Telemetry: integrates memory_measurement with TCMalloc dumps

tools/perf/perf_tools/memory_measurement.py

 # Copyright (c) 2012 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.
 from perf_tools import histogram_metric
 from telemetry.page import page_measurement
 
 MEMORY_HISTOGRAMS = [
     {'name': 'V8.MemoryExternalFragmentationTotal', 'units': 'percent'},
     {'name': 'V8.MemoryHeapSampleTotalCommitted', 'units': 'kb'},
     {'name': 'V8.MemoryHeapSampleTotalUsed', 'units': 'kb'},
     {'name': 'Memory.RendererUsed', 'units': 'kb'}]
 
 BROWSER_MEMORY_HISTOGRAMS =  [
     {'name': 'Memory.BrowserUsed', 'units': 'kb'}]
 
 class MemoryMeasurement(page_measurement.PageMeasurement):
   def __init__(self):
     super(MemoryMeasurement, self).__init__('stress_memory')
     self.histograms = (
         [histogram_metric.HistogramMetric(
             h, histogram_metric.RENDERER_HISTOGRAM)
          for h in MEMORY_HISTOGRAMS] +
         [histogram_metric.HistogramMetric(
             h, histogram_metric.BROWSER_HISTOGRAM)
          for h in BROWSER_MEMORY_HISTOGRAMS])
+    self._is_running_tcmalloc_heap_profiler = False
 
   def DidNavigateToPage(self, page, tab):
     for h in self.histograms:
       h.Start(page, tab)
 
   def CustomizeBrowserOptions(self, options):
     options.AppendExtraBrowserArg('--dom-automation')
     # For a hard-coded set of Google pages (such as GMail), we produce custom
     # memory histograms (V8.Something_gmail) instead of the generic histograms
     # (V8.Something), if we detect that a renderer is only rendering this page
     # and no other pages. For this test, we need to disable histogram
     # customizing, so that we get the same generic histograms produced for all
     # pages.
     options.AppendExtraBrowserArg('--disable-histogram-customizer')
     options.AppendExtraBrowserArg('--memory-metrics')
     options.AppendExtraBrowserArg('--reduce-security-for-dom-automation-tests')
+    self._is_running_tcmalloc_heap_profiler = (options.profiler_tool ==
+        'tcmalloc-heap')
 
   def CanRunForPage(self, page):
     return hasattr(page, 'stress_memory')
 
   def MeasurePage(self, page, tab, results):
     for h in self.histograms:
       h.GetValue(page, tab, results)
+    if self._is_running_tcmalloc_heap_profiler:

[Dai Mikurube (NOT FULLTIME), 2013/05/14 17:02:10]

Ah, I thought that this condition can coexist with chrome && ... &&
chrome.memoryBenchmarking.isHeapProfilerRunning().

+      dumps = eval(tab.EvaluateJavaScript("""

[tonyg, 2013/05/14 18:54:13]

Please use json instead of eval

[bulach, 2013/05/15 10:22:56]

Done.

+        JSON.stringify([
+            chrome.memoryBenchmarking.heapProfilerDump('renderer'),
+            chrome.memoryBenchmarking.heapProfilerDump('browser'),
+        ]);
+      """))
+      if dumps:
+        print 'TCMalloc heap dumps available at ', dumps

[tonyg, 2013/05/14 16:52:44]

I don't like this code in the MemoryMeasurement because it means tcmalloc-heap
doesn't work with other measurements. Is there a way we can push this into the
Profiler?

[bulach, 2013/05/14 17:54:47]

"it's complicated" :)
- Profiler knows nothing about tab, or browser, and that's probably for good.

- tcmalloc will dump the heap at regular intervals, for any measurement, by
using --profiler=tcmalloc-heap.. this is just a "minor" optimization to ensure
that the _last_ dump is actually made, which seems to make sense in
"memory_measurment.py", no?

- at most, other measurements would loose up to the last "interval" dump.. 

so whilst I agree it'd be nice to push this to profiler, would you be ok doing
as a separate patch?

[tonyg, 2013/05/14 18:54:13]

How long is the interval? If it is sufficiently short, should we just remove
this code from memory_measurement? In any case, it seems like we don't need the
print line because the profiler is guaranteed to print that out in collect
profiler. Also, if you decide to keep this, please add a comment explaining that
this is a minor optimization to ensure the last dump is made.

[bulach, 2013/05/15 10:22:56]

the interval is relatively large right now. it takes a couple of secs to dump
the files, so in reality it has to be >=5secs..
I added a comment explaining this.
also, the file names seem indeed irrelevant, so I removed the print, eval, etc..

see more discussion on the underlying layer here:
https://codereview.chromium.org/15082004/