base/threading/thread_perftest.cc

   1 // Copyright 2014 The Chromium Authors. All rights reserved.
   2 // Use of this source code is governed by a BSD-style license that can be
   3 // found in the LICENSE file.
   4
   5 #include "base/base_switches.h"
   6 #include "base/bind.h"
   7 #include "base/command_line.h"
   8 #include "base/memory/scoped_vector.h"
   9 #include "base/strings/stringprintf.h"
  10 #include "base/synchronization/condition_variable.h"
  11 #include "base/synchronization/lock.h"
  12 #include "base/synchronization/waitable_event.h"
  13 #include "base/threading/thread.h"
  14 #include "base/time/time.h"
  15 #include "build/build_config.h"
  16 #include "testing/gtest/include/gtest/gtest.h"
  17 #include "testing/perf/perf_test.h"
  18
  19 #if defined(OS_POSIX)
  20 #include <pthread.h>
  21 #endif
  22
  23 namespace base {
  24
  25 namespace {
  26
  27 const int kNumRuns = 100000;
  28
  29 // Base class for a threading perf-test. This sets up some threads for the
  30 // test and measures the clock-time in addition to time spent on each thread.
  31 class ThreadPerfTest : public testing::Test {
  32  public:
  33   ThreadPerfTest()
  34       : done_(false, false) {
  35     // Disable the task profiler as it adds significant cost!
  36     CommandLine::Init(0, NULL);
  37     CommandLine::ForCurrentProcess()->AppendSwitchASCII(
  38         switches::kProfilerTiming,
  39         switches::kProfilerTimingDisabledValue);
  40   }
  41
  42   // To be implemented by each test. Subclass must uses threads_ such that
  43   // their cpu-time can be measured. Test must return from PingPong() _and_
  44   // call FinishMeasurement from any thread to complete the test.
  45   virtual void Init() {}
  46   virtual void PingPong(int hops) = 0;
  47   virtual void Reset() {}
  48
  49   void TimeOnThread(base::TimeTicks* ticks, base::WaitableEvent* done) {
  50     *ticks = base::TimeTicks::ThreadNow();
  51     done->Signal();
  52   }
  53
  54   base::TimeTicks ThreadNow(base::Thread* thread) {
  55     base::WaitableEvent done(false, false);
  56     base::TimeTicks ticks;
  57     thread->message_loop_proxy()->PostTask(
  58         FROM_HERE,
  59         base::Bind(&ThreadPerfTest::TimeOnThread,
  60                    base::Unretained(this),
  61                    &ticks,
  62                    &done));
  63     done.Wait();
  64     return ticks;
  65   }
  66
  67   void RunPingPongTest(const std::string& name, unsigned num_threads) {
  68     // Create threads and collect starting cpu-time for each thread.
  69     std::vector<base::TimeTicks> thread_starts;
  70     while (threads_.size() < num_threads) {
  71       threads_.push_back(new base::Thread("PingPonger"));
  72       threads_.back()->Start();
  73       if (base::TimeTicks::IsThreadNowSupported())
  74         thread_starts.push_back(ThreadNow(threads_.back()));
  75     }
  76
  77     Init();
  78
  79     base::TimeTicks start = base::TimeTicks::HighResNow();
  80     PingPong(kNumRuns);
  81     done_.Wait();
  82     base::TimeTicks end = base::TimeTicks::HighResNow();
  83
  84     // Gather the cpu-time spent on each thread. This does one extra tasks,
  85     // but that should be in the noise given enough runs.
  86     base::TimeDelta thread_time;
  87     while (threads_.size()) {
  88       if (base::TimeTicks::IsThreadNowSupported()) {
  89         thread_time += ThreadNow(threads_.back()) - thread_starts.back();
  90         thread_starts.pop_back();
  91       }
  92       threads_.pop_back();
  93     }
  94
  95     Reset();
  96
  97     double num_runs = static_cast<double>(kNumRuns);
  98     double us_per_task_clock = (end - start).InMicroseconds() / num_runs;
  99     double us_per_task_cpu = thread_time.InMicroseconds() / num_runs;
 100
 101     // Clock time per task.
