about summary refs log tree commit diff
path: root/absl/synchronization/mutex_benchmark.cc
diff options
context:
space:
mode:
authorAbseil Team <absl-team@google.com>2018-12-04T19·01-0800
committerCJ Johnson <johnsoncj@google.com>2018-12-04T21·27-0500
commit44b0fafc62d9b8f192e8180cbe9c4b806b339d57 (patch)
tree4ada26427deee54333ac271a26c5fe363ba33cc7 /absl/synchronization/mutex_benchmark.cc
parent926bfeb9fff223429c12224b7514243886323e8d (diff)
Export of internal Abseil changes.
--
cd076f55c1fa600131f6dda392533dfe61679fc0 by Abseil Team <absl-team@google.com>:

Internal change

PiperOrigin-RevId: 224008762

--
e05f62b01286d51044ff86ec6ef565749b9faf82 by Abseil Team <absl-team@google.com>:

Create a pow10() test helper function to compute guaranteed-precise double values of 10^x. Not all standard libraries ship bit-accurate pow() functions, causing tests to fail that rely on expected values generated by it.

PiperOrigin-RevId: 223883762

--
fd88e5e3f7ab80f7f5df9fd1488cd58b4573be69 by Abseil Team <absl-team@google.com>:

Remove some absl:: qualifications to work around inline namespace bugs on MSVC 2015.

PiperOrigin-RevId: 223869642

--
6276cfff969d596edd36a2bbaba65ee045808903 by Abseil Team <absl-team@google.com>:

Update absl/memory/CMakeLists.txt to use new functions
i.e. absl_cc_(library|test)

PiperOrigin-RevId: 223854224

--
359de9afc7a34c975fd3e0cbc52afd96637d97bd by Chris Kennelly <ckennelly@google.com>:

Mark spinlock_benchmark_common as alwayslink = 1.

PiperOrigin-RevId: 223844536

--
450cd8cbe2789a6d54ed1eb87170259bb334f8b9 by Abseil Team <absl-team@google.com>:

Support .* (pointer-to-member dereference) expressions in demangle.cc.

PiperOrigin-RevId: 223826797

--
772ca92179c3634f3e31a80bbc272ed8022e3572 by Abseil Team <absl-team@google.com>:

Fix misspellings in absl::variant comments and replace a ' with a `.

PiperOrigin-RevId: 223807911

--
35dcdc2fbf299d195658aac101887f6dcad1de2f by Abseil Team <absl-team@google.com>:

Bug fix in CMakeLists.txt file (SRCS --> HDRS).

The compressed_tuple header-only library is being defined
with the SRCS parameter instead of the HDRS parameter and
this has been observed to cause some builds on some platforms
to attempt to create a static library from it which fails
since there are no .cc sources.

PiperOrigin-RevId: 223805367

--
4a57a3d2045bb137c0c97958e45ce425190b8d3e by Chris Kennelly <ckennelly@google.com>:

Add test that absl::make_unique value initializes memory.

PiperOrigin-RevId: 223801819

--
dfe8289d7f4dcc6bb568a26aaf192a89e896bdfd by Chris Kennelly <ckennelly@google.com>:

SpinLock: Use exchange to avoid missing wakeups.

The default fast path for SpinLock::Unlock does not use an atomic.  If the
SpinLock becomes contended while we are unlocking between lockword_.load and
lockword_.store, we will fail to wake up the new waiter.  This can cause
unexpected latency spikes.

PiperOrigin-RevId: 223800369

--
9b9d35df786482f0016f77dd31691eff81503d23 by Abseil Team <absl-team@google.com>:

Update absl/hash/CMakeLists.txt to use new functions
i.e. absl_cc_(library|test)

PiperOrigin-RevId: 223755819

--
c2014e2704b87e7cdce2d2a0287c7e2397752296 by Abseil Team <absl-team@google.com>:

Update absl/debugging/CMakeLists.txt to use new functions
i.e. absl_cc_(library|test)

PiperOrigin-RevId: 223751986

--
d83a4e09126400e3fd80645cb03ee558f532271e by Derek Mauro <dmauro@google.com>:

Cleanup synchronization benchmarks.

