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:
This diff completes the creation of InlineSkipList<Cmp>, which is like
SkipList<const char*, Cmp> but it always allocates the key contiguously
with the node. This allows us to remove the pointer from the node
to the key. As a result the memory usage of the skip list is reduced
(by 1 to sizeof(void*) bytes depending on the padding required to align
the key storage), cache locality is improved, and we halve the number
of calls to the allocator.
For skip lists whose keys are freshly-allocated const char*,
InlineSkipList is stricly preferrable to SkipList. This diff doesn't
replace SkipList, however, because some of the use cases of SkipList in
RocksDB are either character sequences that are not allocated at the
same time as the skip list node allocation (for example
hash_linklist_rep) or have different key types (for example
write_batch_with_index). Taking advantage of inline allocation for
those cases is left to future work.
The perf win is biggest for small values. For single-threaded CPU-bound
(32M fillrandom operations with no WAL log) with 16 byte keys and 0 byte
values, the db_bench perf goes from ~310k ops/sec to ~410k ops/sec. For
large values the improvement is less pronounced, but seems to be between
5% and 10% on the same configuration.
Test Plan: make check
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D51123
Summary:
Add an option in GetApproximateSize() so that the result will include estimated sizes in mem tables.
To implement it, implement an estimated count from the beginning to a key in skip list. The approach is to count to find the entry, how many Next() is issued from each level, and sum them with a weight that is <branching factor> ^ <level>.
Test Plan: Add a test case
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D40119
Summary:
Introduces a new class for managing write buffer memory across column
families. We supplement ColumnFamilyOptions::write_buffer_size with
ColumnFamilyOptions::write_buffer, a shared pointer to a WriteBuffer
instance that enforces memory limits before flushing out to disk.
Test Plan: Added SharedWriteBuffer unit test to db_test.cc
Reviewers: sdong, rven, ljin, igor
Reviewed By: igor
Subscribers: tnovak, yhchiang, dhruba, xjin, MarkCallaghan, yoshinorim
Differential Revision: https://reviews.facebook.net/D22581
Summary:
This diff introduces the `lookahead` argument to `SkipListFactory()`. This is an
optimization for the tailing use case which includes many seeks. E.g. consider
the following operations on a skip list iterator:
Seek(x), Next(), Next(), Seek(x+2), Next(), Seek(x+3), Next(), Next(), ...
If `lookahead` is positive, `SkipListRep` will return an iterator which also
keeps track of the previously visited node. Seek() then first does a linear
search starting from that node (up to `lookahead` steps). As in the tailing
example above, this may require fewer than ~log(n) comparisons as with regular
skip list search.
Test Plan:
Added a new benchmark (`fillseekseq`) which simulates the usage pattern. It
first writes N records (with consecutive keys), then measures how much time it
takes to read them by calling `Seek()` and `Next()`.
$ time ./db_bench -num 10000000 -benchmarks fillseekseq -prefix_size 1 \
-key_size 8 -write_buffer_size $[1024*1024*1024] -value_size 50 \
-seekseq_next 2 -skip_list_lookahead=0
[...]
DB path: [/dev/shm/rocksdbtest/dbbench]
fillseekseq : 0.389 micros/op 2569047 ops/sec;
real 0m21.806s
user 0m12.106s
sys 0m9.672s
$ time ./db_bench [...] -skip_list_lookahead=2
[...]
DB path: [/dev/shm/rocksdbtest/dbbench]
fillseekseq : 0.153 micros/op 6540684 ops/sec;
real 0m19.469s
user 0m10.192s
sys 0m9.252s
Reviewers: ljin, sdong, igor
Reviewed By: igor
Subscribers: dhruba, leveldb, march, lovro
Differential Revision: https://reviews.facebook.net/D23997
Summary: It seems to me that when ever function MemTableRep::GetIterator(const Slice& slice) is used, we can use MemTableRep::GetDynamicPrefixIterator() instead. Just delete it to simplify the codes.
Test Plan: make all check
Reviewers: yhchiang, ljin
Reviewed By: ljin
Subscribers: xjin, dhruba, haobo, leveldb
Differential Revision: https://reviews.facebook.net/D19281
Summary:
In this patch, try to allocate the whole iterator tree starting from DBIter from an arena
1. ArenaWrappedDBIter is created when serves as the entry point of an iterator tree, with an arena in it.
2. Add an option to create iterator from arena for following iterators: DBIter, MergingIterator, MemtableIterator, all mem table's iterators, all table reader's iterators and two level iterator.
