1505495 - consider geckoprofiling to be run at the end of every normal test run and upload artifacts for every run

Status: RESOLVED DUPLICATE of bug 1786400

(Testing :: Raptor, defect, P3)

Description

[Joel Maher ( :jmaher ) (UTC -8)]

from bug 1459621 comment 17:
The Talos runs with profiling enabled do the same thing, but they output no numbers (because their values are skewed by the profiling overhead), and instead upload a zip file containing a profile of each execution of the test (eg. 10 profiles of the same thing). These 10 profiles are often more confusing than useful, as we currently have no good way to compare them, and the engineers looking at these profiles tend to just randomly pick one out of the 10, and ignore the others.

What I would prefer if it's possible is that in the same job, we run the test 10 times to produce numbers, and an 11th time with the profiler enabled, to upload a single profile.

This would eliminate the difference between "Talos with profiling" and "Talos without profiling", as all Talos runs would generate a profile. If we can move to something like this, I don't think profiling should be limited to mozilla-central. All talos runs on try and integration branches should have profiles. How feasible is this?



Ideally this should be looked at for raptor and talos.

Comment 1

[Alex Thayer [:alexical] (she/her)]

mconley please feel free to correct me, but it feels like this doesn't make sense as an fxperf:p1 right now?