Australis weekly update
dolske at mozilla.com
Tue Aug 6 21:47:06 UTC 2013
This got long, so here's a tl;dr:
* We still want to land as soon as perf regressions are fixed.
* Current perf regressions looking good: down to 3%, only on XP.
* A new test reports a 40% regression but actual impact unclear, could
take months to fix.
1) What's the current land/ship plan?
Six weeks ago I talked about our plans to land as soon as possible in
Firefox 25, and but backout upon uplift and actually try to ship in
Firefox 26. Obviously we're not landing in 25 now! But the same basic
plan remains -- land as soon as we can (i.e., when performance
regressions are fixed), and stay on Nightly until we're ready to ship
(i.e., when quality / UX / bugs count / etc are acceptable to ride a
train to release).
2) The good performance news.
Work continues on fixing our ts_paint and tpaint performance
regressions. We just finished reassessing the current graphs for all
platforms, since the performance numbers are getting into the noise...
You can't just compare any two individual data points, but have to look
at trends. (As one example, tpaint on Ubuntu32 consistently fluctuates
by ~20ms: http://cl.ly/image/003S0p0K0o3d)
The good news is that we're looking much better. On almost every
platform -- OSX (10.6, 10.7, 10.8), Windows (7, 8), and Linux (Ubuntu
32bit, 64bit) -- we are in good shape (no tpaint/ts_paint regression
compared to mozilla-central). Windows 8 is even indicating a slight
(~3%) performance improvement on tpaint.
The only regressing platform is Windows XP. We've slowly been grinding
it down, and a number of improvements landed last week (899608, 898126,
891104, 899608) to claw back gains of a few milliseconds each. But an
unrelated mozilla-central regression (ICU, bug 853301) complicated
matters, by making the baseline higher and noisier. Now that's backed
out, and the results are a little easier to see.
On XP, tpaint (http://cl.ly/image/1P1Y1p0K090H) is now only about 4ms
slower (3%). Ts_paint (http://cl.ly/image/1e242Q311H2r) is now only 14ms
slower (3%). You may notice a weird bimodal distribution since 8/1 in
these graphs for both m-c and UX -- this appears to be a Talos bug
(901713, 901715). Also note that we're looking at non-PGO numbers
because of those bugs, but they've historically been comparable so I
don't expect any PGO-only surprises.
3) The bad performance news.
Back in the beginning of the year, we spent a couple of months working
with the performance, gfx, and layout teams on reimplementing Australis
tabs to fix performance issues (see bugs 738491, 813786). We addressed
the concerns, got tabs animating at 60fps even on low-end hardware, and
everyone was happy.
Unfortunately we've now been told that a new test being written
invalidates the perf team's own previous test, and that Australis tabs
are actually a 40% animation regression. (Even though we're still
running 60fps on the same low-end testing hardware.) We've started
discussing what to do about this. It's unclear what the actual end-user
impact is, or how to balance it against a third reimplementation of
Australis tabs likely blowing out the schedule by another few months. :(
See bug 902024.
More information about the firefox-dev