You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
rocksdb/db/write_thread.cc

796 lines
29 KiB

// Copyright (c) 2011-present, Facebook, Inc. All rights reserved.
// This source code is licensed under both the GPLv2 (found in the
// COPYING file in the root directory) and Apache 2.0 License
// (found in the LICENSE.Apache file in the root directory).
#include "db/write_thread.h"
support for concurrent adds to memtable 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
9 years ago
#include <chrono>
#include <thread>
#include "db/column_family.h"
#include "monitoring/perf_context_imp.h"
support for concurrent adds to memtable 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
9 years ago
#include "port/port.h"
#include "test_util/sync_point.h"
#include "util/random.h"
namespace ROCKSDB_NAMESPACE {
WriteThread::WriteThread(const ImmutableDBOptions& db_options)
: max_yield_usec_(db_options.enable_write_thread_adaptive_yield
? db_options.write_thread_max_yield_usec
: 0),
slow_yield_usec_(db_options.write_thread_slow_yield_usec),
allow_concurrent_memtable_write_(
db_options.allow_concurrent_memtable_write),
enable_pipelined_write_(db_options.enable_pipelined_write),
max_write_batch_group_size_bytes(
db_options.max_write_batch_group_size_bytes),
newest_writer_(nullptr),
newest_memtable_writer_(nullptr),
last_sequence_(0),
write_stall_dummy_(),
stall_mu_(),
stall_cv_(&stall_mu_) {}
support for concurrent adds to memtable 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
9 years ago
uint8_t WriteThread::BlockingAwaitState(Writer* w, uint8_t goal_mask) {
// We're going to block. Lazily create the mutex. We guarantee
// propagation of this construction to the waker via the
// STATE_LOCKED_WAITING state. The waker won't try to touch the mutex
// or the condvar unless they CAS away the STATE_LOCKED_WAITING that
// we install below.
w->CreateMutex();
auto state = w->state.load(std::memory_order_acquire);
assert(state != STATE_LOCKED_WAITING);
if ((state & goal_mask) == 0 &&
w->state.compare_exchange_strong(state, STATE_LOCKED_WAITING)) {
// we have permission (and an obligation) to use StateMutex
std::unique_lock<std::mutex> guard(w->StateMutex());
w->StateCV().wait(guard, [w] {
return w->state.load(std::memory_order_relaxed) != STATE_LOCKED_WAITING;
});
state = w->state.load(std::memory_order_relaxed);
}
// else tricky. Goal is met or CAS failed. In the latter case the waker
// must have changed the state, and compare_exchange_strong has updated
// our local variable with the new one. At the moment WriteThread never
// waits for a transition across intermediate states, so we know that
// since a state change has occurred the goal must have been met.
assert((state & goal_mask) != 0);
return state;
}
uint8_t WriteThread::AwaitState(Writer* w, uint8_t goal_mask,
AdaptationContext* ctx) {
uint8_t state = 0;
support for concurrent adds to memtable 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
9 years ago
Fix the overflow bug in AwaitState Summary: https://github.com/facebook/rocksdb/issues/2559 reports an overflow in AwaitState. nbronson has debugged the issue and presented the fix, which is applied to this patch. Moreover this patch adds more comments to clarify the logic in AwaitState. I tried with both 16 and 64 threads on update benchmark. The fix lowers cpu usage by 1.6 but also lowers the throughput by 1.6 and 2% respectively. Apparently the bug had favored using the spinning more often. Benchmarks: TEST_TMPDIR=/dev/shm/tmpdb time ./db_bench --benchmarks="fillrandom" --threads=16 --num=2000000 TEST_TMPDIR=/dev/shm/tmpdb time ./db_bench --use_existing_db=1 --benchmarks="updaterandom[X3]" --threads=16 --num=2000000 TEST_TMPDIR=/dev/shm/tmpdb time ./db_bench --use_existing_db=1 --benchmarks="updaterandom[X3]" --threads=64 --num=200000 Results $ cat update-16t-bug.txt | tail -4 updaterandom [AVG 3 runs] : 234117 ops/sec; 51.8 MB/sec updaterandom [MEDIAN 3 runs] : 233581 ops/sec; 51.7 MB/sec 3896.42user 1539.12system 6:50.61elapsed 1323%CPU (0avgtext+0avgdata 331308maxresident)k 0inputs+0outputs (0major+1281001minor)pagefaults 0swaps $ cat update-16t-fixed.txt | tail -4 updaterandom [AVG 3 runs] : 230364 ops/sec; 51.0 MB/sec updaterandom [MEDIAN 3 runs] : 226169 ops/sec; 50.0 MB/sec 3865.46user 1568.32system 6:57.63elapsed 1301%CPU (0avgtext+0avgdata 315012maxresident)k 0inputs+0outputs (0major+1342568minor)pagefaults 0swaps $ cat update-64t-bug.txt | tail -4 updaterandom [AVG 3 runs] : 261878 ops/sec; 57.9 MB/sec updaterandom [MEDIAN 3 runs] : 262859 ops/sec; 58.2 MB/sec 926.27user 578.06system 2:27.46elapsed 1020%CPU (0avgtext+0avgdata 475480maxresident)k 0inputs+0outputs (0major+1058728minor)pagefaults 0swaps $ cat update-64t-fixed.txt | tail -4 updaterandom [AVG 3 runs] : 256699 ops/sec; 56.8 MB/sec updaterandom [MEDIAN 3 runs] : 256380 ops/sec; 56.7 MB/sec 933.47user 575.37system 2:30.41elapsed 1003%CPU (0avgtext+0avgdata 482340maxresident)k 0inputs+0outputs (0major+1078557minor)pagefaults 0swaps Closes https://github.com/facebook/rocksdb/pull/2679 Differential Revision: D5553732 Pulled By: maysamyabandeh fbshipit-source-id: 98b72dc3a8e0f22ea29d4f7c7790af10c369c5bb
7 years ago
// 1. Busy loop using "pause" for 1 micro sec
// 2. Else SOMETIMES busy loop using "yield" for 100 micro sec (default)
// 3. Else blocking wait
support for concurrent adds to memtable 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
9 years ago
// On a modern Xeon each loop takes about 7 nanoseconds (most of which
// is the effect of the pause instruction), so 200 iterations is a bit
// more than a microsecond. This is long enough that waits longer than
// this can amortize the cost of accessing the clock and yielding.
for (uint32_t tries = 0; tries < 200; ++tries) {
state = w->state.load(std::memory_order_acquire);
if ((state & goal_mask) != 0) {
return state;
}
port::AsmVolatilePause();
}
// This is below the fast path, so that the stat is zero when all writes are
// from the same thread.
PERF_TIMER_GUARD(write_thread_wait_nanos);
support for concurrent adds to memtable 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
9 years ago
// If we're only going to end up waiting a short period of time,
// it can be a lot more efficient to call std::this_thread::yield()
// in a loop than to block in StateMutex(). For reference, on my 4.0
// SELinux test server with support for syscall auditing enabled, the
// minimum latency between FUTEX_WAKE to returning from FUTEX_WAIT is
// 2.7 usec, and the average is more like 10 usec. That can be a big
// drag on RockDB's single-writer design. Of course, spinning is a
// bad idea if other threads are waiting to run or if we're going to
// wait for a long time. How do we decide?
//
// We break waiting into 3 categories: short-uncontended,
// short-contended, and long. If we had an oracle, then we would always
// spin for short-uncontended, always block for long, and our choice for
// short-contended might depend on whether we were trying to optimize
// RocksDB throughput or avoid being greedy with system resources.
//
// Bucketing into short or long is easy by measuring elapsed time.
// Differentiating short-uncontended from short-contended is a bit
// trickier, but not too bad. We could look for involuntary context
// switches using getrusage(RUSAGE_THREAD, ..), but it's less work
// (portability code and CPU) to just look for yield calls that take
// longer than we expect. sched_yield() doesn't actually result in any
// context switch overhead if there are no other runnable processes
// on the current core, in which case it usually takes less than
// a microsecond.
