2 * Copyright 2012 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>
22 #include <boost/shared_ptr.hpp>
23 #include <boost/thread.hpp>
25 #include <gflags/gflags.h>
26 #include <glog/logging.h>
27 #include "folly/Benchmark.h"
28 #include "folly/ConcurrentSkipList.h"
29 #include "folly/Hash.h"
30 #include "folly/RWSpinLock.h"
33 DEFINE_int32(num_threads, 12, "num concurrent threads to test");
35 // In some case, we may want to test worker threads operating on multiple
36 // lists. For example in search, not all threads are visiting the same posting
37 // list, but for the ones with some popular terms, they do get multiple
38 // visitors at the same time.
39 DEFINE_int32(num_sets, 1, "num of set to operate on");
41 static const int kInitHeadHeight = 10;
42 static const int kMaxValue = 0x1000000;
46 using namespace folly;
48 typedef int ValueType;
49 typedef ConcurrentSkipList<ValueType> SkipListType;
50 typedef SkipListType::Accessor SkipListAccessor;
51 typedef std::set<ValueType> SetType;
53 static std::vector<ValueType> gData;
54 static void initData() {
55 gData.resize(kMaxValue);
56 for (int i = 0; i < kMaxValue; ++i) {
59 std::random_shuffle(gData.begin(), gData.end());
62 // single thread benchmarks
63 void BM_IterateOverSet(int iters, int size) {
68 for (int i = 0; i < size; ++i) {
69 a_set.insert(gData[rand() % kMaxValue]);
74 auto iter = a_set.begin();
75 for (int i = 0; i < iters; ++i) {
77 if (iter == a_set.end()) iter = a_set.begin();
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()) iter = skipList.begin();
101 VLOG(20) << "sum = " << sum;
105 void BM_SetMerge(int iters, int size) {
106 BenchmarkSuspender susp;
109 for (int i = 0; i < iters; ++i) {
110 a_set.insert(rand() % kMaxValue);
112 for (int i = 0; i < size; ++i) {
113 b_set.insert(rand() % kMaxValue);
117 int64_t mergedSum = 0;
118 FOR_EACH(it, a_set) {
119 if (b_set.find(*it) != b_set.end()) mergedSum += *it;
122 VLOG(20) << mergedSum;
126 void BM_CSLMergeLookup(int iters, int size) {
127 BenchmarkSuspender susp;
128 auto skipList = SkipListType::create(kInitHeadHeight);
129 auto skipList2 = SkipListType::create(kInitHeadHeight);
131 for (int i = 0; i < iters; ++i) {
132 skipList.add(rand() % kMaxValue);
134 for (int i = 0; i < size; ++i) {
135 skipList2.add(rand() % kMaxValue);
137 int64_t mergedSum = 0;
140 SkipListType::Skipper skipper(skipList2);
141 FOR_EACH(it, skipList) {
142 if (skipper.to(*it)) mergedSum += *it;
146 VLOG(20) << mergedSum;
150 // merge by two skippers
151 void BM_CSLMergeIntersection(int iters, int size) {
152 BenchmarkSuspender susp;
153 auto skipList = SkipListType::create(kInitHeadHeight);
154 auto skipList2 = SkipListType::create(kInitHeadHeight);
155 for (int i = 0; i < iters; ++i) {
156 skipList.add(rand() % kMaxValue);
158 for (int i = 0; i < size; ++i) {
159 skipList2.add(rand() % kMaxValue);
163 SkipListType::Skipper s1(skipList);
164 SkipListType::Skipper s2(skipList2);
166 int64_t mergedSum = 0;
168 while (s1.good() && s2.good()) {
173 } else if (v1 > v2) {
183 VLOG(20) << mergedSum;
187 void BM_SetContainsNotFound(int iters, int size) {
188 BenchmarkSuspender susp;
190 CHECK_LT(size, kMaxValue);
191 for (int i = 0; i < size; ++i) {
197 for (int i = 0; i < iters; ++i) {
198 sum += (aset.end() == aset.find(2 * i + 1));
206 void BM_SetContainsFound(int iters, int size) {
207 BenchmarkSuspender susp;
209 CHECK_LT(size, kMaxValue);
211 for (int i = 0; i < size; ++i) {
215 std::vector<int> values;
216 for (int i = 0; i < iters; ++i) {
217 values.push_back(rand() % size);
222 for (int i = 0; i < iters; ++i) {
