tools/perf/metrics/timeline.py

   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 import collections
   5
   6 from telemetry.util.statistics import DivideIfPossibleOrZero
   7 from telemetry.value import scalar
   8 from telemetry.web_perf.metrics import timeline_based_metric
   9
  10
  11 class LoadTimesTimelineMetric(timeline_based_metric.TimelineBasedMetric):
  12   def __init__(self):
  13     super(LoadTimesTimelineMetric, self).__init__()
  14     self.report_main_thread_only = True
  15
  16   def AddResults(self, model, renderer_thread, interaction_records, results):
  17     assert model
  18     assert len(interaction_records) == 1, (
  19       'LoadTimesTimelineMetric cannot compute metrics for more than 1 time '
  20       'range.')
  21     interaction_record = interaction_records[0]
  22     if self.report_main_thread_only:
  23       thread_filter = 'CrRendererMain'
  24     else:
  25       thread_filter = None
  26
  27     events_by_name = collections.defaultdict(list)
  28     renderer_process = renderer_thread.parent
  29
  30     for thread in renderer_process.threads.itervalues():
  31
  32       if thread_filter and not thread.name in thread_filter:
  33         continue
  34
  35       thread_name = thread.name.replace('/','_')
  36       for e in thread.IterAllSlicesInRange(interaction_record.start,
  37                                            interaction_record.end):
  38         events_by_name[e.name].append(e)
  39
  40       for event_name, event_group in events_by_name.iteritems():
  41         times = [event.self_time for event in event_group]
  42         total = sum(times)
  43         biggest_jank = max(times)
  44
  45         # Results objects cannot contain the '.' character, so remove that here.
  46         sanitized_event_name = event_name.replace('.', '_')
  47
  48         full_name = thread_name + '|' + sanitized_event_name
  49         results.AddValue(scalar.ScalarValue(
  50             results.current_page, full_name, 'ms', total))
  51         results.AddValue(scalar.ScalarValue(
  52             results.current_page, full_name + '_max', 'ms', biggest_jank))
  53         results.AddValue(scalar.ScalarValue(
  54             results.current_page, full_name + '_avg', 'ms', total / len(times)))
  55
  56     for counter_name, counter in renderer_process.counters.iteritems():
  57       total = sum(counter.totals)
  58
  59       # Results objects cannot contain the '.' character, so remove that here.
  60       sanitized_counter_name = counter_name.replace('.', '_')
  61
  62       results.AddValue(scalar.ScalarValue(
  63           results.current_page, sanitized_counter_name, 'count', total))
  64       results.AddValue(scalar.ScalarValue(
  65           results.current_page, sanitized_counter_name + '_avg', 'count',
  66           total / float(len(counter.totals))))
  67
  68 # We want to generate a consistant picture of our thread usage, despite
  69 # having several process configurations (in-proc-gpu/single-proc).
  70 # Since we can't isolate renderer threads in single-process mode, we
  71 # always sum renderer-process threads' times. We also sum all io-threads
  72 # for simplicity.
  73 TimelineThreadCategories =  {
  74   "Chrome_InProcGpuThread": "GPU",
  75   "CrGpuMain"             : "GPU",
  76   "AsyncTransferThread"   : "GPU_transfer",
  77   "CrBrowserMain"         : "browser",
  78   "Browser Compositor"    : "browser",
  79   "CrRendererMain"        : "renderer_main",
  80   "Compositor"            : "renderer_compositor",
  81   "IOThread"              : "IO",
  82   "CompositorTileWorker"  : "raster",
  83   "DummyThreadName1"      : "other",
  84   "DummyThreadName2"      : "total_fast_path",
  85   "DummyThreadName3"      : "total_all"
  86 }
  87
  88 _MatchBySubString = ["IOThread", "CompositorTileWorker"]
  89
  90 AllThreads = TimelineThreadCategories.values()
  91 NoThreads = []
  92 FastPathThreads = ["GPU", "renderer_compositor", "browser", "IO"]
  93
  94 ReportMainThreadOnly = ["renderer_main"]
  95 ReportSilkDetails = ["renderer_main"]
  96
  97 # TODO(epenner): Thread names above are likely fairly stable but trace names
  98 # could change. We should formalize these traces to keep this robust.
