1014789 - 66.8% svg opacity regression on linux32 from bug 973532 on May 21 (fx32)

Status: RESOLVED WONTFIX

(Firefox :: New Tab Page, defect)

Description

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

oh, this is a bigger regression than ween in bug 1014784.

Here is a graph showing the details:
http://graphs.mozilla.org/graph.html#tests=[[225,132,33]]&sel=1400578923000,1400751723000&displayrange=7&datatype=running

I did some retriggers to ensure this is really the case:
https://tbpl.mozilla.org/?tree=Fx-Team&fromchange=a3a74829a5cf&tochange=a6dcd4c2fb6b&jobname=Ubuntu%20HW%2012.04%20fx-team%20talos%20svgr

here is more information about tsvg opacity:
https://wiki.mozilla.org/Buildbot/Talos/Tests#tsvg-opacity

Comment 1

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

oh, to be clear this is caused by:
http://hg.mozilla.org/integration/fx-team/rev/c3b86a5d52b6

Comment 2

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

I mistyped this, it is bug 937532

Comment 3

[Avi Halachmi (:avih)]

Tim, any reason for you to believe that bug 973532 could have caused this?

Comment 4

[Tim Taubert [:ttaubert] (inactive)]

(In reply to Avi Halachmi (:avih) from comment #3)
> Tim, any reason for you to believe that bug 973532 could have caused this?

I have no clue why that should impact tsvg tests :/ I thought we didn't even hit this code path on talos runs other than TART.

Comment 5

[Tim Taubert [:ttaubert] (inactive)]

Hmmm. Maybe that's related to bug 881590. I just remembered that we had similar tsvg problems when we cancelled a load in the hidden window. Maybe it's the inactive docShells that behaves funny again.

Comment 6

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

very likely.  Anything we could push to try to validate that?

Comment 7

[Tim Taubert [:ttaubert] (inactive)]

I tried starting to preload after the docShell has been set to inactive but that didn't work out.

https://tbpl.mozilla.org/?tree=Try&rev=09c27b193c17
https://tbpl.mozilla.org/?tree=Try&rev=5c924346c60b

http://perf.snarkfest.net/compare-talos/?oldRevs=09c27b193c17&newRev=5c924346c60b&submit=true

No improvement unfortunately. I will try to push something that validates the assumption from comment #5.

Comment 8

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

for reference, this posted a 27.2% regression on the PGO builds of linux32.

Comment 9

[Avi Halachmi (:avih)]

Looking at this test for the past 90 days, I can say that:

- Ubuntu 32 affected sharply
- OS X 10.8 looks like affected (but not as sharp)
- Ubuntu 64 looks like affected very mildly.
- Windows OSs were not affected or affected only very(!) slightly.

Very notable here is the difference between ubuntu 32 and 64.

If I look at the past 90 days:
- Ubuntu 32/64 were very noisy during March. 32 at the same level as this regression.

http://graphs.mozilla.org/graph.html#tests=[[225,64,31],[225,64,24],[225,132,37],[225,132,25],[225,132,35],[225,132,33]]&sel=1393144411512,1400920411512&displayrange=90&datatype=running

The only thing which regressed svg-opacity recently is cache v2 (bug 913806). It was turned on finally on 2014-05-16. Could this somehow be related?

Comment 10

[Avi Halachmi (:avih)]

Just for laughs, can we somehow retrigger/repush with cache v2 disabled?

We need these prefs to disable the new cache:
browser.cache.use_new_backend_temp = false
browser.cache.use_new_backend = 0

(and to enable it, the first one should be true, for some reason it looks like browser.cache.use_new_backend isn't part of controlling the new cache. see bug 913806 comment 15 and 16).

Comment 11

[Tim Taubert [:ttaubert] (inactive)]

I pushed multiple patches:

http://perf.snarkfest.net/compare-talos/?oldRevs=dac70cc0f120&newRev=7b4b70965d6b&submit=true
(cache v2 disabled, no perf gain)

http://perf.snarkfest.net/compare-talos/?oldRevs=dac70cc0f120&newRev=b2e694135ffa&submit=true
(docShell.isActive=false removed, no perf gain)

http://perf.snarkfest.net/compare-talos/?oldRevs=dac70cc0f120&newRev=2e6e9ab381a4&submit=true
(backout of c3b86a5d52b6 just to confirm, brings us back to old numbers indeed)

So what else could it be in that changeset?

Comment 12

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

this is a pretty old bug, the trend of this test is improvements, so this issue (expecially 66%) is not seen overall:
http://graphs.mozilla.org/graph.html#tests=%5B%5B225,132,33%5D,%5B225,53,33%5D%5D&sel=1380646402558,1412182402558&displayrange=365&datatype=running