixes this by using the correct counter for either case.
Closes https://github.com/facebook/rocksdb/pull/2986
Differential Revision: D6016300
Pulled By: miasantreble
fbshipit-source-id: 0aed327522e659450d533d1c47d3a9f568fac65d
Summary:
Add kTypeBlobIndex value type, which will be used by blob db only, to insert a (key, blob_offset) KV pair. The purpose is to
1. Make it possible to open existing rocksdb instance as blob db. Existing value will be of kTypeIndex type, while value inserted by blob db will be of kTypeBlobIndex.
2. Make rocksdb able to detect if the db contains value written by blob db, if so return error.
3. Make it possible to have blob db optionally store value in SST file (with kTypeValue type) or as a blob value (with kTypeBlobIndex type).
The root db (DBImpl) basically pretended kTypeBlobIndex are normal value on write. On Get if is_blob is provided, return whether the value read is of kTypeBlobIndex type, or return Status::NotSupported() status if is_blob is not provided. On scan allow_blob flag is pass and if the flag is true, return wether the value is of kTypeBlobIndex type via iter->IsBlob().
Changes on blob db side will be in a separate patch.
Closes https://github.com/facebook/rocksdb/pull/2886
Differential Revision: D5838431
Pulled By: yiwu-arbug
fbshipit-source-id: 3c5306c62bc13bb11abc03422ec5cbcea1203cca
Summary:
There's no point populating the block cache during this read. The key we read is guaranteed to be overwritten with a new `kValueType` key immediately afterwards, so can't be accessed again. A user was seeing high turnover of data blocks, at least partially due to this.
Closes https://github.com/facebook/rocksdb/pull/2959
Differential Revision: D5961672
Pulled By: ajkr
fbshipit-source-id: e7cb27c156c5db3b32af355c780efb99dbdf087c
Summary:
Recover txns from the WAL. Also added some unit tests.
Closes https://github.com/facebook/rocksdb/pull/2901
Differential Revision: D5859596
Pulled By: maysamyabandeh
fbshipit-source-id: 6424967b231388093b4effffe0a3b1b7ec8caeb0
Summary:
By default the seq number in DB is increased once per written key. WritePrepared txns requires the seq to be increased once per the entire batch so that the seq would be used as the prepare timestamp by which the transaction is identified. Also we need to increase seq for the commit marker since it would give a unique id to the commit timestamp of transactions.
Two unit tests are added to verify our understanding of how the seq should be increased. The recovery path requires much more work and is left to another patch.
Closes https://github.com/facebook/rocksdb/pull/2885
Differential Revision: D5837843
Pulled By: maysamyabandeh
fbshipit-source-id: a08960b93d727e1cf438c254d0c2636fb133cc1c
Summary:
Implement the main body of WritePrepared pseudo code. This includes PrepareInternal and CommitInternal, as well as AddCommitted which updates the commit map. It also provides a IsInSnapshot method that could be later called form the read path to decide if a version is in the read snapshot or it should other be skipped.
This patch lacks unit tests and does not attempt to offer an efficient implementation. The idea is that to have the API specified so that we can work on related tasks in parallel.
Closes https://github.com/facebook/rocksdb/pull/2713
Differential Revision: D5640021
Pulled By: maysamyabandeh
fbshipit-source-id: bfa7a05e8d8498811fab714ce4b9c21530514e1c
Summary:
This reverts the previous commit 1d7048c598, which broke the build.
Did a `git revert 1d7048c`.
Closes https://github.com/facebook/rocksdb/pull/2627
Differential Revision: D5476473
Pulled By: sagar0
fbshipit-source-id: 4756ff5c0dfc88c17eceb00e02c36176de728d06
Summary: This uses `clang-tidy` to comment out unused parameters (in functions, methods and lambdas) in fbcode. Cases that the tool failed to handle are fixed manually.
Reviewed By: igorsugak
Differential Revision: D5454343
fbshipit-source-id: 5dee339b4334e25e963891b519a5aa81fbf627b2
Summary:
Blob db rely on base db returning sequence number through write batch after DB::Write(). However after recent changes to the write path, DB::Writ()e no longer return sequence number in some cases. Fixing it by have WriteBatchInternal::InsertInto() always encode sequence number into write batch.
Stacking on #2375.
Closes https://github.com/facebook/rocksdb/pull/2385
Differential Revision: D5148358
Pulled By: yiwu-arbug
fbshipit-source-id: 8bda0aa07b9334ed03ed381548b39d167dc20c33
Summary:
PipelineWriteImpl is an alternative approach to WriteImpl. In WriteImpl, only one thread is allow to write at the same time. This thread will do both WAL and memtable writes for all write threads in the write group. Pending writers wait in queue until the current writer finishes. In the pipeline write approach, two queue is maintained: one WAL writer queue and one memtable writer queue. All writers (regardless of whether they need to write WAL) will still need to first join the WAL writer queue, and after the house keeping work and WAL writing, they will need to join memtable writer queue if needed. The benefit of this approach is that
1. Writers without memtable writes (e.g. the prepare phase of two phase commit) can exit write thread once WAL write is finish. They don't need to wait for memtable writes in case of group commit.
