2 * Copyright 2016 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 #include <folly/SharedMutex.h>
23 #include <gtest/gtest.h>
24 #include <boost/optional.hpp>
25 #include <boost/thread/shared_mutex.hpp>
27 #include <folly/Benchmark.h>
28 #include <folly/MPMCQueue.h>
29 #include <folly/RWSpinLock.h>
30 #include <folly/Random.h>
31 #include <folly/portability/GFlags.h>
32 #include <folly/test/DeterministicSchedule.h>
34 using namespace folly;
35 using namespace folly::test;
37 using namespace chrono;
39 typedef DeterministicSchedule DSched;
40 typedef SharedMutexImpl<true, void, DeterministicAtomic, true>
41 DSharedMutexReadPriority;
42 typedef SharedMutexImpl<false, void, DeterministicAtomic, true>
43 DSharedMutexWritePriority;
45 template <typename Lock>
48 SharedMutexToken token1;
49 SharedMutexToken token2;
50 SharedMutexToken token3;
52 EXPECT_TRUE(lock.try_lock());
53 EXPECT_FALSE(lock.try_lock());
54 EXPECT_FALSE(lock.try_lock_shared(token1));
57 EXPECT_TRUE(lock.try_lock_shared(token1));
58 EXPECT_FALSE(lock.try_lock());
59 EXPECT_TRUE(lock.try_lock_shared(token2));
60 lock.lock_shared(token3);
61 lock.unlock_shared(token3);
62 lock.unlock_shared(token2);
63 lock.unlock_shared(token1);
68 lock.lock_shared(token1);
69 lock.lock_shared(token2);
70 lock.unlock_shared(token1);
71 lock.unlock_shared(token2);
74 lock.unlock_and_lock_shared(token1);
75 lock.lock_shared(token2);
76 lock.unlock_shared(token2);
77 lock.unlock_shared(token1);
80 TEST(SharedMutex, basic) {
81 runBasicTest<SharedMutexReadPriority>();
82 runBasicTest<SharedMutexWritePriority>();
85 template <typename Lock>
86 void runBasicHoldersTest() {
88 SharedMutexToken token;
91 // create an exclusive write lock via holder
92 typename Lock::WriteHolder holder(lock);
93 EXPECT_FALSE(lock.try_lock());
94 EXPECT_FALSE(lock.try_lock_shared(token));
96 // move ownership to another write holder via move constructor
97 typename Lock::WriteHolder holder2(std::move(holder));
98 EXPECT_FALSE(lock.try_lock());
99 EXPECT_FALSE(lock.try_lock_shared(token));
101 // move ownership to another write holder via assign operator
102 typename Lock::WriteHolder holder3;
103 holder3 = std::move(holder2);
104 EXPECT_FALSE(lock.try_lock());
105 EXPECT_FALSE(lock.try_lock_shared(token));
107 // downgrade from exclusive to upgrade lock via move constructor
108 typename Lock::UpgradeHolder holder4(std::move(holder3));
110 // ensure we can lock from a shared source
111 EXPECT_FALSE(lock.try_lock());
112 EXPECT_TRUE(lock.try_lock_shared(token));
113 lock.unlock_shared(token);
115 // promote from upgrade to exclusive lock via move constructor
116 typename Lock::WriteHolder holder5(std::move(holder4));
117 EXPECT_FALSE(lock.try_lock());
118 EXPECT_FALSE(lock.try_lock_shared(token));
120 // downgrade exclusive to shared lock via move constructor
121 typename Lock::ReadHolder holder6(std::move(holder5));
123 // ensure we can lock from another shared source
124 EXPECT_FALSE(lock.try_lock());
125 EXPECT_TRUE(lock.try_lock_shared(token));
126 lock.unlock_shared(token);
130 typename Lock::WriteHolder holder(lock);
131 EXPECT_FALSE(lock.try_lock());
135 typename Lock::ReadHolder holder(lock);
136 typename Lock::ReadHolder holder2(lock);
137 typename Lock::UpgradeHolder holder3(lock);
141 typename Lock::UpgradeHolder holder(lock);
142 typename Lock::ReadHolder holder2(lock);
143 typename Lock::ReadHolder holder3(std::move(holder));
147 TEST(SharedMutex, basic_holders) {
148 runBasicHoldersTest<SharedMutexReadPriority>();
149 runBasicHoldersTest<SharedMutexWritePriority>();
152 template <typename Lock>
153 void runManyReadLocksTestWithTokens() {
156 vector<SharedMutexToken> tokens;
157 for (int i = 0; i < 1000; ++i) {
158 tokens.emplace_back();
159 EXPECT_TRUE(lock.try_lock_shared(tokens.back()));
161 for (auto& token : tokens) {
162 lock.unlock_shared(token);
164 EXPECT_TRUE(lock.try_lock());
168 TEST(SharedMutex, many_read_locks_with_tokens) {
169 runManyReadLocksTestWithTokens<SharedMutexReadPriority>();
170 runManyReadLocksTestWithTokens<SharedMutexWritePriority>();
173 template <typename Lock>
174 void runManyReadLocksTestWithoutTokens() {
177 for (int i = 0; i < 1000; ++i) {
178 EXPECT_TRUE(lock.try_lock_shared());
180 for (int i = 0; i < 1000; ++i) {
181 lock.unlock_shared();
183 EXPECT_TRUE(lock.try_lock());
187 TEST(SharedMutex, many_read_locks_without_tokens) {
188 runManyReadLocksTestWithoutTokens<SharedMutexReadPriority>();
189 runManyReadLocksTestWithoutTokens<SharedMutexWritePriority>();
192 template <typename Lock>
193 void runTimeoutInPastTest() {
196 EXPECT_TRUE(lock.try_lock_for(milliseconds(0)));
198 EXPECT_TRUE(lock.try_lock_for(milliseconds(-1)));
200 EXPECT_TRUE(lock.try_lock_shared_for(milliseconds(0)));
201 lock.unlock_shared();
202 EXPECT_TRUE(lock.try_lock_shared_for(milliseconds(-1)));
203 lock.unlock_shared();
204 EXPECT_TRUE(lock.try_lock_until(system_clock::now() - milliseconds(1)));
207 lock.try_lock_shared_until(system_clock::now() - milliseconds(1)));
208 lock.unlock_shared();
209 EXPECT_TRUE(lock.try_lock_until(steady_clock::now() - milliseconds(1)));
212 lock.try_lock_shared_until(steady_clock::now() - milliseconds(1)));
213 lock.unlock_shared();
216 TEST(SharedMutex, timeout_in_past) {
217 runTimeoutInPastTest<SharedMutexReadPriority>();
218 runTimeoutInPastTest<SharedMutexWritePriority>();
221 template <class Func>
222 bool funcHasDuration(milliseconds expectedDuration, Func func) {
223 // elapsed time should eventually fall within expectedDuration +- 25%
224 for (int tries = 0; tries < 100; ++tries) {
225 auto start = steady_clock::now();
227 auto elapsed = steady_clock::now() - start;
228 if (elapsed > expectedDuration - expectedDuration / 4 &&
229 elapsed < expectedDuration + expectedDuration / 4) {
236 template <typename Lock>
237 void runFailingTryTimeoutTest() {
240 EXPECT_TRUE(funcHasDuration(milliseconds(10), [&] {
241 EXPECT_FALSE(lock.try_lock_for(milliseconds(10)));
243 EXPECT_TRUE(funcHasDuration(milliseconds(10), [&] {
244 typename Lock::Token token;
245 EXPECT_FALSE(lock.try_lock_shared_for(milliseconds(10), token));
247 EXPECT_TRUE(funcHasDuration(milliseconds(10), [&] {
248 EXPECT_FALSE(lock.try_lock_upgrade_for(milliseconds(10)));
250 EXPECT_TRUE(funcHasDuration(milliseconds(10), [&] {
251 EXPECT_FALSE(lock.try_lock_until(steady_clock::now() + milliseconds(10)));
253 EXPECT_TRUE(funcHasDuration(milliseconds(10), [&] {
254 typename Lock::Token token;
255 EXPECT_FALSE(lock.try_lock_shared_until(
256 steady_clock::now() + milliseconds(10), token));
258 EXPECT_TRUE(funcHasDuration(milliseconds(10), [&] {
260 lock.try_lock_upgrade_until(steady_clock::now() + milliseconds(10)));
262 EXPECT_TRUE(funcHasDuration(milliseconds(10), [&] {
263 EXPECT_FALSE(lock.try_lock_until(system_clock::now() + milliseconds(10)));
265 EXPECT_TRUE(funcHasDuration(milliseconds(10), [&] {
266 typename Lock::Token token;
267 EXPECT_FALSE(lock.try_lock_shared_until(
268 system_clock::now() + milliseconds(10), token));
270 EXPECT_TRUE(funcHasDuration(milliseconds(10), [&] {
272 lock.try_lock_upgrade_until(system_clock::now() + milliseconds(10)));
277 EXPECT_TRUE(funcHasDuration(milliseconds(10), [&] {
278 EXPECT_FALSE(lock.try_lock_for(milliseconds(10)));
280 EXPECT_TRUE(funcHasDuration(milliseconds(10), [&] {
281 EXPECT_FALSE(lock.try_lock_until(steady_clock::now() + milliseconds(10)));
283 EXPECT_TRUE(funcHasDuration(milliseconds(10), [&] {
284 EXPECT_FALSE(lock.try_lock_until(system_clock::now() + milliseconds(10)));
286 lock.unlock_shared();
289 for (int p = 0; p < 8; ++p) {
290 EXPECT_FALSE(lock.try_lock_for(nanoseconds(1 << p)));
294 for (int p = 0; p < 8; ++p) {
295 typename Lock::ReadHolder holder1(lock);
296 typename Lock::ReadHolder holder2(lock);
297 typename Lock::ReadHolder holder3(lock);
298 EXPECT_FALSE(lock.try_lock_for(nanoseconds(1 << p)));
302 TEST(SharedMutex, failing_try_timeout) {
303 runFailingTryTimeoutTest<SharedMutexReadPriority>();
304 runFailingTryTimeoutTest<SharedMutexWritePriority>();
307 template <typename Lock>
308 void runBasicUpgradeTest() {
310 typename Lock::Token token1;
311 typename Lock::Token token2;
314 EXPECT_FALSE(lock.try_lock());
315 EXPECT_TRUE(lock.try_lock_shared(token1));
316 lock.unlock_shared(token1);
317 lock.unlock_upgrade();
320 lock.unlock_upgrade_and_lock();
321 EXPECT_FALSE(lock.try_lock_shared(token1));
325 lock.unlock_upgrade_and_lock_shared(token1);
327 lock.unlock_upgrade_and_lock_shared(token2);
328 lock.unlock_shared(token1);
329 lock.unlock_shared(token2);
332 lock.unlock_and_lock_upgrade();
333 EXPECT_TRUE(lock.try_lock_shared(token1));
334 lock.unlock_upgrade();
335 lock.unlock_shared(token1);
338 TEST(SharedMutex, basic_upgrade_tests) {
339 runBasicUpgradeTest<SharedMutexReadPriority>();
340 runBasicUpgradeTest<SharedMutexWritePriority>();
343 TEST(SharedMutex, read_has_prio) {
344 SharedMutexReadPriority lock;
345 SharedMutexToken token1;
346 SharedMutexToken token2;
347 lock.lock_shared(token1);
348 bool exclusiveAcquired = false;
349 auto writer = thread([&] {
351 exclusiveAcquired = true;
355 // lock() can't complete until we unlock token1, but it should stake
356 // its claim with regards to other exclusive or upgrade locks. We can
357 // use try_lock_upgrade to poll for that eventuality.
