2 * Copyright 2017-present Facebook, Inc.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
16 // @author: Xin Liu <xliux@fb.com>
24 #include <folly/Benchmark.h>
25 #include <folly/ConcurrentSkipList.h>
26 #include <folly/hash/Hash.h>
27 #include <folly/portability/GFlags.h>
28 #include <folly/synchronization/RWSpinLock.h>
29 #include <glog/logging.h>
31 DEFINE_int32(num_threads, 12, "num concurrent threads to test");
33 // In some case, we may want to test worker threads operating on multiple
34 // lists. For example in search, not all threads are visiting the same posting
35 // list, but for the ones with some popular terms, they do get multiple
36 // visitors at the same time.
37 DEFINE_int32(num_sets, 1, "num of set to operate on");
39 static const int kInitHeadHeight = 10;
40 static const int kMaxValue = 0x1000000;
44 using namespace folly;
46 typedef int ValueType;
47 typedef ConcurrentSkipList<ValueType> SkipListType;
48 typedef SkipListType::Accessor SkipListAccessor;
49 typedef std::set<ValueType> SetType;
51 static std::vector<ValueType> gData;
52 static void initData() {
53 gData.resize(kMaxValue);
54 for (int i = 0; i < kMaxValue; ++i) {
57 std::shuffle(gData.begin(), gData.end(), std::mt19937{});
60 // single thread benchmarks
61 void BM_IterateOverSet(int iters, int size) {
66 for (int i = 0; i < size; ++i) {
67 a_set.insert(gData[rand() % kMaxValue]);
72 auto iter = a_set.begin();
73 for (int i = 0; i < iters; ++i) {
75 if (iter == a_set.end()) {
80 // VLOG(20) << "sum = " << sum;
84 void BM_IterateSkipList(int iters, int size) {
85 BenchmarkSuspender susp;
87 auto skipList = SkipListType::create(kInitHeadHeight);
88 for (int i = 0; i < size; ++i) {
89 skipList.add(rand() % kMaxValue);
94 auto iter = skipList.begin();
95 for (int i = 0; i < iters; ++i) {
97 if (iter == skipList.end()) {
98 iter = skipList.begin();
103 // VLOG(20) << "sum = " << sum;
107 void BM_SetMerge(int iters, int size) {
108 BenchmarkSuspender susp;
111 for (int i = 0; i < iters; ++i) {
112 a_set.insert(rand() % kMaxValue);
114 for (int i = 0; i < size; ++i) {
115 b_set.insert(rand() % kMaxValue);
119 int64_t mergedSum = 0;
120 FOR_EACH(it, a_set) {
121 if (b_set.find(*it) != b_set.end()) {
126 // VLOG(20) << mergedSum;
130 void BM_CSLMergeLookup(int iters, int size) {
131 BenchmarkSuspender susp;
132 auto skipList = SkipListType::create(kInitHeadHeight);
133 auto skipList2 = SkipListType::create(kInitHeadHeight);
135 for (int i = 0; i < iters; ++i) {
136 skipList.add(rand() % kMaxValue);
138 for (int i = 0; i < size; ++i) {
139 skipList2.add(rand() % kMaxValue);
141 int64_t mergedSum = 0;
144 SkipListType::Skipper skipper(skipList2);
145 FOR_EACH(it, skipList) {
146 if (skipper.to(*it)) {
152 // VLOG(20) << mergedSum;
156 // merge by two skippers
157 void BM_CSLMergeIntersection(int iters, int size) {
158 BenchmarkSuspender susp;
159 auto skipList = SkipListType::create(kInitHeadHeight);
160 auto skipList2 = SkipListType::create(kInitHeadHeight);
161 for (int i = 0; i < iters; ++i) {
162 skipList.add(rand() % kMaxValue);
164 for (int i = 0; i < size; ++i) {
165 skipList2.add(rand() % kMaxValue);