 102     perf_test::PrintResult(
 103         "task", "", name + "_time ", us_per_task_clock, "us/hop", true);
 104
 105     // Total utilization across threads if available (likely higher).
 106     if (base::TimeTicks::IsThreadNowSupported()) {
 107       perf_test::PrintResult(
 108           "task", "", name + "_cpu ", us_per_task_cpu, "us/hop", true);
 109     }
 110   }
 111
 112  protected:
 113   void FinishMeasurement() { done_.Signal(); }
 114   ScopedVector<base::Thread> threads_;
 115
 116  private:
 117   base::WaitableEvent done_;
 118 };
 119
 120 // Class to test task performance by posting empty tasks back and forth.
 121 class TaskPerfTest : public ThreadPerfTest {
 122   base::Thread* NextThread(int count) {
 123     return threads_[count % threads_.size()];
 124   }
 125
 126   void PingPong(int hops) override {
 127     if (!hops) {
 128       FinishMeasurement();
 129       return;
 130     }
 131     NextThread(hops)->message_loop_proxy()->PostTask(
 132         FROM_HERE,
 133         base::Bind(
 134             &ThreadPerfTest::PingPong, base::Unretained(this), hops - 1));
 135   }
 136 };
 137
 138 // This tries to test the 'best-case' as well as the 'worst-case' task posting
 139 // performance. The best-case keeps one thread alive such that it never yeilds,
 140 // while the worse-case forces a context switch for every task. Four threads are
 141 // used to ensure the threads do yeild (with just two it might be possible for
 142 // both threads to stay awake if they can signal each other fast enough).
 143 TEST_F(TaskPerfTest, TaskPingPong) {
 144   RunPingPongTest("1_Task_Threads", 1);
 145   RunPingPongTest("4_Task_Threads", 4);
 146 }
 147
 148
 149 // Same as above, but add observers to test their perf impact.
 150 class MessageLoopObserver : public base::MessageLoop::TaskObserver {
 151  public:
 152   void WillProcessTask(const base::PendingTask& pending_task) override {}
 153   void DidProcessTask(const base::PendingTask& pending_task) override {}
 154 };
 155 MessageLoopObserver message_loop_observer;
 156
 157 class TaskObserverPerfTest : public TaskPerfTest {
 158  public:
 159   void Init() override {
 160     TaskPerfTest::Init();
 161     for (size_t i = 0; i < threads_.size(); i++) {
 162       threads_[i]->message_loop()->AddTaskObserver(&message_loop_observer);
 163     }
 164   }
 165 };
 166
 167 TEST_F(TaskObserverPerfTest, TaskPingPong) {
 168   RunPingPongTest("1_Task_Threads_With_Observer", 1);
 169   RunPingPongTest("4_Task_Threads_With_Observer", 4);
 170 }
 171
 172 // Class to test our WaitableEvent performance by signaling back and fort.
 173 // WaitableEvent is templated so we can also compare with other versions.
 174 template <typename WaitableEventType>
 175 class EventPerfTest : public ThreadPerfTest {
 176  public:
 177   virtual void Init() override {
 178     for (size_t i = 0; i < threads_.size(); i++)
 179       events_.push_back(new WaitableEventType(false, false));
 180   }
 181
 182   virtual void Reset() override { events_.clear(); }
 183
 184   void WaitAndSignalOnThread(size_t event) {
 185     size_t next_event = (event + 1) % events_.size();
 186     int my_hops = 0;
 187     do {
 188       events_[event]->Wait();
 189       my_hops = --remaining_hops_;  // We own 'hops' between Wait and Signal.
 190       events_[next_event]->Signal();
 191     } while (my_hops > 0);
 192     // Once we are done, all threads will signal as hops passes zero.
 193     // We only signal completion once, on the thread that reaches zero.