//
// There are two primary tunables here: the threshold between "short"
// and "long" waits, and the threshold at which we suspect that a yield
// is slow enough to indicate we should probably block. If these
// thresholds are chosen well then CPU-bound workloads that don't
// have more threads than cores will experience few context switches
// (voluntary or involuntary), and the total number of context switches
// (voluntary and involuntary) will not be dramatically larger (maybe
// 2x) than the number of voluntary context switches that occur when
// --max_yield_wait_micros=0.
//
// There's another constant, which is the number of slow yields we will
// tolerate before reversing our previous decision. Solitary slow
// yields are pretty common (low-priority small jobs ready to run),
// so this should be at least 2. We set this conservatively to 3 so
// that we can also immediately schedule a ctx adaptation, rather than
// waiting for the next update_ctx.
const size_t kMaxSlowYieldsWhileSpinning = 3;
Fix the overflow bug in AwaitState Summary: https://github.com/facebook/rocksdb/issues/2559 reports an overflow in AwaitState. nbronson has debugged the issue and presented the fix, which is applied to this patch. Moreover this patch adds more comments to clarify the logic in AwaitState. I tried with both 16 and 64 threads on update benchmark. The fix lowers cpu usage by 1.6 but also lowers the throughput by 1.6 and 2% respectively. Apparently the bug had favored using the spinning more often. Benchmarks: TEST_TMPDIR=/dev/shm/tmpdb time ./db_bench --benchmarks="fillrandom" --threads=16 --num=2000000 TEST_TMPDIR=/dev/shm/tmpdb time ./db_bench --use_existing_db=1 --benchmarks="updaterandom[X3]" --threads=16 --num=2000000 TEST_TMPDIR=/dev/shm/tmpdb time ./db_bench --use_existing_db=1 --benchmarks="updaterandom[X3]" --threads=64 --num=200000 Results $ cat update-16t-bug.txt | tail -4 updaterandom [AVG 3 runs] : 234117 ops/sec; 51.8 MB/sec updaterandom [MEDIAN 3 runs] : 233581 ops/sec; 51.7 MB/sec 3896.42user 1539.12system 6:50.61elapsed 1323%CPU (0avgtext+0avgdata 331308maxresident)k 0inputs+0outputs (0major+1281001minor)pagefaults 0swaps $ cat update-16t-fixed.txt | tail -4 updaterandom [AVG 3 runs] : 230364 ops/sec; 51.0 MB/sec updaterandom [MEDIAN 3 runs] : 226169 ops/sec; 50.0 MB/sec 3865.46user 1568.32system 6:57.63elapsed 1301%CPU (0avgtext+0avgdata 315012maxresident)k 0inputs+0outputs (0major+1342568minor)pagefaults 0swaps $ cat update-64t-bug.txt | tail -4 updaterandom [AVG 3 runs] : 261878 ops/sec; 57.9 MB/sec updaterandom [MEDIAN 3 runs] : 262859 ops/sec; 58.2 MB/sec 926.27user 578.06system 2:27.46elapsed 1020%CPU (0avgtext+0avgdata 475480maxresident)k 0inputs+0outputs (0major+1058728minor)pagefaults 0swaps $ cat update-64t-fixed.txt | tail -4 updaterandom [AVG 3 runs] : 256699 ops/sec; 56.8 MB/sec updaterandom [MEDIAN 3 runs] : 256380 ops/sec; 56.7 MB/sec 933.47user 575.37system 2:30.41elapsed 1003%CPU (0avgtext+0avgdata 482340maxresident)k 0inputs+0outputs (0major+1078557minor)pagefaults 0swaps Closes https://github.com/facebook/rocksdb/pull/2679 Differential Revision: D5553732 Pulled By: maysamyabandeh fbshipit-source-id: 98b72dc3a8e0f22ea29d4f7c7790af10c369c5bb
7 years ago
// Whether the yield approach has any credit in this context. The credit is
// added by yield being succesfull before timing out, and decreased otherwise.
auto& yield_credit = ctx->value;
// Update the yield_credit based on sample runs or right after a hard failure
support for concurrent adds to memtable 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
9 years ago
bool update_ctx = false;
Fix the overflow bug in AwaitState Summary: https://github.com/facebook/rocksdb/issues/2559 reports an overflow in AwaitState. nbronson has debugged the issue and presented the fix, which is applied to this patch. Moreover this patch adds more comments to clarify the logic in AwaitState. I tried with both 16 and 64 threads on update benchmark. The fix lowers cpu usage by 1.6 but also lowers the throughput by 1.6 and 2% respectively. Apparently the bug had favored using the spinning more often. Benchmarks: TEST_TMPDIR=/dev/shm/tmpdb time ./db_bench --benchmarks="fillrandom" --threads=16 --num=2000000 TEST_TMPDIR=/dev/shm/tmpdb time ./db_bench --use_existing_db=1 --benchmarks="updaterandom[X3]" --threads=16 --num=2000000 TEST_TMPDIR=/dev/shm/tmpdb time ./db_bench --use_existing_db=1 --benchmarks="updaterandom[X3]" --threads=64 --num=200000 Results $ cat update-16t-bug.txt | tail -4 updaterandom [AVG 3 runs] : 234117 ops/sec; 51.8 MB/sec updaterandom [MEDIAN 3 runs] : 233581 ops/sec; 51.7 MB/sec 3896.42user 1539.12system 6:50.61elapsed 1323%CPU (0avgtext+0avgdata 331308maxresident)k 0inputs+0outputs (0major+1281001minor)pagefaults 0swaps $ cat update-16t-fixed.txt | tail -4 updaterandom [AVG 3 runs] : 230364 ops/sec; 51.0 MB/sec updaterandom [MEDIAN 3 runs] : 226169 ops/sec; 50.0 MB/sec 3865.46user 1568.32system 6:57.63elapsed 1301%CPU (0avgtext+0avgdata 315012maxresident)k 0inputs+0outputs (0major+1342568minor)pagefaults 0swaps $ cat update-64t-bug.txt | tail -4 updaterandom [AVG 3 runs] : 261878 ops/sec; 57.9 MB/sec updaterandom [MEDIAN 3 runs] : 262859 ops/sec; 58.2 MB/sec 926.27user 578.06system 2:27.46elapsed 1020%CPU (0avgtext+0avgdata 475480maxresident)k 0inputs+0outputs (0major+1058728minor)pagefaults 0swaps $ cat update-64t-fixed.txt | tail -4 updaterandom [AVG 3 runs] : 256699 ops/sec; 56.8 MB/sec updaterandom [MEDIAN 3 runs] : 256380 ops/sec; 56.7 MB/sec 933.47user 575.37system 2:30.41elapsed 1003%CPU (0avgtext+0avgdata 482340maxresident)k 0inputs+0outputs (0major+1078557minor)pagefaults 0swaps Closes https://github.com/facebook/rocksdb/pull/2679 Differential Revision: D5553732 Pulled By: maysamyabandeh fbshipit-source-id: 98b72dc3a8e0f22ea29d4f7c7790af10c369c5bb
7 years ago
// Should we reinforce the yield credit
support for concurrent adds to memtable 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
9 years ago
bool would_spin_again = false;
Fix the overflow bug in AwaitState Summary: https://github.com/facebook/rocksdb/issues/2559 reports an overflow in AwaitState. nbronson has debugged the issue and presented the fix, which is applied to this patch. Moreover this patch adds more comments to clarify the logic in AwaitState. I tried with both 16 and 64 threads on update benchmark. The fix lowers cpu usage by 1.6 but also lowers the throughput by 1.6 and 2% respectively. Apparently the bug had favored using the spinning more often. Benchmarks: TEST_TMPDIR=/dev/shm/tmpdb time ./db_bench --benchmarks="fillrandom" --threads=16 --num=2000000 TEST_TMPDIR=/dev/shm/tmpdb time ./db_bench --use_existing_db=1 --benchmarks="updaterandom[X3]" --threads=16 --num=2000000 TEST_TMPDIR=/dev/shm/tmpdb time ./db_bench --use_existing_db=1 --benchmarks="updaterandom[X3]" --threads=64 --num=200000 Results $ cat update-16t-bug.txt | tail -4 updaterandom [AVG 3 runs] : 234117 ops/sec; 51.8 MB/sec updaterandom [MEDIAN 3 runs] : 233581 ops/sec; 51.7 MB/sec 3896.42user 1539.12system 6:50.61elapsed 1323%CPU (0avgtext+0avgdata 331308maxresident)k 0inputs+0outputs (0major+1281001minor)pagefaults 0swaps $ cat update-16t-fixed.txt | tail -4 updaterandom [AVG 3 runs] : 230364 ops/sec; 51.0 MB/sec updaterandom [MEDIAN 3 runs] : 226169 ops/sec; 50.0 MB/sec 3865.46user 1568.32system 6:57.63elapsed 1301%CPU (0avgtext+0avgdata 315012maxresident)k 0inputs+0outputs (0major+1342568minor)pagefaults 0swaps $ cat update-64t-bug.txt | tail -4 updaterandom [AVG 3 runs] : 261878 ops/sec; 57.9 MB/sec updaterandom [MEDIAN 3 runs] : 262859 ops/sec; 58.2 MB/sec 926.27user 578.06system 2:27.46elapsed 1020%CPU (0avgtext+0avgdata 475480maxresident)k 0inputs+0outputs (0major+1058728minor)pagefaults 0swaps $ cat update-64t-fixed.txt | tail -4 updaterandom [AVG 3 runs] : 256699 ops/sec; 56.8 MB/sec updaterandom [MEDIAN 3 runs] : 256380 ops/sec; 56.7 MB/sec 933.47user 575.37system 2:30.41elapsed 1003%CPU (0avgtext+0avgdata 482340maxresident)k 0inputs+0outputs (0major+1078557minor)pagefaults 0swaps Closes https://github.com/facebook/rocksdb/pull/2679 Differential Revision: D5553732 Pulled By: maysamyabandeh fbshipit-source-id: 98b72dc3a8e0f22ea29d4f7c7790af10c369c5bb
7 years ago
// The samling base for updating the yeild credit. The sampling rate would be
// 1/sampling_base.
