827956 - Huge tsvg opacity perf regressions on Sep 28, 2012

Status: RESOLVED INCOMPLETE

(Core :: SVG, defect)

Description

[Jonathan Watt [:jwatt]]

This regression is on top of the earlier regression covered by bug 827952, meaning that on some platforms the compounded effect was a regression of up to 800%.

Here's the perf-o-matic graph of the last 365 days of talos history for 'SVG, Opacity Row Major':

m-i: http://goo.gl/YNCof
m-c: http://goo.gl/cqtH0
m-i-non-PGO: http://goo.gl/mLWBy

Select the regression in September and you'll see that on September 28th:

MacOSX 10.7 regressed 91.9%
MacOSX 10.6 regressed 99.5%
WINNT 5.1 regressed 96.1%
WINNT 6.1 regressed 206.2%

The pushlog for the regression range is:

https://hg.mozilla.org/integration/mozilla-inbound/pushloghtml?fromchange=9f476b4ac1e1&tochange=6b58397ad735

The regression also appears to have made the perf results for WINNT 6.1 very erratic.

Comment 1

[Jonathan Watt [:jwatt]]

Note that on October 10/11 and on Nov 6 we gained back some of the lost ground (except on WINNT 6.1), but we're still not close to being back to where we were. It's also not clear that either of those gains were due to partially fixing the regression that resulted in this bug.

Pushlog for October gain:
https://hg.mozilla.org/integration/mozilla-inbound/pushloghtml?fromchange=ebeca12019a2&tochange=7065cefc10dd

Pushlog for November gain:
https://hg.mozilla.org/integration/mozilla-inbound/pushloghtml?fromchange=613bfe02f0c1&tochange=475ae0fac54e

Comment 2

[Martijn Wargers (dead)]

Is the regression still there?

Comment 3

[Jonathan Watt [:jwatt]]

Is it possible to get graph server to produce graphs going back that far?

Comment 4

[Joel Maher ( :jmaher ) (UTC -8)]

It does 365 days by default, I suspect we could get further data, but would have to hack it locally.  You would want 2 years 2 months?  Would it be useful, it seems as though we should just focus on our current work- but it shouldn't be too hard to do something one off for this.