Support total and per-second task/threadtime timeline metrics

tools/perf/metrics/timeline.py

 # Copyright 2014 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.
 import collections
 
 from telemetry.util.statistics import DivideIfPossibleOrZero
 from telemetry.value import scalar
 from telemetry.web_perf.metrics import timeline_based_metric
 
 
@@ skipping 109 matching lines @@
 
 def ThreadCategoryName(thread_name):
   thread_category = "other"
   for substring, category in TimelineThreadCategories.iteritems():
     if substring in _MatchBySubString and substring in thread_name:
       thread_category = category
   if thread_name in TimelineThreadCategories:
     thread_category = TimelineThreadCategories[thread_name]
   return thread_category
 
-def ThreadCpuTimeResultName(thread_category):
+def ThreadCpuTimeResultName(thread_category, measure_per_frame=True):

[jdduke (slow), 2015/04/29 18:07:17]

I'm not married to the |measure_per_frame| name.

[nednguyen, 2015/04/30 05:07:45]

I don't thing these helpers function add any value in making the code more
readable or shorter. We should get rid of these.

[jdduke (slow), 2015/05/01 15:51:44]

Done.

   # This isn't a good name, but I don't want to change it and lose continuity.
-  return "thread_" + thread_category + "_cpu_time_per_frame"
+  if measure_per_frame:
+    return "thread_" + thread_category + "_cpu_time_per_frame"
+  return "thread_" + thread_category + "_cpu_time_total"

[nednguyen, 2015/04/30 05:07:45]

How stable is this thread_X_cpu_time_total metric? Note that we sometime have
situation that page waits on some js condition. Since the waiting logic is
polling, I suspect that this metric has high variance.

[jdduke (slow), 2015/04/30 14:29:39]

In local testing, it's quite stable. Part of that may be that, with each page
measurement (for key_power_cases) we simply wait N seconds. That is, we do all
of the work before the actual measurement is run (load the page, turn off the
screen, wait a couple seconds for stability, then actually start the
measurement).

However, measuring per second will probably scale a lot more nicely in the
future (at least for this case... that won't be useful for the case rbyers@
mentioned with testing taps, but at least if we generalize the normalization
category we'll be ready to do per-event timings or otherwise). I'll make the
change and reupload.

 
-def ThreadTasksResultName(thread_category):
-  return "tasks_per_frame_" + thread_category
+def ThreadTasksResultName(thread_category, measure_per_frame=True):
+  if measure_per_frame:
+    return "tasks_per_frame_" + thread_category
+  return "tasks_" + thread_category

[nednguyen, 2015/04/30 05:07:45]

Same here, I think you should come up with a new ThreadTimeTotal metric that
computes something like tasks/seconds to avoid the timing issue.

[jdduke (slow), 2015/04/30 14:29:39]

Will do.

 
-def ThreadMeanFrameTimeResultName(thread_category):
-  return "mean_frame_time_" + thread_category
+def ThreadFrameTimeResultName(thread_category, measure_per_frame=True):
+  if measure_per_frame:
+    return "mean_frame_time_" + thread_category
+  return "total_frame_time_" + thread_category
 
 def ThreadDetailResultName(thread_category, detail):
   detail_sanitized = detail.replace('.','_')
   return "thread_" + thread_category + "|" + detail_sanitized
 
 
 class ResultsForThread(object):
-  def __init__(self, model, record_ranges, name):
+  def __init__(self, model, record_ranges, name, measure_per_frame):
     self.model = model
     self.toplevel_slices = []
     self.all_slices = []
     self.name = name
     self.record_ranges = record_ranges
     self.all_action_time = \
         sum([record_range.bounds for record_range in self.record_ranges])
+    self.measure_per_frame = measure_per_frame
 
   @property
   def clock_time(self):
     clock_duration = sum([x.duration for x in self.toplevel_slices])
     clock_overhead = sum([ClockOverheadForEvent(x) for x in self.all_slices])
     return clock_duration - clock_overhead
 
   @property
   def cpu_time(self):
     cpu_duration = 0
@@ skipping 20 matching lines @@
           break
     return slices_in_actions
 
   def AppendThreadSlices(self, thread):
     self.all_slices.extend(self.SlicesInActions(thread.all_slices))
     self.toplevel_slices.extend(self.SlicesInActions(thread.toplevel_slices))
 
   # Currently we report cpu-time per frame, tasks per frame, and possibly
   # the mean frame (if there is a trace specified to find it).
   def AddResults(self, num_frames, results):
-    cpu_per_frame = Rate(self.cpu_time, num_frames)
-    tasks_per_frame = Rate(len(self.toplevel_slices), num_frames)
+    num_intervals = num_frames if self.measure_per_frame else 1
+    cpu_per_interval = Rate(self.cpu_time, num_intervals)
+    tasks_per_interval = Rate(len(self.toplevel_slices), num_intervals)
     results.AddValue(scalar.ScalarValue(
-        results.current_page, ThreadCpuTimeResultName(self.name),
-        'ms', cpu_per_frame))
+        results.current_page,
+        ThreadCpuTimeResultName(self.name, self.measure_per_frame), 'ms',
+                                cpu_per_interval))
     results.AddValue(scalar.ScalarValue(
-        results.current_page, ThreadTasksResultName(self.name),
-        'tasks', tasks_per_frame))
+        results.current_page,
+        ThreadTasksResultName(self.name, self.measure_per_frame),
+        'tasks', tasks_per_interval))
     # Report mean frame time if this is the thread we are using for normalizing
     # other results. We could report other frame rates (eg. renderer_main) but
     # this might get confusing.
     if self.name == FrameTraceThreadName:
-      num_frames = self.CountTracesWithName(FrameTraceName)
-      mean_frame_time = Rate(self.all_action_time, num_frames)
+      num_intervals = self.CountTracesWithName(FrameTraceName) \
+          if self.measure_per_frame else 1
+      frame_time = Rate(self.all_action_time, num_intervals)
       results.AddValue(scalar.ScalarValue(
-          results.current_page, ThreadMeanFrameTimeResultName(self.name),
-          'ms', mean_frame_time))
+          results.current_page,
+          ThreadFrameTimeResultName(self.name, self.measure_per_frame),
+          'ms', frame_time))
 