const int sampling_base = 256;
support for concurrent adds to memtable 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
9 years ago
if (max_yield_usec_ > 0) {
Fix the overflow bug in AwaitState Summary: https://github.com/facebook/rocksdb/issues/2559 reports an overflow in AwaitState. nbronson has debugged the issue and presented the fix, which is applied to this patch. Moreover this patch adds more comments to clarify the logic in AwaitState. I tried with both 16 and 64 threads on update benchmark. The fix lowers cpu usage by 1.6 but also lowers the throughput by 1.6 and 2% respectively. Apparently the bug had favored using the spinning more often. Benchmarks: TEST_TMPDIR=/dev/shm/tmpdb time ./db_bench --benchmarks="fillrandom" --threads=16 --num=2000000 TEST_TMPDIR=/dev/shm/tmpdb time ./db_bench --use_existing_db=1 --benchmarks="updaterandom[X3]" --threads=16 --num=2000000 TEST_TMPDIR=/dev/shm/tmpdb time ./db_bench --use_existing_db=1 --benchmarks="updaterandom[X3]" --threads=64 --num=200000 Results $ cat update-16t-bug.txt | tail -4 updaterandom [AVG 3 runs] : 234117 ops/sec; 51.8 MB/sec updaterandom [MEDIAN 3 runs] : 233581 ops/sec; 51.7 MB/sec 3896.42user 1539.12system 6:50.61elapsed 1323%CPU (0avgtext+0avgdata 331308maxresident)k 0inputs+0outputs (0major+1281001minor)pagefaults 0swaps $ cat update-16t-fixed.txt | tail -4 updaterandom [AVG 3 runs] : 230364 ops/sec; 51.0 MB/sec updaterandom [MEDIAN 3 runs] : 226169 ops/sec; 50.0 MB/sec 3865.46user 1568.32system 6:57.63elapsed 1301%CPU (0avgtext+0avgdata 315012maxresident)k 0inputs+0outputs (0major+1342568minor)pagefaults 0swaps $ cat update-64t-bug.txt | tail -4 updaterandom [AVG 3 runs] : 261878 ops/sec; 57.9 MB/sec updaterandom [MEDIAN 3 runs] : 262859 ops/sec; 58.2 MB/sec 926.27user 578.06system 2:27.46elapsed 1020%CPU (0avgtext+0avgdata 475480maxresident)k 0inputs+0outputs (0major+1058728minor)pagefaults 0swaps $ cat update-64t-fixed.txt | tail -4 updaterandom [AVG 3 runs] : 256699 ops/sec; 56.8 MB/sec updaterandom [MEDIAN 3 runs] : 256380 ops/sec; 56.7 MB/sec 933.47user 575.37system 2:30.41elapsed 1003%CPU (0avgtext+0avgdata 482340maxresident)k 0inputs+0outputs (0major+1078557minor)pagefaults 0swaps Closes https://github.com/facebook/rocksdb/pull/2679 Differential Revision: D5553732 Pulled By: maysamyabandeh fbshipit-source-id: 98b72dc3a8e0f22ea29d4f7c7790af10c369c5bb
7 years ago
update_ctx = Random::GetTLSInstance()->OneIn(sampling_base);
support for concurrent adds to memtable 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
9 years ago
Fix the overflow bug in AwaitState Summary: https://github.com/facebook/rocksdb/issues/2559 reports an overflow in AwaitState. nbronson has debugged the issue and presented the fix, which is applied to this patch. Moreover this patch adds more comments to clarify the logic in AwaitState. I tried with both 16 and 64 threads on update benchmark. The fix lowers cpu usage by 1.6 but also lowers the throughput by 1.6 and 2% respectively. Apparently the bug had favored using the spinning more often. Benchmarks: TEST_TMPDIR=/dev/shm/tmpdb time ./db_bench --benchmarks="fillrandom" --threads=16 --num=2000000 TEST_TMPDIR=/dev/shm/tmpdb time ./db_bench --use_existing_db=1 --benchmarks="updaterandom[X3]" --threads=16 --num=2000000 TEST_TMPDIR=/dev/shm/tmpdb time ./db_bench --use_existing_db=1 --benchmarks="updaterandom[X3]" --threads=64 --num=200000 Results $ cat update-16t-bug.txt | tail -4 updaterandom [AVG 3 runs] : 234117 ops/sec; 51.8 MB/sec updaterandom [MEDIAN 3 runs] : 233581 ops/sec; 51.7 MB/sec 3896.42user 1539.12system 6:50.61elapsed 1323%CPU (0avgtext+0avgdata 331308maxresident)k 0inputs+0outputs (0major+1281001minor)pagefaults 0swaps $ cat update-16t-fixed.txt | tail -4 updaterandom [AVG 3 runs] : 230364 ops/sec; 51.0 MB/sec updaterandom [MEDIAN 3 runs] : 226169 ops/sec; 50.0 MB/sec 3865.46user 1568.32system 6:57.63elapsed 1301%CPU (0avgtext+0avgdata 315012maxresident)k 0inputs+0outputs (0major+1342568minor)pagefaults 0swaps $ cat update-64t-bug.txt | tail -4 updaterandom [AVG 3 runs] : 261878 ops/sec; 57.9 MB/sec updaterandom [MEDIAN 3 runs] : 262859 ops/sec; 58.2 MB/sec 926.27user 578.06system 2:27.46elapsed 1020%CPU (0avgtext+0avgdata 475480maxresident)k 0inputs+0outputs (0major+1058728minor)pagefaults 0swaps $ cat update-64t-fixed.txt | tail -4 updaterandom [AVG 3 runs] : 256699 ops/sec; 56.8 MB/sec updaterandom [MEDIAN 3 runs] : 256380 ops/sec; 56.7 MB/sec 933.47user 575.37system 2:30.41elapsed 1003%CPU (0avgtext+0avgdata 482340maxresident)k 0inputs+0outputs (0major+1078557minor)pagefaults 0swaps Closes https://github.com/facebook/rocksdb/pull/2679 Differential Revision: D5553732 Pulled By: maysamyabandeh fbshipit-source-id: 98b72dc3a8e0f22ea29d4f7c7790af10c369c5bb
7 years ago
if (update_ctx || yield_credit.load(std::memory_order_relaxed) >= 0) {
support for concurrent adds to memtable 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
9 years ago
// we're updating the adaptation statistics, or spinning has >
// 50% chance of being shorter than max_yield_usec_ and causing no
// involuntary context switches
auto spin_begin = std::chrono::steady_clock::now();
// this variable doesn't include the final yield (if any) that
// causes the goal to be met
size_t slow_yield_count = 0;
auto iter_begin = spin_begin;
while ((iter_begin - spin_begin) <=
std::chrono::microseconds(max_yield_usec_)) {
std::this_thread::yield();
state = w->state.load(std::memory_order_acquire);
if ((state & goal_mask) != 0) {