3. MergeIteratorBuilder is created to incrementally build the tree of internal iterators. It is passed to mem table list and version set and add iterators to it.
Limitations:
(1) Only DB::NewIterator() without tailing uses the arena. Other cases, including readonly DB and compactions are still from malloc
(2) Two level iterator itself is allocated in arena, but not iterators inside it.
Test Plan: make all check
Reviewers: ljin, haobo
Reviewed By: haobo
Subscribers: leveldb, dhruba, yhchiang, igor
Differential Revision: https://reviews.facebook.net/D18513
Summary:
TLB page allocation errors are now logged to info logs, instead of stderr.
In order to do that, mem table rep's factory functions take a info logger now.
Test Plan: make all check
Reviewers: haobo, igor, yhchiang
Reviewed By: yhchiang
CC: leveldb, yhchiang, dhruba
Differential Revision: https://reviews.facebook.net/D18471
Summary: per sdong's request, this will help processor prefetch on n->key case.
Test Plan: make all check
Reviewers: sdong, haobo, igor
Reviewed By: sdong
CC: leveldb
Differential Revision: https://reviews.facebook.net/D17415
Summary:
(1) Fix SanitizeOptions() to also check HashLinkList. The current
dynamic case just happens to work because the 2 classes have the same
layout.
(2) Do not delete SliceTransform object in HashSkipListFactory and
HashLinkListFactory destructor. Reason: SanitizeOptions() enforces
prefix_extractor and SliceTransform to be the same object when
Hash**Factory is used. This makes the behavior strange: when
Hash**Factory is used, prefix_extractor will be released by RocksDB. If
other memtable factory is used, prefix_extractor should be released by
user.
Test Plan: db_bench && make asan_check
Reviewers: haobo, igor, sdong
Reviewed By: igor
CC: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D16587
Summary:
This patch optimized Get() code paths by avoiding malloc of iterators. Iterator creation is moved to mem table rep implementations, where a callback is called when any key is found. This is the same practice as what we do in (SST) table readers.
db_bench result for readrandom following a writeseq, with no compression, single thread and tmpfs, we see throughput improved to 144958 from 139027, about 3%.
Test Plan: make all check
Reviewers: dhruba, haobo, igor
Reviewed By: haobo
CC: leveldb, yhchiang
Differential Revision: https://reviews.facebook.net/D14685
Summary:
shared_ptr is slower than unique_ptr (which literally comes with no performance cost compare with raw pointers).
In memtable and memtable rep, we use shared_ptr when we'd actually should use unique_ptr.
According to igor's previous work, we are likely to make quite some performance gain from this diff.
Test Plan: make check
Reviewers: dhruba, igor, sdong, haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15213
Summary:
For prefix mem tables, encoding mem table key may be unnecessary if the prefix doesn't have any key. This patch is a little bit hacky but I want to try out the performance gain of removing this lazy initialization.
In longer term, we might want to revisit the way we abstract mem tables implementations.
Test Plan: make all check
Reviewers: haobo, igor, kailiu
Reviewed By: igor
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14265
Summary:
Change namespace from leveldb to rocksdb. This allows a single
application to link in open-source leveldb code as well as
rocksdb code into the same process.
Test Plan: compile rocksdb
Reviewers: emayanke
Reviewed By: emayanke
CC: leveldb
Differential Revision: https://reviews.facebook.net/D13287
Summary: Replace include/leveldb with include/rocksdb.
Test Plan:
make clean; make check
make clean; make release
Differential Revision: https://reviews.facebook.net/D12489
Summary:
This patch adds three new MemTableRep's: UnsortedRep, PrefixHashRep, and VectorRep.
UnsortedRep stores keys in an std::unordered_map of std::sets. When an iterator is requested, it dumps the keys into an std::set and iterates over that.
VectorRep stores keys in an std::vector. When an iterator is requested, it creates a copy of the vector and sorts it using std::sort. The iterator accesses that new vector.
PrefixHashRep stores keys in an unordered_map mapping prefixes to ordered sets.
I also added one API change. I added a function MemTableRep::MarkImmutable. This function is called when the rep is added to the immutable list. It doesn't do anything yet, but it seems like that could be useful. In particular, for the vectorrep, it means we could elide the extra copy and just sort in place. The only reason I haven't done that yet is because the use of the ArenaAllocator complicates things (I can elaborate on this if needed).
Test Plan:
make -j32 check
./db_stress --memtablerep=vector
./db_stress --memtablerep=unsorted
./db_stress --memtablerep=prefixhash --prefix_size=10
Reviewers: dhruba, haobo, emayanke
Reviewed By: dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D12117