169 SkipListType::Skipper s1(skipList);
170 SkipListType::Skipper s2(skipList2);
172 int64_t mergedSum = 0;
174 while (s1.good() && s2.good()) {
179 } else if (v1 > v2) {
189 // VLOG(20) << mergedSum;
193 void BM_SetContainsNotFound(int iters, int size) {
194 BenchmarkSuspender susp;
196 CHECK_LT(size, kMaxValue);
197 for (int i = 0; i < size; ++i) {
203 for (int i = 0; i < iters; ++i) {
204 sum += (aset.end() == aset.find(2 * i + 1));
212 void BM_SetContainsFound(int iters, int size) {
213 BenchmarkSuspender susp;
215 CHECK_LT(size, kMaxValue);
217 for (int i = 0; i < size; ++i) {
221 std::vector<int> values;
222 for (int i = 0; i < iters; ++i) {
223 values.push_back(rand() % size);
228 for (int i = 0; i < iters; ++i) {
229 sum += (aset.end() == aset.find(values[i]));
237 void BM_CSLContainsFound(int iters, int size) {
238 BenchmarkSuspender susp;
239 auto skipList = SkipListType::create(kInitHeadHeight);
240 CHECK_LT(size, kMaxValue);
242 for (int i = 0; i < size; ++i) {
245 std::vector<int> values;
246 for (int i = 0; i < iters; ++i) {
247 values.push_back(rand() % size);
252 for (int i = 0; i < iters; ++i) {
253 sum += skipList.contains(values[i]);
261 void BM_CSLContainsNotFound(int iters, int size) {
262 BenchmarkSuspender susp;
263 auto skipList = SkipListType::create(kInitHeadHeight);
264 CHECK_LT(size, kMaxValue);
266 for (int i = 0; i < size; ++i) {
272 for (int i = 0; i < iters; ++i) {
273 sum += skipList.contains(2 * i + 1);
281 void BM_AddSet(int iters, int size) {
282 BenchmarkSuspender susp;
284 for (int i = 0; i < size; ++i) {
285 aset.insert(gData[i]);
289 for (int i = size; i < size + iters; ++i) {
290 aset.insert(gData[i]);
294 void BM_AddSkipList(int iters, int size) {
295 BenchmarkSuspender susp;
296 auto skipList = SkipListType::create(kInitHeadHeight);
297 for (int i = 0; i < size; ++i) {
298 skipList.add(gData[i]);
302 for (int i = size; i < size + iters; ++i) {
303 skipList.add(gData[i]);
307 BENCHMARK(Accessor, iters) {
308 BenchmarkSuspender susp;
309 auto skiplist = SkipListType::createInstance(kInitHeadHeight);
310 auto sl = skiplist.get();
313 for (size_t i = 0; i < iters; ++i) {
314 SkipListAccessor accessor(sl);
318 // a benchmark to estimate the
319 // low bound of doing a ref counting for an Accessor
320 BENCHMARK(accessorBasicRefcounting, iters) {
321 BenchmarkSuspender susp;
322 auto* value = new std::atomic<int32_t>();
323 auto* dirty = new std::atomic<int32_t>();
325 folly::MicroSpinLock l;
329 for (size_t i = 0; i < iters; ++i) {
330 value->fetch_add(1, std::memory_order_relaxed);
331 if (dirty->load(std::memory_order_acquire) != 0) {
332 folly::MSLGuard g(l);
334 value->fetch_sub(1, std::memory_order_relaxed);
344 // Data For testing contention benchmark
345 class ConcurrentAccessData {
347 explicit ConcurrentAccessData(int size) :
348 skipList_(SkipListType::create(10)),
349 sets_(FLAGS_num_sets), locks_(FLAGS_num_sets) {
351 for (int i = 0; i < size; ++i) {
356 for (int i = 0; i < FLAGS_num_sets; ++i) {
357 locks_[i] = new RWSpinLock();
363 // This requires knowledge of the C++ library internals. Only use it if we're
364 // using the GNU C++ library.