// success
would_spin_again = true;
break;
}
auto now = std::chrono::steady_clock::now();
if (now == iter_begin ||
now - iter_begin >= std::chrono::microseconds(slow_yield_usec_)) {
// conservatively count it as a slow yield if our clock isn't
// accurate enough to measure the yield duration
++slow_yield_count;
if (slow_yield_count >= kMaxSlowYieldsWhileSpinning) {
Fix the overflow bug in AwaitState Summary: https://github.com/facebook/rocksdb/issues/2559 reports an overflow in AwaitState. nbronson has debugged the issue and presented the fix, which is applied to this patch. Moreover this patch adds more comments to clarify the logic in AwaitState. I tried with both 16 and 64 threads on update benchmark. The fix lowers cpu usage by 1.6 but also lowers the throughput by 1.6 and 2% respectively. Apparently the bug had favored using the spinning more often. Benchmarks: TEST_TMPDIR=/dev/shm/tmpdb time ./db_bench --benchmarks="fillrandom" --threads=16 --num=2000000 TEST_TMPDIR=/dev/shm/tmpdb time ./db_bench --use_existing_db=1 --benchmarks="updaterandom[X3]" --threads=16 --num=2000000 TEST_TMPDIR=/dev/shm/tmpdb time ./db_bench --use_existing_db=1 --benchmarks="updaterandom[X3]" --threads=64 --num=200000 Results $ cat update-16t-bug.txt | tail -4 updaterandom [AVG 3 runs] : 234117 ops/sec; 51.8 MB/sec updaterandom [MEDIAN 3 runs] : 233581 ops/sec; 51.7 MB/sec 3896.42user 1539.12system 6:50.61elapsed 1323%CPU (0avgtext+0avgdata 331308maxresident)k 0inputs+0outputs (0major+1281001minor)pagefaults 0swaps $ cat update-16t-fixed.txt | tail -4 updaterandom [AVG 3 runs] : 230364 ops/sec; 51.0 MB/sec updaterandom [MEDIAN 3 runs] : 226169 ops/sec; 50.0 MB/sec 3865.46user 1568.32system 6:57.63elapsed 1301%CPU (0avgtext+0avgdata 315012maxresident)k 0inputs+0outputs (0major+1342568minor)pagefaults 0swaps $ cat update-64t-bug.txt | tail -4 updaterandom [AVG 3 runs] : 261878 ops/sec; 57.9 MB/sec updaterandom [MEDIAN 3 runs] : 262859 ops/sec; 58.2 MB/sec 926.27user 578.06system 2:27.46elapsed 1020%CPU (0avgtext+0avgdata 475480maxresident)k 0inputs+0outputs (0major+1058728minor)pagefaults 0swaps $ cat update-64t-fixed.txt | tail -4 updaterandom [AVG 3 runs] : 256699 ops/sec; 56.8 MB/sec updaterandom [MEDIAN 3 runs] : 256380 ops/sec; 56.7 MB/sec 933.47user 575.37system 2:30.41elapsed 1003%CPU (0avgtext+0avgdata 482340maxresident)k 0inputs+0outputs (0major+1078557minor)pagefaults 0swaps Closes https://github.com/facebook/rocksdb/pull/2679 Differential Revision: D5553732 Pulled By: maysamyabandeh fbshipit-source-id: 98b72dc3a8e0f22ea29d4f7c7790af10c369c5bb
7 years ago
// Not just one ivcsw, but several. Immediately update yield_credit
support for concurrent adds to memtable 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
9 years ago
// and fall back to blocking
update_ctx = true;
break;
}
}
iter_begin = now;
}
}
}
if ((state & goal_mask) == 0) {
TEST_SYNC_POINT_CALLBACK("WriteThread::AwaitState:BlockingWaiting", w);
support for concurrent adds to memtable 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
9 years ago
state = BlockingAwaitState(w, goal_mask);
}
if (update_ctx) {
Fix the overflow bug in AwaitState Summary: https://github.com/facebook/rocksdb/issues/2559 reports an overflow in AwaitState. nbronson has debugged the issue and presented the fix, which is applied to this patch. Moreover this patch adds more comments to clarify the logic in AwaitState. I tried with both 16 and 64 threads on update benchmark. The fix lowers cpu usage by 1.6 but also lowers the throughput by 1.6 and 2% respectively. Apparently the bug had favored using the spinning more often. Benchmarks: TEST_TMPDIR=/dev/shm/tmpdb time ./db_bench --benchmarks="fillrandom" --threads=16 --num=2000000 TEST_TMPDIR=/dev/shm/tmpdb time ./db_bench --use_existing_db=1 --benchmarks="updaterandom[X3]" --threads=16 --num=2000000 TEST_TMPDIR=/dev/shm/tmpdb time ./db_bench --use_existing_db=1 --benchmarks="updaterandom[X3]" --threads=64 --num=200000 Results $ cat update-16t-bug.txt | tail -4 updaterandom [AVG 3 runs] : 234117 ops/sec; 51.8 MB/sec updaterandom [MEDIAN 3 runs] : 233581 ops/sec; 51.7 MB/sec 3896.42user 1539.12system 6:50.61elapsed 1323%CPU (0avgtext+0avgdata 331308maxresident)k 0inputs+0outputs (0major+1281001minor)pagefaults 0swaps $ cat update-16t-fixed.txt | tail -4 updaterandom [AVG 3 runs] : 230364 ops/sec; 51.0 MB/sec updaterandom [MEDIAN 3 runs] : 226169 ops/sec; 50.0 MB/sec 3865.46user 1568.32system 6:57.63elapsed 1301%CPU (0avgtext+0avgdata 315012maxresident)k 0inputs+0outputs (0major+1342568minor)pagefaults 0swaps $ cat update-64t-bug.txt | tail -4 updaterandom [AVG 3 runs] : 261878 ops/sec; 57.9 MB/sec updaterandom [MEDIAN 3 runs] : 262859 ops/sec; 58.2 MB/sec 926.27user 578.06system 2:27.46elapsed 1020%CPU (0avgtext+0avgdata 475480maxresident)k 0inputs+0outputs (0major+1058728minor)pagefaults 0swaps $ cat update-64t-fixed.txt | tail -4 updaterandom [AVG 3 runs] : 256699 ops/sec; 56.8 MB/sec updaterandom [MEDIAN 3 runs] : 256380 ops/sec; 56.7 MB/sec 933.47user 575.37system 2:30.41elapsed 1003%CPU (0avgtext+0avgdata 482340maxresident)k 0inputs+0outputs (0major+1078557minor)pagefaults 0swaps Closes https://github.com/facebook/rocksdb/pull/2679 Differential Revision: D5553732 Pulled By: maysamyabandeh fbshipit-source-id: 98b72dc3a8e0f22ea29d4f7c7790af10c369c5bb
7 years ago
// Since our update is sample based, it is ok if a thread overwrites the
// updates by other threads. Thus the update does not have to be atomic.