223 sum += (aset.end() == aset.find(values[i]));
231 void BM_CSLContainsFound(int iters, int size) {
232 BenchmarkSuspender susp;
233 auto skipList = SkipListType::create(kInitHeadHeight);
234 CHECK_LT(size, kMaxValue);
236 for (int i = 0; i < size; ++i) {
239 std::vector<int> values;
240 for (int i = 0; i < iters; ++i) {
241 values.push_back(rand() % size);
246 for (int i = 0; i < iters; ++i) {
247 sum += skipList.contains(values[i]);
255 void BM_CSLContainsNotFound(int iters, int size) {
256 BenchmarkSuspender susp;
257 auto skipList = SkipListType::create(kInitHeadHeight);
258 CHECK_LT(size, kMaxValue);
260 for (int i = 0; i < size; ++i) {
266 for (int i = 0; i < iters; ++i) {
267 sum += skipList.contains(2 * i + 1);
275 void BM_AddSet(int iters, int size) {
276 BenchmarkSuspender susp;
278 for (int i = 0; i < size; ++i) {
279 aset.insert(gData[i]);
283 for (int i = size; i < size + iters; ++i) {
284 aset.insert(gData[i]);
288 void BM_AddSkipList(int iters, int size) {
289 BenchmarkSuspender susp;
290 auto skipList = SkipListType::create(kInitHeadHeight);
291 for (int i = 0; i < size; ++i) {
292 skipList.add(gData[i]);
296 for (int i = size; i < size + iters; ++i) {
297 skipList.add(gData[i]);
301 BENCHMARK(Accessor, iters) {
302 BenchmarkSuspender susp;
303 auto skiplist = SkipListType::createInstance(kInitHeadHeight);
304 auto sl = skiplist.get();
307 for (int i = 0; i < iters; ++i) {
308 SkipListAccessor accessor(sl);
312 // a benchmark to estimate the
313 // low bound of doing a ref counting for an Accessor
314 BENCHMARK(accessorBasicRefcounting, iters) {
315 BenchmarkSuspender susp;
316 auto* value = new std::atomic<int32_t>();
317 auto* dirty = new std::atomic<int32_t>();
319 folly::MicroSpinLock l;
323 for (int i = 0; i < iters; ++i) {
324 value->fetch_add(1, std::memory_order_relaxed);
325 if (dirty->load(std::memory_order_acquire) != 0) {
326 folly::MSLGuard g(l);
328 value->fetch_sub(1, std::memory_order_relaxed);
338 // Data For testing contention benchmark
339 class ConcurrentAccessData {
341 explicit ConcurrentAccessData(int size) :
342 skipList_(SkipListType::create(10)),
343 sets_(FLAGS_num_sets), locks_(FLAGS_num_sets) {
345 for (int i = 0; i < size; ++i) {
350 for (int i = 0; i < FLAGS_num_sets; ++i) {
351 locks_[i] = new RWSpinLock();
352 if (i > 0) sets_[i] = sets_[0];
356 int64_t setMemorySize = sets_[0].size() * sizeof(*sets_[0].begin()._M_node);
357 int64_t cslMemorySize = 0;
358 for (auto it = skipList_.begin(); it != skipList_.end(); ++it) {
359 cslMemorySize += it.nodeSize();
362 LOG(INFO) << "size=" << sets_[0].size()
363 << "; std::set memory size=" << setMemorySize
364 << "; csl memory size=" << cslMemorySize;
366 readValues_.reserve(size);
367 deleteValues_.reserve(size);
368 writeValues_.reserve(size);
369 for (int i = size; i < 2 * size; ++i) {
370 readValues_.push_back(2 * i);
371 deleteValues_.push_back(2 * i);
373 // half new values and half already in the list
374 writeValues_.push_back((rand() % 2) + 2 * i);
376 std::random_shuffle(readValues_.begin(), readValues_.end());
377 std::random_shuffle(deleteValues_.begin(), deleteValues_.end());
378 std::random_shuffle(writeValues_.begin(), writeValues_.end());
381 ~ConcurrentAccessData() {
382 FOR_EACH(lock, locks_) delete *lock;
385 inline bool skipListFind(int idx, ValueType val) {
386 return skipList_.contains(val);
388 inline void skipListInsert(int idx, ValueType val) {
391 inline void skipListErase(int idx, ValueType val) {
392 skipList_.remove(val);
395 inline bool setFind(int idx, ValueType val) {
