Summary:
InlineSkipList::InsertConcurrently should invalidate the
sequential-insertion cache prev_[] for all inserts of multi-level nodes,
not just those that increase the height of the skip list. The invariant
for prev_ is that prev_[i] (i > 0) is supposed to be the predecessor of
prev_[0] at level i. Before this diff InsertConcurrently could violate
this constraint when inserting a multi-level node after prev_[i] but
before prev_[0].
This diff also reenables kConcurrentSkipList as db_test's
MultiThreaded/MultiThreadedDBTest.MultiThreaded/29.
Test Plan:
1. unit tests
2. temporarily hack kConcurrentSkipList timing so that it is fast but has a 1.5% failure rate on my dev box (1ms stagger on thread launch, 1s test duration, failure rate baseline over 1000 runs)
3. observe 1000 passes post-fix
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: MarkCallaghan, dhruba
Differential Revision: https://reviews.facebook.net/D53751
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:
This diff is 2/3 in a sequence that introduces a skip list optimized
for a key that is a freshly-allocated const char*. The change is broken
into pieces to make it easier to review. This piece removes the Key
template type, introduces the AllocateKey interface, and changes the
unit test from using uint64_t as the Key type to using pointers to an 8
byte blob.
Test Plan: unit test
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D51285
Summary:
This diff is 1/3 in a sequence that introduces a skip list optimized for
a key that is a freshly-allocated const char*. The diff is broken into
pieces to make it easier to review. This piece only introduces the new
type by copying the existing SkipList, with mechanical naming changes
and reformatting.
Test Plan: new unit test
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D51279
Summary:
Using a TLS random instance for skiplist makes it smaller
(useful for hash_skiplist_rep) and prepares skiplist for concurrent
adds. This diff also modifies the branching factor math to avoid an
unnecessary division.
This diff has the effect of changing the sequence of skip list node
height choices made by tests, so it has the potential to cause unit
test failures for tests that implicitly rely on the exact structure
of the skip list. Tests that try to exactly trigger a compaction are
likely suspects for this problem (these tests have always been brittle to
changes in the skiplist details). I've minimizes this risk by reseeding
the main thread's Random at the beginning of each test, increasing the
universal compaction size_ratio limit from 101% to 105% for some tests,
and verifying that the tests pass many times.
Test Plan: for i in `seq 0 9`; do make check; done
Reviewers: sdong, igor
Reviewed By: igor
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D50439
Summary:
Key comparison is the single largest CPU user for CPU-bound
workloads. This diff reduces the number of comparisons in two ways.
The first is that it moves predecessor array gathering from
FindGreaterOrEqual to FindLessThan, so that FindGreaterOrEqual can
return immediately if compare_ returns 0. As part of this change I
moved the sequential insertion optimization into Insert, to remove the
undocumented (and smelly) requirement that prev must be equal to prev_
if it is non-null.
The second optimization is that all of the search functions skip calling
compare_ when moving to a lower level that has the same Next pointer.
With a branching factor of 4 we would expect this to happen 1/4 of
the time.
On a single-threaded CPU-bound workload (-benchmarks=fillrandom -threads=1
-batch_size=1 -memtablerep=skip_list -value_size=0 --num=1600000
-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)
on my dev server this is good for a 7% perf win.
Test Plan: unit tests
Reviewers: rven, ljin, yhchiang, sdong, igor
Reviewed By: igor
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D43233
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: RocksDB already depends on C++11, so we might as well all the goodness that C++11 provides. This means that we don't need AtomicPointer anymore. The less things in port/, the easier it will be to port to other platforms.
Test Plan: make check + careful visual review verifying that NoBarried got memory_order_relaxed, while Acquire/Release methods got memory_order_acquire and memory_order_release
Reviewers: rven, yhchiang, ljin, sdong
Reviewed By: ljin
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D27543
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:
Easy thing goes first. This patch moves arena to internal dir; based
on which, the coming patch will deal with memtable_rep.
