2 * Copyright 2014 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.
17 // @author: Xin Liu <xliux@fb.com>
23 #include <gflags/gflags.h>
24 #include <glog/logging.h>
25 #include <folly/Benchmark.h>
26 #include <folly/ConcurrentSkipList.h>
27 #include <folly/Hash.h>
28 #include <folly/RWSpinLock.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::random_shuffle(gData.begin(), gData.end());
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()) iter = a_set.begin();
78 // VLOG(20) << "sum = " << sum;
82 void BM_IterateSkipList(int iters, int size) {
83 BenchmarkSuspender susp;
85 auto skipList = SkipListType::create(kInitHeadHeight);
86 for (int i = 0; i < size; ++i) {
87 skipList.add(rand() % kMaxValue);
92 auto iter = skipList.begin();
93 for (int i = 0; i < iters; ++i) {
95 if (iter == skipList.end()) iter = skipList.begin();
99 // VLOG(20) << "sum = " << sum;
103 void BM_SetMerge(int iters, int size) {
104 BenchmarkSuspender susp;
107 for (int i = 0; i < iters; ++i) {
108 a_set.insert(rand() % kMaxValue);
110 for (int i = 0; i < size; ++i) {
111 b_set.insert(rand() % kMaxValue);
115 int64_t mergedSum = 0;
116 FOR_EACH(it, a_set) {
117 if (b_set.find(*it) != b_set.end()) mergedSum += *it;
120 // VLOG(20) << mergedSum;
124 void BM_CSLMergeLookup(int iters, int size) {
125 BenchmarkSuspender susp;
126 auto skipList = SkipListType::create(kInitHeadHeight);
127 auto skipList2 = SkipListType::create(kInitHeadHeight);
129 for (int i = 0; i < iters; ++i) {
130 skipList.add(rand() % kMaxValue);
132 for (int i = 0; i < size; ++i) {
133 skipList2.add(rand() % kMaxValue);
135 int64_t mergedSum = 0;
138 SkipListType::Skipper skipper(skipList2);
139 FOR_EACH(it, skipList) {
140 if (skipper.to(*it)) mergedSum += *it;
144 // VLOG(20) << mergedSum;
148 // merge by two skippers
149 void BM_CSLMergeIntersection(int iters, int size) {
150 BenchmarkSuspender susp;
151 auto skipList = SkipListType::create(kInitHeadHeight);
152 auto skipList2 = SkipListType::create(kInitHeadHeight);
153 for (int i = 0; i < iters; ++i) {
154 skipList.add(rand() % kMaxValue);
156 for (int i = 0; i < size; ++i) {
157 skipList2.add(rand() % kMaxValue);
161 SkipListType::Skipper s1(skipList);
162 SkipListType::Skipper s2(skipList2);
164 int64_t mergedSum = 0;
166 while (s1.good() && s2.good()) {
171 } else if (v1 > v2) {
181 // VLOG(20) << mergedSum;
185 void BM_SetContainsNotFound(int iters, int size) {
186 BenchmarkSuspender susp;
188 CHECK_LT(size, kMaxValue);
189 for (int i = 0; i < size; ++i) {
195 for (int i = 0; i < iters; ++i) {
196 sum += (aset.end() == aset.find(2 * i + 1));
204 void BM_SetContainsFound(int iters, int size) {
205 BenchmarkSuspender susp;
207 CHECK_LT(size, kMaxValue);
209 for (int i = 0; i < size; ++i) {
213 std::vector<int> values;
214 for (int i = 0; i < iters; ++i) {
215 values.push_back(rand() % size);
220 for (int i = 0; i < iters; ++i) {
221 sum += (aset.end() == aset.find(values[i]));
229 void BM_CSLContainsFound(int iters, int size) {
230 BenchmarkSuspender susp;
231 auto skipList = SkipListType::create(kInitHeadHeight);
232 CHECK_LT(size, kMaxValue);
234 for (int i = 0; i < size; ++i) {
237 std::vector<int> values;
238 for (int i = 0; i < iters; ++i) {
239 values.push_back(rand() % size);
244 for (int i = 0; i < iters; ++i) {
245 sum += skipList.contains(values[i]);
253 void BM_CSLContainsNotFound(int iters, int size) {
254 BenchmarkSuspender susp;
255 auto skipList = SkipListType::create(kInitHeadHeight);