auto v = yield_credit.load(std::memory_order_relaxed);
support for concurrent adds to memtable 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
9 years ago
// fixed point exponential decay with decay constant 1/1024, with +1
// and -1 scaled to avoid overflow for int32_t
Fix the overflow bug in AwaitState Summary: https://github.com/facebook/rocksdb/issues/2559 reports an overflow in AwaitState. nbronson has debugged the issue and presented the fix, which is applied to this patch. Moreover this patch adds more comments to clarify the logic in AwaitState. I tried with both 16 and 64 threads on update benchmark. The fix lowers cpu usage by 1.6 but also lowers the throughput by 1.6 and 2% respectively. Apparently the bug had favored using the spinning more often. Benchmarks: TEST_TMPDIR=/dev/shm/tmpdb time ./db_bench --benchmarks="fillrandom" --threads=16 --num=2000000 TEST_TMPDIR=/dev/shm/tmpdb time ./db_bench --use_existing_db=1 --benchmarks="updaterandom[X3]" --threads=16 --num=2000000 TEST_TMPDIR=/dev/shm/tmpdb time ./db_bench --use_existing_db=1 --benchmarks="updaterandom[X3]" --threads=64 --num=200000 Results $ cat update-16t-bug.txt | tail -4 updaterandom [AVG 3 runs] : 234117 ops/sec; 51.8 MB/sec updaterandom [MEDIAN 3 runs] : 233581 ops/sec; 51.7 MB/sec 3896.42user 1539.12system 6:50.61elapsed 1323%CPU (0avgtext+0avgdata 331308maxresident)k 0inputs+0outputs (0major+1281001minor)pagefaults 0swaps $ cat update-16t-fixed.txt | tail -4 updaterandom [AVG 3 runs] : 230364 ops/sec; 51.0 MB/sec updaterandom [MEDIAN 3 runs] : 226169 ops/sec; 50.0 MB/sec 3865.46user 1568.32system 6:57.63elapsed 1301%CPU (0avgtext+0avgdata 315012maxresident)k 0inputs+0outputs (0major+1342568minor)pagefaults 0swaps $ cat update-64t-bug.txt | tail -4 updaterandom [AVG 3 runs] : 261878 ops/sec; 57.9 MB/sec updaterandom [MEDIAN 3 runs] : 262859 ops/sec; 58.2 MB/sec 926.27user 578.06system 2:27.46elapsed 1020%CPU (0avgtext+0avgdata 475480maxresident)k 0inputs+0outputs (0major+1058728minor)pagefaults 0swaps $ cat update-64t-fixed.txt | tail -4 updaterandom [AVG 3 runs] : 256699 ops/sec; 56.8 MB/sec updaterandom [MEDIAN 3 runs] : 256380 ops/sec; 56.7 MB/sec 933.47user 575.37system 2:30.41elapsed 1003%CPU (0avgtext+0avgdata 482340maxresident)k 0inputs+0outputs (0major+1078557minor)pagefaults 0swaps Closes https://github.com/facebook/rocksdb/pull/2679 Differential Revision: D5553732 Pulled By: maysamyabandeh fbshipit-source-id: 98b72dc3a8e0f22ea29d4f7c7790af10c369c5bb
7 years ago
//
// On each update the positive credit is decayed by a facor of 1/1024 (i.e.,
// 0.1%). If the sampled yield was successful, the credit is also increased
// by X. Setting X=2^17 ensures that the credit never exceeds
// 2^17*2^10=2^27, which is lower than 2^31 the upperbound of int32_t. Same
// logic applies to negative credits.
v = v - (v / 1024) + (would_spin_again ? 1 : -1) * 131072;
yield_credit.store(v, std::memory_order_relaxed);
support for concurrent adds to memtable 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
9 years ago
}
assert((state & goal_mask) != 0);
return state;
}
support for concurrent adds to memtable 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
9 years ago
void WriteThread::SetState(Writer* w, uint8_t new_state) {
auto state = w->state.load(std::memory_order_acquire);
if (state == STATE_LOCKED_WAITING ||
!w->state.compare_exchange_strong(state, new_state)) {
assert(state == STATE_LOCKED_WAITING);
std::lock_guard<std::mutex> guard(w->StateMutex());
assert(w->state.load(std::memory_order_relaxed) != new_state);
w->state.store(new_state, std::memory_order_relaxed);
w->StateCV().notify_one();
}
}
bool WriteThread::LinkOne(Writer* w, std::atomic<Writer*>* newest_writer) {
assert(newest_writer != nullptr);
support for concurrent adds to memtable 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
9 years ago
assert(w->state == STATE_INIT);
Writer* writers = newest_writer->load(std::memory_order_relaxed);
while (true) {
// If write stall in effect, and w->no_slowdown is not true,
// block here until stall is cleared. If its true, then return
// immediately
if (writers == &write_stall_dummy_) {
if (w->no_slowdown) {
w->status = Status::Incomplete("Write stall");
SetState(w, STATE_COMPLETED);
return false;
}
// Since no_slowdown is false, wait here to be notified of the write
// stall clearing
{
MutexLock lock(&stall_mu_);
writers = newest_writer->load(std::memory_order_relaxed);
if (writers == &write_stall_dummy_) {
stall_cv_.Wait();
// Load newest_writers_ again since it may have changed
writers = newest_writer->load(std::memory_order_relaxed);
continue;
}
}
}
w->link_older = writers;
if (newest_writer->compare_exchange_weak(writers, w)) {
return (writers == nullptr);
}
}
}
bool WriteThread::LinkGroup(WriteGroup& write_group,
std::atomic<Writer*>* newest_writer) {
assert(newest_writer != nullptr);
Writer* leader = write_group.leader;
Writer* last_writer = write_group.last_writer;
Writer* w = last_writer;
while (true) {
// Unset link_newer pointers to make sure when we call
// CreateMissingNewerLinks later it create all missing links.
w->link_newer = nullptr;
w->write_group = nullptr;
if (w == leader) {
break;
}
w = w->link_older;
}
Writer* newest = newest_writer->load(std::memory_order_relaxed);
while (true) {
leader->link_older = newest;
if (newest_writer->compare_exchange_weak(newest, last_writer)) {
return (newest == nullptr);
}
}
}
void WriteThread::CreateMissingNewerLinks(Writer* head) {
while (true) {
Writer* next = head->link_older;
if (next == nullptr || next->link_newer != nullptr) {
assert(next == nullptr || next->link_newer == head);
break;
}
next->link_newer = head;
head = next;
}
}
Fix write get stuck when pipelined write is enabled (#4143) Summary: Fix the issue when pipelined write is enabled, writers can get stuck indefinitely and not able to finish the write. It can show with the following example: Assume there are 4 writers W1, W2, W3, W4 (W1 is the first, W4 is the last). T1: all writers pending in WAL writer queue: WAL writer queue: W1, W2, W3, W4 memtable writer queue: empty T2. W1 finish WAL writer and move to memtable writer queue: WAL writer queue: W2, W3, W4, memtable writer queue: W1 T3. W2 and W3 finish WAL write as a batch group. W2 enter ExitAsBatchGroupLeader and move the group to memtable writer queue, but before wake up next leader. WAL writer queue: W4 memtable writer queue: W1, W2, W3 T4. W1, W2, W3 finish memtable write as a batch group. Note that W2 still in the previous ExitAsBatchGroupLeader, although W1 have done memtable write for W2. WAL writer queue: W4 memtable writer queue: empty T5. The thread corresponding to W3 create another writer W3' with the same address as W3. WAL writer queue: W4, W3' memtable writer queue: empty T6. W2 continue with ExitAsBatchGroupLeader. Because the address of W3' is the same as W3, the last writer in its group, it thinks there are no pending writers, so it reset newest_writer_ to null, emptying the queue. W4 and W3' are deleted from the queue and will never be wake up. The issue exists since pipelined write was introduced in 5.5.0. Closes #3704 Pull Request resolved: https://github.com/facebook/rocksdb/pull/4143 Differential Revision: D8871599 Pulled By: yiwu-arbug fbshipit-source-id: 3502674e51066a954a0660257e24ac588f815e2a
6 years ago
WriteThread::Writer* WriteThread::FindNextLeader(Writer* from,
Writer* boundary) {
assert(from != nullptr && from != boundary);
Writer* current = from;
while (current->link_older != boundary) {
current = current->link_older;
assert(current != nullptr);
}
return current;
}
void WriteThread::CompleteLeader(WriteGroup& write_group) {
assert(write_group.size > 0);
Writer* leader = write_group.leader;
if (write_group.size == 1) {
write_group.leader = nullptr;
write_group.last_writer = nullptr;
} else {
assert(leader->link_newer != nullptr);
leader->link_newer->link_older = nullptr;
write_group.leader = leader->link_newer;
}
write_group.size -= 1;
SetState(leader, STATE_COMPLETED);
}
void WriteThread::CompleteFollower(Writer* w, WriteGroup& write_group) {
assert(write_group.size > 1);
assert(w != write_group.leader);
if (w == write_group.last_writer) {
w->link_older->link_newer = nullptr;
write_group.last_writer = w->link_older;
} else {
w->link_older->link_newer = w->link_newer;
w->link_newer->link_older = w->link_older;
}
write_group.size -= 1;
SetState(w, STATE_COMPLETED);
}
void WriteThread::BeginWriteStall() {
LinkOne(&write_stall_dummy_, &newest_writer_);
// Walk writer list until w->write_group != nullptr. The current write group
// will not have a mix of slowdown/no_slowdown, so its ok to stop at that
// point
Writer* w = write_stall_dummy_.link_older;
Writer* prev = &write_stall_dummy_;
while (w != nullptr && w->write_group == nullptr) {
if (w->no_slowdown) {
prev->link_older = w->link_older;
w->status = Status::Incomplete("Write stall");
SetState(w, STATE_COMPLETED);
// Only update `link_newer` if it's already set.