365 #ifdef _GLIBCXX_SYMVER
367 int64_t setMemorySize = sets_[0].size() * sizeof(*sets_[0].begin()._M_node);
368 int64_t cslMemorySize = 0;
369 for (auto it = skipList_.begin(); it != skipList_.end(); ++it) {
370 cslMemorySize += it.nodeSize();
373 LOG(INFO) << "size=" << sets_[0].size()
374 << "; std::set memory size=" << setMemorySize
375 << "; csl memory size=" << cslMemorySize;
378 readValues_.reserve(size);
379 deleteValues_.reserve(size);
380 writeValues_.reserve(size);
381 for (int i = size; i < 2 * size; ++i) {
382 readValues_.push_back(2 * i);
383 deleteValues_.push_back(2 * i);
385 // half new values and half already in the list
386 writeValues_.push_back((rand() % 2) + 2 * i);
388 std::shuffle(readValues_.begin(), readValues_.end(), std::mt19937{});
389 std::shuffle(deleteValues_.begin(), deleteValues_.end(), std::mt19937{});
390 std::shuffle(writeValues_.begin(), writeValues_.end(), std::mt19937{});
393 ~ConcurrentAccessData() {
394 FOR_EACH(lock, locks_) delete *lock;
397 inline bool skipListFind(int /* idx */, ValueType val) {
398 return skipList_.contains(val);
400 inline void skipListInsert(int /* idx */, ValueType val) {
403 inline void skipListErase(int /* idx */, ValueType val) {
404 skipList_.remove(val);
407 inline bool setFind(int idx, ValueType val) {
408 RWSpinLock::ReadHolder g(locks_[idx]);
409 return sets_[idx].find(val) == sets_[idx].end();
411 inline void setInsert(int idx, ValueType val) {
412 RWSpinLock::WriteHolder g(locks_[idx]);
413 sets_[idx].insert(val);
415 inline void setErase(int idx, ValueType val) {
416 RWSpinLock::WriteHolder g(locks_[idx]);
417 sets_[idx].erase(val);
420 void runSkipList(int id, size_t iters) {
422 for (size_t i = 0; i < iters; ++i) {
423 sum += accessSkipList(id, i);
428 void runSet(size_t id, size_t iters) {
430 for (size_t i = 0; i < iters; ++i) {
431 sum += accessSet(id, i);
436 bool accessSkipList(int64_t id, size_t t) {
437 if (t > readValues_.size()) {
438 t = t % readValues_.size();
440 uint32_t h = folly::hash::twang_32from64(t * id);
443 if ((h & 0x31) == 0) { // 1/4 chance to delete
444 skipListErase(0, deleteValues_[t]);
446 skipListInsert(0, writeValues_[t]);
450 return skipListFind(0, readValues_[t]);
454 bool accessSet(int64_t id, size_t t) {
455 if (t > readValues_.size()) {
456 t = t % readValues_.size();
458 uint32_t h = folly::hash::twang_32from64(t * id);
459 int idx = (h % FLAGS_num_sets);
460 switch (h % 8) { // 1/8 chance to write
462 if ((h & 0x31) == 0) { // 1/32 chance to delete
463 setErase(idx, deleteValues_[t]);
465 setInsert(idx, writeValues_[t]);
469 return setFind(idx, readValues_[t]);
474 SkipListType::Accessor skipList_;
475 std::vector<SetType> sets_;
476 std::vector<RWSpinLock*> locks_;
478 std::vector<ValueType> readValues_;
479 std::vector<ValueType> writeValues_;
480 std::vector<ValueType> deleteValues_;
483 static std::map<int, std::shared_ptr<ConcurrentAccessData> > g_data;
485 static ConcurrentAccessData *mayInitTestData(int size) {