358 while (lock.try_lock_upgrade()) {
359 lock.unlock_upgrade();
360 this_thread::yield();
362 EXPECT_FALSE(exclusiveAcquired);
364 // Even though lock() is stuck we should be able to get token2
365 EXPECT_TRUE(lock.try_lock_shared(token2));
366 lock.unlock_shared(token1);
367 lock.unlock_shared(token2);
369 EXPECT_TRUE(exclusiveAcquired);
372 TEST(SharedMutex, write_has_prio) {
373 SharedMutexWritePriority lock;
374 SharedMutexToken token1;
375 SharedMutexToken token2;
376 lock.lock_shared(token1);
377 auto writer = thread([&] {
382 // eventually lock() should block readers
383 while (lock.try_lock_shared(token2)) {
384 lock.unlock_shared(token2);
385 this_thread::yield();
388 lock.unlock_shared(token1);
393 SharedMutexToken token;
395 template <typename T>
400 template <typename T>
401 void unlock(T* lock) {
405 template <typename T>
406 void lock_shared(T* lock) {
407 lock->lock_shared(token);
410 template <typename T>
411 void unlock_shared(T* lock) {
412 lock->unlock_shared(token);
417 template <typename T>
422 template <typename T>
423 void unlock(T* lock) {
427 template <typename T>
428 void lock_shared(T* lock) {
432 template <typename T>
433 void unlock_shared(T* lock) {
434 lock->unlock_shared();
439 template <typename T>
444 template <typename T>
445 void unlock(T* lock) {
449 template <typename T>
450 void lock_shared(T* lock) {
454 template <typename T>
455 void unlock_shared(T* lock) {
461 pthread_rwlock_t lock_;
463 PosixRWLock() { pthread_rwlock_init(&lock_, nullptr); }
465 ~PosixRWLock() { pthread_rwlock_destroy(&lock_); }
467 void lock() { pthread_rwlock_wrlock(&lock_); }
469 void unlock() { pthread_rwlock_unlock(&lock_); }
471 void lock_shared() { pthread_rwlock_rdlock(&lock_); }
473 void unlock_shared() { pthread_rwlock_unlock(&lock_); }
477 pthread_mutex_t lock_;
479 PosixMutex() { pthread_mutex_init(&lock_, nullptr); }
481 ~PosixMutex() { pthread_mutex_destroy(&lock_); }
483 void lock() { pthread_mutex_lock(&lock_); }
485 void unlock() { pthread_mutex_unlock(&lock_); }
487 void lock_shared() { pthread_mutex_lock(&lock_); }
489 void unlock_shared() { pthread_mutex_unlock(&lock_); }
492 template <template <typename> class Atom, typename Lock, typename Locker>
493 static void runContendedReaders(size_t numOps,
495 bool useSeparateLocks) {
499 int valueProtectedByLock = 10;
502 Atom<bool> go(false);
503 Atom<bool>* goPtr = &go; // workaround for clang bug
504 vector<thread> threads(numThreads);
507 for (size_t t = 0; t < numThreads; ++t) {
508 threads[t] = DSched::thread([&, t, numThreads] {
510 Lock* lock = useSeparateLocks ? &privateLock : &globalLock;
512 while (!goPtr->load()) {
513 this_thread::yield();
515 for (size_t op = t; op < numOps; op += numThreads) {
516 locker.lock_shared(lock);
517 // note: folly::doNotOptimizeAway reads and writes to its arg,
518 // so the following two lines are very different than a call
519 // to folly::doNotOptimizeAway(valueProtectedByLock);
520 auto copy = valueProtectedByLock;
521 folly::doNotOptimizeAway(copy);
522 locker.unlock_shared(lock);
529 for (auto& thr : threads) {
534 static void folly_rwspin_reads(uint32_t numOps,
536 bool useSeparateLocks) {
537 runContendedReaders<atomic, RWSpinLock, Locker>(
538 numOps, numThreads, useSeparateLocks);
541 static void shmtx_wr_pri_reads(uint32_t numOps,
543 bool useSeparateLocks) {
544 runContendedReaders<atomic, SharedMutexWritePriority, TokenLocker>(
545 numOps, numThreads, useSeparateLocks);
548 static void shmtx_w_bare_reads(uint32_t numOps,
550 bool useSeparateLocks) {
551 runContendedReaders<atomic, SharedMutexWritePriority, Locker>(
552 numOps, numThreads, useSeparateLocks);
555 static void shmtx_rd_pri_reads(uint32_t numOps,
557 bool useSeparateLocks) {
558 runContendedReaders<atomic, SharedMutexReadPriority, TokenLocker>(
559 numOps, numThreads, useSeparateLocks);
562 static void shmtx_r_bare_reads(uint32_t numOps,
564 bool useSeparateLocks) {
565 runContendedReaders<atomic, SharedMutexReadPriority, Locker>(
566 numOps, numThreads, useSeparateLocks);
569 static void folly_ticket_reads(uint32_t numOps,
571 bool useSeparateLocks) {
572 runContendedReaders<atomic, RWTicketSpinLock64, Locker>(
573 numOps, numThreads, useSeparateLocks);
576 static void boost_shared_reads(uint32_t numOps,
578 bool useSeparateLocks) {
579 runContendedReaders<atomic, boost::shared_mutex, Locker>(
580 numOps, numThreads, useSeparateLocks);
583 static void pthrd_rwlock_reads(uint32_t numOps,
585 bool useSeparateLocks) {
586 runContendedReaders<atomic, PosixRWLock, Locker>(
587 numOps, numThreads, useSeparateLocks);
590 template <template <typename> class Atom, typename Lock, typename Locker>
591 static void runMixed(size_t numOps,
593 double writeFraction,
594 bool useSeparateLocks) {
598 int valueProtectedByLock = 0;
601 Atom<bool> go(false);
602 Atom<bool>* goPtr = &go; // workaround for clang bug
603 vector<thread> threads(numThreads);
606 for (size_t t = 0; t < numThreads; ++t) {
607 threads[t] = DSched::thread([&, t, numThreads] {
608 struct drand48_data buffer;
609 srand48_r(t, &buffer);
610 long writeThreshold = writeFraction * 0x7fffffff;
612 Lock* lock = useSeparateLocks ? &privateLock : &globalLock;
614 while (!goPtr->load()) {
615 this_thread::yield();
617 for (size_t op = t; op < numOps; op += numThreads) {
619 lrand48_r(&buffer, &randVal);
620 bool writeOp = randVal < writeThreshold;
623 if (!useSeparateLocks) {
624 ++valueProtectedByLock;
628 locker.lock_shared(lock);
629 auto v = valueProtectedByLock;
630 folly::doNotOptimizeAway(v);
631 locker.unlock_shared(lock);
639 for (auto& thr : threads) {
644 static void folly_rwspin(size_t numOps,
646 double writeFraction,
647 bool useSeparateLocks) {
648 runMixed<atomic, RWSpinLock, Locker>(
649 numOps, numThreads, writeFraction, useSeparateLocks);
652 static void shmtx_wr_pri(uint32_t numOps,
654 double writeFraction,
655 bool useSeparateLocks) {
656 runMixed<atomic, SharedMutexWritePriority, TokenLocker>(
657 numOps, numThreads, writeFraction, useSeparateLocks);
660 static void shmtx_w_bare(uint32_t numOps,
662 double writeFraction,
663 bool useSeparateLocks) {
664 runMixed<atomic, SharedMutexWritePriority, Locker>(
665 numOps, numThreads, writeFraction, useSeparateLocks);
668 static void shmtx_rd_pri(uint32_t numOps,
670 double writeFraction,
671 bool useSeparateLocks) {
672 runMixed<atomic, SharedMutexReadPriority, TokenLocker>(
673 numOps, numThreads, writeFraction, useSeparateLocks);
676 static void shmtx_r_bare(uint32_t numOps,
678 double writeFraction,
679 bool useSeparateLocks) {
680 runMixed<atomic, SharedMutexReadPriority, Locker>(
681 numOps, numThreads, writeFraction, useSeparateLocks);
684 static void folly_ticket(size_t numOps,
686 double writeFraction,
687 bool useSeparateLocks) {
688 runMixed<atomic, RWTicketSpinLock64, Locker>(
689 numOps, numThreads, writeFraction, useSeparateLocks);
692 static void boost_shared(size_t numOps,
694 double writeFraction,
695 bool useSeparateLocks) {
696 runMixed<atomic, boost::shared_mutex, Locker>(
697 numOps, numThreads, writeFraction, useSeparateLocks);
700 static void pthrd_rwlock(size_t numOps,
702 double writeFraction,
703 bool useSeparateLocks) {
704 runMixed<atomic, PosixRWLock, Locker>(
705 numOps, numThreads, writeFraction, useSeparateLocks);
708 static void pthrd_mutex_(size_t numOps,
710 double writeFraction,
711 bool useSeparateLocks) {
712 runMixed<atomic, PosixMutex, Locker>(
713 numOps, numThreads, writeFraction, useSeparateLocks);
716 template <typename Lock, template <typename> class Atom>
717 static void runAllAndValidate(size_t numOps, size_t numThreads) {
719 Atom<int> globalExclusiveCount(0);
720 Atom<int> globalUpgradeCount(0);
721 Atom<int> globalSharedCount(0);
723 Atom<bool> go(false);
725 // clang crashes on access to Atom<> captured by ref in closure
726 Atom<int>* globalExclusiveCountPtr = &globalExclusiveCount;
727 Atom<int>* globalUpgradeCountPtr = &globalUpgradeCount;
728 Atom<int>* globalSharedCountPtr = &globalSharedCount;
729 Atom<bool>* goPtr = &go;
731 vector<thread> threads(numThreads);
734 for (size_t t = 0; t < numThreads; ++t) {
735 threads[t] = DSched::thread([&, t, numThreads] {
736 struct drand48_data buffer;
737 srand48_r(t, &buffer);
739 bool exclusive = false;
740 bool upgrade = false;
742 bool ourGlobalTokenUsed = false;
743 SharedMutexToken ourGlobalToken;
746 vector<SharedMutexToken> privateTokens;
748 while (!goPtr->load()) {
749 this_thread::yield();
751 for (size_t op = t; op < numOps; op += numThreads) {
752 // randVal in [0,1000)
754 lrand48_r(&buffer, &randVal);
755 randVal = (long)((randVal * (uint64_t)1000) / 0x7fffffff);
757 // make as many assertions as possible about the global state
759 EXPECT_EQ(1, globalExclusiveCountPtr->load(memory_order_acquire));
760 EXPECT_EQ(0, globalUpgradeCountPtr->load(memory_order_acquire));
761 EXPECT_EQ(0, globalSharedCountPtr->load(memory_order_acquire));
764 EXPECT_EQ(0, globalExclusiveCountPtr->load(memory_order_acquire));
765 EXPECT_EQ(1, globalUpgradeCountPtr->load(memory_order_acquire));
768 EXPECT_EQ(0, globalExclusiveCountPtr->load(memory_order_acquire));
769 EXPECT_TRUE(globalSharedCountPtr->load(memory_order_acquire) > 0);
771 EXPECT_FALSE(ourGlobalTokenUsed);
774 // independent 20% chance we do something to the private lock
776 // it's okay to take multiple private shared locks because
777 // we never take an exclusive lock, so reader versus writer
778 // priority doesn't cause deadlocks
779 if (randVal < 100 && privateTokens.size() > 0) {
780 auto i = randVal % privateTokens.size();
781 privateLock.unlock_shared(privateTokens[i]);
782 privateTokens.erase(privateTokens.begin() + i);
784 SharedMutexToken token;
785 privateLock.lock_shared(token);
786 privateTokens.push_back(token);
791 // if we've got a lock, the only thing we can do is release it
792 // or transform it into a different kind of lock
795 --*globalExclusiveCountPtr;
798 } else if (randVal < 700) {
799 globalLock.unlock_and_lock_shared();
800 ++*globalSharedCountPtr;
802 } else if (randVal < 900) {
803 globalLock.unlock_and_lock_shared(ourGlobalToken);
804 ++*globalSharedCountPtr;
806 ourGlobalTokenUsed = true;
808 globalLock.unlock_and_lock_upgrade();
809 ++*globalUpgradeCountPtr;
812 } else if (upgrade) {
814 --*globalUpgradeCountPtr;
816 globalLock.unlock_upgrade();
817 } else if (randVal < 700) {
818 globalLock.unlock_upgrade_and_lock_shared();
819 ++*globalSharedCountPtr;
821 } else if (randVal < 900) {
822 globalLock.unlock_upgrade_and_lock_shared(ourGlobalToken);
823 ++*globalSharedCountPtr;
825 ourGlobalTokenUsed = true;
827 globalLock.unlock_upgrade_and_lock();
828 ++*globalExclusiveCountPtr;
833 --*globalSharedCountPtr;
834 if (ourGlobalTokenUsed) {
835 globalLock.unlock_shared(ourGlobalToken);
836 ourGlobalTokenUsed = false;
838 globalLock.unlock_shared();
840 } else if (randVal < 400) {
841 // 40% chance of shared lock with token, 5 ways to get it
843 // delta t goes from -1 millis to 7 millis
844 auto dt = microseconds(10 * (randVal - 100));
847 globalLock.lock_shared(ourGlobalToken);
849 } else if (randVal < 500) {
850 shared = globalLock.try_lock_shared(ourGlobalToken);
851 } else if (randVal < 600) {
852 shared = globalLock.try_lock_shared_for(dt, ourGlobalToken);
853 } else if (randVal < 800) {
854 shared = globalLock.try_lock_shared_until(
855 system_clock::now() + dt, ourGlobalToken);
858 ourGlobalTokenUsed = true;
859 ++*globalSharedCountPtr;
861 } else if (randVal < 800) {
862 // 40% chance of shared lock without token
863 auto dt = microseconds(10 * (randVal - 100));
865 globalLock.lock_shared();
867 } else if (randVal < 500) {
868 shared = globalLock.try_lock_shared();
869 } else if (randVal < 600) {
870 shared = globalLock.try_lock_shared_for(dt);
871 } else if (randVal < 800) {
872 shared = globalLock.try_lock_shared_until(
873 system_clock::now() + dt);
876 ++*globalSharedCountPtr;
878 } else if (randVal < 900) {
879 // 10% change of upgrade lock
880 globalLock.lock_upgrade();
882 ++*globalUpgradeCountPtr;
884 // 10% chance of exclusive lock, 5 ways to get it
886 // delta t goes from -1 millis to 9 millis
887 auto dt = microseconds(100 * (randVal - 910));
892 } else if (randVal < 500) {
893 exclusive = globalLock.try_lock();
894 } else if (randVal < 600) {
895 exclusive = globalLock.try_lock_for(dt);
896 } else if (randVal < 700) {
897 exclusive = globalLock.try_lock_until(steady_clock::now() + dt);
899 exclusive = globalLock.try_lock_until(system_clock::now() + dt);
902 ++*globalExclusiveCountPtr;
908 --*globalExclusiveCountPtr;
912 --*globalUpgradeCountPtr;
913 globalLock.unlock_upgrade();
916 --*globalSharedCountPtr;
917 if (ourGlobalTokenUsed) {