2. Pending writers only need to wait for previous WAL writer finish to be able to join the write thread, instead of wait also for previous memtable writes.
Merging #2056 and #2058 into this PR.
Closes https://github.com/facebook/rocksdb/pull/2286
Differential Revision: D5054606
Pulled By: yiwu-arbug
fbshipit-source-id: ee5b11efd19d3e39d6b7210937b11cefdd4d1c8d
Summary:
Extend TransactionOptions to include max_write_batch_size which determines the maximum size of the writebatch representation. If memory limit is exceeded, the operation will abort with subcode kMemoryLimit.
Closes https://github.com/facebook/rocksdb/pull/2124
Differential Revision: D4861842
Pulled By: lth
fbshipit-source-id: 46fd172ea67cc90bbba829bf0d70cfab2261c161
Summary:
also did minor refactoring
Closes https://github.com/facebook/rocksdb/pull/2115
Differential Revision: D4855818
Pulled By: maysamyabandeh
fbshipit-source-id: fbca6ac57e5c6677fffe8354f7291e596a50cb77
Summary:
Move some files under util/ to new directories env/, monitoring/ options/ and cache/
Closes https://github.com/facebook/rocksdb/pull/2090
Differential Revision: D4833681
Pulled By: siying
fbshipit-source-id: 2fd8bef
Summary:
MemTableInserter default constructs Post processing info
std::map. However, on Windows with 2015 STL the default
constructed map still dynamically allocates one node
which shows up on a profiler and we loose ~40% throughput
on fillrandom benchmark.
Solution: declare a map as std::aligned storage and optionally
construct.
This addresses https://github.com/facebook/rocksdb/issues/1976
Before:
-------------------------------------------------------------------
Initializing RocksDB Options from command-line flags
DB path: [k:\data\BulkLoadRandom_10M_fillonly]
fillrandom : 2.775 micros/op 360334 ops/sec; 280.4 MB/s
Microseconds per write:
Count: 10000000 Average: 2.7749 StdDev: 39.92
Min: 1 Median: 2.0826 Max: 26051
Percentiles: P50: 2.08 P75: 2.55 P99: 3.55 P99.9: 9.58 P99.99: 51.5**6
------------------------------------------------------
After:
Initializing RocksDB Options from command-line flags
DB path: [k:\data\BulkLoadRandom_10M_fillon
Closes https://github.com/facebook/rocksdb/pull/2011
Differential Revision: D4740823
Pulled By: siying
fbshipit-source-id: 1daaa2c
Summary:
Return an error from DeleteRange() (or Write() if the user is using the
low-level WriteBatch API) if an unsupported table type is configured.
Closes https://github.com/facebook/rocksdb/pull/1519
Differential Revision: D4185933
Pulled By: ajkr
fbshipit-source-id: abcdf84
Summary:
This is a previous fix that has a typo
Closes https://github.com/facebook/rocksdb/pull/1487
Differential Revision: D4157381
Pulled By: lightmark
fbshipit-source-id: f079be8
Summary:
When constructing a write batch a client may now call MarkWalTerminationPoint() on that batch. No batch operations after this call will be added written to the WAL but will still be inserted into the Memtable. This facility is used to remove one of the three WriteImpl calls in 2PC transactions. This produces a ~1% perf improvement.
```
RocksDB - unoptimized 2pc, sync_binlog=1, disable_2pc=off
INFO 2016-08-31 14:30:38,814 [main]: REQUEST PHASE COMPLETED. 75000000 requests done in 2619 seconds. Requests/second = 28628
RocksDB - optimized 2pc , sync_binlog=1, disable_2pc=off
INFO 2016-08-31 16:26:59,442 [main]: REQUEST PHASE COMPLETED. 75000000 requests done in 2581 seconds. Requests/second = 29054
```
Test Plan: Two unit tests added.
Reviewers: sdong, yiwu, IslamAbdelRahman
Reviewed By: yiwu
Subscribers: hermanlee4, dhruba, andrewkr
Differential Revision: https://reviews.facebook.net/D64599
Summary:
Mitigate regression bug of options.max_successive_merges hit during DB Recovery
For https://reviews.facebook.net/D62625
Test Plan: make all check
Reviewers: horuff, sdong
Reviewed By: sdong
Subscribers: andrewkr, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D62655
Summary:
Add API to WriteBatch to store range deletions in its buffer
which are later added to memtable. In the WriteBatch buffer, a range
deletion is encoded as "<optype><CF ID (optional)><begin key><end key>".