// `CreateMissingNewerLinks()` will update the nullptr `link_newer` later,
// which assumes the the first non-nullptr `link_newer` is the last
// nullptr link in the writer list.
// If `link_newer` is set here, `CreateMissingNewerLinks()` may stop
// updating the whole list when it sees the first non nullptr link.
if (prev->link_older && prev->link_older->link_newer) {
prev->link_older->link_newer = prev;
}
w = prev->link_older;
} else {
prev = w;
w = w->link_older;
}
}
}
void WriteThread::EndWriteStall() {
MutexLock lock(&stall_mu_);
// Unlink write_stall_dummy_ from the write queue. This will unblock
// pending write threads to enqueue themselves
assert(newest_writer_.load(std::memory_order_relaxed) == &write_stall_dummy_);
assert(write_stall_dummy_.link_older != nullptr);
write_stall_dummy_.link_older->link_newer = write_stall_dummy_.link_newer;
newest_writer_.exchange(write_stall_dummy_.link_older);
// Wake up writers
stall_cv_.SignalAll();
}
static WriteThread::AdaptationContext jbg_ctx("JoinBatchGroup");
void WriteThread::JoinBatchGroup(Writer* w) {
TEST_SYNC_POINT_CALLBACK("WriteThread::JoinBatchGroup:Start", w);
assert(w->batch != nullptr);
bool linked_as_leader = LinkOne(w, &newest_writer_);
if (linked_as_leader) {
SetState(w, STATE_GROUP_LEADER);
}
TEST_SYNC_POINT_CALLBACK("WriteThread::JoinBatchGroup:Wait", w);
support for concurrent adds to memtable 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
9 years ago
if (!linked_as_leader) {
/**
* Wait util:
* 1) An existing leader pick us as the new leader when it finishes
* 2) An existing leader pick us as its follewer and
* 2.1) finishes the memtable writes on our behalf
* 2.2) Or tell us to finish the memtable writes in pralallel
* 3) (pipelined write) An existing leader pick us as its follower and
* finish book-keeping and WAL write for us, enqueue us as pending
* memtable writer, and
* 3.1) we become memtable writer group leader, or
* 3.2) an existing memtable writer group leader tell us to finish memtable
* writes in parallel.
*/
TEST_SYNC_POINT_CALLBACK("WriteThread::JoinBatchGroup:BeganWaiting", w);
AwaitState(w, STATE_GROUP_LEADER | STATE_MEMTABLE_WRITER_LEADER |
STATE_PARALLEL_MEMTABLE_WRITER | STATE_COMPLETED,
&jbg_ctx);
TEST_SYNC_POINT_CALLBACK("WriteThread::JoinBatchGroup:DoneWaiting", w);
}
}
size_t WriteThread::EnterAsBatchGroupLeader(Writer* leader,
WriteGroup* write_group) {
assert(leader->link_older == nullptr);
assert(leader->batch != nullptr);
assert(write_group != nullptr);
size_t size = WriteBatchInternal::ByteSize(leader->batch);
// Allow the group to grow up to a maximum size, but if the
// original write is small, limit the growth so we do not slow
// down the small write too much.
size_t max_size = max_write_batch_group_size_bytes;
const uint64_t min_batch_size_bytes = max_write_batch_group_size_bytes / 8;
if (size <= min_batch_size_bytes) {
max_size = size + min_batch_size_bytes;
}
leader->write_group = write_group;
write_group->leader = leader;
write_group->last_writer = leader;
write_group->size = 1;
Writer* newest_writer = newest_writer_.load(std::memory_order_acquire);
// This is safe regardless of any db mutex status of the caller. Previous
// calls to ExitAsGroupLeader either didn't call CreateMissingNewerLinks
// (they emptied the list and then we added ourself as leader) or had to
support for concurrent adds to memtable 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
9 years ago
// explicitly wake us up (the list was non-empty when we added ourself,
// so we have already received our MarkJoined).
CreateMissingNewerLinks(newest_writer);
// Tricky. Iteration start (leader) is exclusive and finish
// (newest_writer) is inclusive. Iteration goes from old to new.
Writer* w = leader;
while (w != newest_writer) {
assert(w->link_newer);
w = w->link_newer;
if (w->sync && !leader->sync) {
// Do not include a sync write into a batch handled by a non-sync write.
break;
}
if (w->no_slowdown != leader->no_slowdown) {
// Do not mix writes that are ok with delays with the ones that
// request fail on delays.
break;
}
if (w->disable_wal != leader->disable_wal) {
// Do not mix writes that enable WAL with the ones whose
// WAL disabled.
break;
}
if (w->batch == nullptr) {
// Do not include those writes with nullptr batch. Those are not writes,
// those are something else. They want to be alone
break;
}
if (w->callback != nullptr && !w->callback->AllowWriteBatching()) {
// don't batch writes that don't want to be batched
break;
}
auto batch_size = WriteBatchInternal::ByteSize(w->batch);
if (size + batch_size > max_size) {
// Do not make batch too big
break;
}
w->write_group = write_group;
size += batch_size;
write_group->last_writer = w;
write_group->size++;
}
TEST_SYNC_POINT_CALLBACK("WriteThread::EnterAsBatchGroupLeader:End", w);
return size;
}
void WriteThread::EnterAsMemTableWriter(Writer* leader,
WriteGroup* write_group) {
assert(leader != nullptr);
assert(leader->link_older == nullptr);
assert(leader->batch != nullptr);
assert(write_group != nullptr);
size_t size = WriteBatchInternal::ByteSize(leader->batch);
// Allow the group to grow up to a maximum size, but if the
// original write is small, limit the growth so we do not slow
// down the small write too much.