486 auto it = g_data.find(size);
487 if (it == g_data.end()) {
488 auto ptr = std::make_shared<ConcurrentAccessData>(size);
492 return it->second.get();
495 void BM_ContentionCSL(int iters, int size) {
496 BenchmarkSuspender susp;
497 auto data = mayInitTestData(size);
498 std::vector<std::thread> threads;
501 for (int i = 0; i < FLAGS_num_threads; ++i) {
502 threads.push_back(std::thread(
503 &ConcurrentAccessData::runSkipList, data, i, iters));
505 FOR_EACH(t, threads) {
510 void BM_ContentionStdSet(int iters, int size) {
511 BenchmarkSuspender susp;
512 auto data = mayInitTestData(size);
513 std::vector<std::thread> threads;
516 for (int i = 0; i < FLAGS_num_threads; ++i) {
517 threads.push_back(std::thread(
518 &ConcurrentAccessData::runSet, data, i, iters));
520 FOR_EACH(t, threads) {
527 // Single-thread benchmarking
529 BENCHMARK_DRAW_LINE();
531 BENCHMARK_PARAM(BM_IterateOverSet, 1000);
532 BENCHMARK_PARAM(BM_IterateSkipList, 1000);
533 BENCHMARK_DRAW_LINE();
534 BENCHMARK_PARAM(BM_IterateOverSet, 1000000);
535 BENCHMARK_PARAM(BM_IterateSkipList, 1000000);
536 BENCHMARK_DRAW_LINE();
538 // find with keys in the set
539 BENCHMARK_PARAM(BM_SetContainsFound, 1000);
540 BENCHMARK_PARAM(BM_CSLContainsFound, 1000);
541 BENCHMARK_DRAW_LINE();
542 BENCHMARK_PARAM(BM_SetContainsFound, 100000);
543 BENCHMARK_PARAM(BM_CSLContainsFound, 100000);
544 BENCHMARK_DRAW_LINE();
545 BENCHMARK_PARAM(BM_SetContainsFound, 1000000);
546 BENCHMARK_PARAM(BM_CSLContainsFound, 1000000);
547 BENCHMARK_DRAW_LINE();
548 BENCHMARK_PARAM(BM_SetContainsFound, 10000000);
549 BENCHMARK_PARAM(BM_CSLContainsFound, 10000000);
550 BENCHMARK_DRAW_LINE();
553 // find with keys not in the set
554 BENCHMARK_PARAM(BM_SetContainsNotFound, 1000);
555 BENCHMARK_PARAM(BM_CSLContainsNotFound, 1000);
556 BENCHMARK_DRAW_LINE();
557 BENCHMARK_PARAM(BM_SetContainsNotFound, 100000);
558 BENCHMARK_PARAM(BM_CSLContainsNotFound, 100000);
559 BENCHMARK_DRAW_LINE();
560 BENCHMARK_PARAM(BM_SetContainsNotFound, 1000000);
561 BENCHMARK_PARAM(BM_CSLContainsNotFound, 1000000);
562 BENCHMARK_DRAW_LINE();
565 BENCHMARK_PARAM(BM_AddSet, 1000);
566 BENCHMARK_PARAM(BM_AddSkipList, 1000);
567 BENCHMARK_DRAW_LINE();
569 BENCHMARK_PARAM(BM_AddSet, 65536);
570 BENCHMARK_PARAM(BM_AddSkipList, 65536);
571 BENCHMARK_DRAW_LINE();
573 BENCHMARK_PARAM(BM_AddSet, 1000000);
574 BENCHMARK_PARAM(BM_AddSkipList, 1000000);
575 BENCHMARK_DRAW_LINE();
577 BENCHMARK_PARAM(BM_SetMerge, 1000);
578 BENCHMARK_PARAM(BM_CSLMergeIntersection, 1000);
579 BENCHMARK_PARAM(BM_CSLMergeLookup, 1000);
580 BENCHMARK_DRAW_LINE();
582 BENCHMARK_PARAM(BM_SetMerge, 65536);
583 BENCHMARK_PARAM(BM_CSLMergeIntersection, 65536);
584 BENCHMARK_PARAM(BM_CSLMergeLookup, 65536);
585 BENCHMARK_DRAW_LINE();
587 BENCHMARK_PARAM(BM_SetMerge, 1000000);
588 BENCHMARK_PARAM(BM_CSLMergeIntersection, 1000000);
589 BENCHMARK_PARAM(BM_CSLMergeLookup, 1000000);
590 BENCHMARK_DRAW_LINE();
593 // multithreaded benchmarking
595 BENCHMARK_PARAM(BM_ContentionStdSet, 1024);