With this diff, the range tombstones are stored inline with the data in
the memtable. It's useful for now because the test cases rely on the
data being accessible via memtable. My next step is to store range
tombstones in a separate area in the memtable.
Test Plan: unit tests
Reviewers: IslamAbdelRahman, sdong, wanning
Reviewed By: wanning
Subscribers: andrewkr, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D61401
Summary:
Stale log files can be deleted out of order. This can happen for various reasons. One of the reason is that no data is ever inserted to a column family and we have an optimization to update its log number, but not all the old log files are cleaned up (the case shown in the unit tests added). It can also happen when we simply delete multiple log files out of order.
This causes data corruption because we simply increase seqID after processing the next row and we may end up with writing data with smaller seqID than what is already flushed to memtables.
In DB recovery, for the oldest files we are replaying, if there it contains no data for any column family, we ignore the sequence IDs in the file.
Test Plan: Add two unit tests that fail without the fix.
Reviewers: IslamAbdelRahman, igor, yiwu
Reviewed By: yiwu
Subscribers: hermanlee4, yoshinorim, leveldb, andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D60891
Summary: In concurrent memtable insert case, updating counters in MemTable::Add() can count for 5% CPU usage. By batch all the counters and update in the end of the write batch, the CPU overheads are overhead in the use cases where more than one key is updated in one write batch.
Test Plan:
Write throughput increases 12% with this benchmark setting:
TEST_TMPDIR=/dev/shm/ ./db_bench --benchmarks=fillrandom -disable_auto_compactions -level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999 -num=10000000 --writes=1000000 -max_background_flushes=16 -max_write_buffer_number=16 --threads=64 --batch_size=128 -allow_concurrent_memtable_write -enable_write_thread_adaptive_yield
Reviewers: andrewkr, IslamAbdelRahman, ngbronson, igor
Reviewed By: ngbronson
Subscribers: ngbronson, leveldb, andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D60495
Summary: filter_deltes is not a frequently used feature. Remove it.
Test Plan: Run all test suites.
Reviewers: igor, yhchiang, IslamAbdelRahman
Reviewed By: IslamAbdelRahman
Subscribers: leveldb, andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D59427
Summary:
We have alot of code duplication whenever we call FullMerge we keep duplicating the instrumentation and statistics code
This is a simple diff to refactor the code to use TimedFullMerge instead of FullMerge
Test Plan: COMPILE_WITH_ASAN=1 make check -j64
Reviewers: andrewkr, yhchiang, sdong
Reviewed By: sdong
Subscribers: andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D59577
Summary:
1. prepare()
2. crash
3. recover
4. commit()
5. crash
6. data is lost
This is due to the transaction data still only residing in the WAL but because the logs were flushed on the first recovery the data is ignored on the second recovery. We must scan all logs found on recovery and only ignore redundant data at the time of replay. It is not possible to know which logs still contain relevant data at time of recovery. We cannot simply ignore a log because all of the non-2pc data it contains has already been written to L0.
The changes made to MemTableInserter are to ensure that prepared sections are still recovered even if all of the non-2pc data in that log has already been flushed to L0.
Test Plan: Provided test.
Reviewers: sdong
Subscribers: andrewkr, hermanlee4, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D57729
Summary:
Consider the following WAL with 4 batch entries prefixed with their sequence at time of memtable insert.
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(a)]
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(b)]
[4: COMMIT(a)]
[7: COMMIT(b)]
The first two batches do not consume any sequence numbers so are both prefixed with seq=1.
For 2pc commit, memtable insertion takes place before COMMIT batch is written to WAL.
We can see that sequence number consumption takes place between WAL entries giving us the seemingly sparse sequence prefix for WAL entries.
This is a valid WAL.
Because with 2PC markers one WriteBatch points to another batch containing its inserts a writebatch can consume more or less sequence numbers than the number of sequence consuming entries that it contains.
We can see that, given the entries in the WAL, 6 sequence ids were consumed. Yet on recovery the maximum sequence consumed would be 7 + 3 (the number of sequence numbers consumed by COMMIT(b))
So, now upon recovery we must track the actual consumption of sequence numbers.
In the provided scenario there will be no sequence gaps, but it is possible to produce a sequence gap. This should not be a problem though. correct?
Test Plan: provided test.
Reviewers: sdong
Subscribers: andrewkr, leveldb, dhruba, hermanlee4
Differential Revision: https://reviews.facebook.net/D57645
Summary:
This diff is built on top of WriteBatch modification: https://reviews.facebook.net/D54093 and adds the required functionality to rocksdb core necessary for rocksdb to support 2PC.
modfication of DBImpl::WriteImpl()
- added two arguments *uint64_t log_used = nullptr, uint64_t log_ref = 0;
- *log_used is an output argument which will return the log number which the incoming batch was inserted into, 0 if no WAL insert took place.