396 RWSpinLock::ReadHolder g(locks_[idx]);
397 return sets_[idx].find(val) == sets_[idx].end();
399 inline void setInsert(int idx, ValueType val) {
400 RWSpinLock::WriteHolder g(locks_[idx]);
401 sets_[idx].insert(val);
403 inline void setErase(int idx, ValueType val) {
404 RWSpinLock::WriteHolder g(locks_[idx]);
405 sets_[idx].erase(val);
408 void runSkipList(int id, int iters) {
410 for (int i = 0; i < iters; ++i) {
411 sum += accessSkipList(id, i);
416 void runSet(int id, int iters) {
418 for (int i = 0; i < iters; ++i) {
419 sum += accessSet(id, i);
424 bool accessSkipList(int64_t id, int t) {
425 if (t > readValues_.size()) {
426 t = t % readValues_.size();
428 uint32_t h = folly::hash::twang_32from64(t * id);
431 if ((h & 0x31) == 0) { // 1/4 chance to delete
432 skipListErase(0, deleteValues_[t]);
434 skipListInsert(0, writeValues_[t]);
438 return skipListFind(0, readValues_[t]);
442 bool accessSet(int64_t id, int t) {
443 if (t > readValues_.size()) {
444 t = t % readValues_.size();
446 uint32_t h = folly::hash::twang_32from64(t * id);
447 int idx = (h % FLAGS_num_sets);
448 switch (h % 8) { // 1/8 chance to write
450 if ((h & 0x31) == 0) { // 1/32 chance to delete
451 setErase(idx, deleteValues_[t]);
453 setInsert(idx, writeValues_[t]);
457 return setFind(idx, readValues_[t]);
462 SkipListType::Accessor skipList_;
463 std::vector<SetType> sets_;
464 std::vector<RWSpinLock*> locks_;
466 std::vector<ValueType> readValues_;
467 std::vector<ValueType> writeValues_;
468 std::vector<ValueType> deleteValues_;
471 static std::map<int, boost::shared_ptr<ConcurrentAccessData> > g_data;
473 static ConcurrentAccessData *mayInitTestData(int size) {
474 auto it = g_data.find(size);
475 if (it == g_data.end()) {
476 auto ptr = boost::shared_ptr<ConcurrentAccessData>(
477 new ConcurrentAccessData(size));
481 return it->second.get();
484 void BM_ContentionCSL(int iters, int size) {
485 BenchmarkSuspender susp;
486 auto data = mayInitTestData(size);
487 std::vector<boost::thread> threads;
490 for (int i = 0; i < FLAGS_num_threads; ++i) {
491 threads.push_back(boost::thread(
492 &ConcurrentAccessData::runSkipList, data, i, iters));
494 FOR_EACH(t, threads) {
499 void BM_ContentionStdSet(int iters, int size) {
500 BenchmarkSuspender susp;
501 auto data = mayInitTestData(size);
502 std::vector<boost::thread> threads;
505 for (int i = 0; i < FLAGS_num_threads; ++i) {
506 threads.push_back(boost::thread(
507 &ConcurrentAccessData::runSet, data, i, iters));
509 FOR_EACH(t, threads) {
516 // Single-thread benchmarking
518 BENCHMARK_DRAW_LINE();
520 BENCHMARK_PARAM(BM_IterateOverSet, 1000);
521 BENCHMARK_PARAM(BM_IterateSkipList, 1000);
522 BENCHMARK_DRAW_LINE();
523 BENCHMARK_PARAM(BM_IterateOverSet, 1000000);
524 BENCHMARK_PARAM(BM_IterateSkipList, 1000000);
525 BENCHMARK_DRAW_LINE();
527 // find with keys in the set
528 BENCHMARK_PARAM(BM_SetContainsFound, 1000);
529 BENCHMARK_PARAM(BM_CSLContainsFound, 1000);
530 BENCHMARK_DRAW_LINE();
531 BENCHMARK_PARAM(BM_SetContainsFound, 100000);
532 BENCHMARK_PARAM(BM_CSLContainsFound, 100000);
533 BENCHMARK_DRAW_LINE();
534 BENCHMARK_PARAM(BM_SetContainsFound, 1000000);
535 BENCHMARK_PARAM(BM_CSLContainsFound, 1000000);
536 BENCHMARK_DRAW_LINE();
537 BENCHMARK_PARAM(BM_SetContainsFound, 10000000);
538 BENCHMARK_PARAM(BM_CSLContainsFound, 10000000);
539 BENCHMARK_DRAW_LINE();
542 // find with keys not in the set
543 BENCHMARK_PARAM(BM_SetContainsNotFound, 1000);
544 BENCHMARK_PARAM(BM_CSLContainsNotFound, 1000);
545 BENCHMARK_DRAW_LINE();
546 BENCHMARK_PARAM(BM_SetContainsNotFound, 100000);
