1850231 - Disabling background threads in the GC improves Speedometer3 by about 1.7% on Android

Status: RESOLVED INVALID

(Core :: JavaScript: GC, defect, P2)

Description

[Denis Palmeiro [:denispal]]

Try results here from returning false in ShouldUseBackgroundThreads: https://treeherder.mozilla.org/perfherder/comparesubtest?originalProject=try&newProject=try&newRevision=b499289b3c7a470f5f10130e279b1caf4ee2d5b5&originalSignature=4590278&newSignature=4590278&framework=13&application=geckoview&originalRevision=23c3983bdb65ebde5dd7448024d72368054ad09e&page=1&showOnlyImportant=1&replicates=1

Desktop results were much more mixed, so it seems to be an android only problem. Likely related to problem seen with stylo in bug 1848766.

Comment 1

[Jon Coppeard (:jonco)]

I'm not sure why this makes things so much worse, as we don't normally block on background threads in an incremental GC. Perhaps it causes GCs to extend into the measured period though.

Regardless, we should try adding performance hints like was done for stylo.

Comment 2

[Jamie Nicol [:jnicol]]

It's perhaps also interesting that disabling the background threads actually shows an improvement on try. The stylo issue used to show up on try, but since bug 1812589 landed it no longer did. IIRC after that landed neither disabling stylo threads, using performance hints, or setting the thread CPU affinities makes a noticeable difference to try results

Comment 3

[Denis Palmeiro [:denispal]]

This is likely just noise. It's not reproducible on any local devices, and a recent run by Jamie showed no improvement.