  99 OverheadTraceCategory = "trace_event_overhead"
 100 OverheadTraceName = "overhead"
 101 FrameTraceName = "::SwapBuffers"
 102 FrameTraceThreadName = "renderer_compositor"
 103
 104 def Rate(numerator, denominator):
 105   return DivideIfPossibleOrZero(numerator, denominator)
 106
 107 def ClockOverheadForEvent(event):
 108   if (event.category == OverheadTraceCategory and
 109       event.name == OverheadTraceName):
 110     return event.duration
 111   else:
 112     return 0
 113
 114 def CpuOverheadForEvent(event):
 115   if (event.category == OverheadTraceCategory and
 116       event.thread_duration):
 117     return event.thread_duration
 118   else:
 119     return 0
 120
 121 def ThreadCategoryName(thread_name):
 122   thread_category = "other"
 123   for substring, category in TimelineThreadCategories.iteritems():
 124     if substring in _MatchBySubString and substring in thread_name:
 125       thread_category = category
 126   if thread_name in TimelineThreadCategories:
 127     thread_category = TimelineThreadCategories[thread_name]
 128   return thread_category
 129
 130 def ThreadCpuTimeResultName(thread_category):
 131   # This isn't a good name, but I don't want to change it and lose continuity.
 132   return "thread_" + thread_category + "_cpu_time_per_frame"
 133
 134 def ThreadTasksResultName(thread_category):
 135   return "tasks_per_frame_" + thread_category
 136
 137 def ThreadMeanFrameTimeResultName(thread_category):
 138   return "mean_frame_time_" + thread_category
 139
 140 def ThreadDetailResultName(thread_category, detail):
 141   detail_sanitized = detail.replace('.','_')
 142   return "thread_" + thread_category + "|" + detail_sanitized
 143
 144
 145 class ResultsForThread(object):
 146   def __init__(self, model, record_ranges, name):
 147     self.model = model
 148     self.toplevel_slices = []
 149     self.all_slices = []
 150     self.name = name
 151     self.record_ranges = record_ranges
 152     self.all_action_time = \
 153         sum([record_range.bounds for record_range in self.record_ranges])
 154
 155   @property
 156   def clock_time(self):
 157     clock_duration = sum([x.duration for x in self.toplevel_slices])
 158     clock_overhead = sum([ClockOverheadForEvent(x) for x in self.all_slices])
 159     return clock_duration - clock_overhead
 160
 161   @property
 162   def cpu_time(self):
 163     cpu_duration = 0
 164     cpu_overhead = sum([CpuOverheadForEvent(x) for x in self.all_slices])
 165     for x in self.toplevel_slices:
 166       # Only report thread-duration if we have it for all events.
 167       #
 168       # A thread_duration of 0 is valid, so this only returns 0 if it is None.
 169       if x.thread_duration == None:
 170         if not x.duration:
 171           continue
 172         else:
 173           return 0
 174       else:
 175         cpu_duration += x.thread_duration
 176     return cpu_duration - cpu_overhead
 177
 178   def SlicesInActions(self, slices):
 179     slices_in_actions = []
 180     for event in slices:
 181       for record_range in self.record_ranges:
 182         if record_range.ContainsInterval(event.start, event.end):
 183           slices_in_actions.append(event)
 184           break
 185     return slices_in_actions
 186
 187   def AppendThreadSlices(self, thread):
 188     self.all_slices.extend(self.SlicesInActions(thread.all_slices))
 189     self.toplevel_slices.extend(self.SlicesInActions(thread.toplevel_slices))
 190
 191   # Currently we report cpu-time per frame, tasks per frame, and possibly
 192   # the mean frame (if there is a trace specified to find it).