547 BENCHMARK_PARAM(BM_CSLContainsNotFound, 100000);
548 BENCHMARK_DRAW_LINE();
549 BENCHMARK_PARAM(BM_SetContainsNotFound, 1000000);
550 BENCHMARK_PARAM(BM_CSLContainsNotFound, 1000000);
551 BENCHMARK_DRAW_LINE();
554 BENCHMARK_PARAM(BM_AddSet, 1000);
555 BENCHMARK_PARAM(BM_AddSkipList, 1000);
556 BENCHMARK_DRAW_LINE();
558 BENCHMARK_PARAM(BM_AddSet, 65536);
559 BENCHMARK_PARAM(BM_AddSkipList, 65536);
560 BENCHMARK_DRAW_LINE();
562 BENCHMARK_PARAM(BM_AddSet, 1000000);
563 BENCHMARK_PARAM(BM_AddSkipList, 1000000);
564 BENCHMARK_DRAW_LINE();
566 BENCHMARK_PARAM(BM_SetMerge, 1000);
567 BENCHMARK_PARAM(BM_CSLMergeIntersection, 1000);
568 BENCHMARK_PARAM(BM_CSLMergeLookup, 1000);
569 BENCHMARK_DRAW_LINE();
571 BENCHMARK_PARAM(BM_SetMerge, 65536);
572 BENCHMARK_PARAM(BM_CSLMergeIntersection, 65536);
573 BENCHMARK_PARAM(BM_CSLMergeLookup, 65536);
574 BENCHMARK_DRAW_LINE();
576 BENCHMARK_PARAM(BM_SetMerge, 1000000);
577 BENCHMARK_PARAM(BM_CSLMergeIntersection, 1000000);
578 BENCHMARK_PARAM(BM_CSLMergeLookup, 1000000);
579 BENCHMARK_DRAW_LINE();
582 // multi-thread benchmarking
584 BENCHMARK_PARAM(BM_ContentionStdSet, 1024);
585 BENCHMARK_PARAM(BM_ContentionCSL, 1024);
586 BENCHMARK_DRAW_LINE();
588 BENCHMARK_PARAM(BM_ContentionStdSet, 65536);
589 BENCHMARK_PARAM(BM_ContentionCSL, 65536);
590 BENCHMARK_DRAW_LINE();
592 BENCHMARK_PARAM(BM_ContentionStdSet, 1048576);
593 BENCHMARK_PARAM(BM_ContentionCSL, 1048576);
594 BENCHMARK_DRAW_LINE();
598 int main(int argc, char** argv) {
599 google::InitGoogleLogging(argv[0]);
600 google::ParseCommandLineFlags(&argc, &argv, true);
609 Benchmark on Intel(R) Xeon(R) CPU X5650 @2.67GHz
611 ==============================================================================
612 1 thread Benchmark Iters Total t t/iter iter/sec
613 ------------------------------------------------------------------------------
614 +37.0% BM_Accessor 100000 1.958 ms 19.58 ns 48.71 M
615 * BM_AccessorBasicRefcounting 100000 1.429 ms 14.29 ns 66.74 M
616 ------------------------------------------------------------------------------
617 + 603% BM_IterateOverSet/1000 100000 1.589 ms 15.89 ns 60.02 M
618 * BM_IterateSkipList/1000 100000 226 us 2.26 ns 422 M
619 ------------------------------------------------------------------------------
620 + 107% BM_IterateOverSet/976.6k 100000 8.324 ms 83.24 ns 11.46 M
621 * BM_IterateSkipList/976.6k 100000 4.016 ms 40.16 ns 23.75 M
622 ------------------------------------------------------------------------------
623 * BM_SetContainsFound/1000 100000 7.082 ms 70.82 ns 13.47 M
624 +39.9% BM_CSLContainsFound/1000 100000 9.908 ms 99.08 ns 9.625 M
625 ------------------------------------------------------------------------------
626 * BM_SetContainsFound/97.66k 100000 23.8 ms 238 ns 4.006 M
627 +5.97% BM_CSLContainsFound/97.66k 100000 25.23 ms 252.3 ns 3.781 M
628 ------------------------------------------------------------------------------
629 +33.6% BM_SetContainsFound/976.6k 100000 64.3 ms 643 ns 1.483 M
630 * BM_CSLContainsFound/976.6k 100000 48.13 ms 481.3 ns 1.981 M
631 ------------------------------------------------------------------------------
632 +30.3% BM_SetContainsFound/9.537M 100000 115.1 ms 1.151 us 848.6 k
633 * BM_CSLContainsFound/9.537M 100000 88.33 ms 883.3 ns 1.08 M
634 ------------------------------------------------------------------------------
635 * BM_SetContainsNotFound/1000 100000 2.081 ms 20.81 ns 45.83 M
636 +76.2% BM_CSLContainsNotFound/1000 100000 3.667 ms 36.67 ns 26.01 M
637 ------------------------------------------------------------------------------