918 globalLock.unlock_shared(ourGlobalToken);
919 ourGlobalTokenUsed = false;
921 globalLock.unlock_shared();
924 for (auto& token : privateTokens) {
925 privateLock.unlock_shared(token);
932 for (auto& thr : threads) {
937 TEST(SharedMutex, deterministic_concurrent_readers_of_one_lock_read_prio) {
938 for (int pass = 0; pass < 3; ++pass) {
939 DSched sched(DSched::uniform(pass));
940 runContendedReaders<DeterministicAtomic,
941 DSharedMutexReadPriority,
942 Locker>(1000, 3, false);
946 TEST(SharedMutex, deterministic_concurrent_readers_of_one_lock_write_prio) {
947 for (int pass = 0; pass < 3; ++pass) {
948 DSched sched(DSched::uniform(pass));
949 runContendedReaders<DeterministicAtomic,
950 DSharedMutexWritePriority,
951 Locker>(1000, 3, false);
955 TEST(SharedMutex, concurrent_readers_of_one_lock_read_prio) {
956 for (int pass = 0; pass < 10; ++pass) {
957 runContendedReaders<atomic, SharedMutexReadPriority, Locker>(
962 TEST(SharedMutex, concurrent_readers_of_one_lock_write_prio) {
963 for (int pass = 0; pass < 10; ++pass) {
964 runContendedReaders<atomic, SharedMutexWritePriority, Locker>(
969 TEST(SharedMutex, deterministic_readers_of_concurrent_locks_read_prio) {
970 for (int pass = 0; pass < 3; ++pass) {
971 DSched sched(DSched::uniform(pass));
972 runContendedReaders<DeterministicAtomic,
973 DSharedMutexReadPriority,
974 Locker>(1000, 3, true);
978 TEST(SharedMutex, deterministic_readers_of_concurrent_locks_write_prio) {
979 for (int pass = 0; pass < 3; ++pass) {
980 DSched sched(DSched::uniform(pass));
981 runContendedReaders<DeterministicAtomic,
982 DSharedMutexWritePriority,
983 Locker>(1000, 3, true);
987 TEST(SharedMutex, readers_of_concurrent_locks_read_prio) {
988 for (int pass = 0; pass < 10; ++pass) {
989 runContendedReaders<atomic, SharedMutexReadPriority, TokenLocker>(
994 TEST(SharedMutex, readers_of_concurrent_locks_write_prio) {
995 for (int pass = 0; pass < 10; ++pass) {
996 runContendedReaders<atomic, SharedMutexWritePriority, TokenLocker>(
1001 TEST(SharedMutex, deterministic_mixed_mostly_read_read_prio) {
1002 for (int pass = 0; pass < 3; ++pass) {
1003 DSched sched(DSched::uniform(pass));
1004 runMixed<DeterministicAtomic, DSharedMutexReadPriority, Locker>(
1005 1000, 3, 0.1, false);
1009 TEST(SharedMutex, deterministic_mixed_mostly_read_write_prio) {
1010 for (int pass = 0; pass < 3; ++pass) {
1011 DSched sched(DSched::uniform(pass));
1012 runMixed<DeterministicAtomic, DSharedMutexWritePriority, Locker>(
1013 1000, 3, 0.1, false);
1017 TEST(SharedMutex, mixed_mostly_read_read_prio) {
1018 for (int pass = 0; pass < 5; ++pass) {
1019 runMixed<atomic, SharedMutexReadPriority, TokenLocker>(
1020 10000, 32, 0.1, false);
1024 TEST(SharedMutex, mixed_mostly_read_write_prio) {
1025 for (int pass = 0; pass < 5; ++pass) {
1026 runMixed<atomic, SharedMutexWritePriority, TokenLocker>(
1027 10000, 32, 0.1, false);
1031 TEST(SharedMutex, deterministic_mixed_mostly_write_read_prio) {
1032 for (int pass = 0; pass < 1; ++pass) {
1033 DSched sched(DSched::uniform(pass));
1034 runMixed<DeterministicAtomic, DSharedMutexReadPriority, TokenLocker>(
1035 1000, 10, 0.9, false);
1039 TEST(SharedMutex, deterministic_mixed_mostly_write_write_prio) {
1040 for (int pass = 0; pass < 1; ++pass) {
1041 DSched sched(DSched::uniform(pass));
1042 runMixed<DeterministicAtomic, DSharedMutexWritePriority, TokenLocker>(
1043 1000, 10, 0.9, false);
1047 TEST(SharedMutex, deterministic_lost_wakeup_write_prio) {
1048 for (int pass = 0; pass < 10; ++pass) {
1049 DSched sched(DSched::uniformSubset(pass, 2, 200));
1050 runMixed<DeterministicAtomic, DSharedMutexWritePriority, TokenLocker>(
1051 1000, 3, 1.0, false);
1055 TEST(SharedMutex, mixed_mostly_write_read_prio) {
1056 for (int pass = 0; pass < 5; ++pass) {
1057 runMixed<atomic, SharedMutexReadPriority, TokenLocker>(
1058 50000, 300, 0.9, false);
1062 TEST(SharedMutex, mixed_mostly_write_write_prio) {
1063 for (int pass = 0; pass < 5; ++pass) {
1064 runMixed<atomic, SharedMutexWritePriority, TokenLocker>(
1065 50000, 300, 0.9, false);
1069 TEST(SharedMutex, deterministic_all_ops_read_prio) {
1070 for (int pass = 0; pass < 5; ++pass) {
1071 DSched sched(DSched::uniform(pass));
1072 runAllAndValidate<DSharedMutexReadPriority, DeterministicAtomic>(1000, 8);
1076 TEST(SharedMutex, deterministic_all_ops_write_prio) {
1077 for (int pass = 0; pass < 5; ++pass) {
1078 DSched sched(DSched::uniform(pass));
1079 runAllAndValidate<DSharedMutexWritePriority, DeterministicAtomic>(1000, 8);
1083 TEST(SharedMutex, all_ops_read_prio) {
1084 for (int pass = 0; pass < 5; ++pass) {
1085 runAllAndValidate<SharedMutexReadPriority, atomic>(100000, 32);
1089 TEST(SharedMutex, all_ops_write_prio) {
1090 for (int pass = 0; pass < 5; ++pass) {
1091 runAllAndValidate<SharedMutexWritePriority, atomic>(100000, 32);
1095 FOLLY_ASSUME_FBVECTOR_COMPATIBLE(
1096 boost::optional<boost::optional<SharedMutexToken>>)
1098 // Setup is a set of threads that either grab a shared lock, or exclusive
1099 // and then downgrade it, or upgrade then upgrade and downgrade, then
1100 // enqueue the shared lock to a second set of threads that just performs
1101 // unlocks. Half of the shared locks use tokens, the others don't.
1102 template <typename Lock, template <typename> class Atom>
1103 static void runRemoteUnlock(size_t numOps,
1104 double preWriteFraction,
1105 double preUpgradeFraction,
1106 size_t numSendingThreads,
1107 size_t numReceivingThreads) {
1109 MPMCQueue<boost::optional<boost::optional<SharedMutexToken>>, Atom>
1111 auto queuePtr = &queue; // workaround for clang crash
1113 Atom<bool> go(false);
1114 auto goPtr = &go; // workaround for clang crash
1115 Atom<int> pendingSenders(numSendingThreads);
1116 auto pendingSendersPtr = &pendingSenders; // workaround for clang crash
1117 vector<thread> threads(numSendingThreads + numReceivingThreads);
1120 for (size_t t = 0; t < threads.size(); ++t) {
1121 threads[t] = DSched::thread([&, t, numSendingThreads] {
1122 if (t >= numSendingThreads) {
1124 typename decltype(queue)::value_type elem;
1126 queuePtr->blockingRead(elem);
1128 // EOF, pass the EOF token
1129 queuePtr->blockingWrite(std::move(elem));
1133 globalLock.unlock_shared(**elem);
1135 globalLock.unlock_shared();
1140 // else we're a sender
1142 struct drand48_data buffer;
1143 srand48_r(t, &buffer);
1145 while (!goPtr->load()) {
1146 this_thread::yield();
1148 for (size_t op = t; op < numOps; op += numSendingThreads) {
1149 long unscaledRandVal;
1150 lrand48_r(&buffer, &unscaledRandVal);
1153 double randVal = ((double)unscaledRandVal) / 0x7fffffff;
1155 // extract a bit and rescale
1156 bool useToken = randVal >= 0.5;
1157 randVal = (randVal - (useToken ? 0.5 : 0.0)) * 2;
1159 boost::optional<SharedMutexToken> maybeToken;
1162 SharedMutexToken token;
1163 if (randVal < preWriteFraction) {
1165 globalLock.unlock_and_lock_shared(token);
1166 } else if (randVal < preWriteFraction + preUpgradeFraction / 2) {
1167 globalLock.lock_upgrade();
1168 globalLock.unlock_upgrade_and_lock_shared(token);
1169 } else if (randVal < preWriteFraction + preUpgradeFraction) {
1170 globalLock.lock_upgrade();
1171 globalLock.unlock_upgrade_and_lock();
1172 globalLock.unlock_and_lock_shared(token);
1174 globalLock.lock_shared(token);
1178 if (randVal < preWriteFraction) {
1180 globalLock.unlock_and_lock_shared();
1181 } else if (randVal < preWriteFraction + preUpgradeFraction / 2) {
1182 globalLock.lock_upgrade();
1183 globalLock.unlock_upgrade_and_lock_shared();
1184 } else if (randVal < preWriteFraction + preUpgradeFraction) {
1185 globalLock.lock_upgrade();
1186 globalLock.unlock_upgrade_and_lock();
1187 globalLock.unlock_and_lock_shared();
1189 globalLock.lock_shared();
1193 // blockingWrite is emplace-like, so this automatically adds
1194 // another level of wrapping
1195 queuePtr->blockingWrite(maybeToken);
1197 if (--*pendingSendersPtr == 0) {
1198 queuePtr->blockingWrite(boost::none);
1205 for (auto& thr : threads) {
1210 TEST(SharedMutex, deterministic_remote_write_prio) {
1211 for (int pass = 0; pass < 1; ++pass) {
1212 DSched sched(DSched::uniform(pass));
1213 runRemoteUnlock<DSharedMutexWritePriority, DeterministicAtomic>(
1214 500, 0.1, 0.1, 5, 5);
1218 TEST(SharedMutex, deterministic_remote_read_prio) {
1219 for (int pass = 0; pass < 1; ++pass) {
1220 DSched sched(DSched::uniform(pass));
1221 runRemoteUnlock<DSharedMutexReadPriority, DeterministicAtomic>(
1222 500, 0.1, 0.1, 5, 5);
1226 TEST(SharedMutex, remote_write_prio) {
1227 for (int pass = 0; pass < 10; ++pass) {
1228 runRemoteUnlock<SharedMutexWritePriority, atomic>(100000, 0.1, 0.1, 5, 5);
1232 TEST(SharedMutex, remote_read_prio) {
1233 for (int pass = 0; pass < 100; ++pass) {
1234 runRemoteUnlock<SharedMutexReadPriority, atomic>(100000, 0.1, 0.1, 5, 5);
1238 static void burn(size_t n) {
1239 for (size_t i = 0; i < n; ++i) {
1240 folly::doNotOptimizeAway(i);
1244 // Two threads and three locks, arranged so that they have to proceed
1245 // in turn with reader/writer conflict
1246 template <typename Lock, template <typename> class Atom = atomic>
1247 static void runPingPong(size_t numRounds, size_t burnCount) {
1250 pair<Lock, char[56]> locks[3];
1255 auto availPtr = &avail; // workaround for clang crash
1256 Atom<bool> go(false);
1257 auto goPtr = &go; // workaround for clang crash
1258 vector<thread> threads(2);
1260 locks[0].first.lock();
1261 locks[1].first.lock();
1262 locks[2].first.lock_shared();
1265 threads[0] = DSched::thread([&] {
1267 while (!goPtr->load()) {
1268 this_thread::yield();
1270 for (size_t i = 0; i < numRounds; ++i) {
1271 locks[i % 3].first.unlock();
1272 locks[(i + 2) % 3].first.lock();
1276 threads[1] = DSched::thread([&] {
1278 while (!goPtr->load()) {
1279 this_thread::yield();
1281 for (size_t i = 0; i < numRounds; ++i) {
1282 locks[i % 3].first.lock_shared();
1284 locks[(i + 2) % 3].first.unlock_shared();
1288 while (avail.load() < 2) {
1289 this_thread::yield();
1294 for (auto& thr : threads) {
1297 locks[numRounds % 3].first.unlock();
1298 locks[(numRounds + 1) % 3].first.unlock();
1299 locks[(numRounds + 2) % 3].first.unlock_shared();
1302 static void folly_rwspin_ping_pong(size_t n, size_t scale, size_t burnCount) {
1303 runPingPong<RWSpinLock>(n / scale, burnCount);
1306 static void shmtx_w_bare_ping_pong(size_t n, size_t scale, size_t burnCount) {
1307 runPingPong<SharedMutexWritePriority>(n / scale, burnCount);
1310 static void shmtx_r_bare_ping_pong(size_t n, size_t scale, size_t burnCount) {
1311 runPingPong<SharedMutexReadPriority>(n / scale, burnCount);
1314 static void folly_ticket_ping_pong(size_t n, size_t scale, size_t burnCount) {
1315 runPingPong<RWTicketSpinLock64>(n / scale, burnCount);
1318 static void boost_shared_ping_pong(size_t n, size_t scale, size_t burnCount) {
1319 runPingPong<boost::shared_mutex>(n / scale, burnCount);
1322 static void pthrd_rwlock_ping_pong(size_t n, size_t scale, size_t burnCount) {
1323 runPingPong<PosixRWLock>(n / scale, burnCount);
1326 TEST(SharedMutex, deterministic_ping_pong_write_prio) {
1327 for (int pass = 0; pass < 1; ++pass) {
1328 DSched sched(DSched::uniform(pass));
1329 runPingPong<DSharedMutexWritePriority, DeterministicAtomic>(500, 0);
1333 TEST(SharedMutex, deterministic_ping_pong_read_prio) {
1334 for (int pass = 0; pass < 1; ++pass) {
1335 DSched sched(DSched::uniform(pass));
1336 runPingPong<DSharedMutexReadPriority, DeterministicAtomic>(500, 0);
1340 TEST(SharedMutex, ping_pong_write_prio) {
1341 for (int pass = 0; pass < 1; ++pass) {
1342 runPingPong<SharedMutexWritePriority, atomic>(50000, 0);
1346 TEST(SharedMutex, ping_pong_read_prio) {
1347 for (int pass = 0; pass < 1; ++pass) {
1348 runPingPong<SharedMutexReadPriority, atomic>(50000, 0);
1352 // This is here so you can tell how much of the runtime reported by the
1353 // more complex harnesses is due to the harness, although due to the
1354 // magic of compiler optimization it may also be slower
1355 BENCHMARK(single_thread_lock_shared_unlock_shared, iters) {
1357 for (size_t n = 0; n < iters; ++n) {
1358 SharedMutex::Token token;
1359 lock.lock_shared(token);
1360 folly::doNotOptimizeAway(0);
1361 lock.unlock_shared(token);
1365 BENCHMARK(single_thread_lock_unlock, iters) {
1367 for (size_t n = 0; n < iters; ++n) {
1369 folly::doNotOptimizeAway(0);
1374 #define BENCH_BASE(...) FB_VA_GLUE(BENCHMARK_NAMED_PARAM, (__VA_ARGS__))
1375 #define BENCH_REL(...) FB_VA_GLUE(BENCHMARK_RELATIVE_NAMED_PARAM, (__VA_ARGS__))
1377 // 100% reads. Best-case scenario for deferred locks. Lock is colocated
1378 // with read data, so inline lock takes cache miss every time but deferred
1379 // lock has only cache hits and local access.
