1243080 - Use --chunk-by-runtime in order to rebalance runtimes for mochitest-plain chunks

Status: NEW

(Testing :: General, defect)

Description

[Armen [:armenzg]]

Spinning from bug 1242502 because mozreview doesn't let me push the second change.

Comment 1

[Armen [:armenzg]]

https://treeherder.mozilla.org/#/jobs?repo=try&revision=ea4dfe1cc337

(In reply to Andrew Halberstadt [:ahal] from comment #6)
> I suspect it is turned on. They look closer than before to me, at least for
> buildbot linux opt. On central I saw the following job times:
> 30 67 31 26 38
> 
> In your try push, I saw:
> 22 58 27 21 24
> 
> That obviously isn't perfect, but it's an improvement. I bet that 58 is from
> a really large test manifest, the algorithm won't split up any manifests.
> You'd probably also get better results by changing the PERCENTILE [1] to
> something lower. Doing this comes at the cost of larger runtimes file sizes.
> Try lowering that number to 0, and re-generate the files and do a new try
> push. You should see a better result and that way you'll know it's enabled.
> 
> [1]
> https://dxr.mozilla.org/mozilla-central/source/testing/runtimes/
> writeruntimes.py#12

(In reply to Armen Zambrano [:armenzg] - Engineering productivity from comment #9)
> Thanks ahal.
> I was focusing on the longest running task and didn't realized that it was
> actually working.
> 
> I'm trying now with 0.01 percentile.

Comment 2

[Armen [:armenzg]]

0.01 made chunk 4 even run longer than 90 minutes while in the previous run actual run under 50 minutes.

It probably makes sense to invest in bug 1242682.

https://treeherder.mozilla.org/#/jobs?repo=try&revision=d342f527e10f,65afcfc167ad&group_state=expanded&filter-searchStr=tc&selectedJob=15899552

Unrelated, I also now see a crash for mochitest e10s 4 (since we bumped the maxruntime)

Comment 3

[Armen [:armenzg]]

Not working on this.