256 CHECK_LT(size, kMaxValue);
258 for (int i = 0; i < size; ++i) {
264 for (int i = 0; i < iters; ++i) {
265 sum += skipList.contains(2 * i + 1);
273 void BM_AddSet(int iters, int size) {
274 BenchmarkSuspender susp;
276 for (int i = 0; i < size; ++i) {
277 aset.insert(gData[i]);
281 for (int i = size; i < size + iters; ++i) {
282 aset.insert(gData[i]);
286 void BM_AddSkipList(int iters, int size) {
287 BenchmarkSuspender susp;
288 auto skipList = SkipListType::create(kInitHeadHeight);
289 for (int i = 0; i < size; ++i) {
290 skipList.add(gData[i]);
294 for (int i = size; i < size + iters; ++i) {
295 skipList.add(gData[i]);
299 BENCHMARK(Accessor, iters) {
300 BenchmarkSuspender susp;
301 auto skiplist = SkipListType::createInstance(kInitHeadHeight);
302 auto sl = skiplist.get();
305 for (size_t i = 0; i < iters; ++i) {
306 SkipListAccessor accessor(sl);
310 // a benchmark to estimate the
311 // low bound of doing a ref counting for an Accessor
312 BENCHMARK(accessorBasicRefcounting, iters) {
313 BenchmarkSuspender susp;
314 auto* value = new std::atomic<int32_t>();
315 auto* dirty = new std::atomic<int32_t>();
317 folly::MicroSpinLock l;
321 for (size_t i = 0; i < iters; ++i) {
322 value->fetch_add(1, std::memory_order_relaxed);
323 if (dirty->load(std::memory_order_acquire) != 0) {
324 folly::MSLGuard g(l);
326 value->fetch_sub(1, std::memory_order_relaxed);
336 // Data For testing contention benchmark
337 class ConcurrentAccessData {
339 explicit ConcurrentAccessData(int size) :
340 skipList_(SkipListType::create(10)),
341 sets_(FLAGS_num_sets), locks_(FLAGS_num_sets) {
343 for (int i = 0; i < size; ++i) {
348 for (int i = 0; i < FLAGS_num_sets; ++i) {
349 locks_[i] = new RWSpinLock();
350 if (i > 0) sets_[i] = sets_[0];
353 // This requires knowledge of the C++ library internals. Only use it if we're
354 // using the GNU C++ library.
355 #ifdef _GLIBCXX_SYMVER
357 int64_t setMemorySize = sets_[0].size() * sizeof(*sets_[0].begin()._M_node);
358 int64_t cslMemorySize = 0;
359 for (auto it = skipList_.begin(); it != skipList_.end(); ++it) {
360 cslMemorySize += it.nodeSize();
363 LOG(INFO) << "size=" << sets_[0].size()
364 << "; std::set memory size=" << setMemorySize
365 << "; csl memory size=" << cslMemorySize;
368 readValues_.reserve(size);
369 deleteValues_.reserve(size);
370 writeValues_.reserve(size);
371 for (int i = size; i < 2 * size; ++i) {
372 readValues_.push_back(2 * i);
373 deleteValues_.push_back(2 * i);
375 // half new values and half already in the list
376 writeValues_.push_back((rand() % 2) + 2 * i);
378 std::random_shuffle(readValues_.begin(), readValues_.end());
379 std::random_shuffle(deleteValues_.begin(), deleteValues_.end());
380 std::random_shuffle(writeValues_.begin(), writeValues_.end());
383 ~ConcurrentAccessData() {
384 FOR_EACH(lock, locks_) delete *lock;
387 inline bool skipListFind(int idx, ValueType val) {
388 return skipList_.contains(val);
390 inline void skipListInsert(int idx, ValueType val) {
393 inline void skipListErase(int idx, ValueType val) {
394 skipList_.remove(val);
397 inline bool setFind(int idx, ValueType val) {
398 RWSpinLock::ReadHolder g(locks_[idx]);
399 return sets_[idx].find(val) == sets_[idx].end();
401 inline void setInsert(int idx, ValueType val) {
402 RWSpinLock::WriteHolder g(locks_[idx]);
403 sets_[idx].insert(val);
405 inline void setErase(int idx, ValueType val) {
406 RWSpinLock::WriteHolder g(locks_[idx]);
407 sets_[idx].erase(val);
