99th_percentile_frame_rasterizer_time_millis regression on engine roll...

[kf6gpe]

https://flutter-flutter-perf.skia.org/e/?begin=1608245265&end=1608571746&keys=Xf5c75aaa34c194d39513caee2051940e&num_commits=50&request_type=1&xbaroffset=22916

194711e

/cc @zanderso This week's engine sheriff.

[zanderso]

Comparing before/after that engine roll, I'm not able to reproduce this locally on my Nexus 5x. The benchmark appears bimodal, occasionally having a ~5-10% larger 99th %-ile frame rasterization. @CareF for advice.

[CareF]

multi_widget_construction_perf__e2e_summary test is designed to rebuild many widgets in the widget tree for every frame. The original motivation is to observe the cost of rebuilding the widget tree, but as a side effect there are many GCs during the test. If the larger number of the bimodal distribution comes from some random peaks in rasterize time/build time, this may be the issue.
Previously we also see bimodal behavior #68620 where the app becomes slower as the test runs: this may be because of device thermal issue.

[jason-simmons]

I can reproduce this on a Moto G4 and bisected it to google/skia@1d89532.

Not sure why that change would affect the rasterizer metrics in this benchmark.

[kf6gpe]

@herbderby, this looks like something you committed --- any thoughts? Thanks!

[herbderby]

This would be very surprising. Looking at the trace, you don't have much history. The trace starts on Dec 16th, and this was flagged on Dec 18th. I suspect you are measuring something environmental, and not the code change.

[chinmaygarde]

Thanks for the clarification. Let's re-baseline and move on.

[github-actions]

This thread has been automatically locked since there has not been any recent activity after it was closed. If you are still experiencing a similar issue, please open a new bug, including the output of flutter doctor -v and a minimal reproduction of the issue.