folly/test/ThreadLocalBenchmark.cpp

   1 /*
   2  * Copyright 2017 Facebook, Inc.
   3  *
   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
   7  *
   8  *   http://www.apache.org/licenses/LICENSE-2.0
   9  *
  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.
  15  */
  16
  17 #include <folly/ThreadLocal.h>
  18
  19 #include <sys/types.h>
  20
  21 #include <array>
  22 #include <atomic>
  23 #include <condition_variable>
  24 #include <map>
  25 #include <mutex>
  26 #include <set>
  27 #include <thread>
  28
  29 #include <boost/thread/tss.hpp>
  30 #include <glog/logging.h>
  31
  32 #include <folly/Benchmark.h>
  33 #include <folly/experimental/io/FsUtil.h>
  34 #include <folly/portability/GFlags.h>
  35
  36 using namespace folly;
  37
  38 // Simple reference implementation using pthread_get_specific
  39 template <typename T>
  40 class PThreadGetSpecific {
  41  public:
  42   PThreadGetSpecific() : key_(0) { pthread_key_create(&key_, OnThreadExit); }
  43
  44   T* get() const { return static_cast<T*>(pthread_getspecific(key_)); }
  45
  46   void reset(T* t) {
  47     delete get();
  48     pthread_setspecific(key_, t);
  49   }
  50   static void OnThreadExit(void* obj) { delete static_cast<T*>(obj); }
  51
  52  private:
  53   pthread_key_t key_;
  54 };
  55
  56 DEFINE_int32(numThreads, 8, "Number simultaneous threads for benchmarks.");
  57
  58 #define REG(var)                                         \
  59   BENCHMARK(FB_CONCATENATE(BM_mt_, var), iters) {        \
  60     const int itersPerThread = iters / FLAGS_numThreads; \
  61     std::vector<std::thread> threads;                    \
  62     for (int i = 0; i < FLAGS_numThreads; ++i) {         \
  63       threads.push_back(std::thread([&]() {              \
  64         var.reset(new int(0));                           \
  65         for (int j = 0; j < itersPerThread; ++j) {       \
  66           ++(*var.get());                                \
  67         }                                                \
  68       }));                                               \
  69     }                                                    \
  70     for (auto& t : threads) {                            \
  71       t.join();                                          \
  72     }                                                    \
  73   }
  74
  75 ThreadLocalPtr<int> tlp;
  76 REG(tlp);
  77 PThreadGetSpecific<int> pthread_get_specific;
  78 REG(pthread_get_specific);
  79 boost::thread_specific_ptr<int> boost_tsp;
  80 REG(boost_tsp);
  81 BENCHMARK_DRAW_LINE();
  82
  83 struct foo {
  84   int a{0};
  85   int b{0};
  86 };
  87
  88 template <typename TL>
  89 void run_multi(uint32_t iters) {
  90   const int itersPerThread = iters / FLAGS_numThreads;
  91   std::vector<std::thread> threads;
  92   TL var;
  93   for (int i = 0; i < FLAGS_numThreads; ++i) {
  94     threads.push_back(std::thread([&]() {
  95       var.reset(new foo);
  96       for (int j = 0; j < itersPerThread; ++j) {
  97         ++var.get()->a;
  98         var.get()->b += var.get()->a;
  99         --var.get()->a;
 100         var.get()->b += var.get()->a;
 101       }
 102     }));
 103   }
 104   for (auto& t : threads) {
 105     t.join();
 106   }
 107 }
 108
 109 BENCHMARK(BM_mt_tlp_multi, iters) {
 110   run_multi<ThreadLocalPtr<foo>>(iters);
 111 }
 112 BENCHMARK(BM_mt_pthread_get_specific_multi, iters) {
 113   run_multi<PThreadGetSpecific<foo>>(iters);
 114 }
 115 BENCHMARK(BM_mt_boost_tsp_multi, iters) {
 116   run_multi<boost::thread_specific_ptr<foo>>(iters);
 117 }
 118 BENCHMARK_DRAW_LINE();
 119
 120 int main(int argc, char** argv) {
 121   gflags::ParseCommandLineFlags(&argc, &argv, true);
 122   gflags::SetCommandLineOptionWithMode(
 123       "bm_max_iters", "100000000", gflags::SET_FLAG_IF_DEFAULT);
 124   folly::runBenchmarks();
 125   return 0;
 126 }
 127
 128 /*
 129 ./buck-out/gen/folly/test/thread_local_benchmark --bm_min_iters=10000000
 130 --numThreads=1
 131
 132 ============================================================================
 133 folly/test/ThreadLocalBenchmark.cpp             relative  time/iter  iters/s
 134 ============================================================================
 135 BM_mt_tlp                                                    1.92ns  520.02M
 136 BM_mt_pthread_get_specific                                   2.69ns  372.15M
 137 BM_mt_boost_tsp                                             11.81ns   84.67M
 138 ----------------------------------------------------------------------------
 139 BM_mt_tlp_multi                                              7.53ns  132.79M
 140 BM_mt_pthread_get_specific_multi                            15.80ns   63.29M
 141 BM_mt_boost_tsp_multi                                       71.70ns   13.95M
 142 ----------------------------------------------------------------------------
 143 ============================================================================
 144 */