PiperOrigin-RevId: 223589416

--
fad140b473586531b5b12843f942ec27dfcf5e93 by CJ Johnson <johnsoncj@google.com>:

Makes unifies the order of forward_iterator and input_iterator overloads

PiperOrigin-RevId: 223580660

--
6cd7c96faa7cc5f79f574e35a1b13837ef187d05 by Abseil Team <absl-team@google.com>:

Internal Change.

PiperOrigin-RevId: 223561629

--
bd2e545356b0f548af0e3c14bb2f7f0e712e49d0 by Shaindel Schwartz <shaindel@google.com>:

Remove misleading comments. try_emplace() does not exist for the hash_set containers.

PiperOrigin-RevId: 223543089

--
0cd380a53b587eb7aacc4003a4a3bbb6c78d7c10 by Derek Mauro <dmauro@google.com>:

Internal change

PiperOrigin-RevId: 223512551

--
7156dfee599cb72e9adddfe0e6ae07a95ddf10bb by Greg Miller <jgm@google.com>:

Fixes UB that would result from constructing, multiplying, or dividing a
Duration with a double "NaN" value. This CL changes the absl::Duration
*implementation* to return an InfiniteDuration value that has the same sign as
the given NaN.

PiperOrigin-RevId: 223407499

--
196b7d18609958267951882baf7f9429e49bcafa by CJ Johnson <johnsoncj@google.com>:

Addresses NVCC+MSVC compilation bug where `inlined_capacity()` was not considered valid in constexpr

PiperOrigin-RevId: 223397718
GitOrigin-RevId: cd076f55c1fa600131f6dda392533dfe61679fc0
Change-Id: I5423ca6470f661a7c6f73aa8fee49990446f157f
Diffstat (limited to 'absl/synchronization/mutex_benchmark.cc')
-rw-r--r--absl/synchronization/mutex_benchmark.cc151
1 files changed, 140 insertions, 11 deletions
diff --git a/absl/synchronization/mutex_benchmark.cc b/absl/synchronization/mutex_benchmark.cc
index 1e019e001ae4..2652bb974e99 100644
--- a/absl/synchronization/mutex_benchmark.cc
+++ b/absl/synchronization/mutex_benchmark.cc
@@ -12,16 +12,154 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#include <cstdint>
+#include <mutex>  // NOLINT(build/c++11)
 #include <vector>
 
-#include "benchmark/benchmark.h"
-#include "absl/base/internal/sysinfo.h"
+#include "absl/base/internal/cycleclock.h"
+#include "absl/base/internal/spinlock.h"
 #include "absl/synchronization/blocking_counter.h"
 #include "absl/synchronization/internal/thread_pool.h"
 #include "absl/synchronization/mutex.h"
+#include "benchmark/benchmark.h"
 
