Open Bug 1640949 Opened 5 years ago Updated 5 months ago

Evaluate performance of a build with the profiler disabled

Categories

(Core :: Performance, task)

task

Tracking

()

Performance Impact none

People

(Reporter: bas.schouten, Unassigned)

References

(Blocks 1 open bug)

Details

We should evaluate firefox performance without the profiler compiled in. This should be fairly easy to do and give us an idea of the 'total overhead' the profiler is adding when not being used. Hopefully this won't amount to much but we should make sure.

Gerald, could you take this? I think we just need 1 try push with the profiler and 1 without, and trigger all the talos jobs.

Blocks: 1635350
Flags: needinfo?(gsquelart)

Looking into it.

Assignee: nobody → gsquelart
Flags: needinfo?(gsquelart)

I've managed to build without the profiler, but I had to fix a couple of things on my local Windows build, and there are still a few failures elsewhere, I should investigate them...
Also, quite a few tests just try to use the profiler without testing if it's there!

But here are some comparisons:
https://treeherder.mozilla.org/perf.html#/compare?originalProject=try&originalRevision=2e9f87b20d3b398c80e5e49c902c7d7e911b17aa&newProject=try&newRevision=913d0f3598f83606217ee4cf5649aca227bc674e&framework=1
Base = default, with profiler; New = without profiler. So we don't want the right side to get noticeably better.
Click "Hide uncomparable results" to remove 1-sided tests (because of the aforementioned build failures).
I did multiple runs of only a few tests, so the confidence is still low overall.

I'm not used to analyzing these (what units are used?), so I'm not sure how to interpret the results, but AFAICS they didn't seem to consistently go one way.

Can someone knowledgeable please have a look? Feel free to re-trigger more tests as needed, and please let me know if I should try and fix the no-profiler build further to collect more data on more platforms&tests.

NI:Bas for follow-up.

(And whatever we prove today, this seems like a good kind of test that the profiler team should keep running when updating the profiler code, and adding more markers&labels, etc. Thank you for suggesting this.)

Flags: needinfo?(bas)
Whiteboard: [qf-]
Performance Impact: --- → -
Whiteboard: [qf-]
Assignee: mozbugz → nobody

I failed to look at this in time. Sorry about that Gerald! It's interesting that this lead to failures. And I'm still interested in this.

Flags: needinfo?(bas)
You need to log in before you can comment on or make changes to this bug.