596 BENCHMARK_PARAM(BM_ContentionCSL, 1024);
597 BENCHMARK_DRAW_LINE();
599 BENCHMARK_PARAM(BM_ContentionStdSet, 65536);
600 BENCHMARK_PARAM(BM_ContentionCSL, 65536);
601 BENCHMARK_DRAW_LINE();
603 BENCHMARK_PARAM(BM_ContentionStdSet, 1048576);
604 BENCHMARK_PARAM(BM_ContentionCSL, 1048576);
605 BENCHMARK_DRAW_LINE();
609 int main(int argc, char** argv) {
610 google::InitGoogleLogging(argv[0]);
611 gflags::ParseCommandLineFlags(&argc, &argv, true);
620 Benchmark on Intel(R) Xeon(R) CPU X5650 @2.67GHz
622 ==============================================================================
623 1 thread Benchmark Iters Total t t/iter iter/sec
624 ------------------------------------------------------------------------------
625 +37.0% BM_Accessor 100000 1.958 ms 19.58 ns 48.71 M
626 * BM_AccessorBasicRefcounting 100000 1.429 ms 14.29 ns 66.74 M
627 ------------------------------------------------------------------------------
628 + 603% BM_IterateOverSet/1000 100000 1.589 ms 15.89 ns 60.02 M
629 * BM_IterateSkipList/1000 100000 226 us 2.26 ns 422 M
630 ------------------------------------------------------------------------------
631 + 107% BM_IterateOverSet/976.6k 100000 8.324 ms 83.24 ns 11.46 M
632 * BM_IterateSkipList/976.6k 100000 4.016 ms 40.16 ns 23.75 M
633 ------------------------------------------------------------------------------
634 * BM_SetContainsFound/1000 100000 7.082 ms 70.82 ns 13.47 M
635 +39.9% BM_CSLContainsFound/1000 100000 9.908 ms 99.08 ns 9.625 M
636 ------------------------------------------------------------------------------
637 * BM_SetContainsFound/97.66k 100000 23.8 ms 238 ns 4.006 M
638 +5.97% BM_CSLContainsFound/97.66k 100000 25.23 ms 252.3 ns 3.781 M
639 ------------------------------------------------------------------------------
640 +33.6% BM_SetContainsFound/976.6k 100000 64.3 ms 643 ns 1.483 M
641 * BM_CSLContainsFound/976.6k 100000 48.13 ms 481.3 ns 1.981 M
642 ------------------------------------------------------------------------------
643 +30.3% BM_SetContainsFound/9.537M 100000 115.1 ms 1.151 us 848.6 k
644 * BM_CSLContainsFound/9.537M 100000 88.33 ms 883.3 ns 1.08 M
645 ------------------------------------------------------------------------------
646 * BM_SetContainsNotFound/1000 100000 2.081 ms 20.81 ns 45.83 M
647 +76.2% BM_CSLContainsNotFound/1000 100000 3.667 ms 36.67 ns 26.01 M
648 ------------------------------------------------------------------------------
649 * BM_SetContainsNotFound/97.66k 100000 6.049 ms 60.49 ns 15.77 M
650 +32.7% BM_CSLContainsNotFound/97.66k 100000 8.025 ms 80.25 ns 11.88 M
651 ------------------------------------------------------------------------------
652 * BM_SetContainsNotFound/976.6k 100000 7.464 ms 74.64 ns 12.78 M
653 +12.8% BM_CSLContainsNotFound/976.6k 100000 8.417 ms 84.17 ns 11.33 M
654 ------------------------------------------------------------------------------
655 * BM_AddSet/1000 100000 29.26 ms 292.6 ns 3.259 M
656 +70.0% BM_AddSkipList/1000 100000 49.75 ms 497.5 ns 1.917 M
