Open Bug 1252768 Opened 4 years ago Updated 1 year ago

2–7% e10s CART regression (with APZ disabled)

Categories

(Firefox :: General, defect, P3)

defect

Tracking

()

Tracking Status
e10s + ---
firefox46 --- ?
firefox47 --- affected

People

(Reporter: cpeterson, Unassigned)

References

(Blocks 1 open bug)

Details

(Keywords: perf)

@ mconley: do you own the CART test?

https://treeherder.allizom.org/perf.html#/e10s?filter=cart&showOnlyBlockers=1

cart summary 	non-e10s 		e10s 	Delta
linux64 pgo     20.71 ± 0.54% 	< 	23.51 ± 0.36% 	13.56%
osx-10-10 opt   35.01 ± 1.06% 	< 	37.28 ± 1.52% 	6.48%
windows7-32 pgo 23.62 ± 0.81% 	< 	24.99 ± 0.63% 	5.82%
windows8-64 pgo 22.39 ± 0.42% 	< 	23.47 ± 1.04% 	4.85%
Flags: needinfo?(mconley)
Blocks: 1251547
(In reply to Chris Peterson [:cpeterson] from comment #0)
> @ mconley: do you own the CART test?
> 
> https://treeherder.allizom.org/perf.html#/e10s?filter=cart&showOnlyBlockers=1
> 
> cart summary 	non-e10s 		e10s 	Delta
> linux64 pgo     20.71 ± 0.54% 	< 	23.51 ± 0.36% 	13.56%
> osx-10-10 opt   35.01 ± 1.06% 	< 	37.28 ± 1.52% 	6.48%
> windows7-32 pgo 23.62 ± 0.81% 	< 	24.99 ± 0.63% 	5.82%
> windows8-64 pgo 22.39 ± 0.42% 	< 	23.47 ± 1.04% 	4.85%

Currently, no. I'm assigned to bug 1251547, which covers TART, tpaint, tresize and tps.

According to https://wiki.mozilla.org/Buildbot/Talos/Tests#TART.2FCART, the points of contact for that test are avih, jmaher, and MattN.
Flags: needinfo?(mconley)
(In reply to Mike Conley (:mconley) - Needinfo me! from comment #1)
> (In reply to Chris Peterson [:cpeterson] from comment #0)
> > @ mconley: do you own the CART test?
> > 
> > https://treeherder.allizom.org/perf.html#/e10s?filter=cart&showOnlyBlockers=1
> > 
> > cart summary 	non-e10s 		e10s 	Delta
> > linux64 pgo     20.71 ± 0.54% 	< 	23.51 ± 0.36% 	13.56%
> > osx-10-10 opt   35.01 ± 1.06% 	< 	37.28 ± 1.52% 	6.48%
> > windows7-32 pgo 23.62 ± 0.81% 	< 	24.99 ± 0.63% 	5.82%
> > windows8-64 pgo 22.39 ± 0.42% 	< 	23.47 ± 1.04% 	4.85%
> 
> Currently, no. I'm assigned to bug 1251547, which covers TART, tpaint,
> tresize and tps.
> 

I'm basing that off of https://wiki.mozilla.org/Electrolysis/Release_Criteria/UI_Smoothness. If we need to add CART to that list, well, I guess I own it too then.
This might go away with APZ disabled. Needinfo'ing myself to confirm.
Flags: needinfo?(mconley)
Here's some data with APZ disabled:

https://treeherder.mozilla.org/perf.html#/compare?originalProject=mozilla-central&originalRevision=a4929411c0aa&newProject=try&newRevision=850e39c8f668&framework=1&filter=cart&showOnlyImportant=0

Conclusions:

1) APZ causes a ~2% - ~5% CART regression with e10s enabled. This takes care of a good chunk of the regression across the board.

2) Doing some napkin math, I can re-write the e10s-only regressions (without APZ enabled) as:

           non-e10s (*)     e10s           Diff (%)
-----------------------------------------------------
Linux64:    26.32          28.10            +6.76%
OS X 10:    35.01          36.30            +3.68%
Win 7  :    33.91          35.28            +4.04%
Win 8  :    28.09          28.66            +2.02%

All but Linux64 are below the 5% threshold.

So the next steps for this bug:

1) File a new bug for the APZ releated CART regressions and make sure the APZ team sees it.
2) See if I we can bring the Linux64 regression below 5%, or try to argue that 6.76% isn't enough to fight for. I actually have a theory that the cost is us shutting down and starting up content processes multiple times (since about:customizing isn't remote, and it takes over the pre-existing remote about:blank tab, it causes the content process to repeatedly shutdown and restart for each test run, and that's where the cost is going).

(*) I chose the worst case, since I have two sets of samples here given that disabling APZ has no effect on non-e10s builds.
Flags: needinfo?(mconley)
Summary: 5–14% e10s CART regression → 2–6% e10s CART regression (with APZ disabled)
To be fair, 6.76 is close to 7 than it is to 6.
Summary: 2–6% e10s CART regression (with APZ disabled) → 2–7% e10s CART regression (with APZ disabled)
See Also: → 1253380
Filed bug 1253380 for the regression attributed to APZ.
Priority: -- → P2
Moving to p3 because no activity for at least 1 year(s).
See https://github.com/mozilla/bug-handling/blob/master/policy/triage-bugzilla.md#how-do-you-triage for more information
Priority: P2 → P3
You need to log in before you can comment on or make changes to this bug.