1380 BENCHMARK_DRAW_LINE()
1381 BENCHMARK_DRAW_LINE()
1382 BENCH_BASE(folly_rwspin_reads, 1thread, 1, false)
1383 BENCH_REL (shmtx_wr_pri_reads, 1thread, 1, false)
1384 BENCH_REL (shmtx_w_bare_reads, 1thread, 1, false)
1385 BENCH_REL (shmtx_rd_pri_reads, 1thread, 1, false)
1386 BENCH_REL (shmtx_r_bare_reads, 1thread, 1, false)
1387 BENCH_REL (folly_ticket_reads, 1thread, 1, false)
1388 BENCH_REL (boost_shared_reads, 1thread, 1, false)
1389 BENCH_REL (pthrd_rwlock_reads, 1thread, 1, false)
1390 BENCHMARK_DRAW_LINE()
1391 BENCH_BASE(folly_rwspin_reads, 2thread, 2, false)
1392 BENCH_REL (shmtx_wr_pri_reads, 2thread, 2, false)
1393 BENCH_REL (shmtx_w_bare_reads, 2thread, 2, false)
1394 BENCH_REL (shmtx_rd_pri_reads, 2thread, 2, false)
1395 BENCH_REL (shmtx_r_bare_reads, 2thread, 2, false)
1396 BENCH_REL (folly_ticket_reads, 2thread, 2, false)
1397 BENCH_REL (boost_shared_reads, 2thread, 2, false)
1398 BENCH_REL (pthrd_rwlock_reads, 2thread, 2, false)
1399 BENCHMARK_DRAW_LINE()
1400 BENCH_BASE(folly_rwspin_reads, 4thread, 4, false)
1401 BENCH_REL (shmtx_wr_pri_reads, 4thread, 4, false)
1402 BENCH_REL (shmtx_w_bare_reads, 4thread, 4, false)
1403 BENCH_REL (shmtx_rd_pri_reads, 4thread, 4, false)
1404 BENCH_REL (shmtx_r_bare_reads, 4thread, 4, false)
1405 BENCH_REL (folly_ticket_reads, 4thread, 4, false)
1406 BENCH_REL (boost_shared_reads, 4thread, 4, false)
1407 BENCH_REL (pthrd_rwlock_reads, 4thread, 4, false)
1408 BENCHMARK_DRAW_LINE()
1409 BENCH_BASE(folly_rwspin_reads, 8thread, 8, false)
1410 BENCH_REL (shmtx_wr_pri_reads, 8thread, 8, false)
1411 BENCH_REL (shmtx_w_bare_reads, 8thread, 8, false)
1412 BENCH_REL (shmtx_rd_pri_reads, 8thread, 8, false)
1413 BENCH_REL (shmtx_r_bare_reads, 8thread, 8, false)
1414 BENCH_REL (folly_ticket_reads, 8thread, 8, false)
1415 BENCH_REL (boost_shared_reads, 8thread, 8, false)
1416 BENCH_REL (pthrd_rwlock_reads, 8thread, 8, false)
1417 BENCHMARK_DRAW_LINE()
1418 BENCH_BASE(folly_rwspin_reads, 16thread, 16, false)
1419 BENCH_REL (shmtx_wr_pri_reads, 16thread, 16, false)
1420 BENCH_REL (shmtx_w_bare_reads, 16thread, 16, false)
1421 BENCH_REL (shmtx_rd_pri_reads, 16thread, 16, false)
1422 BENCH_REL (shmtx_r_bare_reads, 16thread, 16, false)
1423 BENCH_REL (folly_ticket_reads, 16thread, 16, false)
1424 BENCH_REL (boost_shared_reads, 16thread, 16, false)
1425 BENCH_REL (pthrd_rwlock_reads, 16thread, 16, false)
1426 BENCHMARK_DRAW_LINE()
1427 BENCH_BASE(folly_rwspin_reads, 32thread, 32, false)
1428 BENCH_REL (shmtx_wr_pri_reads, 32thread, 32, false)
1429 BENCH_REL (shmtx_w_bare_reads, 32thread, 32, false)
1430 BENCH_REL (shmtx_rd_pri_reads, 32thread, 32, false)
1431 BENCH_REL (shmtx_r_bare_reads, 32thread, 32, false)
1432 BENCH_REL (folly_ticket_reads, 32thread, 32, false)
1433 BENCH_REL (boost_shared_reads, 32thread, 32, false)
1434 BENCH_REL (pthrd_rwlock_reads, 32thread, 32, false)
1435 BENCHMARK_DRAW_LINE()
1436 BENCH_BASE(folly_rwspin_reads, 64thread, 64, false)
1437 BENCH_REL (shmtx_wr_pri_reads, 64thread, 64, false)
1438 BENCH_REL (shmtx_w_bare_reads, 64thread, 64, false)
1439 BENCH_REL (shmtx_rd_pri_reads, 64thread, 64, false)
1440 BENCH_REL (shmtx_r_bare_reads, 64thread, 64, false)
1441 BENCH_REL (folly_ticket_reads, 64thread, 64, false)
1442 BENCH_REL (boost_shared_reads, 64thread, 64, false)
1443 BENCH_REL (pthrd_rwlock_reads, 64thread, 64, false)
1445 // 1 lock used by everybody, 100% writes. Threads only hurt, but it is
1446 // good to not fail catastrophically. Compare to single_thread_lock_unlock
1447 // to see the overhead of the generic driver (and its pseudo-random number
1448 // generator). pthrd_mutex_ is a pthread_mutex_t (default, not adaptive),
1449 // which is better than any of the reader-writer locks for this scenario.
1450 BENCHMARK_DRAW_LINE()
1451 BENCHMARK_DRAW_LINE()
1452 BENCH_BASE(folly_rwspin, 1thread_all_write, 1, 1.0, false)
1453 BENCH_REL (shmtx_wr_pri, 1thread_all_write, 1, 1.0, false)
1454 BENCH_REL (shmtx_rd_pri, 1thread_all_write, 1, 1.0, false)
1455 BENCH_REL (folly_ticket, 1thread_all_write, 1, 1.0, false)
1456 BENCH_REL (boost_shared, 1thread_all_write, 1, 1.0, false)
1457 BENCH_REL (pthrd_rwlock, 1thread_all_write, 1, 1.0, false)
1458 BENCH_REL (pthrd_mutex_, 1thread_all_write, 1, 1.0, false)
1459 BENCHMARK_DRAW_LINE()
1460 BENCH_BASE(folly_rwspin, 2thread_all_write, 2, 1.0, false)
1461 BENCH_REL (shmtx_wr_pri, 2thread_all_write, 2, 1.0, false)
1462 BENCH_REL (shmtx_rd_pri, 2thread_all_write, 2, 1.0, false)
1463 BENCH_REL (folly_ticket, 2thread_all_write, 2, 1.0, false)
1464 BENCH_REL (boost_shared, 2thread_all_write, 2, 1.0, false)
1465 BENCH_REL (pthrd_rwlock, 2thread_all_write, 2, 1.0, false)
1466 BENCH_REL (pthrd_mutex_, 2thread_all_write, 2, 1.0, false)
1467 BENCHMARK_DRAW_LINE()
1468 BENCH_BASE(folly_rwspin, 4thread_all_write, 4, 1.0, false)
1469 BENCH_REL (shmtx_wr_pri, 4thread_all_write, 4, 1.0, false)
1470 BENCH_REL (shmtx_rd_pri, 4thread_all_write, 4, 1.0, false)
1471 BENCH_REL (folly_ticket, 4thread_all_write, 4, 1.0, false)
1472 BENCH_REL (boost_shared, 4thread_all_write, 4, 1.0, false)
1473 BENCH_REL (pthrd_rwlock, 4thread_all_write, 4, 1.0, false)
1474 BENCH_REL (pthrd_mutex_, 4thread_all_write, 4, 1.0, false)
1475 BENCHMARK_DRAW_LINE()
1476 BENCH_BASE(folly_rwspin, 8thread_all_write, 8, 1.0, false)
1477 BENCH_REL (shmtx_wr_pri, 8thread_all_write, 8, 1.0, false)
1478 BENCH_REL (shmtx_rd_pri, 8thread_all_write, 8, 1.0, false)
1479 BENCH_REL (folly_ticket, 8thread_all_write, 8, 1.0, false)
1480 BENCH_REL (boost_shared, 8thread_all_write, 8, 1.0, false)
1481 BENCH_REL (pthrd_rwlock, 8thread_all_write, 8, 1.0, false)
1482 BENCH_REL (pthrd_mutex_, 8thread_all_write, 8, 1.0, false)
1483 BENCHMARK_DRAW_LINE()
1484 BENCH_BASE(folly_rwspin, 16thread_all_write, 16, 1.0, false)
1485 BENCH_REL (shmtx_wr_pri, 16thread_all_write, 16, 1.0, false)
1486 BENCH_REL (shmtx_rd_pri, 16thread_all_write, 16, 1.0, false)
1487 BENCH_REL (folly_ticket, 16thread_all_write, 16, 1.0, false)
1488 BENCH_REL (boost_shared, 16thread_all_write, 16, 1.0, false)
1489 BENCH_REL (pthrd_rwlock, 16thread_all_write, 16, 1.0, false)
1490 BENCH_REL (pthrd_mutex_, 16thread_all_write, 16, 1.0, false)
1491 BENCHMARK_DRAW_LINE()
1492 BENCH_BASE(folly_rwspin, 32thread_all_write, 32, 1.0, false)
1493 BENCH_REL (shmtx_wr_pri, 32thread_all_write, 32, 1.0, false)
1494 BENCH_REL (shmtx_rd_pri, 32thread_all_write, 32, 1.0, false)
1495 BENCH_REL (folly_ticket, 32thread_all_write, 32, 1.0, false)
1496 BENCH_REL (boost_shared, 32thread_all_write, 32, 1.0, false)
1497 BENCH_REL (pthrd_rwlock, 32thread_all_write, 32, 1.0, false)
1498 BENCH_REL (pthrd_mutex_, 32thread_all_write, 32, 1.0, false)
1499 BENCHMARK_DRAW_LINE()
1500 BENCH_BASE(folly_rwspin, 64thread_all_write, 64, 1.0, false)
1501 BENCH_REL (shmtx_wr_pri, 64thread_all_write, 64, 1.0, false)
1502 BENCH_REL (shmtx_rd_pri, 64thread_all_write, 64, 1.0, false)
1503 BENCH_REL (folly_ticket, 64thread_all_write, 64, 1.0, false)
1504 BENCH_REL (boost_shared, 64thread_all_write, 64, 1.0, false)
1505 BENCH_REL (pthrd_rwlock, 64thread_all_write, 64, 1.0, false)
1506 BENCH_REL (pthrd_mutex_, 64thread_all_write, 64, 1.0, false)
1508 // 1 lock used by everybody, 10% writes. Not much scaling to be had. Perf
1509 // is best at 1 thread, once you've got multiple threads > 8 threads hurts.
