1918170 - Investigate improving speed of xpcshell-test builders by adjusting the number of threads

Status: NEW

(Release Engineering :: General, task)

Description

[Mark Banner (:standard8)]

In bug 1917833 I found that changing the number of threads (processes) that xpcshell-tests use on my Mac M1 Max significantly improved the run time of xpcshell-tests and the performance of the machine whilst running the tests.

We should consider if we want the same for CI or not.

A couple of try pushes:

• With no changes, the test runners use 48 cores
• With changes, the test runners use 12 cores

:padenot suggested that there appears to be a 2x improvement in time reduction of the xpcshell-test part of the builder run.

I'll attach a patch for this, but if more in-depth investigations are required, I'd prefer to pass those off to someone working on CI.

Comment 1

[Mark Banner (:standard8)]

Attachment: Bug 1918170 - Reduce the number of threads when running for xpcshell-test Mac CI builders to improve performance. r?jcristau!

Comment 2

[Paul Adenot (:padenot)]

"up to 2x", but more importantly tighter stddev, without having enough retries for the data to be significant.

Comment 3

[Mark Banner (:standard8)]

Having had some testing of this on Intel silicon on local machines, there's no clear advantage to reducing the threads (for Intel Silicon at least) and may actually be a regression point.

I do think that as this hasn't been revisited for a while, that it may be worth revisiting and doing some analysis on the builders across all platforms. However I'm not best positioned to do that, so un-assigning myself.