Closed Bug 1042186 Opened 10 years ago Closed 10 years ago

7% SVGX + 14% TScrollX linux regressions on inbound (fx33) seen on Jul 19th from bug 1022612

Categories

(Core :: Layout, defect)

Product:
Core
Component:
Layout
Platform:
All
Linux
Type:
defect
Priority:
Not set
Severity:
normal

Tracking

()

Status:
RESOLVED FIXED
Tracking Flags:
Tracking Status
firefox34 + fixed

People

(Reporter: jmaher, Assigned: roc)

References

Details

(Keywords: perf, regression, Whiteboard: [talos_regression]) 

Oh, the largest set of patches I have seen land for a given bug in a long while has made it in!  This is great news.  Our graph server and Talos tests actually enjoy this for the most part!

If there were no performance regressions, this wouldn't be a good patch, congrats- there are a few, here is the summary (http://54.215.155.53:8080/alerts.html?rev=24a69de91baa):
+---------------------+-------------------------------+---------+
| platform            | test                          | percent |
+---------------------+-------------------------------+---------+
| WINNT 5.1 (ix)      | Customization Animation Tests | +3.7%   |
| WINNT 5.1 (ix)      | SVG-ASAP                      | -11%    |
| Ubuntu HW 12.04     | tscroll-ASAP                  | -6.85%  |
| Ubuntu HW 12.04     | SVG-ASAP                      | -5.6%   |
| Ubuntu HW 12.04     | SVG, Opacity Row Major        | +42.6%  |
| Ubuntu HW 12.04 x64 | tscroll-ASAP                  | -11.4%  |
| Ubuntu HW 12.04 x64 | SVG, Opacity Row Major        | +3.3%   |
| Ubuntu HW 12.04 x64 | SVG-ASAP                      | -5.83%  |
| Ubuntu HW 12.04 x64 | Tp5 Optimized (XRes)          | -16.6%  |
| Ubuntu HW 12.04     | Tp5 Optimized (XRes)          | -16.9%  |
+---------------------+-------------------------------+---------+

We had a few improvments, but SVG-ASAP and tscroll-ASAP took a hit.
these have surfaced on aurora now that we uplifted.

:roc, can you help us figure out what might be going on here?
Flags: needinfo?(roc)
I backed 1022612 out on Aurora.
Flags: needinfo?(roc)
oh, verified this is backed on out Aurora, so we just have to resolve this regression on Trunk (v.34)
In tscroll-ASAP, tiled-fixed.html and tiled-fixed-downscale.html obviously regressed about 30%.

In SVG-ASAP, some tests got a little bit worse, and hixie-005.xml and hixie-006.xml got a lot worse (about 20%).

It looks like the XRes data is not broken down by page so I'll ignore that for now.
Filed bug 1050096 on an obvious bug triggered by tiled-fixed.html.
Should we back this out on mozilla-central as well until the regressions are fixed?
Flags: needinfo?(roc)
No. At least I need to recheck after landing 1050096 to see what impact that had.
Flags: needinfo?(roc)
the aggregated tscrollx regression still shows up on inbound (now v.34):
http://graphs.mozilla.org/graph.html#tests=[[287,131,33]]&sel=none&displayrange=90&datatype=running

Here is the aggregated view for tsvgx:
http://graphs.mozilla.org/graph.html#tests=[[281,131,33]]&sel=none&displayrange=90&datatype=running

You can see the regression on July 19th, and it has stayed regressed since then.
this is back on aurora after the uplift.
> > Should we back this out on mozilla-central as well until the regressions are
> > fixed?
> (In reply to Robert O'Callahan (:roc) (Mozilla Corporation) (on vacation Sep 27 to Oct 12) from comment #7)
> No. At least I need to recheck after landing 1050096 to see what impact that
> had.

I really think we should back out this change if we don't expect the regressions will be fixed soon. We wouldn't want these regressions making it to Release
Flags: needinfo?(roc)
The TscrollX regression seems to have been fixed.
http://graphs.mozilla.org/graph.html#tests=[[287,131,33]]&sel=none&displayrange=90&datatype=running
I believe that's due to bug 1062100 being fixed properly.

The Tsvg regression hasn't:
http://graphs.mozilla.org/graph.html#tests=[[281,131,33]]&sel=none&displayrange=90&datatype=running
But that's a smaller and less important regression.

We can't back out 1022612. A lot of stuff has been built on it, including features promised for FxOS.
Flags: needinfo?(roc)
the problem is tscrollx is still highly regressed on Aurora, can we apply whatever fix (apparently http://hg.mozilla.org/integration/mozilla-inbound/pushloghtml?changeset=f782e395482d) to Aurora?
roc - I set you as the owner as it looks like you've been handling the fixes. See Joel's question in comment 12 for you as well.
Assignee: nobody → roc
Flags: needinfo?(roc)
Requested Aurora approval for uplift of bug 1062100.
Flags: needinfo?(roc)
I have approved bug 1062100 for Aurora. 

Joel - Once the fix lands, can you please confirm that the regression has been addressed and, if so, resolve this bug?
Flags: needinfo?(jmaher)
tscrollX has been resolved with this uplift-

we still have the svgx regression:
http://graphs.mozilla.org/graph.html#tests=%5B%5B281,131,33%5D,%5B281,52,33%5D%5D&sel=1405428698877,1413204698877&displayrange=90&datatype=running

I suspect part of the regression will be fixed when we uplift 35 to aurora.  If it doesn't make sense to fix the svgx refgression and we know a future revision will be [partially] fixed, I am fine resolving this bug.  I will let you guys make that call.
Flags: needinfo?(jmaher)
So is this something that affects 34, but not other versions right now? Can you explain what this affects a bit more clearly?  Thanks!
Flags: needinfo?(jmaher)
right now we have a svg regression which is seen in 34, 35 and 36.  Of course it was introduced July 19th on trunk:
http://graphs.mozilla.org/graph.html#tests=%5B%5B281,52,33%5D,%5B281,1,33%5D,%5B281,53,33%5D%5D&sel=none&displayrange=90&datatype=running (linux32 historical view)

This affects svg rendering performance.  There are some slight adjustments to this over time, but we are looking at a ~7% regression in svg rendering (at max speed, probably much smaller in reality) shipping in v.34 and every version thereafter.  

this affects linux32, linux64 and Windows XP:
http://graphs.mozilla.org/graph.html#tests=%5B%5B281,53,33%5D,%5B281,53,37%5D,%5B281,53,35%5D%5D&sel=1406393480080,1414169480080&displayrange=90&datatype=running

I have no clue how commonplace svg is on the open web, I just know we care about it enough to have 2 talos regression tests for it.
Flags: needinfo?(jmaher)
With tscrollx fixed, I think we can resolve this bug for 34. 

Roc - Do you think the Tsvg regression warrants a follow up bug or should we just live with the regression?
Flags: needinfo?(roc)
I'm planning to live with it. Thanks for tracking this.
Flags: needinfo?(roc)
Status: NEW → RESOLVED
Closed: 10 years ago
status-firefox34: affected → fixed
Resolution: --- → FIXED
You need to log in before you can comment on or make changes to this bug.