 namespace {
 
+void BM_Mutex(benchmark::State& state) {
+  static absl::Mutex* mu = new absl::Mutex;
+  for (auto _ : state) {
+    absl::MutexLock lock(mu);
+  }
+}
+BENCHMARK(BM_Mutex)->UseRealTime()->Threads(1)->ThreadPerCpu();
+
+static void DelayNs(int64_t ns, int* data) {
+  int64_t end = absl::base_internal::CycleClock::Now() +
+                ns * absl::base_internal::CycleClock::Frequency() / 1e9;
+  while (absl::base_internal::CycleClock::Now() < end) {
+    ++(*data);
+    benchmark::DoNotOptimize(*data);
+  }
+}
+
+template <typename MutexType>
+class RaiiLocker {
+ public:
+  explicit RaiiLocker(MutexType* mu) : mu_(mu) { mu_->Lock(); }
+  ~RaiiLocker() { mu_->Unlock(); }
+ private:
+  MutexType* mu_;
+};
+
+template <>
+class RaiiLocker<std::mutex> {
+ public:
+  explicit RaiiLocker(std::mutex* mu) : mu_(mu) { mu_->lock(); }
+  ~RaiiLocker() { mu_->unlock(); }
+ private:
+  std::mutex* mu_;
+};
+
+template <typename MutexType>
+void BM_Contended(benchmark::State& state) {
+  struct Shared {
+    MutexType mu;
+    int data = 0;
+  };
+  static auto* shared = new Shared;
+  int local = 0;
+  for (auto _ : state) {
+    // Here we model both local work outside of the critical section as well as
+    // some work inside of the critical section. The idea is to capture some
+    // more or less realisitic contention levels.
+    // If contention is too low, the benchmark won't measure anything useful.
+    // If contention is unrealistically high, the benchmark will favor
+    // bad mutex implementations that block and otherwise distract threads
+    // from the mutex and shared state for as much as possible.
+    // To achieve this amount of local work is multiplied by number of threads
+    // to keep ratio between local work and critical section approximately
+    // equal regardless of number of threads.
+    DelayNs(100 * state.threads, &local);
+    RaiiLocker<MutexType> locker(&shared->mu);
+    DelayNs(state.range(0), &shared->data);
+  }
+}
+
+BENCHMARK_TEMPLATE(BM_Contended, absl::Mutex)
+    ->UseRealTime()
+    // ThreadPerCpu poorly handles non-power-of-two CPU counts.
+    ->Threads(1)
+    ->Threads(2)
+    ->Threads(4)
+    ->Threads(6)
+    ->Threads(8)
+    ->Threads(12)
+    ->Threads(16)
+    ->Threads(24)
+    ->Threads(32)
+    ->Threads(48)
+    ->Threads(64)
+    ->Threads(96)
+    ->Threads(128)
+    ->Threads(192)
+    ->Threads(256)
+    // Some empirically chosen amounts of work in critical section.
+    // 1 is low contention, 200 is high contention and few values in between.
+    ->Arg(1)
+    ->Arg(20)
+    ->Arg(50)
+    ->Arg(200);
+
+BENCHMARK_TEMPLATE(BM_Contended, absl::base_internal::SpinLock)
+    ->UseRealTime()
+    // ThreadPerCpu poorly handles non-power-of-two CPU counts.
+    ->Threads(1)
+    ->Threads(2)
+    ->Threads(4)
+    ->Threads(6)
+    ->Threads(8)
+    ->Threads(12)
+    ->Threads(16)
+    ->Threads(24)
+    ->Threads(32)
+    ->Threads(48)
+    ->Threads(64)
+    ->Threads(96)
+    ->Threads(128)
+    ->Threads(192)
+    ->Threads(256)
+    // Some empirically chosen amounts of work in critical section.
+    // 1 is low contention, 200 is high contention and few values in between.
+    ->Arg(1)
+    ->Arg(20)
+    ->Arg(50)
+    ->Arg(200);
+
+BENCHMARK_TEMPLATE(BM_Contended, std::mutex)
+    ->UseRealTime()
+    // ThreadPerCpu poorly handles non-power-of-two CPU counts.
+    ->Threads(1)
+    ->Threads(2)
+    ->Threads(4)
+    ->Threads(6)
+    ->Threads(8)
+    ->Threads(12)
+    ->Threads(16)
+    ->Threads(24)
+    ->Threads(32)
+    ->Threads(48)
+    ->Threads(64)
+    ->Threads(96)
+    ->Threads(128)
+    ->Threads(192)
+    ->Threads(256)
+    // Some empirically chosen amounts of work in critical section.
+    // 1 is low contention, 200 is high contention and few values in between.
+    ->Arg(1)
+    ->Arg(20)
+    ->Arg(50)
+    ->Arg(200);
+
 // Measure the overhead of conditions on mutex release (when they must be
 // evaluated).  Mutex has (some) support for equivalence classes allowing
 // Conditions with the same function/argument to potentially not be multiply
@@ -82,13 +220,4 @@ constexpr int kMaxConditionWaiters = 1024;
 #endif
 BENCHMARK(BM_ConditionWaiters)->RangePair(0, 2, 1, kMaxConditionWaiters);
 
-void BM_ContendedMutex(benchmark::State& state) {
-  static absl::Mutex* mu = new absl::Mutex;
-  for (auto _ : state) {
-    absl::MutexLock lock(mu);
-  }
-}
-BENCHMARK(BM_ContendedMutex)->Threads(1);
-BENCHMARK(BM_ContendedMutex)->ThreadPerCpu();
-
 }  // namespace