   def AddDetailedResults(self, num_frames, results):
+    num_intervals = num_frames if self.measure_per_frame else 1
     slices_by_category = collections.defaultdict(list)
     for s in self.all_slices:
       slices_by_category[s.category].append(s)
     all_self_times = []
     for category, slices_in_category in slices_by_category.iteritems():
       self_time = sum([x.self_time for x in slices_in_category])
       all_self_times.append(self_time)
-      self_time_result = (float(self_time) / num_frames) if num_frames else 0
+      self_time_result = (
+          (float(self_time) / num_intervals) if num_intervals else 0)
       results.AddValue(scalar.ScalarValue(
           results.current_page, ThreadDetailResultName(self.name, category),
           'ms', self_time_result))
     all_measured_time = sum(all_self_times)
     idle_time = max(0, self.all_action_time - all_measured_time)
-    idle_time_result = (float(idle_time) / num_frames) if num_frames else 0
+    idle_time_result = (
+        (float(idle_time) / num_intervals) if num_intervals else 0)
     results.AddValue(scalar.ScalarValue(
         results.current_page, ThreadDetailResultName(self.name, "idle"),
         'ms', idle_time_result))
 
   def CountTracesWithName(self, substring):
     count = 0
     for event in self.all_slices:
       if substring in event.name:
         count += 1
     return count
 
+
 class ThreadTimesTimelineMetric(timeline_based_metric.TimelineBasedMetric):
-  def __init__(self):
+  def __init__(self, measure_per_frame=True):

[nednguyen, 2015/04/30 05:07:45]

Can you refactor this differently so we have two different classes instead of
one class with param like this?

The idea here is later, we gonna port thread_times benchmark to use timeline
based measurement instead. And in TBM model, all the timeline_based_metric
should be directly constructible so we can do things like:

for all subclass S of timeline based metrics:
   S.AddResults(...)

[jdduke (slow), 2015/04/30 14:29:39]

Sure, what about a
PerFrameThreadTimesTimelineMetric/PerSecondThreadTimesTimelineMetric that
essentially just expose the denominator to the base class?

     super(ThreadTimesTimelineMetric, self).__init__()
     # Minimal traces, for minimum noise in CPU-time measurements.
     self.results_to_report = AllThreads
     self.details_to_report = NoThreads
+    self.measure_per_frame = measure_per_frame
 
   def AddResults(self, model, _, interaction_records, results):
     # Set up each thread category for consistant results.
     thread_category_results = {}
     for name in TimelineThreadCategories.values():
       thread_category_results[name] = ResultsForThread(
-        model, [r.GetBounds() for r in interaction_records], name)
+          model, [r.GetBounds() for r in interaction_records], name,
+          self.measure_per_frame)
 
     # Group the slices by their thread category.
     for thread in model.GetAllThreads():
       thread_category = ThreadCategoryName(thread.name)
       thread_category_results[thread_category].AppendThreadSlices(thread)
 
     # Group all threads.
     for thread in model.GetAllThreads():
       thread_category_results['total_all'].AppendThreadSlices(thread)
 
     # Also group fast-path threads.
     for thread in model.GetAllThreads():
       if ThreadCategoryName(thread.name) in FastPathThreads:
         thread_category_results['total_fast_path'].AppendThreadSlices(thread)
 
     # Calculate the number of frames.
     frame_rate_thread = thread_category_results[FrameTraceThreadName]
     num_frames = frame_rate_thread.CountTracesWithName(FrameTraceName)
 
     # Report the desired results and details.
     for thread_results in thread_category_results.values():
       if thread_results.name in self.results_to_report:
         thread_results.AddResults(num_frames, results)
       # TOOD(nduca): When generic results objects are done, this special case
       # can be replaced with a generic UI feature.
       if thread_results.name in self.details_to_report:
         thread_results.AddDetailedResults(num_frames, results)
+
+  def ThreadDetailResultName(self, thread_category, detail):
+    return ThreadDetailResultName(thread_category, detail)
+
+  def ThreadCpuTimeResultName(self, thread_category):
+    return ThreadCpuTimeResultName(thread_category, self.measure_per_frame)
+
+  def ThreadTasksResultName(self, thread_category):
+    return ThreadTasksResultName(thread_category, self.measure_per_frame)
+
+  def ThreadFrameTimeResultName(self, thread_category):
+    return ThreadFrameTimeResultName(thread_category, self.measure_per_frame)