1510 BENCHMARK_DRAW_LINE()
1511 BENCHMARK_DRAW_LINE()
1512 BENCH_BASE(folly_rwspin, 1thread_10pct_write, 1, 0.10, false)
1513 BENCH_REL (shmtx_wr_pri, 1thread_10pct_write, 1, 0.10, false)
1514 BENCH_REL (shmtx_rd_pri, 1thread_10pct_write, 1, 0.10, false)
1515 BENCH_REL (folly_ticket, 1thread_10pct_write, 1, 0.10, false)
1516 BENCH_REL (boost_shared, 1thread_10pct_write, 1, 0.10, false)
1517 BENCH_REL (pthrd_rwlock, 1thread_10pct_write, 1, 0.10, false)
1518 BENCHMARK_DRAW_LINE()
1519 BENCH_BASE(folly_rwspin, 2thread_10pct_write, 2, 0.10, false)
1520 BENCH_REL (shmtx_wr_pri, 2thread_10pct_write, 2, 0.10, false)
1521 BENCH_REL (shmtx_rd_pri, 2thread_10pct_write, 2, 0.10, false)
1522 BENCH_REL (folly_ticket, 2thread_10pct_write, 2, 0.10, false)
1523 BENCH_REL (boost_shared, 2thread_10pct_write, 2, 0.10, false)
1524 BENCH_REL (pthrd_rwlock, 2thread_10pct_write, 2, 0.10, false)
1525 BENCHMARK_DRAW_LINE()
1526 BENCH_BASE(folly_rwspin, 4thread_10pct_write, 4, 0.10, false)
1527 BENCH_REL (shmtx_wr_pri, 4thread_10pct_write, 4, 0.10, false)
1528 BENCH_REL (shmtx_rd_pri, 4thread_10pct_write, 4, 0.10, false)
1529 BENCH_REL (folly_ticket, 4thread_10pct_write, 4, 0.10, false)
1530 BENCH_REL (boost_shared, 4thread_10pct_write, 4, 0.10, false)
1531 BENCH_REL (pthrd_rwlock, 4thread_10pct_write, 4, 0.10, false)
1532 BENCHMARK_DRAW_LINE()
1533 BENCH_BASE(folly_rwspin, 8thread_10pct_write, 8, 0.10, false)
1534 BENCH_REL (shmtx_wr_pri, 8thread_10pct_write, 8, 0.10, false)
1535 BENCH_REL (shmtx_rd_pri, 8thread_10pct_write, 8, 0.10, false)
1536 BENCH_REL (folly_ticket, 8thread_10pct_write, 8, 0.10, false)
1537 BENCH_REL (boost_shared, 8thread_10pct_write, 8, 0.10, false)
1538 BENCH_REL (pthrd_rwlock, 8thread_10pct_write, 8, 0.10, false)
1539 BENCHMARK_DRAW_LINE()
1540 BENCH_BASE(folly_rwspin, 16thread_10pct_write, 16, 0.10, false)
1541 BENCH_REL (shmtx_wr_pri, 16thread_10pct_write, 16, 0.10, false)
1542 BENCH_REL (shmtx_rd_pri, 16thread_10pct_write, 16, 0.10, false)
1543 BENCH_REL (folly_ticket, 16thread_10pct_write, 16, 0.10, false)
1544 BENCH_REL (boost_shared, 16thread_10pct_write, 16, 0.10, false)
1545 BENCH_REL (pthrd_rwlock, 16thread_10pct_write, 16, 0.10, false)
1546 BENCHMARK_DRAW_LINE()
1547 BENCH_BASE(folly_rwspin, 32thread_10pct_write, 32, 0.10, false)
1548 BENCH_REL (shmtx_wr_pri, 32thread_10pct_write, 32, 0.10, false)
1549 BENCH_REL (shmtx_rd_pri, 32thread_10pct_write, 32, 0.10, false)
1550 BENCH_REL (folly_ticket, 32thread_10pct_write, 32, 0.10, false)
1551 BENCH_REL (boost_shared, 32thread_10pct_write, 32, 0.10, false)
1552 BENCH_REL (pthrd_rwlock, 32thread_10pct_write, 32, 0.10, false)
1553 BENCHMARK_DRAW_LINE()
1554 BENCH_BASE(folly_rwspin, 64thread_10pct_write, 64, 0.10, false)
1555 BENCH_REL (shmtx_wr_pri, 64thread_10pct_write, 64, 0.10, false)
1556 BENCH_REL (shmtx_rd_pri, 64thread_10pct_write, 64, 0.10, false)
1557 BENCH_REL (folly_ticket, 64thread_10pct_write, 64, 0.10, false)
1558 BENCH_REL (boost_shared, 64thread_10pct_write, 64, 0.10, false)
1559 BENCH_REL (pthrd_rwlock, 64thread_10pct_write, 64, 0.10, false)
1561 // 1 lock used by everybody, 1% writes. This is a more realistic example
1562 // than the concurrent_*_reads benchmark, but still shows SharedMutex locks
1563 // winning over all of the others
1564 BENCHMARK_DRAW_LINE()
1565 BENCHMARK_DRAW_LINE()
1566 BENCH_BASE(folly_rwspin, 1thread_1pct_write, 1, 0.01, false)
1567 BENCH_REL (shmtx_wr_pri, 1thread_1pct_write, 1, 0.01, false)
1568 BENCH_REL (shmtx_w_bare, 1thread_1pct_write, 1, 0.01, false)
1569 BENCH_REL (shmtx_rd_pri, 1thread_1pct_write, 1, 0.01, false)
1570 BENCH_REL (shmtx_r_bare, 1thread_1pct_write, 1, 0.01, false)
1571 BENCH_REL (folly_ticket, 1thread_1pct_write, 1, 0.01, false)
1572 BENCH_REL (boost_shared, 1thread_1pct_write, 1, 0.01, false)
1573 BENCH_REL (pthrd_rwlock, 1thread_1pct_write, 1, 0.01, false)
1574 BENCHMARK_DRAW_LINE()
1575 BENCH_BASE(folly_rwspin, 2thread_1pct_write, 2, 0.01, false)
1576 BENCH_REL (shmtx_wr_pri, 2thread_1pct_write, 2, 0.01, false)
1577 BENCH_REL (shmtx_w_bare, 2thread_1pct_write, 2, 0.01, false)
1578 BENCH_REL (shmtx_rd_pri, 2thread_1pct_write, 2, 0.01, false)
1579 BENCH_REL (shmtx_r_bare, 2thread_1pct_write, 2, 0.01, false)
1580 BENCH_REL (folly_ticket, 2thread_1pct_write, 2, 0.01, false)
1581 BENCH_REL (boost_shared, 2thread_1pct_write, 2, 0.01, false)
1582 BENCH_REL (pthrd_rwlock, 2thread_1pct_write, 2, 0.01, false)
1583 BENCHMARK_DRAW_LINE()
1584 BENCH_BASE(folly_rwspin, 4thread_1pct_write, 4, 0.01, false)
1585 BENCH_REL (shmtx_wr_pri, 4thread_1pct_write, 4, 0.01, false)
1586 BENCH_REL (shmtx_w_bare, 4thread_1pct_write, 4, 0.01, false)
1587 BENCH_REL (shmtx_rd_pri, 4thread_1pct_write, 4, 0.01, false)
1588 BENCH_REL (shmtx_r_bare, 4thread_1pct_write, 4, 0.01, false)
1589 BENCH_REL (folly_ticket, 4thread_1pct_write, 4, 0.01, false)
1590 BENCH_REL (boost_shared, 4thread_1pct_write, 4, 0.01, false)
1591 BENCH_REL (pthrd_rwlock, 4thread_1pct_write, 4, 0.01, false)
1592 BENCHMARK_DRAW_LINE()
1593 BENCH_BASE(folly_rwspin, 8thread_1pct_write, 8, 0.01, false)
1594 BENCH_REL (shmtx_wr_pri, 8thread_1pct_write, 8, 0.01, false)
1595 BENCH_REL (shmtx_w_bare, 8thread_1pct_write, 8, 0.01, false)
1596 BENCH_REL (shmtx_rd_pri, 8thread_1pct_write, 8, 0.01, false)
1597 BENCH_REL (shmtx_r_bare, 8thread_1pct_write, 8, 0.01, false)
1598 BENCH_REL (folly_ticket, 8thread_1pct_write, 8, 0.01, false)
1599 BENCH_REL (boost_shared, 8thread_1pct_write, 8, 0.01, false)
1600 BENCH_REL (pthrd_rwlock, 8thread_1pct_write, 8, 0.01, false)
1601 BENCHMARK_DRAW_LINE()
1602 BENCH_BASE(folly_rwspin, 16thread_1pct_write, 16, 0.01, false)
1603 BENCH_REL (shmtx_wr_pri, 16thread_1pct_write, 16, 0.01, false)
1604 BENCH_REL (shmtx_w_bare, 16thread_1pct_write, 16, 0.01, false)
1605 BENCH_REL (shmtx_rd_pri, 16thread_1pct_write, 16, 0.01, false)
1606 BENCH_REL (shmtx_r_bare, 16thread_1pct_write, 16, 0.01, false)
1607 BENCH_REL (folly_ticket, 16thread_1pct_write, 16, 0.01, false)
1608 BENCH_REL (boost_shared, 16thread_1pct_write, 16, 0.01, false)
1609 BENCH_REL (pthrd_rwlock, 16thread_1pct_write, 16, 0.01, false)
1610 BENCHMARK_DRAW_LINE()
1611 BENCH_BASE(folly_rwspin, 32thread_1pct_write, 32, 0.01, false)
1612 BENCH_REL (shmtx_wr_pri, 32thread_1pct_write, 32, 0.01, false)
1613 BENCH_REL (shmtx_w_bare, 32thread_1pct_write, 32, 0.01, false)
1614 BENCH_REL (shmtx_rd_pri, 32thread_1pct_write, 32, 0.01, false)
1615 BENCH_REL (shmtx_r_bare, 32thread_1pct_write, 32, 0.01, false)
1616 BENCH_REL (folly_ticket, 32thread_1pct_write, 32, 0.01, false)
1617 BENCH_REL (boost_shared, 32thread_1pct_write, 32, 0.01, false)
1618 BENCH_REL (pthrd_rwlock, 32thread_1pct_write, 32, 0.01, false)
1619 BENCHMARK_DRAW_LINE()
1620 BENCH_BASE(folly_rwspin, 64thread_1pct_write, 64, 0.01, false)
1621 BENCH_REL (shmtx_wr_pri, 64thread_1pct_write, 64, 0.01, false)
1622 BENCH_REL (shmtx_w_bare, 64thread_1pct_write, 64, 0.01, false)
1623 BENCH_REL (shmtx_rd_pri, 64thread_1pct_write, 64, 0.01, false)
1624 BENCH_REL (shmtx_r_bare, 64thread_1pct_write, 64, 0.01, false)
1625 BENCH_REL (folly_ticket, 64thread_1pct_write, 64, 0.01, false)
1626 BENCH_REL (boost_shared, 64thread_1pct_write, 64, 0.01, false)
1627 BENCH_REL (pthrd_rwlock, 64thread_1pct_write, 64, 0.01, false)
1629 // Worst case scenario for deferred locks. No actual sharing, likely that
1630 // read operations will have to first set the kDeferredReadersPossibleBit,
1631 // and likely that writers will have to scan deferredReaders[].
