Adds a couple of simple micro benchmarks for the TaskQueueManager

content/renderer/scheduler/task_queue_manager_perftest.cc

+// Copyright 2015 The Chromium Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#include "content/renderer/scheduler/task_queue_manager.h"
+
+#include "base/bind.h"
+#include "base/threading/thread.h"
+#include "content/renderer/scheduler/task_queue_selector.h"
+#include "testing/gtest/include/gtest/gtest.h"
+#include "testing/perf/perf_test.h"
+
+namespace content {
+
+namespace {
+
+class SelectorForTest : public TaskQueueSelector {
+ public:
+  SelectorForTest() {}
+
+  void RegisterWorkQueues(
+      const std::vector<const base::TaskQueue*>& work_queues) override {
+    work_queues_ = work_queues;
+  }
+
+  bool SelectWorkQueueToService(size_t* out_queue_index) override {
+    // Choose the oldest task, if any.
+    bool found_one = false;
+    for (size_t i = 0; i < work_queues_.size(); i++) {
+      if (work_queues_[i]->empty())
+        continue;
+      // Note: the < comparison is correct due to the fact that the PendingTask
+      // operator inverts its comparison operation in order to work well in a
+      // heap based priority queue.
+      if (!found_one ||
+          work_queues_[*out_queue_index]->front() < work_queues_[i]->front())
+        *out_queue_index = i;
+      found_one = true;
+    }
+    CHECK(found_one);
+    return found_one;
+  }
+
+  void AsValueInto(base::trace_event::TracedValue* state) const override {}
+
+ private:
+  std::vector<const base::TaskQueue*> work_queues_;
+
+  DISALLOW_COPY_AND_ASSIGN(SelectorForTest);
+};
+
+}  // namespace
+
+class TaskQueueManagerPerfTest : public testing::Test {
+ protected:
+  void Initialize(size_t num_queues) {
+    num_queues_ = num_queues;
+    message_loop_.reset(new base::MessageLoop());
+    selector_ = make_scoped_ptr(new SelectorForTest);
+    manager_ = make_scoped_ptr(new TaskQueueManager(
+        num_queues, message_loop_->task_runner(), selector_.get()));
+  }
+
+  void TestTask() {

[Sami, 2015/03/04 13:00:09]

nit: TestDelayedTask() since that's what this is mostly posting.

[alex clarke (OOO till 29th), 2015/03/04 16:34:20]

Done.

+    num_tasks_in_flight_--;
+    unsigned int max_tasks_to_post = num_tasks_to_run_ % 2 ? 1 : 10;

[Sami, 2015/03/04 13:00:09]

How deep queues do you see in DoWork() with this and how does that compare to
real sites? It feels like the the fan-out factor is still pretty large.

[alex clarke (OOO till 29th), 2015/03/04 16:34:20]

The number of delayed tasks tasks is capped to max_tasks_in_flight_.

I'm only ever seeing 1 task on the queue in
TaskQueueManager::ProcessTaskFromWorkQueue.

[Sami, 2015/03/04 18:33:29]

I more meant how many tasks there are pending in each of the incoming and work
queues. I think if we post most of the tasks up front instead of getting them as
a steady trickle, then we get away with just a few reloads which may not be
realistic.

[alex clarke (OOO till 29th), 2015/03/05 17:44:39]

There is a maximum of 200 tasks in flight.  It was occasionally dipping down to
1 but was mostly at two hundred.  The change I just added makes the number of
tasks in flight much more variable.

+    for (unsigned int i = 0;
+         i < max_tasks_to_post && num_tasks_in_flight_ < max_tasks_in_flight_ &&
+             num_tasks_to_run_ > 0;
+         i++) {
+      // Choose a queue weighted towards queue 0.
+      unsigned int queue = num_tasks_to_run_ % (num_queues_ + 1);
+      if (queue == num_queues_) {
+        queue = 0;
+      }
+      // Simulate a mix of short and longer delays.
+      unsigned int delay =
+          num_tasks_to_run_ % 2 ? 1 : (10 + num_tasks_to_run_ % 10);
+      scoped_refptr<base::SingleThreadTaskRunner> runner =
+          manager_->TaskRunnerForQueue(queue);
+      runner->PostDelayedTask(FROM_HERE,
+                              base::Bind(&TaskQueueManagerPerfTest::TestTask,
+                                         base::Unretained(this)),
+                              base::TimeDelta::FromMicroseconds(delay));
+      num_tasks_in_flight_++;
+      num_tasks_to_run_--;
+    }
+
+    if (num_tasks_to_run_ == 0)
+      message_loop_->Quit();
+  }
+
+  void Benchmark(const std::string& trace, unsigned int num_tasks_to_run) {

[Sami, 2015/03/04 13:00:09]

I guess you could make this a template if you don't want to hardcode the test
function here.

[alex clarke (OOO till 29th), 2015/03/04 16:34:20]

Good point, but I think base::Closure is more readable than templating it.

[Sami, 2015/03/04 18:33:29]

Works for me.

+    base::TimeTicks start = base::TimeTicks::Now();
+    base::TimeTicks now;
+    unsigned long long num_iterations = 0;
+    do {
+      num_tasks_in_flight_ = 1;
+      num_tasks_to_run_ = num_tasks_to_run;
+      TestTask();
+      message_loop_->Run();
+      now = base::TimeTicks::Now();
+      num_iterations++;
+    } while (now - start < base::TimeDelta::FromSeconds(5));
+    perf_test::PrintResult(
+        "task", "", trace,
+        (now - start).InMicroseconds() / static_cast<double>(num_iterations),
+        "us/run", true);
+  }
+
+  size_t num_queues_;
+  unsigned int max_tasks_in_flight_;
+  unsigned int num_tasks_in_flight_;
+  unsigned int num_tasks_to_run_;
+  scoped_ptr<SelectorForTest> selector_;
+  scoped_ptr<TaskQueueManager> manager_;
+  scoped_ptr<base::MessageLoop> message_loop_;
+};
+
+TEST_F(TaskQueueManagerPerfTest, RunTenThousandDelayedTasks_OneQueue) {
+  Initialize(1u);
+
+  max_tasks_in_flight_ = 200;
+  Benchmark("post 10000 delayed tasks with one queue", 10000);
+}
+
+TEST_F(TaskQueueManagerPerfTest, RunTenThousandDelayedTasks_FourQueues) {
+  Initialize(4u);
+
+  max_tasks_in_flight_ = 200;
+  Benchmark("post 10000 delayed tasks with four queues", 10000);
+}
+
+TEST_F(TaskQueueManagerPerfTest, RunTenThousandDelayedTasks_EightQueues) {
+  Initialize(8u);
+
+  max_tasks_in_flight_ = 200;
+  Benchmark("post 10000 delayed tasks with eight queues", 10000);
+}
+
+}  // namespace content