Austin Schuh | 36244a1 | 2019-09-21 17:52:38 -0700 | [diff] [blame^] | 1 | // Copyright 2017 The Abseil Authors. |
| 2 | // |
| 3 | // Licensed under the Apache License, Version 2.0 (the "License"); |
| 4 | // you may not use this file except in compliance with the License. |
| 5 | // You may obtain a copy of the License at |
| 6 | // |
| 7 | // https://www.apache.org/licenses/LICENSE-2.0 |
| 8 | // |
| 9 | // Unless required by applicable law or agreed to in writing, software |
| 10 | // distributed under the License is distributed on an "AS IS" BASIS, |
| 11 | // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 12 | // See the License for the specific language governing permissions and |
| 13 | // limitations under the License. |
| 14 | |
| 15 | #include <cstdint> |
| 16 | #include <mutex> // NOLINT(build/c++11) |
| 17 | #include <vector> |
| 18 | |
| 19 | #include "absl/base/internal/cycleclock.h" |
| 20 | #include "absl/base/internal/spinlock.h" |
| 21 | #include "absl/synchronization/blocking_counter.h" |
| 22 | #include "absl/synchronization/internal/thread_pool.h" |
| 23 | #include "absl/synchronization/mutex.h" |
| 24 | #include "benchmark/benchmark.h" |
| 25 | |
| 26 | namespace { |
| 27 | |
| 28 | void BM_Mutex(benchmark::State& state) { |
| 29 | static absl::Mutex* mu = new absl::Mutex; |
| 30 | for (auto _ : state) { |
| 31 | absl::MutexLock lock(mu); |
| 32 | } |
| 33 | } |
| 34 | BENCHMARK(BM_Mutex)->UseRealTime()->Threads(1)->ThreadPerCpu(); |
| 35 | |
| 36 | static void DelayNs(int64_t ns, int* data) { |
| 37 | int64_t end = absl::base_internal::CycleClock::Now() + |
| 38 | ns * absl::base_internal::CycleClock::Frequency() / 1e9; |
| 39 | while (absl::base_internal::CycleClock::Now() < end) { |
| 40 | ++(*data); |
| 41 | benchmark::DoNotOptimize(*data); |
| 42 | } |
| 43 | } |
| 44 | |
| 45 | template <typename MutexType> |
| 46 | class RaiiLocker { |
| 47 | public: |
| 48 | explicit RaiiLocker(MutexType* mu) : mu_(mu) { mu_->Lock(); } |
| 49 | ~RaiiLocker() { mu_->Unlock(); } |
| 50 | private: |
| 51 | MutexType* mu_; |
| 52 | }; |
| 53 | |
| 54 | template <> |
| 55 | class RaiiLocker<std::mutex> { |
| 56 | public: |
| 57 | explicit RaiiLocker(std::mutex* mu) : mu_(mu) { mu_->lock(); } |
| 58 | ~RaiiLocker() { mu_->unlock(); } |
| 59 | private: |
| 60 | std::mutex* mu_; |
| 61 | }; |
| 62 | |
| 63 | template <typename MutexType> |
| 64 | void BM_Contended(benchmark::State& state) { |
| 65 | struct Shared { |
| 66 | MutexType mu; |
| 67 | int data = 0; |
| 68 | }; |
| 69 | static auto* shared = new Shared; |
| 70 | int local = 0; |
| 71 | for (auto _ : state) { |
| 72 | // Here we model both local work outside of the critical section as well as |
| 73 | // some work inside of the critical section. The idea is to capture some |
| 74 | // more or less realisitic contention levels. |
| 75 | // If contention is too low, the benchmark won't measure anything useful. |
| 76 | // If contention is unrealistically high, the benchmark will favor |
| 77 | // bad mutex implementations that block and otherwise distract threads |
| 78 | // from the mutex and shared state for as much as possible. |
| 79 | // To achieve this amount of local work is multiplied by number of threads |
| 80 | // to keep ratio between local work and critical section approximately |
| 81 | // equal regardless of number of threads. |
| 82 | DelayNs(100 * state.threads, &local); |
| 83 | RaiiLocker<MutexType> locker(&shared->mu); |
| 84 | DelayNs(state.range(0), &shared->data); |
| 85 | } |
| 86 | } |
| 87 | |
| 88 | BENCHMARK_TEMPLATE(BM_Contended, absl::Mutex) |
| 89 | ->UseRealTime() |
| 90 | // ThreadPerCpu poorly handles non-power-of-two CPU counts. |
| 91 | ->Threads(1) |
| 92 | ->Threads(2) |
| 93 | ->Threads(4) |
| 94 | ->Threads(6) |
| 95 | ->Threads(8) |
| 96 | ->Threads(12) |
| 97 | ->Threads(16) |
| 98 | ->Threads(24) |
| 99 | ->Threads(32) |
| 100 | ->Threads(48) |
| 101 | ->Threads(64) |
| 102 | ->Threads(96) |
| 103 | ->Threads(128) |
| 104 | ->Threads(192) |
| 105 | ->Threads(256) |
