[Android] Support idle wakeup measurement with /proc/timer_stats

[Android] Support idle wakeup measurement with /proc/timer_stats

Add support for using /proc/timer_stats to measure idle wakeups, as
tracked by power metrics. This involves piping begin/end markers to
/proc/timer_stats to collect the necessary stats, and mapping the
resulting per-thread wakeup data to a given process id.

Unfortunately, the way Mac measures idle wakeups is very different,
which makes the current measurement glue somewhat misleading for all
platforms. Longer term, idle wakeup measurements should be rolled
into a standalone PowerMonitor.

Based on the original change by thestig@chromium.org at
https://codereview.chromium.org/675843004/.

BUG=391309, 422170
CQ_EXTRA_TRYBOTS=tryserver.chromium.perf:linux_perf_bisect;tryserver.chromium.perf:android_nexus5_perf_bisect

[jdduke (slow), 2015-07-22 21:57:19]

Patchset #2 (id:20001) has been deleted

[jdduke (slow), 2015-07-23 19:04:38]

jdduke@chromium.org changed reviewers:
+ jeremy@chromium.org, thestig@chromium.org

[jdduke (slow), 2015-07-23 19:04:39]

thestig@, jeremy@: I rebased thestig's patch from a while ago, and managed to
get the basic accounting working (the problem before was that /proc/timer_stats
uses *thread* "pids", so we have to map from the thread pid to the process pid
to account for all process wakeups).

However, I'm seeing more noise in the numbers than I'd like, and I expect part
of this has to do with how we're pulling data from the device. That is, we make
separate pulls to ps & proc/timer_stats for each process, when we really ought
to be pulling it just once for all processes, then mapping to pids.

What I'd really like to do is make idle wakeup measurement its own PowerMonitor
instance. We'd have one for Mac, using ps, and one for Android/Linux, using
/proc/timer_stats. Support for multiple PowerMonitors is in the works, and I
think this would eliminate some of the awkwardness of trying to make the Linux
idle measurement approach match that of Mac. Thoughts?

If you guys agree about the PowerMonitor approach, I'll wait until the necessary
support lands (https://codereview.chromium.org/1235943004/), factor out Mac's
implementation into a PowerMonitor type, then add this code as another
PowerMonitor type.

[jdduke (slow), 2015-07-23 19:29:49]

jdduke@chromium.org changed reviewers:
+ sullivan@chromium.org
- jeremy@chromium.org

[jdduke (slow), 2015-07-23 19:29:50]

s/jeremy/sullivan

Annie, see my initial comment about trying to do this instead as a PowerMonitor
instance. The only reason I'd push for this approach is that it means we don't
have to fiddle with the Mac wakeup stats, but I'd feel more comfortable
abstracting Linux/Mac approaches into standalone PowerMonitor types.

[sullivan, 2015-07-27 16:24:25]

sullivan@chromium.org changed reviewers:
+ nednguyen@google.com

[sullivan, 2015-07-27 16:24:26]

I'm okay with this but would like to get Ned's opinion as well.

[nednguyen, 2015-07-29 22:22:45]

nednguyen@google.com changed reviewers:
+ rnephew@google.com

[nednguyen, 2015-07-29 22:22:45]

+Randy

https://codereview.chromium.org/1242943005/diff/50001/tools/perf/measurements...
File tools/perf/measurements/smoothness_unittest.py (right):

https://codereview.chromium.org/1242943005/diff/50001/tools/perf/measurements...
tools/perf/measurements/smoothness_unittest.py:29: def
CanMeasureIdleWakeUps(self):
Why do you need to add this? IIRC, we removed power_metric from smoothness
measurement

https://codereview.chromium.org/1242943005/diff/50001/tools/telemetry/telemet...
File tools/telemetry/telemetry/core/platform.py (right):

https://codereview.chromium.org/1242943005/diff/50001/tools/telemetry/telemet...
tools/telemetry/telemetry/core/platform.py:297: """Stops monitoring idle
wakeups."""
This does not return any thing? How do the callsites get the idea wakeups data?

[jdduke (slow), 2015-07-29 22:30:03]

https://codereview.chromium.org/1242943005/diff/50001/tools/telemetry/telemet...
File tools/telemetry/telemetry/core/platform.py (right):

https://codereview.chromium.org/1242943005/diff/50001/tools/telemetry/telemet...
tools/telemetry/telemetry/core/platform.py:297: """Stops monitoring idle
wakeups."""
On 2015/07/29 22:22:45, nednguyen wrote:
> This does not return any thing? How do the callsites get the idea wakeups
data?

Exactly =/. That's why I prefaced this CL with:

"What I'd really like to do is make idle wakeup measurement its own PowerMonitor
instance. We'd have one for Mac, using ps, and one for Android/Linux, using
/proc/timer_stats. Support for multiple PowerMonitors is in the works, and I
think this would eliminate some of the awkwardness of trying to make the Linux
idle measurement approach match that of Mac."

Now, what I could do as an intermediate step, is move the cpu_stats-based delta
measurement for Mac into the Mac backend, and have the /proc/timer_stats
measurement as a separate implementation for the Android backend.

Basically, it seems like the original change from thestig@ was designed to
minimize disruption/changes to how idle wakeups are measured on Mac.

[rnephew (Reviews Here), 2015-07-29 22:42:25]

Support for multiple power monitors is in with
https://codereview.chromium.org/1255673002/

[rnephew (Reviews Here), 2015-07-30 15:49:18]

On 2015/07/29 22:42:25, rnephew1 wrote:
> Support for multiple power monitors is in with
> https://codereview.chromium.org/1255673002/

I may have spoke too soon. There is a chance I'll have to revert it today.
crbug.com/515207

[jdduke (slow), 2015-10-21 19:30:02]

Tabling this for now until I can prioritize it properly.