- log_ref is a supplied log_number which all memtables inserted into will reference after the batch insert takes place. This number will reside in 'FindMinPrepLogReferencedByMemTable()' until all Memtables insertinto have flushed.
- Recovery/writepath is now aware of prepared batches and commit and rollback markers.
Test Plan: There is currently no test on this diff. All testing of this functionality takes place in the Transaction layer/diff but I will add some testing.
Reviewers: IslamAbdelRahman, sdong
Subscribers: leveldb, santoshb, andrewkr, vasilep, dhruba, hermanlee4
Differential Revision: https://reviews.facebook.net/D56919
Summary: Adds three new WriteBatch data types: Prepare(xid), Commit(xid), Rollback(xid). Prepare(xid) should precede the (single) operation to which is applies. There can obviously be multiple Prepare(xid) markers. There should only be one Rollback(xid) or Commit(xid) marker yet not both. None of this logic is currently enforced and will most likely be implemented further up such as in the memtableinserter. All three markers are similar to PutLogData in that they are writebatch meta-data, ie stored but not counted. All three markers differ from PutLogData in that they will actually be written to disk. As for WriteBatchWithIndex, Prepare, Commit, Rollback are all implemented just as PutLogData and none are tested just as PutLogData.
Test Plan: single unit test in write_batch_test.
Reviewers: hermanlee4, sdong, anthony
Subscribers: leveldb, dhruba, vasilep, andrewkr
Differential Revision: https://reviews.facebook.net/D57867
Summary: Adds three new WriteBatch data types: Prepare(xid), Commit(xid), Rollback(xid). Prepare(xid) should precede the (single) operation to which is applies. There can obviously be multiple Prepare(xid) markers. There should only be one Rollback(xid) or Commit(xid) marker yet not both. None of this logic is currently enforced and will most likely be implemented further up such as in the memtableinserter. All three markers are similar to PutLogData in that they are writebatch meta-data, ie stored but not counted. All three markers differ from PutLogData in that they will actually be written to disk. As for WriteBatchWithIndex, Prepare, Commit, Rollback are all implemented just as PutLogData and none are tested just as PutLogData.
Test Plan: single unit test in write_batch_test.
Reviewers: hermanlee4, sdong, anthony
Subscribers: andrewkr, vasilep, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D54093
Summary:
- Put key offset and key size in WriteBatchIndexEntry
- Use vector for comparators in WriteBatchEntryComparator
I use a slightly modified version of @yoshinorim code to benchmark
https://gist.github.com/IslamAbdelRahman/b120f4fba8d6ff7d58d2
For Put I create a transaction that put a 1000000 keys and measure the time spent without commit.
For GetForUpdate I read the keys that I added in the Put transaction.
Original time:
```
rm -rf /dev/shm/rocksdb-example/
./txn_bench put 1000000
1000000 OK Ops | took 3.679 seconds
./txn_bench get_for_update 1000000
1000000 OK Ops | took 3.940 seconds
```
New Time
```
rm -rf /dev/shm/rocksdb-example/
./txn_bench put 1000000
1000000 OK Ops | took 2.727 seconds
./txn_bench get_for_update 1000000
1000000 OK Ops | took 3.880 seconds
```
It looks like there is no significant improvement in GetForUpdate() but we can see ~30% improvement in Put()
Test Plan: unittests
Reviewers: yhchiang, anthony, sdong
Reviewed By: sdong
Subscribers: andrewkr, dhruba, yoshinorim
Differential Revision: https://reviews.facebook.net/D55539
record.size() should not be less than 12.
This "magic number" seems to be the WriteBatch header (8 byte sequence
and 4 byte count). Replaced all the places where "12" was used
by WriteBatchInternal::kHeader.
Summary:
copy from task 8196669:
1) Optimistic transactions do not support batching writes from different threads.
2) Pessimistic transactions do not support batching writes if an expiration time is set.
In these 2 cases, we currently do not do any write batching in DBImpl::WriteImpl() because there is a WriteCallback that could decide at the last minute to abort the write. But we could support batching write operations with callbacks if we make sure to process the callbacks correctly.
To do this, we would first need to modify write_thread.cc to stop preventing writes with callbacks from being batched together. Then we would need to change DBImpl::WriteImpl() to call all WriteCallback's in a batch, only write the batches that succeed, and correctly set the state of each batch's WriteThread::Writer.
Test Plan: Added test WriteWithCallbackTest to write_callback_test.cc which creates multiple client threads and verifies that writes are batched and executed properly.
Reviewers: hermanlee4, anthony, ngbronson
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D52863
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
Prashant D