 194     if (!my_hops)
 195       FinishMeasurement();
 196   }
 197
 198   virtual void PingPong(int hops) override {
 199     remaining_hops_ = hops;
 200     for (size_t i = 0; i < threads_.size(); i++) {
 201       threads_[i]->message_loop_proxy()->PostTask(
 202           FROM_HERE,
 203           base::Bind(&EventPerfTest::WaitAndSignalOnThread,
 204                      base::Unretained(this),
 205                      i));
 206     }
 207
 208     // Kick off the Signal ping-ponging.
 209     events_.front()->Signal();
 210   }
 211
 212   int remaining_hops_;
 213   ScopedVector<WaitableEventType> events_;
 214 };
 215
 216 // Similar to the task posting test, this just tests similar functionality
 217 // using WaitableEvents. We only test four threads (worst-case), but we
 218 // might want to craft a way to test the best-case (where the thread doesn't
 219 // end up blocking because the event is already signalled).
 220 typedef EventPerfTest<base::WaitableEvent> WaitableEventPerfTest;
 221 TEST_F(WaitableEventPerfTest, EventPingPong) {
 222   RunPingPongTest("4_WaitableEvent_Threads", 4);
 223 }
 224
 225 // Build a minimal event using ConditionVariable.
 226 class ConditionVariableEvent {
 227  public:
 228   ConditionVariableEvent(bool manual_reset, bool initially_signaled)
 229       : cond_(&lock_), signaled_(false) {
 230     DCHECK(!manual_reset);
 231     DCHECK(!initially_signaled);
 232   }
 233
 234   void Signal() {
 235     {
 236       base::AutoLock scoped_lock(lock_);
 237       signaled_ = true;
 238     }
 239     cond_.Signal();
 240   }
 241
 242   void Wait() {
 243     base::AutoLock scoped_lock(lock_);
 244     while (!signaled_)
 245       cond_.Wait();
 246     signaled_ = false;
 247   }
 248
 249  private:
 250   base::Lock lock_;
 251   base::ConditionVariable cond_;
 252   bool signaled_;
 253 };
 254
 255 // This is meant to test the absolute minimal context switching time
 256 // using our own base synchronization code.
 257 typedef EventPerfTest<ConditionVariableEvent> ConditionVariablePerfTest;
 258 TEST_F(ConditionVariablePerfTest, EventPingPong) {
 259   RunPingPongTest("4_ConditionVariable_Threads", 4);
 260 }
 261 #if defined(OS_POSIX)
 262
 263 // Absolutely 100% minimal posix waitable event. If there is a better/faster
 264 // way to force a context switch, we should use that instead.
 265 class PthreadEvent {
 266  public:
 267   PthreadEvent(bool manual_reset, bool initially_signaled) {
 268     DCHECK(!manual_reset);
 269     DCHECK(!initially_signaled);
 270     pthread_mutex_init(&mutex_, 0);
 271     pthread_cond_init(&cond_, 0);
 272     signaled_ = false;
 273   }
 274
 275   ~PthreadEvent() {
 276     pthread_cond_destroy(&cond_);
 277     pthread_mutex_destroy(&mutex_);
 278   }
 279
 280   void Signal() {
 281     pthread_mutex_lock(&mutex_);
 282     signaled_ = true;
 283     pthread_mutex_unlock(&mutex_);
 284     pthread_cond_signal(&cond_);
 285   }
 286
 287   void Wait() {
 288     pthread_mutex_lock(&mutex_);
 289     while (!signaled_)
 290       pthread_cond_wait(&cond_, &mutex_);
 291     signaled_ = false;
 292     pthread_mutex_unlock(&mutex_);
 293   }
 294
 295  private:
 296   bool signaled_;
 297   pthread_mutex_t mutex_;
 298   pthread_cond_t cond_;
 299 };
 300
 301 // This is meant to test the absolute minimal context switching time.
 302 // If there is any faster way to do this we should substitute it in.
 303 typedef EventPerfTest<PthreadEvent> PthreadEventPerfTest;
 304 TEST_F(PthreadEventPerfTest, EventPingPong) {
 305   RunPingPongTest("4_PthreadCondVar_Threads", 4);
 306 }
 307
 308 #endif
 309
 310 }  // namespace
 311
 312 }  // namespace base