Telemetry: adds CPU frequency stats.

tools/telemetry/telemetry/core/platform/proc_supporting_platform_backend.py

Index: tools/telemetry/telemetry/core/platform/proc_supporting_platform_backend.py
diff --git a/tools/telemetry/telemetry/core/platform/proc_supporting_platform_backend.py b/tools/telemetry/telemetry/core/platform/proc_supporting_platform_backend.py
index 8a7ddd0d81593fd16f8825b97c2a98c684106e4b..12162e096dbfc3b145e19739f83117b085b75dfc 100644
--- a/tools/telemetry/telemetry/core/platform/proc_supporting_platform_backend.py
+++ b/tools/telemetry/telemetry/core/platform/proc_supporting_platform_backend.py
@@ -2,6 +2,8 @@
 # Use of this source code is governed by a BSD-style license that can be
 # found in the LICENSE file.
 
+import collections
+
 try:
   import resource  # pylint: disable=F0401
 except ImportError:
@@ -32,6 +34,22 @@ class ProcSupportingPlatformBackend(platform_backend.PlatformBackend):
     meminfo = self._GetProcFileDict(meminfo_contents)
     return self._ConvertKbToByte(meminfo['MemTotal'])
 
+  def GetGlobalCpuStats(self):
+    num_cpus = int(self._GetFileContents(
+        '/sys/devices/system/cpu/kernel_max'))
+    cpu_frequencies = collections.defaultdict(int)
+    for cpu in range(num_cpus):
+      try:
+        stats = self._GetFileContents(
+            '/sys/devices/system/cpu/cpu%d/cpufreq/stats/time_in_state' %
+            cpu)
+        for e in stats.splitlines():
+          k, v = e.split()
+          cpu_frequencies[int(k)] += int(v)

[pasko, 2014/04/22 14:06:30]

I'm concerned with the story towards two vectors:
1. CPU hotplug
2. scaling governor set to 'performance'

More on (1):
Depending on CPU hotplug state (on/off) at the time of collecting the stats (in
page_cycler case when the page has already loaded), a few CPUs can be already
switched off. The decision to switch CPUs on/off are quite coarse-grained, so we
may end up loosing large chunks of cpufreq data occasionally. This would lead to
very noisy stats especially for "time spent at freq X".

More on (2):
Wouldn't the frequency be maxed out most of the time with scaling governor set
to 'performance'? Do we want to detect thermal frequency scaling this way?

Ways to reduce noise would be:
1. hotpug ON all cores all the time (just like
https://codereview.chromium.org/73173005/), this does not lead to realistic
power usage, however
2. special power-predictability bots with number of online cores restricted to
one at all times (infra maintenance troubles)
3. don't restrict anything, measure only cpu0, it would be less noisy this way
and may actually correlate to power usage much better, however, correlation
coefficients may have to me gathered experimentally for each test in separation
(i.e. data analysis would be tricky)

+      except IOError:
+        continue
+    return {'GlobalCpuFrequencyStats': cpu_frequencies}
+
   def GetCpuStats(self, pid):
     stats = self._GetProcFileForPid(pid, 'stat')
     if not stats: