1714184 - Fission CONTENT_FRAME_TIME_VSYNC regression vs e10s

Status: NEW

(Core :: Graphics, defect, P3)

Description

[Chris Peterson [:cpeterson]]

Attachment: Screenshot of Fission Beta 89 experiment's CONTENT_FRAME_TIME_VSYNC regression.png

In a recent Fission Beta 89 experiment [1] , we saw CHECKERBOARDING_SEVERITY and CONTENT_FRAME_TIME_VSYNC were worse with Fission than e10s.

I've attached a screenshot of the experiment dashboard's CONTENT_FRAME_TIME_VSYNC results.

We saw CONTENT_FRAME_TIME_VSYNC regress about 10-15% (about 50–100 vsync interval percentage points) for users who were already seeing slow CONTENT_FRAME_TIME_VSYNC (500–750 vsync interval percentage points) with e10s. Users who had fast CONTENT_FRAME_TIME_VSYNC with e10s saw almost no regression with Fission.

• Is increasing from 500–750 vsync interval percentage points to 600–850 a big regression? What is the visible user impact?

• Do you have any suggestions for diagnosing the CONTENT_FRAME_TIME_VSYNC regression? I see bug 1517980 added a CONTENT_FRAME_TIME_REASON field. Perhaps we can query the CONTENT_FRAME_TIME_REASONs for Fission Beta 89 users to see if there are any differences compared to e10s users?

[1] https://protosaur.dev/partybal/bug_1706428_pref_fission_m7_beta_experiment_with_memory_filter_beta_89_91.html

Comment 1

[Jim Mathies [:jimm]]

Is increasing from 500–750 vsync interval percentage points to 600–850 a big regression? What is the visible user impact?

Pretty minor. Worth investigating to understand more but not worth blocking on. It might turn out to be general slow down on lower end systems simply trying to keep up with more active child processes.

Comment 2

[Chris Peterson [:cpeterson]]

(In reply to Jim Mathies [:jimm] from comment #1)

Worth investigating to understand more but not worth blocking on. It might turn out to be general slow down on lower end systems simply trying to keep up with more active child processes.

In that case, I will move this bug to Fission Milestone "Future" so it doesn't block shipping Fission MVP. We will still keep an eye on CONTENT_FRAME_TIME_VSYNC in our Beta 90–91 experiments and Release experiments.

Comparing our Beta 89 experiment's user segments for high/low CPU and high/low memory, it looks like CONTENT_FRAME_TIME_VSYNC is generally better with Fission, but users with low memory (<= 4 GB RAM) are more likely to have longer long tails. The CONTENT_FRAME_TIME_VSYNC regressions seem more correlated with low memory than low CPU.

https://protosaur.dev/partybal/bug_1706428_pref_fission_m7_beta_experiment_with_memory_filter_beta_89_91.html#content_frame_time_vsync373