1632 BENCHMARK_DRAW_LINE()
1633 BENCH_BASE(folly_rwspin, 2thr_2lock_50pct_write, 2, 0.50, true)
1634 BENCH_REL (shmtx_wr_pri, 2thr_2lock_50pct_write, 2, 0.50, true)
1635 BENCH_REL (shmtx_rd_pri, 2thr_2lock_50pct_write, 2, 0.50, true)
1636 BENCH_BASE(folly_rwspin, 4thr_4lock_50pct_write, 4, 0.50, true)
1637 BENCH_REL (shmtx_wr_pri, 4thr_4lock_50pct_write, 4, 0.50, true)
1638 BENCH_REL (shmtx_rd_pri, 4thr_4lock_50pct_write, 4, 0.50, true)
1639 BENCH_BASE(folly_rwspin, 8thr_8lock_50pct_write, 8, 0.50, true)
1640 BENCH_REL (shmtx_wr_pri, 8thr_8lock_50pct_write, 8, 0.50, true)
1641 BENCH_REL (shmtx_rd_pri, 8thr_8lock_50pct_write, 8, 0.50, true)
1642 BENCH_BASE(folly_rwspin, 16thr_16lock_50pct_write, 16, 0.50, true)
1643 BENCH_REL (shmtx_wr_pri, 16thr_16lock_50pct_write, 16, 0.50, true)
1644 BENCH_REL (shmtx_rd_pri, 16thr_16lock_50pct_write, 16, 0.50, true)
1645 BENCH_BASE(folly_rwspin, 32thr_32lock_50pct_write, 32, 0.50, true)
1646 BENCH_REL (shmtx_wr_pri, 32thr_32lock_50pct_write, 32, 0.50, true)
1647 BENCH_REL (shmtx_rd_pri, 32thr_32lock_50pct_write, 32, 0.50, true)
1648 BENCH_BASE(folly_rwspin, 64thr_64lock_50pct_write, 64, 0.50, true)
1649 BENCH_REL (shmtx_wr_pri, 64thr_64lock_50pct_write, 64, 0.50, true)
1650 BENCH_REL (shmtx_rd_pri, 64thr_64lock_50pct_write, 64, 0.50, true)
1651 BENCHMARK_DRAW_LINE()
1652 BENCH_BASE(folly_rwspin, 2thr_2lock_10pct_write, 2, 0.10, true)
1653 BENCH_REL (shmtx_wr_pri, 2thr_2lock_10pct_write, 2, 0.10, true)
1654 BENCH_REL (shmtx_rd_pri, 2thr_2lock_10pct_write, 2, 0.10, true)
1655 BENCH_BASE(folly_rwspin, 4thr_4lock_10pct_write, 4, 0.10, true)
1656 BENCH_REL (shmtx_wr_pri, 4thr_4lock_10pct_write, 4, 0.10, true)
1657 BENCH_REL (shmtx_rd_pri, 4thr_4lock_10pct_write, 4, 0.10, true)
1658 BENCH_BASE(folly_rwspin, 8thr_8lock_10pct_write, 8, 0.10, true)
1659 BENCH_REL (shmtx_wr_pri, 8thr_8lock_10pct_write, 8, 0.10, true)
1660 BENCH_REL (shmtx_rd_pri, 8thr_8lock_10pct_write, 8, 0.10, true)
1661 BENCH_BASE(folly_rwspin, 16thr_16lock_10pct_write, 16, 0.10, true)
1662 BENCH_REL (shmtx_wr_pri, 16thr_16lock_10pct_write, 16, 0.10, true)
1663 BENCH_REL (shmtx_rd_pri, 16thr_16lock_10pct_write, 16, 0.10, true)
1664 BENCH_BASE(folly_rwspin, 32thr_32lock_10pct_write, 32, 0.10, true)
1665 BENCH_REL (shmtx_wr_pri, 32thr_32lock_10pct_write, 32, 0.10, true)
1666 BENCH_REL (shmtx_rd_pri, 32thr_32lock_10pct_write, 32, 0.10, true)
1667 BENCH_BASE(folly_rwspin, 64thr_64lock_10pct_write, 64, 0.10, true)
1668 BENCH_REL (shmtx_wr_pri, 64thr_64lock_10pct_write, 64, 0.10, true)
1669 BENCH_REL (shmtx_rd_pri, 64thr_64lock_10pct_write, 64, 0.10, true)
1670 BENCHMARK_DRAW_LINE()
1671 BENCH_BASE(folly_rwspin, 2thr_2lock_1pct_write, 2, 0.01, true)
1672 BENCH_REL (shmtx_wr_pri, 2thr_2lock_1pct_write, 2, 0.01, true)
1673 BENCH_REL (shmtx_rd_pri, 2thr_2lock_1pct_write, 2, 0.01, true)
1674 BENCH_BASE(folly_rwspin, 4thr_4lock_1pct_write, 4, 0.01, true)
1675 BENCH_REL (shmtx_wr_pri, 4thr_4lock_1pct_write, 4, 0.01, true)
1676 BENCH_REL (shmtx_rd_pri, 4thr_4lock_1pct_write, 4, 0.01, true)
1677 BENCH_BASE(folly_rwspin, 8thr_8lock_1pct_write, 8, 0.01, true)
1678 BENCH_REL (shmtx_wr_pri, 8thr_8lock_1pct_write, 8, 0.01, true)
1679 BENCH_REL (shmtx_rd_pri, 8thr_8lock_1pct_write, 8, 0.01, true)
1680 BENCH_BASE(folly_rwspin, 16thr_16lock_1pct_write, 16, 0.01, true)
1681 BENCH_REL (shmtx_wr_pri, 16thr_16lock_1pct_write, 16, 0.01, true)
1682 BENCH_REL (shmtx_rd_pri, 16thr_16lock_1pct_write, 16, 0.01, true)
1683 BENCH_BASE(folly_rwspin, 32thr_32lock_1pct_write, 32, 0.01, true)
1684 BENCH_REL (shmtx_wr_pri, 32thr_32lock_1pct_write, 32, 0.01, true)
1685 BENCH_REL (shmtx_rd_pri, 32thr_32lock_1pct_write, 32, 0.01, true)
1686 BENCH_BASE(folly_rwspin, 64thr_64lock_1pct_write, 64, 0.01, true)
1687 BENCH_REL (shmtx_wr_pri, 64thr_64lock_1pct_write, 64, 0.01, true)
1688 BENCH_REL (shmtx_rd_pri, 64thr_64lock_1pct_write, 64, 0.01, true)
1690 // Ping-pong tests have a scaled number of iterations, because their burn
1691 // loop would make them too slow otherwise. Ping-pong with burn count of
1692 // 100k or 300k shows the advantage of soft-spin, reducing the cost of
1693 // each wakeup by about 20 usec. (Take benchmark reported difference,
1694 // ~400 nanos, multiply by the scale of 100, then divide by 2 because
1695 // each round has two wakeups.)
1696 BENCHMARK_DRAW_LINE()
1697 BENCHMARK_DRAW_LINE()
1698 BENCH_BASE(folly_rwspin_ping_pong, burn0, 1, 0)
1699 BENCH_REL (shmtx_w_bare_ping_pong, burn0, 1, 0)
1700 BENCH_REL (shmtx_r_bare_ping_pong, burn0, 1, 0)
1701 BENCH_REL (folly_ticket_ping_pong, burn0, 1, 0)
1702 BENCH_REL (boost_shared_ping_pong, burn0, 1, 0)
1703 BENCH_REL (pthrd_rwlock_ping_pong, burn0, 1, 0)
1704 BENCHMARK_DRAW_LINE()
1705 BENCH_BASE(folly_rwspin_ping_pong, burn100k, 100, 100000)
1706 BENCH_REL (shmtx_w_bare_ping_pong, burn100k, 100, 100000)
1707 BENCH_REL (shmtx_r_bare_ping_pong, burn100k, 100, 100000)
1708 BENCH_REL (folly_ticket_ping_pong, burn100k, 100, 100000)
1709 BENCH_REL (boost_shared_ping_pong, burn100k, 100, 100000)
1710 BENCH_REL (pthrd_rwlock_ping_pong, burn100k, 100, 100000)
1711 BENCHMARK_DRAW_LINE()
1712 BENCH_BASE(folly_rwspin_ping_pong, burn300k, 100, 300000)
1713 BENCH_REL (shmtx_w_bare_ping_pong, burn300k, 100, 300000)
1714 BENCH_REL (shmtx_r_bare_ping_pong, burn300k, 100, 300000)
1715 BENCH_REL (folly_ticket_ping_pong, burn300k, 100, 300000)
1716 BENCH_REL (boost_shared_ping_pong, burn300k, 100, 300000)
1717 BENCH_REL (pthrd_rwlock_ping_pong, burn300k, 100, 300000)
1718 BENCHMARK_DRAW_LINE()
1719 BENCH_BASE(folly_rwspin_ping_pong, burn1M, 1000, 1000000)
1720 BENCH_REL (shmtx_w_bare_ping_pong, burn1M, 1000, 1000000)
1721 BENCH_REL (shmtx_r_bare_ping_pong, burn1M, 1000, 1000000)
1722 BENCH_REL (folly_ticket_ping_pong, burn1M, 1000, 1000000)
1723 BENCH_REL (boost_shared_ping_pong, burn1M, 1000, 1000000)
1724 BENCH_REL (pthrd_rwlock_ping_pong, burn1M, 1000, 1000000)
1726 // Reproduce with 10 minutes and
1728 // shared_mutex_test --benchmark --bm_min_iters=1000000
1730 // Comparison use folly::RWSpinLock as the baseline, with the
1731 // following row being the default SharedMutex (using *Holder or
1732 // Token-ful methods).
1733 // ============================================================================
1734 // folly/experimental/test/SharedMutexTest.cpp relative time/iter iters/s
1735 // ============================================================================
1736 // single_thread_lock_shared_unlock_shared 22.78ns 43.89M
1737 // single_thread_lock_unlock 26.01ns 38.45M
1738 // ----------------------------------------------------------------------------
1739 // ----------------------------------------------------------------------------
1740 // folly_rwspin_reads(1thread) 15.09ns 66.25M
1741 // shmtx_wr_pri_reads(1thread) 69.89% 21.60ns 46.30M
1742 // shmtx_w_bare_reads(1thread) 58.25% 25.91ns 38.59M
1743 // shmtx_rd_pri_reads(1thread) 72.50% 20.82ns 48.03M
1744 // shmtx_r_bare_reads(1thread) 58.27% 25.91ns 38.60M
1745 // folly_ticket_reads(1thread) 54.80% 27.55ns 36.30M
1746 // boost_shared_reads(1thread) 10.88% 138.80ns 7.20M
1747 // pthrd_rwlock_reads(1thread) 40.68% 37.11ns 26.95M
1748 // ----------------------------------------------------------------------------
1749 // folly_rwspin_reads(2thread) 92.63ns 10.80M
1750 // shmtx_wr_pri_reads(2thread) 462.86% 20.01ns 49.97M
1751 // shmtx_w_bare_reads(2thread) 430.53% 21.51ns 46.48M
1752 // shmtx_rd_pri_reads(2thread) 487.13% 19.01ns 52.59M
1753 // shmtx_r_bare_reads(2thread) 433.35% 21.37ns 46.79M
1754 // folly_ticket_reads(2thread) 69.82% 132.67ns 7.54M
1755 // boost_shared_reads(2thread) 36.66% 252.63ns 3.96M
1756 // pthrd_rwlock_reads(2thread) 127.76% 72.50ns 13.79M
1757 // ----------------------------------------------------------------------------
1758 // folly_rwspin_reads(4thread) 97.45ns 10.26M
1759 // shmtx_wr_pri_reads(4thread) 978.22% 9.96ns 100.38M
1760 // shmtx_w_bare_reads(4thread) 908.35% 10.73ns 93.21M
1761 // shmtx_rd_pri_reads(4thread) 1032.29% 9.44ns 105.93M
1762 // shmtx_r_bare_reads(4thread) 912.38% 10.68ns 93.63M
1763 // folly_ticket_reads(4thread) 46.08% 211.46ns 4.73M
1764 // boost_shared_reads(4thread) 25.00% 389.74ns 2.57M
1765 // pthrd_rwlock_reads(4thread) 47.53% 205.01ns 4.88M
1766 // ----------------------------------------------------------------------------
1767 // folly_rwspin_reads(8thread) 147.24ns 6.79M
1768 // shmtx_wr_pri_reads(8thread) 2915.66% 5.05ns 198.02M
1769 // shmtx_w_bare_reads(8thread) 2699.32% 5.45ns 183.32M
1770 // shmtx_rd_pri_reads(8thread) 3092.58% 4.76ns 210.03M
1771 // shmtx_r_bare_reads(8thread) 2744.63% 5.36ns 186.40M
1772 // folly_ticket_reads(8thread) 54.84% 268.47ns 3.72M
1773 // boost_shared_reads(8thread) 42.40% 347.30ns 2.88M
1774 // pthrd_rwlock_reads(8thread) 78.90% 186.63ns 5.36M
1775 // ----------------------------------------------------------------------------
1776 // folly_rwspin_reads(16thread) 166.25ns 6.02M
1777 // shmtx_wr_pri_reads(16thread) 6133.03% 2.71ns 368.91M
1778 // shmtx_w_bare_reads(16thread) 5936.05% 2.80ns 357.06M
1779 // shmtx_rd_pri_reads(16thread) 6786.57% 2.45ns 408.22M
1780 // shmtx_r_bare_reads(16thread) 5995.54% 2.77ns 360.64M
1781 // folly_ticket_reads(16thread) 56.35% 295.01ns 3.39M
1782 // boost_shared_reads(16thread) 51.62% 322.08ns 3.10M
1783 // pthrd_rwlock_reads(16thread) 92.47% 179.79ns 5.56M
1784 // ----------------------------------------------------------------------------