Test Plan: make check
Reviewers: haobo, sdong, dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15615
Summary: By removing some includes form options.h and reply on forward declaration, we can more easily reason the dependencies.
Test Plan: make all check
Reviewers: kailiu, haobo, igor, dhruba
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15411
Summary: As title. Especially, HashSkipListRepFactory will be able to specify a relatively small height, to reduce the memory overhead of one skiplist per bucket.
Test Plan: make check and test it on leaf4
Reviewers: dhruba, sdong, kailiu
CC: reconnect.grayhat, leveldb
Differential Revision: https://reviews.facebook.net/D14307
Summary: What @haobo done with TransformRep, now in TransformRepNoLock. Similar implementation, except that I made DynamicIterator a subclass of Iterator which makes me have less iterator initializations.
Test Plan: ./prefix_test. Seeing huge savings vs. TransformRep again!
Reviewers: dhruba, haobo, sdong, kailiu
Reviewed By: haobo
CC: leveldb, haobo
Differential Revision: https://reviews.facebook.net/D13953
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: The original optimization missed updating links other than the lowest level.
Test Plan: make check; perf_context_test
Reviewers: dhruba
Reviewed By: dhruba
CC: leveldb, adsharma
Differential Revision: https://reviews.facebook.net/D13119
Summary:
Add an option for arena block size, default value 4096 bytes. Arena will allocate blocks with such size.
I am not sure about passing parameter to skiplist in the new virtualized framework, though I talked to Jim a bit. So add Jim as reviewer.
Test Plan:
new unit test, I am running db_test.
For passing paramter from configured option to Arena, I tried tests like:
TEST(DBTest, Arena_Option) {
std::string dbname = test::TmpDir() + "/db_arena_option_test";
DestroyDB(dbname, Options());
DB* db = nullptr;
Options opts;
opts.create_if_missing = true;
opts.arena_block_size = 1000000; // tested 99, 999999
Status s = DB::Open(opts, dbname, &db);
db->Put(WriteOptions(), "a", "123");
}
and printed some debug info. The results look good. Any suggestion for such a unit-test?
Reviewers: haobo, dhruba, emayanke, jpaton
Reviewed By: dhruba
CC: leveldb, zshao
Differential Revision: https://reviews.facebook.net/D11799
published in https://reviews.facebook.net/D5997.
Summary:
This patch allows compaction to occur in multiple background threads
concurrently.
If a manual compaction is issued, the system falls back to a
single-compaction-thread model. This is done to ensure correctess
and simplicity of code. When the manual compaction is finished,
the system resumes its concurrent-compaction mode automatically.
The updates to the manifest are done via group-commit approach.
Test Plan: run db_bench
Summary:
skiplist doesn't cache the location of the last insert and becomes
CPU bound when the input data has sequential keys.
Notes on thread safety: ::Insert() already requires external
synchronization. So this change is not making it any worse.
Test Plan: skiplist_test
Reviewers: dhruba
Reviewed By: dhruba
Differential Revision: https://reviews.facebook.net/D3129
- Replace raw slice comparison with a call to user comparator.
Added test for custom comparators.
- Fix end of namespace comments.
- Fixed bug in picking inputs for a level-0 compaction.
When finding overlapping files, the covered range may expand
as files are added to the input set. We now correctly expand
the range when this happens instead of continuing to use the
old range. For example, suppose L0 contains files with the
following ranges:
F1: a .. d
F2: c .. g
F3: f .. j
and the initial compaction target is F3. We used to search
for range f..j which yielded {F2,F3}. However we now expand
the range as soon as another file is added. In this case,
when F2 is added, we expand the range to c..j and restart the
search. That picks up file F1 as well.
This change fixes a bug related to deleted keys showing up
incorrectly after a compaction as described in Issue 44.
(Sync with upstream @25072954)
Nathan Bronson