 193   def AddResults(self, num_frames, results):
 194     cpu_per_frame = Rate(self.cpu_time, num_frames)
 195     tasks_per_frame = Rate(len(self.toplevel_slices), num_frames)
 196     results.AddValue(scalar.ScalarValue(
 197         results.current_page, ThreadCpuTimeResultName(self.name),
 198         'ms', cpu_per_frame))
 199     results.AddValue(scalar.ScalarValue(
 200         results.current_page, ThreadTasksResultName(self.name),
 201         'tasks', tasks_per_frame))
 202     # Report mean frame time if this is the thread we are using for normalizing
 203     # other results. We could report other frame rates (eg. renderer_main) but
 204     # this might get confusing.
 205     if self.name == FrameTraceThreadName:
 206       num_frames = self.CountTracesWithName(FrameTraceName)
 207       mean_frame_time = Rate(self.all_action_time, num_frames)
 208       results.AddValue(scalar.ScalarValue(
 209           results.current_page, ThreadMeanFrameTimeResultName(self.name),
 210           'ms', mean_frame_time))
 211
 212   def AddDetailedResults(self, num_frames, results):
 213     slices_by_category = collections.defaultdict(list)
 214     for s in self.all_slices:
 215       slices_by_category[s.category].append(s)
 216     all_self_times = []
 217     for category, slices_in_category in slices_by_category.iteritems():
 218       self_time = sum([x.self_time for x in slices_in_category])
 219       all_self_times.append(self_time)
 220       self_time_result = (float(self_time) / num_frames) if num_frames else 0
 221       results.AddValue(scalar.ScalarValue(
 222           results.current_page, ThreadDetailResultName(self.name, category),
 223           'ms', self_time_result))
 224     all_measured_time = sum(all_self_times)
 225     idle_time = max(0, self.all_action_time - all_measured_time)
 226     idle_time_result = (float(idle_time) / num_frames) if num_frames else 0
 227     results.AddValue(scalar.ScalarValue(
 228         results.current_page, ThreadDetailResultName(self.name, "idle"),
 229         'ms', idle_time_result))
 230
 231   def CountTracesWithName(self, substring):
 232     count = 0
 233     for event in self.all_slices:
 234       if substring in event.name:
 235         count += 1
 236     return count
 237
 238 class ThreadTimesTimelineMetric(timeline_based_metric.TimelineBasedMetric):
 239   def __init__(self):
 240     super(ThreadTimesTimelineMetric, self).__init__()
 241     # Minimal traces, for minimum noise in CPU-time measurements.
 242     self.results_to_report = AllThreads
 243     self.details_to_report = NoThreads
 244
 245   def AddResults(self, model, _, interaction_records, results):
 246     # Set up each thread category for consistant results.
 247     thread_category_results = {}
 248     for name in TimelineThreadCategories.values():
 249       thread_category_results[name] = ResultsForThread(
 250         model, [r.GetBounds() for r in interaction_records], name)
 251
 252     # Group the slices by their thread category.
 253     for thread in model.GetAllThreads():
 254       thread_category = ThreadCategoryName(thread.name)
 255       thread_category_results[thread_category].AppendThreadSlices(thread)
 256
 257     # Group all threads.
 258     for thread in model.GetAllThreads():
 259       thread_category_results['total_all'].AppendThreadSlices(thread)
 260
 261     # Also group fast-path threads.
 262     for thread in model.GetAllThreads():
 263       if ThreadCategoryName(thread.name) in FastPathThreads:
 264         thread_category_results['total_fast_path'].AppendThreadSlices(thread)
 265
 266     # Calculate the number of frames.
 267     frame_rate_thread = thread_category_results[FrameTraceThreadName]
 268     num_frames = frame_rate_thread.CountTracesWithName(FrameTraceName)
 269
 270     # Report the desired results and details.
 271     for thread_results in thread_category_results.values():
 272       if thread_results.name in self.results_to_report:
 273         thread_results.AddResults(num_frames, results)
 274       # TOOD(nduca): When generic results objects are done, this special case
 275       # can be replaced with a generic UI feature.
 276       if thread_results.name in self.details_to_report:
 277         thread_results.AddDetailedResults(num_frames, results)