410 void runSkipList(int id, size_t iters) {
412 for (size_t i = 0; i < iters; ++i) {
413 sum += accessSkipList(id, i);
418 void runSet(size_t id, size_t iters) {
420 for (size_t i = 0; i < iters; ++i) {
421 sum += accessSet(id, i);
426 bool accessSkipList(int64_t id, size_t t) {
427 if (t > readValues_.size()) {
428 t = t % readValues_.size();
430 uint32_t h = folly::hash::twang_32from64(t * id);
433 if ((h & 0x31) == 0) { // 1/4 chance to delete
434 skipListErase(0, deleteValues_[t]);
436 skipListInsert(0, writeValues_[t]);
440 return skipListFind(0, readValues_[t]);
444 bool accessSet(int64_t id, size_t t) {
445 if (t > readValues_.size()) {
446 t = t % readValues_.size();
448 uint32_t h = folly::hash::twang_32from64(t * id);
449 int idx = (h % FLAGS_num_sets);
450 switch (h % 8) { // 1/8 chance to write
452 if ((h & 0x31) == 0) { // 1/32 chance to delete
453 setErase(idx, deleteValues_[t]);
455 setInsert(idx, writeValues_[t]);
459 return setFind(idx, readValues_[t]);
464 SkipListType::Accessor skipList_;
465 std::vector<SetType> sets_;
466 std::vector<RWSpinLock*> locks_;
468 std::vector<ValueType> readValues_;
469 std::vector<ValueType> writeValues_;
470 std::vector<ValueType> deleteValues_;
473 static std::map<int, std::shared_ptr<ConcurrentAccessData> > g_data;
475 static ConcurrentAccessData *mayInitTestData(int size) {
476 auto it = g_data.find(size);
477 if (it == g_data.end()) {
478 auto ptr = std::shared_ptr<ConcurrentAccessData>(
479 new ConcurrentAccessData(size));
483 return it->second.get();
486 void BM_ContentionCSL(int iters, int size) {
487 BenchmarkSuspender susp;
488 auto data = mayInitTestData(size);
489 std::vector<std::thread> threads;
492 for (int i = 0; i < FLAGS_num_threads; ++i) {
493 threads.push_back(std::thread(
494 &ConcurrentAccessData::runSkipList, data, i, iters));
496 FOR_EACH(t, threads) {
501 void BM_ContentionStdSet(int iters, int size) {
502 BenchmarkSuspender susp;
503 auto data = mayInitTestData(size);
504 std::vector<std::thread> threads;
507 for (int i = 0; i < FLAGS_num_threads; ++i) {
508 threads.push_back(std::thread(
509 &ConcurrentAccessData::runSet, data, i, iters));
511 FOR_EACH(t, threads) {
518 // Single-thread benchmarking
520 BENCHMARK_DRAW_LINE();
522 BENCHMARK_PARAM(BM_IterateOverSet, 1000);
523 BENCHMARK_PARAM(BM_IterateSkipList, 1000);
524 BENCHMARK_DRAW_LINE();
525 BENCHMARK_PARAM(BM_IterateOverSet, 1000000);
526 BENCHMARK_PARAM(BM_IterateSkipList, 1000000);
527 BENCHMARK_DRAW_LINE();
529 // find with keys in the set
530 BENCHMARK_PARAM(BM_SetContainsFound, 1000);
531 BENCHMARK_PARAM(BM_CSLContainsFound, 1000);
532 BENCHMARK_DRAW_LINE();
533 BENCHMARK_PARAM(BM_SetContainsFound, 100000);
534 BENCHMARK_PARAM(BM_CSLContainsFound, 100000);
535 BENCHMARK_DRAW_LINE();
536 BENCHMARK_PARAM(BM_SetContainsFound, 1000000);
537 BENCHMARK_PARAM(BM_CSLContainsFound, 1000000);
538 BENCHMARK_DRAW_LINE();
539 BENCHMARK_PARAM(BM_SetContainsFound, 10000000);
540 BENCHMARK_PARAM(BM_CSLContainsFound, 10000000);
541 BENCHMARK_DRAW_LINE();
544 // find with keys not in the set
545 BENCHMARK_PARAM(BM_SetContainsNotFound, 1000);
546 BENCHMARK_PARAM(BM_CSLContainsNotFound, 1000);
547 BENCHMARK_DRAW_LINE();
548 BENCHMARK_PARAM(BM_SetContainsNotFound, 100000);
549 BENCHMARK_PARAM(BM_CSLContainsNotFound, 100000);
550 BENCHMARK_DRAW_LINE();
551 BENCHMARK_PARAM(BM_SetContainsNotFound, 1000000);
552 BENCHMARK_PARAM(BM_CSLContainsNotFound, 1000000);
553 BENCHMARK_DRAW_LINE();