1785 // folly_rwspin_reads(32thread) 107.72ns 9.28M
1786 // shmtx_wr_pri_reads(32thread) 6772.80% 1.59ns 628.77M
1787 // shmtx_w_bare_reads(32thread) 6236.13% 1.73ns 578.94M
1788 // shmtx_rd_pri_reads(32thread) 8143.32% 1.32ns 756.00M
1789 // shmtx_r_bare_reads(32thread) 6485.18% 1.66ns 602.06M
1790 // folly_ticket_reads(32thread) 35.12% 306.73ns 3.26M
1791 // boost_shared_reads(32thread) 28.19% 382.17ns 2.62M
1792 // pthrd_rwlock_reads(32thread) 65.29% 164.99ns 6.06M
1793 // ----------------------------------------------------------------------------
1794 // folly_rwspin_reads(64thread) 119.46ns 8.37M
1795 // shmtx_wr_pri_reads(64thread) 6744.92% 1.77ns 564.60M
1796 // shmtx_w_bare_reads(64thread) 6268.50% 1.91ns 524.72M
1797 // shmtx_rd_pri_reads(64thread) 7508.56% 1.59ns 628.52M
1798 // shmtx_r_bare_reads(64thread) 6299.53% 1.90ns 527.32M
1799 // folly_ticket_reads(64thread) 37.42% 319.26ns 3.13M
1800 // boost_shared_reads(64thread) 32.58% 366.70ns 2.73M
1801 // pthrd_rwlock_reads(64thread) 73.64% 162.24ns 6.16M
1802 // ----------------------------------------------------------------------------
1803 // ----------------------------------------------------------------------------
1804 // folly_rwspin(1thread_all_write) 25.51ns 39.19M
1805 // shmtx_wr_pri(1thread_all_write) 97.38% 26.20ns 38.17M
1806 // shmtx_rd_pri(1thread_all_write) 97.55% 26.16ns 38.23M
1807 // folly_ticket(1thread_all_write) 90.98% 28.04ns 35.66M
1808 // boost_shared(1thread_all_write) 16.80% 151.89ns 6.58M
1809 // pthrd_rwlock(1thread_all_write) 63.86% 39.96ns 25.03M
1810 // pthrd_mutex_(1thread_all_write) 82.05% 31.09ns 32.16M
1811 // ----------------------------------------------------------------------------
1812 // folly_rwspin(2thread_all_write) 100.70ns 9.93M
1813 // shmtx_wr_pri(2thread_all_write) 40.83% 246.61ns 4.05M
1814 // shmtx_rd_pri(2thread_all_write) 40.53% 248.44ns 4.03M
1815 // folly_ticket(2thread_all_write) 58.49% 172.17ns 5.81M
1816 // boost_shared(2thread_all_write) 24.26% 415.00ns 2.41M
1817 // pthrd_rwlock(2thread_all_write) 41.35% 243.49ns 4.11M
1818 // pthrd_mutex_(2thread_all_write) 146.91% 68.55ns 14.59M
1819 // ----------------------------------------------------------------------------
1820 // folly_rwspin(4thread_all_write) 199.52ns 5.01M
1821 // shmtx_wr_pri(4thread_all_write) 51.71% 385.86ns 2.59M
1822 // shmtx_rd_pri(4thread_all_write) 49.43% 403.62ns 2.48M
1823 // folly_ticket(4thread_all_write) 117.88% 169.26ns 5.91M
1824 // boost_shared(4thread_all_write) 9.81% 2.03us 491.48K
1825 // pthrd_rwlock(4thread_all_write) 28.23% 706.69ns 1.42M
1826 // pthrd_mutex_(4thread_all_write) 111.54% 178.88ns 5.59M
1827 // ----------------------------------------------------------------------------
1828 // folly_rwspin(8thread_all_write) 304.61ns 3.28M
1829 // shmtx_wr_pri(8thread_all_write) 69.77% 436.59ns 2.29M
1830 // shmtx_rd_pri(8thread_all_write) 66.58% 457.51ns 2.19M
1831 // folly_ticket(8thread_all_write) 141.00% 216.03ns 4.63M
1832 // boost_shared(8thread_all_write) 6.11% 4.99us 200.59K
1833 // pthrd_rwlock(8thread_all_write) 38.03% 800.88ns 1.25M
1834 // pthrd_mutex_(8thread_all_write) 177.66% 171.45ns 5.83M
1835 // ----------------------------------------------------------------------------
1836 // folly_rwspin(16thread_all_write) 576.97ns 1.73M
1837 // shmtx_wr_pri(16thread_all_write) 105.72% 545.77ns 1.83M
1838 // shmtx_rd_pri(16thread_all_write) 105.13% 548.83ns 1.82M
1839 // folly_ticket(16thread_all_write) 161.70% 356.82ns 2.80M
1840 // boost_shared(16thread_all_write) 7.73% 7.46us 134.03K
1841 // pthrd_rwlock(16thread_all_write) 96.88% 595.54ns 1.68M
1842 // pthrd_mutex_(16thread_all_write) 330.44% 174.61ns 5.73M
1843 // ----------------------------------------------------------------------------
1844 // folly_rwspin(32thread_all_write) 1.41us 707.76K
1845 // shmtx_wr_pri(32thread_all_write) 240.46% 587.58ns 1.70M
1846 // shmtx_rd_pri(32thread_all_write) 393.71% 358.87ns 2.79M
1847 // folly_ticket(32thread_all_write) 325.07% 434.65ns 2.30M
1848 // boost_shared(32thread_all_write) 18.57% 7.61us 131.43K
1849 // pthrd_rwlock(32thread_all_write) 266.78% 529.62ns 1.89M
1850 // pthrd_mutex_(32thread_all_write) 877.89% 160.94ns 6.21M
1851 // ----------------------------------------------------------------------------
1852 // folly_rwspin(64thread_all_write) 1.76us 566.94K
1853 // shmtx_wr_pri(64thread_all_write) 255.67% 689.91ns 1.45M
1854 // shmtx_rd_pri(64thread_all_write) 468.82% 376.23ns 2.66M
1855 // folly_ticket(64thread_all_write) 294.72% 598.49ns 1.67M
1856 // boost_shared(64thread_all_write) 23.39% 7.54us 132.58K
1857 // pthrd_rwlock(64thread_all_write) 321.39% 548.83ns 1.82M
1858 // pthrd_mutex_(64thread_all_write) 1165.04% 151.40ns 6.61M
1859 // ----------------------------------------------------------------------------
1860 // ----------------------------------------------------------------------------
1861 // folly_rwspin(1thread_10pct_write) 19.51ns 51.26M
1862 // shmtx_wr_pri(1thread_10pct_write) 83.25% 23.43ns 42.67M
1863 // shmtx_rd_pri(1thread_10pct_write) 83.31% 23.42ns 42.71M
1864 // folly_ticket(1thread_10pct_write) 70.88% 27.52ns 36.34M
1865 // boost_shared(1thread_10pct_write) 13.09% 148.99ns 6.71M
1866 // pthrd_rwlock(1thread_10pct_write) 47.41% 41.15ns 24.30M
1867 // ----------------------------------------------------------------------------
1868 // folly_rwspin(2thread_10pct_write) 159.42ns 6.27M
1869 // shmtx_wr_pri(2thread_10pct_write) 188.44% 84.60ns 11.82M
1870 // shmtx_rd_pri(2thread_10pct_write) 188.29% 84.67ns 11.81M
1871 // folly_ticket(2thread_10pct_write) 140.28% 113.64ns 8.80M
1872 // boost_shared(2thread_10pct_write) 42.09% 378.81ns 2.64M
1873 // pthrd_rwlock(2thread_10pct_write) 103.86% 153.49ns 6.51M
1874 // ----------------------------------------------------------------------------
1875 // folly_rwspin(4thread_10pct_write) 193.35ns 5.17M
1876 // shmtx_wr_pri(4thread_10pct_write) 184.30% 104.91ns 9.53M
1877 // shmtx_rd_pri(4thread_10pct_write) 163.76% 118.07ns 8.47M
1878 // folly_ticket(4thread_10pct_write) 124.07% 155.84ns 6.42M
1879 // boost_shared(4thread_10pct_write) 16.32% 1.18us 843.92K
1880 // pthrd_rwlock(4thread_10pct_write) 48.59% 397.94ns 2.51M
1881 // ----------------------------------------------------------------------------
1882 // folly_rwspin(8thread_10pct_write) 373.17ns 2.68M
1883 // shmtx_wr_pri(8thread_10pct_write) 252.02% 148.08ns 6.75M
1884 // shmtx_rd_pri(8thread_10pct_write) 203.59% 183.30ns 5.46M
1885 // folly_ticket(8thread_10pct_write) 184.37% 202.40ns 4.94M
1886 // boost_shared(8thread_10pct_write) 15.85% 2.35us 424.72K
1887 // pthrd_rwlock(8thread_10pct_write) 83.03% 449.45ns 2.22M
1888 // ----------------------------------------------------------------------------
1889 // folly_rwspin(16thread_10pct_write) 742.87ns 1.35M
1890 // shmtx_wr_pri(16thread_10pct_write) 344.27% 215.78ns 4.63M
1891 // shmtx_rd_pri(16thread_10pct_write) 287.04% 258.80ns 3.86M
1892 // folly_ticket(16thread_10pct_write) 277.25% 267.94ns 3.73M
1893 // boost_shared(16thread_10pct_write) 15.33% 4.85us 206.30K
1894 // pthrd_rwlock(16thread_10pct_write) 158.34% 469.16ns 2.13M
1895 // ----------------------------------------------------------------------------
1896 // folly_rwspin(32thread_10pct_write) 799.97ns 1.25M
1897 // shmtx_wr_pri(32thread_10pct_write) 351.40% 227.65ns 4.39M
1898 // shmtx_rd_pri(32thread_10pct_write) 341.71% 234.11ns 4.27M
1899 // folly_ticket(32thread_10pct_write) 245.91% 325.31ns 3.07M
1900 // boost_shared(32thread_10pct_write) 7.72% 10.36us 96.56K
1901 // pthrd_rwlock(32thread_10pct_write) 165.87% 482.30ns 2.07M
1902 // ----------------------------------------------------------------------------
1903 // folly_rwspin(64thread_10pct_write) 1.12us 892.01K
1904 // shmtx_wr_pri(64thread_10pct_write) 429.84% 260.81ns 3.83M
1905 // shmtx_rd_pri(64thread_10pct_write) 456.93% 245.35ns 4.08M
1906 // folly_ticket(64thread_10pct_write) 219.21% 511.42ns 1.96M
1907 // boost_shared(64thread_10pct_write) 5.43% 20.65us 48.44K
1908 // pthrd_rwlock(64thread_10pct_write) 233.93% 479.23ns 2.09M
1909 // ----------------------------------------------------------------------------
1910 // ----------------------------------------------------------------------------
1911 // folly_rwspin(1thread_1pct_write) 18.88ns 52.98M
1912 // shmtx_wr_pri(1thread_1pct_write) 81.53% 23.15ns 43.19M
1913 // shmtx_w_bare(1thread_1pct_write) 67.90% 27.80ns 35.97M
1914 // shmtx_rd_pri(1thread_1pct_write) 81.50% 23.16ns 43.18M
1915 // shmtx_r_bare(1thread_1pct_write) 67.74% 27.86ns 35.89M
1916 // folly_ticket(1thread_1pct_write) 68.68% 27.48ns 36.39M
1917 // boost_shared(1thread_1pct_write) 12.80% 147.51ns 6.78M
1918 // pthrd_rwlock(1thread_1pct_write) 45.81% 41.20ns 24.27M
1919 // ----------------------------------------------------------------------------
1920 // folly_rwspin(2thread_1pct_write) 125.85ns 7.95M
1921 // shmtx_wr_pri(2thread_1pct_write) 359.04% 35.05ns 28.53M
1922 // shmtx_w_bare(2thread_1pct_write) 475.60% 26.46ns 37.79M
1923 // shmtx_rd_pri(2thread_1pct_write) 332.75% 37.82ns 26.44M
1924 // shmtx_r_bare(2thread_1pct_write) 115.64% 108.83ns 9.19M
1925 // folly_ticket(2thread_1pct_write) 140.24% 89.74ns 11.14M
1926 // boost_shared(2thread_1pct_write) 40.62% 309.82ns 3.23M
1927 // pthrd_rwlock(2thread_1pct_write) 134.67% 93.45ns 10.70M