| 106 | // Some empirically chosen amounts of work in critical section. |
| 107 | // 1 is low contention, 200 is high contention and few values in between. |
| 108 | ->Arg(1) |
| 109 | ->Arg(20) |
| 110 | ->Arg(50) |
| 111 | ->Arg(200); |
| 112 | |
| 113 | BENCHMARK_TEMPLATE(BM_Contended, absl::base_internal::SpinLock) |
| 114 | ->UseRealTime() |
| 115 | // ThreadPerCpu poorly handles non-power-of-two CPU counts. |
| 116 | ->Threads(1) |
| 117 | ->Threads(2) |
| 118 | ->Threads(4) |
| 119 | ->Threads(6) |
| 120 | ->Threads(8) |
| 121 | ->Threads(12) |
| 122 | ->Threads(16) |
| 123 | ->Threads(24) |
| 124 | ->Threads(32) |
| 125 | ->Threads(48) |
| 126 | ->Threads(64) |
| 127 | ->Threads(96) |
| 128 | ->Threads(128) |
| 129 | ->Threads(192) |
| 130 | ->Threads(256) |
| 131 | // Some empirically chosen amounts of work in critical section. |
| 132 | // 1 is low contention, 200 is high contention and few values in between. |
| 133 | ->Arg(1) |
| 134 | ->Arg(20) |
| 135 | ->Arg(50) |
| 136 | ->Arg(200); |
| 137 | |
| 138 | BENCHMARK_TEMPLATE(BM_Contended, std::mutex) |
| 139 | ->UseRealTime() |
| 140 | // ThreadPerCpu poorly handles non-power-of-two CPU counts. |
| 141 | ->Threads(1) |
| 142 | ->Threads(2) |
| 143 | ->Threads(4) |
| 144 | ->Threads(6) |
| 145 | ->Threads(8) |
| 146 | ->Threads(12) |
| 147 | ->Threads(16) |
| 148 | ->Threads(24) |
| 149 | ->Threads(32) |
| 150 | ->Threads(48) |
| 151 | ->Threads(64) |
| 152 | ->Threads(96) |
| 153 | ->Threads(128) |
| 154 | ->Threads(192) |
| 155 | ->Threads(256) |
| 156 | // Some empirically chosen amounts of work in critical section. |
| 157 | // 1 is low contention, 200 is high contention and few values in between. |
| 158 | ->Arg(1) |
| 159 | ->Arg(20) |
| 160 | ->Arg(50) |
| 161 | ->Arg(200); |
| 162 | |
| 163 | // Measure the overhead of conditions on mutex release (when they must be |
| 164 | // evaluated). Mutex has (some) support for equivalence classes allowing |
| 165 | // Conditions with the same function/argument to potentially not be multiply |
| 166 | // evaluated. |
| 167 | // |
| 168 | // num_classes==0 is used for the special case of every waiter being distinct. |
| 169 | void BM_ConditionWaiters(benchmark::State& state) { |
| 170 | int num_classes = state.range(0); |
| 171 | int num_waiters = state.range(1); |
| 172 | |
| 173 | struct Helper { |
| 174 | static void Waiter(absl::BlockingCounter* init, absl::Mutex* m, int* p) { |
| 175 | init->DecrementCount(); |
| 176 | m->LockWhen(absl::Condition( |
| 177 | static_cast<bool (*)(int*)>([](int* v) { return *v == 0; }), p)); |
| 178 | m->Unlock(); |
| 179 | } |
| 180 | }; |
| 181 | |
| 182 | if (num_classes == 0) { |
| 183 | // No equivalence classes. |
| 184 | num_classes = num_waiters; |
| 185 | } |
| 186 | |
| 187 | absl::BlockingCounter init(num_waiters); |
| 188 | absl::Mutex mu; |
| 189 | std::vector<int> equivalence_classes(num_classes, 1); |
| 190 | |
| 191 | // Must be declared last to be destroyed first. |
| 192 | absl::synchronization_internal::ThreadPool pool(num_waiters); |
| 193 | |
| 194 | for (int i = 0; i < num_waiters; i++) { |
| 195 | // Mutex considers Conditions with the same function and argument |
| 196 | // to be equivalent. |
| 197 | pool.Schedule([&, i] { |
| 198 | Helper::Waiter(&init, &mu, &equivalence_classes[i % num_classes]); |
| 199 | }); |
| 200 | } |
| 201 | init.Wait(); |
| 202 | |
| 203 | for (auto _ : state) { |
| 204 | mu.Lock(); |
| 205 | mu.Unlock(); // Each unlock requires Condition evaluation for our waiters. |
| 206 | } |
| 207 | |
| 208 | mu.Lock(); |
| 209 | for (int i = 0; i < num_classes; i++) { |
| 210 | equivalence_classes[i] = 0; |
| 211 | } |
| 212 | mu.Unlock(); |
| 213 | } |
| 214 | |
| 215 | // Some configurations have higher thread limits than others. |
| 216 | #if defined(__linux__) && !defined(THREAD_SANITIZER) |
| 217 | constexpr int kMaxConditionWaiters = 8192; |
| 218 | #else |
| 219 | constexpr int kMaxConditionWaiters = 1024; |
| 220 | #endif |
| 221 | BENCHMARK(BM_ConditionWaiters)->RangePair(0, 2, 1, kMaxConditionWaiters); |
| 222 | |
| 223 | } // namespace |