556 BENCHMARK_PARAM(BM_AddSet, 1000);
557 BENCHMARK_PARAM(BM_AddSkipList, 1000);
558 BENCHMARK_DRAW_LINE();
560 BENCHMARK_PARAM(BM_AddSet, 65536);
561 BENCHMARK_PARAM(BM_AddSkipList, 65536);
562 BENCHMARK_DRAW_LINE();
564 BENCHMARK_PARAM(BM_AddSet, 1000000);
565 BENCHMARK_PARAM(BM_AddSkipList, 1000000);
566 BENCHMARK_DRAW_LINE();
568 BENCHMARK_PARAM(BM_SetMerge, 1000);
569 BENCHMARK_PARAM(BM_CSLMergeIntersection, 1000);
570 BENCHMARK_PARAM(BM_CSLMergeLookup, 1000);
571 BENCHMARK_DRAW_LINE();
573 BENCHMARK_PARAM(BM_SetMerge, 65536);
574 BENCHMARK_PARAM(BM_CSLMergeIntersection, 65536);
575 BENCHMARK_PARAM(BM_CSLMergeLookup, 65536);
576 BENCHMARK_DRAW_LINE();
578 BENCHMARK_PARAM(BM_SetMerge, 1000000);
579 BENCHMARK_PARAM(BM_CSLMergeIntersection, 1000000);
580 BENCHMARK_PARAM(BM_CSLMergeLookup, 1000000);
581 BENCHMARK_DRAW_LINE();
584 // multithreaded benchmarking
586 BENCHMARK_PARAM(BM_ContentionStdSet, 1024);
587 BENCHMARK_PARAM(BM_ContentionCSL, 1024);
588 BENCHMARK_DRAW_LINE();
590 BENCHMARK_PARAM(BM_ContentionStdSet, 65536);
591 BENCHMARK_PARAM(BM_ContentionCSL, 65536);
592 BENCHMARK_DRAW_LINE();
594 BENCHMARK_PARAM(BM_ContentionStdSet, 1048576);
595 BENCHMARK_PARAM(BM_ContentionCSL, 1048576);
596 BENCHMARK_DRAW_LINE();
600 int main(int argc, char** argv) {
601 google::InitGoogleLogging(argv[0]);
602 gflags::ParseCommandLineFlags(&argc, &argv, true);
611 Benchmark on Intel(R) Xeon(R) CPU X5650 @2.67GHz
613 ==============================================================================
614 1 thread Benchmark Iters Total t t/iter iter/sec
615 ------------------------------------------------------------------------------
616 +37.0% BM_Accessor 100000 1.958 ms 19.58 ns 48.71 M
617 * BM_AccessorBasicRefcounting 100000 1.429 ms 14.29 ns 66.74 M
618 ------------------------------------------------------------------------------
619 + 603% BM_IterateOverSet/1000 100000 1.589 ms 15.89 ns 60.02 M
620 * BM_IterateSkipList/1000 100000 226 us 2.26 ns 422 M
621 ------------------------------------------------------------------------------
622 + 107% BM_IterateOverSet/976.6k 100000 8.324 ms 83.24 ns 11.46 M
623 * BM_IterateSkipList/976.6k 100000 4.016 ms 40.16 ns 23.75 M
624 ------------------------------------------------------------------------------
625 * BM_SetContainsFound/1000 100000 7.082 ms 70.82 ns 13.47 M
626 +39.9% BM_CSLContainsFound/1000 100000 9.908 ms 99.08 ns 9.625 M
627 ------------------------------------------------------------------------------
628 * BM_SetContainsFound/97.66k 100000 23.8 ms 238 ns 4.006 M
629 +5.97% BM_CSLContainsFound/97.66k 100000 25.23 ms 252.3 ns 3.781 M
630 ------------------------------------------------------------------------------
631 +33.6% BM_SetContainsFound/976.6k 100000 64.3 ms 643 ns 1.483 M
632 * BM_CSLContainsFound/976.6k 100000 48.13 ms 481.3 ns 1.981 M
633 ------------------------------------------------------------------------------
634 +30.3% BM_SetContainsFound/9.537M 100000 115.1 ms 1.151 us 848.6 k
635 * BM_CSLContainsFound/9.537M 100000 88.33 ms 883.3 ns 1.08 M
636 ------------------------------------------------------------------------------
637 * BM_SetContainsNotFound/1000 100000 2.081 ms 20.81 ns 45.83 M
638 +76.2% BM_CSLContainsNotFound/1000 100000 3.667 ms 36.67 ns 26.01 M
639 ------------------------------------------------------------------------------
640 * BM_SetContainsNotFound/97.66k 100000 6.049 ms 60.49 ns 15.77 M