1928 // ----------------------------------------------------------------------------
1929 // folly_rwspin(4thread_1pct_write) 126.70ns 7.89M
1930 // shmtx_wr_pri(4thread_1pct_write) 422.20% 30.01ns 33.32M
1931 // shmtx_w_bare(4thread_1pct_write) 403.52% 31.40ns 31.85M
1932 // shmtx_rd_pri(4thread_1pct_write) 282.50% 44.85ns 22.30M
1933 // shmtx_r_bare(4thread_1pct_write) 66.30% 191.10ns 5.23M
1934 // folly_ticket(4thread_1pct_write) 91.93% 137.83ns 7.26M
1935 // boost_shared(4thread_1pct_write) 22.74% 557.10ns 1.80M
1936 // pthrd_rwlock(4thread_1pct_write) 55.66% 227.62ns 4.39M
1937 // ----------------------------------------------------------------------------
1938 // folly_rwspin(8thread_1pct_write) 169.42ns 5.90M
1939 // shmtx_wr_pri(8thread_1pct_write) 567.81% 29.84ns 33.51M
1940 // shmtx_w_bare(8thread_1pct_write) 519.18% 32.63ns 30.64M
1941 // shmtx_rd_pri(8thread_1pct_write) 172.36% 98.30ns 10.17M
1942 // shmtx_r_bare(8thread_1pct_write) 75.56% 224.21ns 4.46M
1943 // folly_ticket(8thread_1pct_write) 104.03% 162.85ns 6.14M
1944 // boost_shared(8thread_1pct_write) 22.01% 769.73ns 1.30M
1945 // pthrd_rwlock(8thread_1pct_write) 71.79% 235.99ns 4.24M
1946 // ----------------------------------------------------------------------------
1947 // folly_rwspin(16thread_1pct_write) 385.88ns 2.59M
1948 // shmtx_wr_pri(16thread_1pct_write) 1039.03% 37.14ns 26.93M
1949 // shmtx_w_bare(16thread_1pct_write) 997.26% 38.69ns 25.84M
1950 // shmtx_rd_pri(16thread_1pct_write) 263.60% 146.39ns 6.83M
1951 // shmtx_r_bare(16thread_1pct_write) 173.16% 222.85ns 4.49M
1952 // folly_ticket(16thread_1pct_write) 179.37% 215.13ns 4.65M
1953 // boost_shared(16thread_1pct_write) 26.95% 1.43us 698.42K
1954 // pthrd_rwlock(16thread_1pct_write) 166.70% 231.48ns 4.32M
1955 // ----------------------------------------------------------------------------
1956 // folly_rwspin(32thread_1pct_write) 382.49ns 2.61M
1957 // shmtx_wr_pri(32thread_1pct_write) 1046.64% 36.54ns 27.36M
1958 // shmtx_w_bare(32thread_1pct_write) 922.87% 41.45ns 24.13M
1959 // shmtx_rd_pri(32thread_1pct_write) 251.93% 151.82ns 6.59M
1960 // shmtx_r_bare(32thread_1pct_write) 176.44% 216.78ns 4.61M
1961 // folly_ticket(32thread_1pct_write) 131.07% 291.82ns 3.43M
1962 // boost_shared(32thread_1pct_write) 12.77% 2.99us 333.95K
1963 // pthrd_rwlock(32thread_1pct_write) 173.43% 220.55ns 4.53M
1964 // ----------------------------------------------------------------------------
1965 // folly_rwspin(64thread_1pct_write) 510.54ns 1.96M
1966 // shmtx_wr_pri(64thread_1pct_write) 1378.27% 37.04ns 27.00M
1967 // shmtx_w_bare(64thread_1pct_write) 1178.24% 43.33ns 23.08M
1968 // shmtx_rd_pri(64thread_1pct_write) 325.29% 156.95ns 6.37M
1969 // shmtx_r_bare(64thread_1pct_write) 247.82% 206.02ns 4.85M
1970 // folly_ticket(64thread_1pct_write) 117.87% 433.13ns 2.31M
1971 // boost_shared(64thread_1pct_write) 9.45% 5.40us 185.09K
1972 // pthrd_rwlock(64thread_1pct_write) 236.72% 215.68ns 4.64M
1973 // ----------------------------------------------------------------------------
1974 // folly_rwspin(2thr_2lock_50pct_write) 10.85ns 92.15M
1975 // shmtx_wr_pri(2thr_2lock_50pct_write) 81.73% 13.28ns 75.32M
1976 // shmtx_rd_pri(2thr_2lock_50pct_write) 81.82% 13.26ns 75.40M
1977 // folly_rwspin(4thr_4lock_50pct_write) 5.29ns 188.90M
1978 // shmtx_wr_pri(4thr_4lock_50pct_write) 80.89% 6.54ns 152.80M
1979 // shmtx_rd_pri(4thr_4lock_50pct_write) 81.07% 6.53ns 153.14M
1980 // folly_rwspin(8thr_8lock_50pct_write) 2.63ns 380.57M
1981 // shmtx_wr_pri(8thr_8lock_50pct_write) 80.56% 3.26ns 306.57M
1982 // shmtx_rd_pri(8thr_8lock_50pct_write) 80.29% 3.27ns 305.54M
1983 // folly_rwspin(16thr_16lock_50pct_write) 1.31ns 764.70M
1984 // shmtx_wr_pri(16thr_16lock_50pct_write) 79.32% 1.65ns 606.54M
1985 // shmtx_rd_pri(16thr_16lock_50pct_write) 79.62% 1.64ns 608.84M
1986 // folly_rwspin(32thr_32lock_50pct_write) 1.20ns 836.75M
1987 // shmtx_wr_pri(32thr_32lock_50pct_write) 91.67% 1.30ns 767.07M
1988 // shmtx_rd_pri(32thr_32lock_50pct_write) 92.00% 1.30ns 769.82M
1989 // folly_rwspin(64thr_64lock_50pct_write) 1.39ns 717.80M
1990 // shmtx_wr_pri(64thr_64lock_50pct_write) 93.21% 1.49ns 669.08M
1991 // shmtx_rd_pri(64thr_64lock_50pct_write) 92.49% 1.51ns 663.89M
1992 // ----------------------------------------------------------------------------
1993 // folly_rwspin(2thr_2lock_10pct_write) 10.24ns 97.70M
1994 // shmtx_wr_pri(2thr_2lock_10pct_write) 76.46% 13.39ns 74.70M
1995 // shmtx_rd_pri(2thr_2lock_10pct_write) 76.35% 13.41ns 74.60M
1996 // folly_rwspin(4thr_4lock_10pct_write) 5.02ns 199.03M
1997 // shmtx_wr_pri(4thr_4lock_10pct_write) 75.83% 6.63ns 150.91M
1998 // shmtx_rd_pri(4thr_4lock_10pct_write) 76.10% 6.60ns 151.46M
1999 // folly_rwspin(8thr_8lock_10pct_write) 2.47ns 405.50M
2000 // shmtx_wr_pri(8thr_8lock_10pct_write) 74.54% 3.31ns 302.27M
2001 // shmtx_rd_pri(8thr_8lock_10pct_write) 74.85% 3.29ns 303.52M
2002 // folly_rwspin(16thr_16lock_10pct_write) 1.22ns 818.68M
2003 // shmtx_wr_pri(16thr_16lock_10pct_write) 73.35% 1.67ns 600.47M
2004 // shmtx_rd_pri(16thr_16lock_10pct_write) 73.38% 1.66ns 600.73M
2005 // folly_rwspin(32thr_32lock_10pct_write) 1.21ns 827.95M
2006 // shmtx_wr_pri(32thr_32lock_10pct_write) 96.13% 1.26ns 795.89M
2007 // shmtx_rd_pri(32thr_32lock_10pct_write) 96.01% 1.26ns 794.95M
2008 // folly_rwspin(64thr_64lock_10pct_write) 1.40ns 716.17M
2009 // shmtx_wr_pri(64thr_64lock_10pct_write) 96.91% 1.44ns 694.03M
2010 // shmtx_rd_pri(64thr_64lock_10pct_write) 96.85% 1.44ns 693.64M
2011 // ----------------------------------------------------------------------------
2012 // folly_rwspin(2thr_2lock_1pct_write) 10.11ns 98.91M
2013 // shmtx_wr_pri(2thr_2lock_1pct_write) 75.07% 13.47ns 74.25M
2014 // shmtx_rd_pri(2thr_2lock_1pct_write) 74.98% 13.48ns 74.16M
2015 // folly_rwspin(4thr_4lock_1pct_write) 4.96ns 201.77M
2016 // shmtx_wr_pri(4thr_4lock_1pct_write) 74.59% 6.64ns 150.49M
2017 // shmtx_rd_pri(4thr_4lock_1pct_write) 74.60% 6.64ns 150.51M
2018 // folly_rwspin(8thr_8lock_1pct_write) 2.44ns 410.42M
2019 // shmtx_wr_pri(8thr_8lock_1pct_write) 73.68% 3.31ns 302.41M
2020 // shmtx_rd_pri(8thr_8lock_1pct_write) 73.38% 3.32ns 301.16M
2021 // folly_rwspin(16thr_16lock_1pct_write) 1.21ns 827.53M
2022 // shmtx_wr_pri(16thr_16lock_1pct_write) 72.11% 1.68ns 596.74M
2023 // shmtx_rd_pri(16thr_16lock_1pct_write) 72.23% 1.67ns 597.73M
2024 // folly_rwspin(32thr_32lock_1pct_write) 1.22ns 819.53M
2025 // shmtx_wr_pri(32thr_32lock_1pct_write) 98.17% 1.24ns 804.50M
2026 // shmtx_rd_pri(32thr_32lock_1pct_write) 98.21% 1.24ns 804.86M
2027 // folly_rwspin(64thr_64lock_1pct_write) 1.41ns 710.26M
2028 // shmtx_wr_pri(64thr_64lock_1pct_write) 97.81% 1.44ns 694.71M
2029 // shmtx_rd_pri(64thr_64lock_1pct_write) 99.44% 1.42ns 706.28M
2030 // ----------------------------------------------------------------------------
2031 // ----------------------------------------------------------------------------
2032 // folly_rwspin_ping_pong(burn0) 641.24ns 1.56M
2033 // shmtx_w_bare_ping_pong(burn0) 91.07% 704.12ns 1.42M
2034 // shmtx_r_bare_ping_pong(burn0) 78.70% 814.84ns 1.23M
2035 // folly_ticket_ping_pong(burn0) 85.67% 748.53ns 1.34M
2036 // boost_shared_ping_pong(burn0) 5.58% 11.50us 86.96K
2037 // pthrd_rwlock_ping_pong(burn0) 8.81% 7.28us 137.40K
2038 // ----------------------------------------------------------------------------
2039 // folly_rwspin_ping_pong(burn100k) 678.97ns 1.47M
2040 // shmtx_w_bare_ping_pong(burn100k) 99.73% 680.78ns 1.47M
2041 // shmtx_r_bare_ping_pong(burn100k) 98.67% 688.13ns 1.45M
2042 // folly_ticket_ping_pong(burn100k) 99.31% 683.68ns 1.46M
2043 // boost_shared_ping_pong(burn100k) 58.23% 1.17us 857.64K
2044 // pthrd_rwlock_ping_pong(burn100k) 57.43% 1.18us 845.86K
2045 // ----------------------------------------------------------------------------
2046 // folly_rwspin_ping_pong(burn300k) 2.03us 492.99K
2047 // shmtx_w_bare_ping_pong(burn300k) 99.98% 2.03us 492.88K
2048 // shmtx_r_bare_ping_pong(burn300k) 99.94% 2.03us 492.68K
2049 // folly_ticket_ping_pong(burn300k) 99.88% 2.03us 492.40K
2050 // boost_shared_ping_pong(burn300k) 81.43% 2.49us 401.47K
2051 // pthrd_rwlock_ping_pong(burn300k) 83.22% 2.44us 410.29K
2052 // ----------------------------------------------------------------------------
2053 // folly_rwspin_ping_pong(burn1M) 677.07ns 1.48M
2054 // shmtx_w_bare_ping_pong(burn1M) 100.50% 673.74ns 1.48M
2055 // shmtx_r_bare_ping_pong(burn1M) 100.14% 676.12ns 1.48M
2056 // folly_ticket_ping_pong(burn1M) 100.44% 674.14ns 1.48M
2057 // boost_shared_ping_pong(burn1M) 93.04% 727.72ns 1.37M
2058 // pthrd_rwlock_ping_pong(burn1M) 94.52% 716.30ns 1.40M
2059 // ============================================================================
2061 int main(int argc, char** argv) {
2062 (void)folly_rwspin_reads;
2063 (void)shmtx_wr_pri_reads;
2064 (void)shmtx_w_bare_reads;
2065 (void)shmtx_rd_pri_reads;
2066 (void)shmtx_r_bare_reads;
2067 (void)folly_ticket_reads;
2068 (void)boost_shared_reads;
2069 (void)pthrd_rwlock_reads;
2079 (void)folly_rwspin_ping_pong;
2080 (void)shmtx_w_bare_ping_pong;
2081 (void)shmtx_r_bare_ping_pong;
2082 (void)folly_ticket_ping_pong;
2083 (void)boost_shared_ping_pong;
2084 (void)pthrd_rwlock_ping_pong;
2086 testing::InitGoogleTest(&argc, argv);
2087 gflags::ParseCommandLineFlags(&argc, &argv, true);
2088 int rv = RUN_ALL_TESTS();
2089 folly::runBenchmarksOnFlag();