1638045 - Run Windows jit-tests in parallel again

Status: RESOLVED FIXED

(Core :: JavaScript Engine, task, P3)

Description

[Steve Fink [:sfink] [:s:]]

Way back in bug 1391877, we restricted Windows jit-tests jobs to a single test process due to mysterious timeouts. It would be much better to relax that restriction -- perhaps a max of 16 would work, as with bug 1557130?

Comment 1

[Andrew Halberstadt [:ahal]]

This looks like a change to autospider.py which according to ./mach file-info bugzilla-component belongs here.

Comment 2

[Jon Coppeard (:jonco)]

Is this why CGC tests are so much slower on Windows? I thought there was a bug for this but I can't find it.

Comment 3

[Steve Fink [:sfink] [:s:]]

Attachment: Bug 1638045 - Allow parallel jit-tests to run up to 8-way on Windows

Comment 4

[Steve Fink [:sfink] [:s:]]

I first tried a max of 16, and got an error OSError: [WinError 1455] The paging file is too small for this operation to complete. We could probably do better by flagging memory-intensive tests and running them one at a time. But 8 worked, and worked very well: Win x64 opt SM(p) was halved from 58→27 minutes, Win x64 debug SM(cgc) was 88→32, and Win x64 debug SM(p) was 118→47.

So right now, yes, it seems like this is a large reason why cgc is so slow on Windows. But I remember a different bug that was a lot more mysterious. I looked at individual test durations and things. (Or maybe I did and didn't find a difference, and forgot that I singlethreaded them 8 years ago?)

Hm. But I just looked at a linux x64 debug SM(cgc), and it's only 17 minutes. And according to the variants file, SM(cgc) skips the jstests entirely on Windows. (And I checked, the linux and windows log files agree with this.) So there's still an enormous difference.

Comment 5

[Steve Fink [:sfink] [:s:]]

Ah right, bug 1622824 comment 10. Still a mystery.

Comment 6

[Jon Coppeard (:jonco)]

(In reply to Steve Fink [:sfink] [:s:] from comment #4)

We could probably do better by flagging memory-intensive tests and running them one at a time.

This is a good idea. There's a test called memory-fill.js that fails for me locally if I run with --tbpl.

Comment 7

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/214891cc1234
Allow parallel jit-tests to run up to 8-way on Windows r=spidermonkey-reviewers,jonco,mgaudet

Comment 8

[Atila Butkovits]

https://hg.mozilla.org/mozilla-central/rev/214891cc1234