657 ------------------------------------------------------------------------------
658 * BM_AddSet/64k 100000 38.73 ms 387.3 ns 2.462 M
659 +55.7% BM_AddSkipList/64k 100000 60.3 ms 603 ns 1.581 M
660 ------------------------------------------------------------------------------
661 * BM_AddSet/976.6k 100000 75.71 ms 757.1 ns 1.26 M
662 +33.6% BM_AddSkipList/976.6k 100000 101.2 ms 1.012 us 965.3 k
663 ------------------------------------------------------------------------------
664 + 716% BM_SetMerge/1000 100000 6.872 ms 68.72 ns 13.88 M
665 * BM_CSLMergeIntersection/1000 100000 842 us 8.42 ns 113.3 M
666 + 268% BM_CSLMergeLookup/1000 100000 3.1 ms 31 ns 30.76 M
667 ------------------------------------------------------------------------------
668 +36.3% BM_SetMerge/64k 100000 14.03 ms 140.3 ns 6.798 M
669 +39.4% BM_CSLMergeIntersection/64k 100000 14.35 ms 143.5 ns 6.645 M
670 * BM_CSLMergeLookup/64k 100000 10.29 ms 102.9 ns 9.266 M
671 ------------------------------------------------------------------------------
672 +10.3% BM_SetMerge/976.6k 100000 46.24 ms 462.4 ns 2.062 M
673 +25.1% BM_CSLMergeIntersection/976.6k 100000 52.47 ms 524.7 ns 1.818 M
674 * BM_CSLMergeLookup/976.6k 100000 41.94 ms 419.3 ns 2.274 M
675 ------------------------------------------------------------------------------
678 ==============================================================================
679 Contention benchmark 7/8 find, 3/32 insert, 1/32 erase
681 4 threads Benchmark Iters Total t t/iter iter/sec
682 ------------------------------------------------------------------------------
683 + 269% BM_ContentionStdSet/1k 100000 75.66 ms 756.6 ns 1.26 M
684 * BM_ContentionCSL/1k 100000 20.47 ms 204.7 ns 4.658 M
685 ------------------------------------------------------------------------------
686 + 228% BM_ContentionStdSet/64k 100000 105.6 ms 1.056 us 924.9 k
687 * BM_ContentionCSL/64k 100000 32.18 ms 321.8 ns 2.963 M
688 ------------------------------------------------------------------------------
689 + 224% BM_ContentionStdSet/1M 100000 117.4 ms 1.174 us 832.2 k
690 * BM_ContentionCSL/1M 100000 36.18 ms 361.8 ns 2.636 M
691 ------------------------------------------------------------------------------
694 12 threads Benchmark Iters Total t t/iter iter/sec
695 ------------------------------------------------------------------------------
696 + 697% BM_ContentionStdSet/1k 100000 455.3 ms 4.553 us 214.5 k
697 * BM_ContentionCSL/1k 100000 57.12 ms 571.2 ns 1.67 M
698 ------------------------------------------------------------------------------
699 +1257% BM_ContentionStdSet/64k 100000 654.9 ms 6.549 us 149.1 k
700 * BM_ContentionCSL/64k 100000 48.24 ms 482.4 ns 1.977 M
701 ------------------------------------------------------------------------------
702 +1262% BM_ContentionStdSet/1M 100000 657.3 ms 6.573 us 148.6 k
703 * BM_ContentionCSL/1M 100000 48.25 ms 482.5 ns 1.977 M
704 ------------------------------------------------------------------------------