1156798 - MSVC PGO is slowing us down again

Status: RESOLVED WONTFIX

(Core :: JavaScript Engine, defect)

Description

[Jan de Mooij [:jandem]]

Last year I noticed MSVC 2010 with PGO was slowing us down on some workloads like asm.js compilation, see bug 1030706, and we ended up disabling MSVC 2010 PGO for js/src.

About 6 months ago, we switched to Visual Studio 2013 and re-enabled PGO for js/src. VS 2013's PGO seemed less buggy and more predictable.

Back then, I also tried to add more benchmarks (Octane and Kraken) to the PGO profile (bug 1096280), but that didn't help much and introduced some minor regressions/noise, so I backed it out again.

Today Hannes fixed the PGO build on AWFY and it turns out PGO is helping us a bit on Sunspider but nowadays it's slowing us down on Octane and Kraken, see http://arewefastyet.com/#machine=17 and especially Octane-raytrace is *much* slower:

http://arewefastyet.com/#machine=17&view=single&suite=octane&subtest=RayTrace

Also, a few weeks ago Nicolas ran into an Aurora-only Windows PGO bug (bug 1149377) that took a lot of time to track down.

So, based on all this, I think it's reasonable to disable PGO in js/src again and see how it affects the benchmarks.

Comment 1

[Jan de Mooij [:jandem]]

Try push disabling PGO in js/src, so we can do some measurements:

https://treeherder.mozilla.org/#/jobs?repo=try&revision=45a7ac5299bf

Comment 2

[Hannes Verschore [:h4writer]]

Attachment: results disabled pgo in js/src

Comment 3

[Hannes Verschore [:h4writer]]

Best way to compare is to compare with the numbers at:
http://arewefastyet.com/overview/#/machine/17

Comment 4

[Hannes Verschore [:h4writer]]

Attachment: results disabled pgo in js/src (second run)

Comment 5

[Jan de Mooij [:jandem]]

Attachment: Patch

The results Hannes attached show this is a win on most benchmarks.

Comment 6

[Hannes Verschore [:h4writer]]

I just saw that the numbers AWFY is reporting for PGO are actually incorrect.

We have some extra debug checks that we disable using "JSGC_DISABLE_POISONING=1". But that env. variable isn't set for the PGO line. So I'll enable that now.

After we get new numbers we should look again how those numbers compare with the numbers I posted here (which are with those check disabled).

(So please wait with landing this!)

Comment 7

[Jan de Mooij [:jandem]]

(In reply to Hannes Verschore [:h4writer] from comment #6)
> We have some extra debug checks that we disable using
> "JSGC_DISABLE_POISONING=1". But that env. variable isn't set for the PGO
> line. So I'll enable that now.

Also, my patch to disable PGO in moz.build no longer works since bug 1084162. I'll just close this for now.