1739996 - 28.84 - 7.32% glterrain / glterrain + 7 more (Linux, OSX, Windows) regression on Thu November 4 2021

Status: RESOLVED FIXED

(Core :: Graphics, defect)

Description

[Andra Esanu (needinfo me)]

Perfherder has detected a talos performance regression from push 9cae29707d33f1455303c0d8a90ee49698b455dc. As author of one of the patches included in that push, we need your help to address this regression.

Regressions:

Ratio	Test	Platform	Options	Absolute values (old vs new)
29%	glterrain	windows10-64-qr	e10s stylo webgl-ipc webrender	1.81 -> 2.33
26%	glterrain	windows10-64-shippable-qr	e10s stylo webrender	1.77 -> 2.24
25%	glterrain	windows10-64-qr	e10s stylo webgl-ipc webrender	1.82 -> 2.28
24%	glterrain	windows10-64-shippable-qr	e10s stylo webrender-sw	1.42 -> 1.76
23%	glterrain	windows10-64-shippable-qr	e10s stylo webgl-ipc webrender	1.78 -> 2.18
10%	glterrain	linux1804-64-shippable-qr	e10s stylo webgl-ipc webrender	12.24 -> 13.49
8%	glterrain	macosx1015-64-shippable-qr	e10s stylo webgl-ipc webrender	1.84 -> 1.98
8%	glterrain	macosx1015-64-shippable-qr	e10s stylo webrender	1.83 -> 1.97
7%	glterrain	macosx1015-64-shippable-qr	e10s stylo webrender-sw	3.47 -> 3.72

Details of the alert can be found in the alert summary, including links to graphs and comparisons for each of the affected tests. Please follow our guide to handling regression bugs and let us know your plans within 3 business days, or the offending patch(es) will be backed out in accordance with our regression policy.

For more information on performance sheriffing please see our FAQ.

Comment 1

[Andrew Osmond [:aosmond] (he/him)]

Well this is fun. I see the webgl-ipc variant sets webgl.out-of-process to true, which explains why Linux is on this list (no change for the users otherwise).

That's a pretty steep regression on Windows for presumably little gain. Should I consider making it adaptable, to select the compositor thread if THREADSAFE_GL is allowed, and select the renderer thread otherwise?

Comment 2

[Andrew Osmond [:aosmond] (he/him)]

Attachment: Bug 1739996 - Select compositor or renderer thread for WebGL OOP depending on threadsafe GL status.

If GL is threadsafe, we can run on the compositor thread. This appears
to have performance benefits, possibly because the renderer thread is
too busy. If GL is not threadsafe, we must run the WebGL OOP instances
on the renderer thread.

At the time of writing, only the nouveau drivers on Linux are considered
to be not threadsafe, so most users will see WebGL running on the
compositor thread. This patch also adds prefers to override the
blocklist to either assume GL is threadsafe
(webgl.threadsafe-gl.force-enabled) and not threadsafe
(webgl.threadsafe-gl.force-disabled).

Comment 3

[Andrew Osmond [:aosmond] (he/him)]

try (default, should be threadsafe GL): https://treeherder.mozilla.org/jobs?repo=try&revision=1b3a1a8f38b623246e26fb5e4b45843751e31268
try (force not threadsafe GL): https://treeherder.mozilla.org/#/jobs?repo=try&revision=32f68fb8fb9c89016a585952be708190c63e30d5

Comment 4

[Kelsey Gilbert [:jgilbert]]

This is low severity per se, actually. It's only the glterrain test that has issues, and it's already a poor test (1.8->2.2ms) is likely overwhelmed by factors other than user-noticeable perf. The motionmark-webgl test looks flat.
It's probably uncommon for the webgl and webrender to be busy at the same time, I would imagine? (at least in a way that affected eachother)

Comment 5

[BugBot [:suhaib / :marco/ :calixte]]

Set release status flags based on info from the regressing bug 1734649

Comment 6

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/971876accc44
Select compositor or renderer thread for WebGL OOP depending on threadsafe GL status. r=jgilbert

Comment 7

[Cristina Cozmuta (:CrissCozmuta)]

https://hg.mozilla.org/mozilla-central/rev/971876accc44

Comment 8

[Andra Esanu (needinfo me)]

(In reply to Pulsebot from comment #6)

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/971876accc44
Select compositor or renderer thread for WebGL OOP depending on threadsafe
GL status. r=jgilbert

== Change summary for alert #32342 (as of Thu, 11 Nov 2021 06:13:33 GMT) ==

Improvements:

Ratio	Test	Platform	Options	Absolute values (old vs new)
21%	glterrain	windows10-64-shippable-qr	e10s stylo webrender-sw	1.76 -> 1.38
20%	glterrain	windows10-64-shippable-qr	e10s stylo webrender	2.22 -> 1.79
19%	glterrain	windows10-64-shippable-qr	e10s stylo webgl-ipc webrender	2.19 -> 1.78
18%	glterrain	windows10-64-qr	e10s stylo webgl-ipc webrender	2.19 -> 1.80
18%	glterrain	windows10-64-qr	e10s stylo webgl-ipc webrender	2.27 -> 1.87
...	...	...	...	...
5%	glterrain	linux1804-64-qr	e10s stylo webgl-ipc webrender	13.22 -> 12.53

For up to date results, see: https://treeherder.mozilla.org/perfherder/alerts?id=32342