641 +32.7% BM_CSLContainsNotFound/97.66k 100000 8.025 ms 80.25 ns 11.88 M
642 ------------------------------------------------------------------------------
643 * BM_SetContainsNotFound/976.6k 100000 7.464 ms 74.64 ns 12.78 M
644 +12.8% BM_CSLContainsNotFound/976.6k 100000 8.417 ms 84.17 ns 11.33 M
645 ------------------------------------------------------------------------------
646 * BM_AddSet/1000 100000 29.26 ms 292.6 ns 3.259 M
647 +70.0% BM_AddSkipList/1000 100000 49.75 ms 497.5 ns 1.917 M
648 ------------------------------------------------------------------------------
649 * BM_AddSet/64k 100000 38.73 ms 387.3 ns 2.462 M
650 +55.7% BM_AddSkipList/64k 100000 60.3 ms 603 ns 1.581 M
651 ------------------------------------------------------------------------------
652 * BM_AddSet/976.6k 100000 75.71 ms 757.1 ns 1.26 M
653 +33.6% BM_AddSkipList/976.6k 100000 101.2 ms 1.012 us 965.3 k
654 ------------------------------------------------------------------------------
655 + 716% BM_SetMerge/1000 100000 6.872 ms 68.72 ns 13.88 M
656 * BM_CSLMergeIntersection/1000 100000 842 us 8.42 ns 113.3 M
657 + 268% BM_CSLMergeLookup/1000 100000 3.1 ms 31 ns 30.76 M
658 ------------------------------------------------------------------------------
659 +36.3% BM_SetMerge/64k 100000 14.03 ms 140.3 ns 6.798 M
660 +39.4% BM_CSLMergeIntersection/64k 100000 14.35 ms 143.5 ns 6.645 M
661 * BM_CSLMergeLookup/64k 100000 10.29 ms 102.9 ns 9.266 M
662 ------------------------------------------------------------------------------
663 +10.3% BM_SetMerge/976.6k 100000 46.24 ms 462.4 ns 2.062 M
664 +25.1% BM_CSLMergeIntersection/976.6k 100000 52.47 ms 524.7 ns 1.818 M
665 * BM_CSLMergeLookup/976.6k 100000 41.94 ms 419.3 ns 2.274 M
666 ------------------------------------------------------------------------------
669 ==============================================================================
670 Contention benchmark 7/8 find, 3/32 insert, 1/32 erase
672 4 threads Benchmark Iters Total t t/iter iter/sec
673 ------------------------------------------------------------------------------
674 + 269% BM_ContentionStdSet/1k 100000 75.66 ms 756.6 ns 1.26 M
675 * BM_ContentionCSL/1k 100000 20.47 ms 204.7 ns 4.658 M
676 ------------------------------------------------------------------------------
677 + 228% BM_ContentionStdSet/64k 100000 105.6 ms 1.056 us 924.9 k
678 * BM_ContentionCSL/64k 100000 32.18 ms 321.8 ns 2.963 M
679 ------------------------------------------------------------------------------
680 + 224% BM_ContentionStdSet/1M 100000 117.4 ms 1.174 us 832.2 k
681 * BM_ContentionCSL/1M 100000 36.18 ms 361.8 ns 2.636 M
682 ------------------------------------------------------------------------------
685 12 threads Benchmark Iters Total t t/iter iter/sec
686 ------------------------------------------------------------------------------
687 + 697% BM_ContentionStdSet/1k 100000 455.3 ms 4.553 us 214.5 k
688 * BM_ContentionCSL/1k 100000 57.12 ms 571.2 ns 1.67 M
689 ------------------------------------------------------------------------------
690 +1257% BM_ContentionStdSet/64k 100000 654.9 ms 6.549 us 149.1 k
691 * BM_ContentionCSL/64k 100000 48.24 ms 482.4 ns 1.977 M
692 ------------------------------------------------------------------------------
693 +1262% BM_ContentionStdSet/1M 100000 657.3 ms 6.573 us 148.6 k
694 * BM_ContentionCSL/1M 100000 48.25 ms 482.5 ns 1.977 M
695 ------------------------------------------------------------------------------