1681047 - 2.61 - 3.24% sessionrestore (linux64-shippable) regression on push 13a85de368dec8d88d4a8fb022f3f3ef1e1e3da8 (Thu December 3 2020)

Status: RESOLVED WONTFIX

(Core :: Graphics: WebRender, defect)

Description

[Alexandru Ionescu (needinfo me) [:alexandrui]]

Perfherder has detected a talos performance regression from push 13a85de368dec8d88d4a8fb022f3f3ef1e1e3da8. As author of one of the patches included in that push, we need your help to address this regression.

Regressions:

Ratio	Suite	Test	Platform	Options	Absolute values (old vs new)
3%	sessionrestore	sessionrestore	linux64-shippable	e10s stylo	511.58 -> 528.17
3%	sessionrestore	sessionrestore	linux64-shippable	e10s stylo	515.00 -> 528.42

Improvements:

Ratio	Suite	Test	Platform	Options	Absolute values (old vs new)
11%	startup_about_home_paint_realworld_webextensions	startup_about_home_paint_realworld_webextensions	linux64-shippable	e10s stylo	575.42 -> 514.83

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

[Martin Stránský [:stransky] (ni? me)]

Is there any log available from the run? Looks like the tests use non-compositing screens which does not match real-life scenario. We disabled WebRender for non-compositing screens due to missing shape mask implementation, but non-compositing screens should not be generally used with HW acceleration / WebRender.

Comment 2

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

Set release status flags based on info from the regressing bug 1479135

Comment 3

[Alexandru Ionescu (needinfo me) [:alexandrui]]

(In reply to Martin Stránský [:stransky] from comment #1)

Is there any log available from the run? Looks like the tests use non-compositing screens which does not match real-life scenario. We disabled WebRender for non-compositing screens due to missing shape mask implementation, but non-compositing screens should not be generally used with HW acceleration / WebRender.

Those are the jobs that regressed the tests from your patch. You can find : https://treeherder.mozilla.org/jobs?repo=autoland&revision=13a85de368dec8d88d4a8fb022f3f3ef1e1e3da8&group_state=expanded&selectedTaskRun=QEyfUFsiRdSd5AKEJKdYQg.0&searchStr=Linux%2Cx64%2Cshippable%2Copt%2CTalos%2Cperformance%2Ctests%2Ctest-linux64-shippable%2Fopt-talos-other-e10s%2Co

I'm not sure I understand your statement, as I'm not really familiar with the terms nor test. In the artifact tab of each job you can find the log that you're looking for.

Comment 4

[Martin Stránský [:stransky] (ni? me)]

Okay, the regression here is expected due to the test nature.

Comment 5

[Dave Hunt [:davehunt] [he/him] ⌚BST]

(In reply to Martin Stránský [:stransky] from comment #4)

Okay, the regression here is expected due to the test nature.

These tests are running with WebRender disabled. Should we expect such a regression in this case? Looking at the webrender and sw-webrender variants of this test there looks like there might be a similar regression, but no alerts were generated either due to the noise or the change being within the alert threshold.

Comment 6

[Martin Stránský [:stransky] (ni? me)]

(In reply to Dave Hunt [:davehunt] [he/him] ⌚GMT from comment #5)

(In reply to Martin Stránský [:stransky] from comment #4)

Okay, the regression here is expected due to the test nature.

These tests are running with WebRender disabled. Should we expect such a regression in this case?

It depends if you run the test on compositing or non-compositing backend. There should not be any change on compositing backend.
On non-compositing backends we use shape masks to draw transparent areas and that's done by basic compositor no matter if Basic/WebRender/GL compositor is used.

Comment 7

[Dave Hunt [:davehunt] [he/him] ⌚BST]

The hardware running the performance tests is documented here. How can we tell if these are running with compositing?

Comment 8

[Martin Stránský [:stransky] (ni? me)]

I did some investigation here. The tests are running under virtual X11 server (xvfb) which does not do compositing. The tests are also running with sw-wr (according the logs) so the regression here is definitely expected as sw-wr is disabled for popups (sw-wr contains the same bug 1479135 as wr).