All graphs for this bug:
  https://chromeperf.appspot.com/group_report?bug_id=659771

Original alerts at time of bug-filing:
  https://chromeperf.appspot.com/group_report?keys=agxzfmNocm9tZXBlcmZyFAsSB0Fub21hbHkYgICgjZbHvAkM,agxzfmNocm9tZXBlcmZyFAsSB0Fub21hbHkYgICgje_rogoM,agxzfmNocm9tZXBlcmZyFAsSB0Fub21hbHkYgICgje-_swoM,agxzfmNocm9tZXBlcmZyFAsSB0Fub21hbHkYgICgjeqHpwsM


Bot(s) for this bug's original alert(s):

webrtc-mac-large-tests
webrtc-win-large-tests

Sorry, no immediate ideas. My guess is that there is something wrong in the test rather than with what is being tested.

Can we be fully certain that it is this refactoring CL that caused the regression?

It's even more surprising given that the affected tests - as far as I can tell - don't use any of the objects that were touched by the refactoring, they create the AudioProcessingModule directly, and the APM does not use RtcEventLog.

What about memory footprint of the binaries, or something related to the cache? 
The tests only process 150 frames, and if the duration for the first frame is very high, that could perhaps offset the results. 

You could output the processing time per frame as perf data as well, into an array, like this:

RESULT stream_offset: synchronization= [-48,-73,-56,-82,-65,-89,-70,-55,-81,-62,-90,-71,-53,-77,-63,-86,-84,-53,-79,-76,-100,-70,-66,-91,-74,-99,-80,-107,-75,-58,-82,-65,-92,-84,-70,-50,-75,-59,-86,-68,-64,-89,-59,-83,-74,-55,-82,-63,-89,-71,-70,-94,-64,-88,-71,-55,-81,-78,-103,-84,-109,-78,-61,-88,-70,-69,-93,-62,-86,-70,-94,-76,-73,-99,-124,-149,-75,-57,-83,-66,-92,-74,-72,-96,-64,-90,-73,-56,-81,-63,-91,-74,-54,-82,-63,-91,-86,-69,-93,-75,-103,-70,-68,-94,-62,-87,-83,-109,-79,-75,-102,-83,-107,-77,-62,-84,-67,-93,-77,-73,-100,-80,-105,-75,-57,-83,-66,-92,-74,-57,-84,-66,-91,-114,-57,-79,-58,-85,-68,-92,-77,-60,-86,-66,-94,-75,-70,-96,-66,-91,-77,-72,-96,-76,-45] ms

Then you'd be able to see such data.

I had a look at a longer period of time and I can see that the regressions caused by Per's CLs later on ( bug 659771  and  bug 659768 ) have more significance than this refactoring. I think these metrics are just a bit too noisy, so let's close this bug.