dolske at mozilla.com
Wed Nov 13 00:22:51 UTC 2013
It's been a while since my last Australis update. We've been busy
slogging through our performance blockers, and now the end appears to be
in sight. I'd like to give a specific landing date, but the last few
days have been filled with some delightful last-minute surprises that
we're still working though. But we've also had questions about what's
up, so let's talk about where we currently stand.
As of a week ago, the three performance regressions we've been tracking
have all been cleared. The tpaint regression (bug 889758) was ground to
zero by 8/22, the ts_paint (bug 880611) regression was effectively
ground to zero by 10/16, and the TART regression (bug 902024 and deps)
-- the main focus of recent work -- was cleared on 11/4. That put us
into a prepare-for-landing mode. Checking in with the other involved
parties (Release, QA, PR, Marketing, etc.) revealed no other outstanding
issues, and Jared has begin work on spinning up our "backout branch" --
note here, in case you're wondering, the landing and post-landing plans
for Australis are the same as before. I'll post a reminder/summary when
we're officially ready to go.)
We then made one last pass through all the performance test data, just
to make sure we understood all the results. And that's made the last few
days interesting. Of the 3 new issues that have popped up, the first 2
are now resolved...
1) Bug 936469 ("Investigate TART ".error" regressions on UX branch on
OSX 10.8"). TART results can be tricky to interpret. It's comprised of
many subtests, varying in importance, which can (and do) show both
improvements and regressions. OS X, in particular, is complicated
because TART and OMTC don't play well together. We thought the perf team
was ok with the results, but upon further review this bug was filed to
look at things again. And then upon further further-review it was
cleared as being OK.
2) tsvgx/tscrollx regression (no bug). These two tests are purely
checking content-area performance, which Australis shouldn't impact. But
after looking at the graph data is was a clear but tiny regression. For
example, on XP (http://is.gd/Vq0xt9) Australis is consistently 1.5%
slower (50 microseconds). The cause is delightful: there's a little bit
more content area -- same window size, but Australis chrome is smaller.
So there are an extra 16 rows of pixels to scroll, and amazingly this
test manages to detect that. (We verified this by a couple different
means of making the content area the original height, and the regression
goes away.) Scrolling the bigger area takes an extra 50 microseconds
(3.534 milliseconds instead of 3.434ms). Ergo, expected. No stone left
unturned! Aside: the accuracy and precision of this test is surprising
-- compare it to the static in other tests like http://is.gd/un2Iwv,
which randomly varies by many orders of magnitude.
3) Bug 937519 ("UX (Australis) branch Windows 7/8 Tab Animation
Regression Test (TART) regression on 2013-11-07"). Last night MattN
found a merge from mozilla-central that introduced a large TART
regression on UX (~40%) but not on m-c. We know the layout bug that
caused it (934860), and we (frontend, layout, gfx) are currently
discussing what to do about it... A backout is possible, but it's not
yet settled because it's also seen as a correctness fix to a major
text-rendering bug (812695).
So, that's where we are. As soon as we know the plan for the last item
(bug 937519), we should be able to figure out when we're clear to land.
If it's simple to resolve, it could be as soon as next week. If not, it
could take longer.
More information about the firefox-dev