638 * BM_SetContainsNotFound/97.66k 100000 6.049 ms 60.49 ns 15.77 M
639 +32.7% BM_CSLContainsNotFound/97.66k 100000 8.025 ms 80.25 ns 11.88 M
640 ------------------------------------------------------------------------------
641 * BM_SetContainsNotFound/976.6k 100000 7.464 ms 74.64 ns 12.78 M
642 +12.8% BM_CSLContainsNotFound/976.6k 100000 8.417 ms 84.17 ns 11.33 M
643 ------------------------------------------------------------------------------
644 * BM_AddSet/1000 100000 29.26 ms 292.6 ns 3.259 M
645 +70.0% BM_AddSkipList/1000 100000 49.75 ms 497.5 ns 1.917 M
646 ------------------------------------------------------------------------------
647 * BM_AddSet/64k 100000 38.73 ms 387.3 ns 2.462 M
648 +55.7% BM_AddSkipList/64k 100000 60.3 ms 603 ns 1.581 M
649 ------------------------------------------------------------------------------
650 * BM_AddSet/976.6k 100000 75.71 ms 757.1 ns 1.26 M
651 +33.6% BM_AddSkipList/976.6k 100000 101.2 ms 1.012 us 965.3 k
652 ------------------------------------------------------------------------------
653 + 716% BM_SetMerge/1000 100000 6.872 ms 68.72 ns 13.88 M
654 * BM_CSLMergeIntersection/1000 100000 842 us 8.42 ns 113.3 M
655 + 268% BM_CSLMergeLookup/1000 100000 3.1 ms 31 ns 30.76 M
656 ------------------------------------------------------------------------------
657 +36.3% BM_SetMerge/64k 100000 14.03 ms 140.3 ns 6.798 M
658 +39.4% BM_CSLMergeIntersection/64k 100000 14.35 ms 143.5 ns 6.645 M
659 * BM_CSLMergeLookup/64k 100000 10.29 ms 102.9 ns 9.266 M
660 ------------------------------------------------------------------------------
661 +10.3% BM_SetMerge/976.6k 100000 46.24 ms 462.4 ns 2.062 M
662 +25.1% BM_CSLMergeIntersection/976.6k 100000 52.47 ms 524.7 ns 1.818 M
663 * BM_CSLMergeLookup/976.6k 100000 41.94 ms 419.3 ns 2.274 M
664 ------------------------------------------------------------------------------
667 ==============================================================================
668 Contention benchmark 7/8 find, 3/32 insert, 1/32 erase
670 4 threads Benchmark Iters Total t t/iter iter/sec
671 ------------------------------------------------------------------------------
672 + 269% BM_ContentionStdSet/1k 100000 75.66 ms 756.6 ns 1.26 M
673 * BM_ContentionCSL/1k 100000 20.47 ms 204.7 ns 4.658 M
674 ------------------------------------------------------------------------------
675 + 228% BM_ContentionStdSet/64k 100000 105.6 ms 1.056 us 924.9 k
676 * BM_ContentionCSL/64k 100000 32.18 ms 321.8 ns 2.963 M
677 ------------------------------------------------------------------------------
678 + 224% BM_ContentionStdSet/1M 100000 117.4 ms 1.174 us 832.2 k
679 * BM_ContentionCSL/1M 100000 36.18 ms 361.8 ns 2.636 M
680 ------------------------------------------------------------------------------
683 12 threads Benchmark Iters Total t t/iter iter/sec
684 ------------------------------------------------------------------------------
685 + 697% BM_ContentionStdSet/1k 100000 455.3 ms 4.553 us 214.5 k
686 * BM_ContentionCSL/1k 100000 57.12 ms 571.2 ns 1.67 M
687 ------------------------------------------------------------------------------
688 +1257% BM_ContentionStdSet/64k 100000 654.9 ms 6.549 us 149.1 k
689 * BM_ContentionCSL/64k 100000 48.24 ms 482.4 ns 1.977 M
690 ------------------------------------------------------------------------------
691 +1262% BM_ContentionStdSet/1M 100000 657.3 ms 6.573 us 148.6 k
692 * BM_ContentionCSL/1M 100000 48.25 ms 482.5 ns 1.977 M
693 ------------------------------------------------------------------------------