cc: Add useful TaskGraphRunner performance tests.

cc: Add useful TaskGraphRunner performance tests.

Current perf tests are not useful for profiling as output heavily
depends on kernel scheduling. An improvement in CPU time is not
guaranteed to result in a better result from these tests as they
mostly measure context switching cost.

This includes a small refactor to TaskGraphRunner that allow us
to run tests on a single thread. The result is more predictable
results and useful profiling output.

The number of tasks used in tests have also been adjusted to more
realistic values.

BUG=338355
Committed: https://src.chromium.org/viewvc/chrome?view=rev&revision=247625

[reveman, 2014-01-28 23:09:18]

Please take a look.

[vmpstr, 2014-01-28 23:16:53]

Can you re-upload the patch, side-by-side diff gives me an old chunk mismatch

https://codereview.chromium.org/147883003/diff/1/cc/resources/task_graph_runn...
File cc/resources/task_graph_runner.h (right):

https://codereview.chromium.org/147883003/diff/1/cc/resources/task_graph_runn...
cc/resources/task_graph_runner.h:147: 
*ForTesting seems to be far more common in cc/

[reveman, 2014-01-28 23:21:04]

On 2014/01/28 23:16:53, vmpstr wrote:
> Can you re-upload the patch, side-by-side diff gives me an old chunk mismatch
> 
>
https://codereview.chromium.org/147883003/diff/1/cc/resources/task_graph_runn...
> File cc/resources/task_graph_runner.h (right):
> 
>
https://codereview.chromium.org/147883003/diff/1/cc/resources/task_graph_runn...
> cc/resources/task_graph_runner.h:147: 
> *ForTesting seems to be far more common in cc/

Done.

PTAL.

[vmpstr, 2014-01-28 23:45:07]

lgtm

https://codereview.chromium.org/147883003/diff/40001/cc/resources/task_graph_...
File cc/resources/task_graph_runner_perftest.cc (right):

https://codereview.chromium.org/147883003/diff/40001/cc/resources/task_graph_...
cc/resources/task_graph_runner_perftest.cc:181: } while
(task_graph_runner_->RunTaskForTesting());
nit: can you change this to while (...) {} instead?

[reveman, 2014-01-29 00:17:07]

https://codereview.chromium.org/147883003/diff/40001/cc/resources/task_graph_...
File cc/resources/task_graph_runner_perftest.cc (right):

https://codereview.chromium.org/147883003/diff/40001/cc/resources/task_graph_...
cc/resources/task_graph_runner_perftest.cc:181: } while
(task_graph_runner_->RunTaskForTesting());
On 2014/01/28 23:45:07, vmpstr wrote:
> nit: can you change this to while (...) {} instead?

git cl format doesn't allow that.

I can do:
while (task_graph_runner_->RunTaskForTesting()) {
}

or:

while (true) {
  if (!task_graph_runner_->RunTaskForTesting())
    break;
}

or:

while (task_graph_runner_->RunTaskForTesting())
  continue;

picked the last as it's listed as a good example in the google c++ style guide.

[commit-bot: I haz the power, 2014-01-29 00:18:58]

CQ is trying da patch. Follow status at
https://chromium-status.appspot.com/cq/reveman@chromium.org/147883003/60001

[commit-bot: I haz the power, 2014-01-29 05:21:39]

Change committed as 247625

[tomhudson, 2014-01-29 10:59:01]

Particularly with the change to realistic (I assume smaller) numbers of tasks,
Do we have a CPU profile showing that the majority of the time in the test is
actually spent in the targeted code? One problem with some of the
microbenchmarks I've been looking at is that that they're dominated by test
setup/teardown or file IO costs.

[reveman, 2014-01-29 16:35:18]

On 2014/01/29 10:59:01, tomhudson wrote:
> Particularly with the change to realistic (I assume smaller) numbers of tasks,
> Do we have a CPU profile showing that the majority of the time in the test is
> actually spent in the targeted code? One problem with some of the
> microbenchmarks I've been looking at is that that they're dominated by test
> setup/teardown or file IO costs.

Time spent in setup/teardown was improved by this patch but is still a problem.
ie. ScheduleTasks also exercise the task graph building code but it would
ideally just exercise the task scheduling code. I'm hoping to fix this in follow
up patches.

FYI, I considered changing the timing code to not include setup/teardown but
decided that that was a bad idea as profiling would no longer give data that is
representative of the test results.