size_t max_size = max_write_batch_group_size_bytes;
const uint64_t min_batch_size_bytes = max_write_batch_group_size_bytes / 8;
if (size <= min_batch_size_bytes) {
max_size = size + min_batch_size_bytes;
}
leader->write_group = write_group;
write_group->leader = leader;
write_group->size = 1;
Writer* last_writer = leader;
if (!allow_concurrent_memtable_write_ || !leader->batch->HasMerge()) {
Writer* newest_writer = newest_memtable_writer_.load();
CreateMissingNewerLinks(newest_writer);
Writer* w = leader;
while (w != newest_writer) {
assert(w->link_newer);
w = w->link_newer;
if (w->batch == nullptr) {
break;
}
if (w->batch->HasMerge()) {
break;
}
support for concurrent adds to memtable 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
9 years ago
if (!allow_concurrent_memtable_write_) {
auto batch_size = WriteBatchInternal::ByteSize(w->batch);
if (size + batch_size > max_size) {
// Do not make batch too big
break;
}
size += batch_size;
}
support for concurrent adds to memtable 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
9 years ago
w->write_group = write_group;
last_writer = w;
write_group->size++;
}
}
write_group->last_writer = last_writer;
write_group->last_sequence =
last_writer->sequence + WriteBatchInternal::Count(last_writer->batch) - 1;
}
support for concurrent adds to memtable 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
9 years ago
void WriteThread::ExitAsMemTableWriter(Writer* /*self*/,
WriteGroup& write_group) {
Writer* leader = write_group.leader;
Writer* last_writer = write_group.last_writer;
Writer* newest_writer = last_writer;
if (!newest_memtable_writer_.compare_exchange_strong(newest_writer,
nullptr)) {
CreateMissingNewerLinks(newest_writer);
Writer* next_leader = last_writer->link_newer;
assert(next_leader != nullptr);
next_leader->link_older = nullptr;
SetState(next_leader, STATE_MEMTABLE_WRITER_LEADER);
}
Writer* w = leader;
while (true) {
if (!write_group.status.ok()) {
w->status = write_group.status;
}
Writer* next = w->link_newer;
if (w != leader) {
SetState(w, STATE_COMPLETED);
}
if (w == last_writer) {
break;
}
assert(next);
w = next;
}
// Note that leader has to exit last, since it owns the write group.
SetState(leader, STATE_COMPLETED);
}
support for concurrent adds to memtable 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
9 years ago
void WriteThread::LaunchParallelMemTableWriters(WriteGroup* write_group) {
assert(write_group != nullptr);
write_group->running.store(write_group->size);
for (auto w : *write_group) {
SetState(w, STATE_PARALLEL_MEMTABLE_WRITER);
support for concurrent adds to memtable 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
9 years ago
}
}
static WriteThread::AdaptationContext cpmtw_ctx("CompleteParallelMemTableWriter");
// This method is called by both the leader and parallel followers
bool WriteThread::CompleteParallelMemTableWriter(Writer* w) {
support for concurrent adds to memtable 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
9 years ago
auto* write_group = w->write_group;
support for concurrent adds to memtable 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
9 years ago
if (!w->status.ok()) {
std::lock_guard<std::mutex> guard(write_group->leader->StateMutex());
write_group->status = w->status;
support for concurrent adds to memtable 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
9 years ago
}
if (write_group->running-- > 1) {
support for concurrent adds to memtable 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
9 years ago
// we're not the last one
AwaitState(w, STATE_COMPLETED, &cpmtw_ctx);
support for concurrent adds to memtable 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
9 years ago
return false;
}
// else we're the last parallel worker and should perform exit duties.
w->status = write_group->status;
// Callers of this function must ensure w->status is checked.
write_group->status.PermitUncheckedError();
return true;
support for concurrent adds to memtable 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
9 years ago
}
void WriteThread::ExitAsBatchGroupFollower(Writer* w) {
auto* write_group = w->write_group;
support for concurrent adds to memtable 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
9 years ago
assert(w->state == STATE_PARALLEL_MEMTABLE_WRITER);
assert(write_group->status.ok());
ExitAsBatchGroupLeader(*write_group, write_group->status);
support for concurrent adds to memtable 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
9 years ago
assert(w->status.ok());
assert(w->state == STATE_COMPLETED);
SetState(write_group->leader, STATE_COMPLETED);
support for concurrent adds to memtable 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
9 years ago
}
static WriteThread::AdaptationContext eabgl_ctx("ExitAsBatchGroupLeader");
void WriteThread::ExitAsBatchGroupLeader(WriteGroup& write_group,
Status& status) {
Writer* leader = write_group.leader;
Writer* last_writer = write_group.last_writer;
assert(leader->link_older == nullptr);
// If status is non-ok already, then write_group.status won't have the chance
// of being propagated to caller.
if (!status.ok()) {
write_group.status.PermitUncheckedError();
}
// Propagate memtable write error to the whole group.
if (status.ok() && !write_group.status.ok()) {
status = write_group.status;
}
if (enable_pipelined_write_) {
// Notify writers don't write to memtable to exit.
for (Writer* w = last_writer; w != leader;) {
Writer* next = w->link_older;
w->status = status;
if (!w->ShouldWriteToMemtable()) {
CompleteFollower(w, write_group);
}
w = next;
}
if (!leader->ShouldWriteToMemtable()) {
CompleteLeader(write_group);
}
Fix write get stuck when pipelined write is enabled (#4143) Summary: Fix the issue when pipelined write is enabled, writers can get stuck indefinitely and not able to finish the write. It can show with the following example: Assume there are 4 writers W1, W2, W3, W4 (W1 is the first, W4 is the last). T1: all writers pending in WAL writer queue: WAL writer queue: W1, W2, W3, W4 memtable writer queue: empty T2. W1 finish WAL writer and move to memtable writer queue: WAL writer queue: W2, W3, W4, memtable writer queue: W1 T3. W2 and W3 finish WAL write as a batch group. W2 enter ExitAsBatchGroupLeader and move the group to memtable writer queue, but before wake up next leader. WAL writer queue: W4 memtable writer queue: W1, W2, W3 T4. W1, W2, W3 finish memtable write as a batch group. Note that W2 still in the previous ExitAsBatchGroupLeader, although W1 have done memtable write for W2. WAL writer queue: W4 memtable writer queue: empty T5. The thread corresponding to W3 create another writer W3' with the same address as W3. WAL writer queue: W4, W3' memtable writer queue: empty T6. W2 continue with ExitAsBatchGroupLeader. Because the address of W3' is the same as W3, the last writer in its group, it thinks there are no pending writers, so it reset newest_writer_ to null, emptying the queue. W4 and W3' are deleted from the queue and will never be wake up. The issue exists since pipelined write was introduced in 5.5.0. Closes #3704 Pull Request resolved: https://github.com/facebook/rocksdb/pull/4143 Differential Revision: D8871599 Pulled By: yiwu-arbug fbshipit-source-id: 3502674e51066a954a0660257e24ac588f815e2a
6 years ago
Writer* next_leader = nullptr;
// Look for next leader before we call LinkGroup. If there isn't
// pending writers, place a dummy writer at the tail of the queue
// so we know the boundary of the current write group.
Writer dummy;
Writer* expected = last_writer;
bool has_dummy = newest_writer_.compare_exchange_strong(expected, &dummy);
if (!has_dummy) {
// We find at least one pending writer when we insert dummy. We search
// for next leader from there.
next_leader = FindNextLeader(expected, last_writer);
assert(next_leader != nullptr && next_leader != last_writer);
}
// Link the ramaining of the group to memtable writer list.
Fix write get stuck when pipelined write is enabled (#4143) Summary: Fix the issue when pipelined write is enabled, writers can get stuck indefinitely and not able to finish the write. It can show with the following example: Assume there are 4 writers W1, W2, W3, W4 (W1 is the first, W4 is the last). T1: all writers pending in WAL writer queue: WAL writer queue: W1, W2, W3, W4 memtable writer queue: empty T2. W1 finish WAL writer and move to memtable writer queue: WAL writer queue: W2, W3, W4, memtable writer queue: W1 T3. W2 and W3 finish WAL write as a batch group. W2 enter ExitAsBatchGroupLeader and move the group to memtable writer queue, but before wake up next leader. WAL writer queue: W4 memtable writer queue: W1, W2, W3 T4. W1, W2, W3 finish memtable write as a batch group. Note that W2 still in the previous ExitAsBatchGroupLeader, although W1 have done memtable write for W2. WAL writer queue: W4 memtable writer queue: empty T5. The thread corresponding to W3 create another writer W3' with the same address as W3. WAL writer queue: W4, W3' memtable writer queue: empty T6. W2 continue with ExitAsBatchGroupLeader. Because the address of W3' is the same as W3, the last writer in its group, it thinks there are no pending writers, so it reset newest_writer_ to null, emptying the queue. W4 and W3' are deleted from the queue and will never be wake up. The issue exists since pipelined write was introduced in 5.5.0. Closes #3704 Pull Request resolved: https://github.com/facebook/rocksdb/pull/4143 Differential Revision: D8871599 Pulled By: yiwu-arbug fbshipit-source-id: 3502674e51066a954a0660257e24ac588f815e2a
6 years ago
//
// We have to link our group to memtable writer queue before wake up the
// next leader or set newest_writer_ to null, otherwise the next leader
// can run ahead of us and link to memtable writer queue before we do.
if (write_group.size > 0) {
if (LinkGroup(write_group, &newest_memtable_writer_)) {
// The leader can now be different from current writer.
