1201960 - Investigate regression in COMPOSITE_TIME since 2015-09-02

Status: RESOLVED FIXED

(Core :: Graphics, defect)

Description

[u279076]

I was taking a look at the Telemetry dashboard for COMPOSITE_TIME and noticed a significant regression starting with the 2015-09-02 Nightly build. See http://mzl.la/1Qel2DU.

Based on the data it looks like Nightly 43 was averaging about 0.16 since merge with a peak of 0.2 and minimum of 0.1. That is until 2015-09-02 when it spiked to 0.4 and is up to 0.86 as of yesterday.

I've filed this bug to track the investigation.

[Tracking Requested - why for this release]: this is a regression in time spent compositing. I think this should be tracked until we have a better idea of the cause.

Comment 1

[u279076]

Digging a bit further into the data it looks like this is a Windows issue. There is not a significant change in Linux and Mac OS X compositing time. When looking at just the Windows data the spike is evident.

Windows XP goes from 1.53 on September 1 to 2.18 on September 2 to 5.51 on September 3. (+260%)
Windows 7 goes from 0.31 on September 1 to 0.44 on September 2 to 1.00 on September 3. (+223%)
Windows 8 goes from 0.03 on September 1 to 0.17 on September 2 to 2.64 on September 3. (+8700%)
Windows 10 goes from 0.16 on September 1 to 0.30 on September 2 to 0.54 on September 3. (+238%)

Looking at the numbers it seems the regression is most pronounced on Windows 8.

Comment 2

[u279076]

Pushlog for mozilla-central between 2015-09-01 and 2015-09-02:
http://hg.mozilla.org/mozilla-central/pushloghtml?fromchange=cafb1c90f794a73100a8f0afb9fe3301df0f2bde&tochange=fb720c90eb49590ba55bf52a8a4826ffff9f528b

Comment 3

[u279076]

Another factor here is that non-e10s results do not show the spike so this might be an e10s bug.

Comment 4

[Jim Mathies [:jimm]]

should be fixed by the backout of 5bb965937e24.

Comment 5

[Liz Henry (:lizzard) (relman/hg->git project)]

Tracking for 43. Anthony, does that look better? Or still a problem?

Comment 6

[u279076]

Data by Nightly build date:
===========================
2015-09-01: 0.12
2015-09-02: 0.15
2015-09-03: 0.15
2015-09-04: 0.18
2015-09-05: 0.21
2015-09-06: 0.15
2015-09-07: 0.16
2015-09-08: 0.20
2015-09-09: 0.15
2015-09-10: 0.14
2015-09-11: 0.22
2015-09-12: 0.26
2015-09-13: 0.37
2015-09-14: 0.57
===========================
Source: http://mzl.la/1Qel2DU

I think I need someone more familiar with Telemetry to interpret the data. Looking at it now it seems like the spike on September 2 has shifted to September 11.