SetState(write_group.leader, STATE_MEMTABLE_WRITER_LEADER);
}
}
Fix write get stuck when pipelined write is enabled (#4143) Summary: Fix the issue when pipelined write is enabled, writers can get stuck indefinitely and not able to finish the write. It can show with the following example: Assume there are 4 writers W1, W2, W3, W4 (W1 is the first, W4 is the last). T1: all writers pending in WAL writer queue: WAL writer queue: W1, W2, W3, W4 memtable writer queue: empty T2. W1 finish WAL writer and move to memtable writer queue: WAL writer queue: W2, W3, W4, memtable writer queue: W1 T3. W2 and W3 finish WAL write as a batch group. W2 enter ExitAsBatchGroupLeader and move the group to memtable writer queue, but before wake up next leader. WAL writer queue: W4 memtable writer queue: W1, W2, W3 T4. W1, W2, W3 finish memtable write as a batch group. Note that W2 still in the previous ExitAsBatchGroupLeader, although W1 have done memtable write for W2. WAL writer queue: W4 memtable writer queue: empty T5. The thread corresponding to W3 create another writer W3' with the same address as W3. WAL writer queue: W4, W3' memtable writer queue: empty T6. W2 continue with ExitAsBatchGroupLeader. Because the address of W3' is the same as W3, the last writer in its group, it thinks there are no pending writers, so it reset newest_writer_ to null, emptying the queue. W4 and W3' are deleted from the queue and will never be wake up. The issue exists since pipelined write was introduced in 5.5.0. Closes #3704 Pull Request resolved: https://github.com/facebook/rocksdb/pull/4143 Differential Revision: D8871599 Pulled By: yiwu-arbug fbshipit-source-id: 3502674e51066a954a0660257e24ac588f815e2a
6 years ago
// If we have inserted dummy in the queue, remove it now and check if there
// are pending writer join the queue since we insert the dummy. If so,
// look for next leader again.
if (has_dummy) {
assert(next_leader == nullptr);
expected = &dummy;
bool has_pending_writer =
!newest_writer_.compare_exchange_strong(expected, nullptr);
if (has_pending_writer) {
next_leader = FindNextLeader(expected, &dummy);
assert(next_leader != nullptr && next_leader != &dummy);
}
Fix write get stuck when pipelined write is enabled (#4143) Summary: Fix the issue when pipelined write is enabled, writers can get stuck indefinitely and not able to finish the write. It can show with the following example: Assume there are 4 writers W1, W2, W3, W4 (W1 is the first, W4 is the last). T1: all writers pending in WAL writer queue: WAL writer queue: W1, W2, W3, W4 memtable writer queue: empty T2. W1 finish WAL writer and move to memtable writer queue: WAL writer queue: W2, W3, W4, memtable writer queue: W1 T3. W2 and W3 finish WAL write as a batch group. W2 enter ExitAsBatchGroupLeader and move the group to memtable writer queue, but before wake up next leader. WAL writer queue: W4 memtable writer queue: W1, W2, W3 T4. W1, W2, W3 finish memtable write as a batch group. Note that W2 still in the previous ExitAsBatchGroupLeader, although W1 have done memtable write for W2. WAL writer queue: W4 memtable writer queue: empty T5. The thread corresponding to W3 create another writer W3' with the same address as W3. WAL writer queue: W4, W3' memtable writer queue: empty T6. W2 continue with ExitAsBatchGroupLeader. Because the address of W3' is the same as W3, the last writer in its group, it thinks there are no pending writers, so it reset newest_writer_ to null, emptying the queue. W4 and W3' are deleted from the queue and will never be wake up. The issue exists since pipelined write was introduced in 5.5.0. Closes #3704 Pull Request resolved: https://github.com/facebook/rocksdb/pull/4143 Differential Revision: D8871599 Pulled By: yiwu-arbug fbshipit-source-id: 3502674e51066a954a0660257e24ac588f815e2a
6 years ago
}
if (next_leader != nullptr) {
next_leader->link_older = nullptr;
SetState(next_leader, STATE_GROUP_LEADER);
}
AwaitState(leader, STATE_MEMTABLE_WRITER_LEADER |
STATE_PARALLEL_MEMTABLE_WRITER | STATE_COMPLETED,
&eabgl_ctx);
} else {
Writer* head = newest_writer_.load(std::memory_order_acquire);
if (head != last_writer ||
!newest_writer_.compare_exchange_strong(head, nullptr)) {
// Either w wasn't the head during the load(), or it was the head
// during the load() but somebody else pushed onto the list before
// we did the compare_exchange_strong (causing it to fail). In the
// latter case compare_exchange_strong has the effect of re-reading
// its first param (head). No need to retry a failing CAS, because
// only a departing leader (which we are at the moment) can remove
// nodes from the list.
assert(head != last_writer);
// After walking link_older starting from head (if not already done)
// we will be able to traverse w->link_newer below. This function
// can only be called from an active leader, only a leader can
// clear newest_writer_, we didn't, and only a clear newest_writer_
// could cause the next leader to start their work without a call
// to MarkJoined, so we can definitely conclude that no other leader
// work is going on here (with or without db mutex).
CreateMissingNewerLinks(head);
assert(last_writer->link_newer->link_older == last_writer);
last_writer->link_newer->link_older = nullptr;
// Next leader didn't self-identify, because newest_writer_ wasn't
// nullptr when they enqueued (we were definitely enqueued before them
// and are still in the list). That means leader handoff occurs when
// we call MarkJoined
SetState(last_writer->link_newer, STATE_GROUP_LEADER);
}
// else nobody else was waiting, although there might already be a new
// leader now
while (last_writer != leader) {
assert(last_writer);
last_writer->status = status;
// we need to read link_older before calling SetState, because as soon
// as it is marked committed the other thread's Await may return and
// deallocate the Writer.
auto next = last_writer->link_older;
SetState(last_writer, STATE_COMPLETED);
last_writer = next;
}
}
}
static WriteThread::AdaptationContext eu_ctx("EnterUnbatched");
void WriteThread::EnterUnbatched(Writer* w, InstrumentedMutex* mu) {
assert(w != nullptr && w->batch == nullptr);
mu->Unlock();
bool linked_as_leader = LinkOne(w, &newest_writer_);
support for concurrent adds to memtable 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
9 years ago
if (!linked_as_leader) {
TEST_SYNC_POINT("WriteThread::EnterUnbatched:Wait");
// Last leader will not pick us as a follower since our batch is nullptr
AwaitState(w, STATE_GROUP_LEADER, &eu_ctx);
}
if (enable_pipelined_write_) {
WaitForMemTableWriters();
}
mu->Lock();
}
void WriteThread::ExitUnbatched(Writer* w) {
assert(w != nullptr);
Writer* newest_writer = w;
if (!newest_writer_.compare_exchange_strong(newest_writer, nullptr)) {
CreateMissingNewerLinks(newest_writer);
Writer* next_leader = w->link_newer;
assert(next_leader != nullptr);
next_leader->link_older = nullptr;
SetState(next_leader, STATE_GROUP_LEADER);
}
}
static WriteThread::AdaptationContext wfmw_ctx("WaitForMemTableWriters");
void WriteThread::WaitForMemTableWriters() {
assert(enable_pipelined_write_);
if (newest_memtable_writer_.load() == nullptr) {
return;
}
Writer w;
if (!LinkOne(&w, &newest_memtable_writer_)) {
AwaitState(&w, STATE_MEMTABLE_WRITER_LEADER, &wfmw_ctx);
}
newest_memtable_writer_.store(nullptr);
}
} // namespace ROCKSDB_NAMESPACE