1689702 - Fission visual metrics regression on Mac for Netflix

Status: RESOLVED FIXED

(Core :: Performance, defect, P2)

Description

[Randell Jesup [:jesup] (needinfo me)]

You can see a ~15-30% visualmetric regression for Netflix on MacOS here: https://vismet.arewefastyet.com/mac/cold-page-load/overview?numDays=60

Comment 1

[Neha Kochar [:neha]]

Let's target fixing this in M7.

Comment 2

[Randell Jesup [:jesup] (needinfo me)]

From a run against a local recording (no network delay), median SI of 15 runs. median SI regression was ~17%. Note that on CSI it was a median 5% gain, and PSI was a 15% regression. Netflix has a moderate amount of noise even against a network recording.

e10s: https://share.firefox.dev/3qMWqub
fission: https://share.firefox.dev/3qRAyOE

Comment 3

[Randell Jesup [:jesup] (needinfo me)]

Attachment: 004-www.netflix.com_title_80117263.mp4

Comment 5

[Randell Jesup [:jesup] (needinfo me)]

For the Live test, here are matched-up profiles showing the difference (median SI of 15 runs, median regression: 21.6% (SI))
e10s: https://share.firefox.dev/37JzpAS
fission: https://share.firefox.dev/3aJpnlf

Note two things: one, the network request was slower for fission, both in connection establishment and transfer. Two, even factoring out the slower network, it still takes considerably longer to paint after the initial load is completed. There's no direct reason I can think of for the network to be slower in fission, unless it's fighting for CPU resources perhaps or something has a lock somewhere (due to process creation???)

Comment 6

[Randell Jesup [:jesup] (needinfo me)]

Interestingly the order their scripts run is is quite different; the cadium-playercore runs first for e10s (and with all the other scripts); it runs last after a gap for fission

Comment 7

[Randell Jesup [:jesup] (needinfo me)]

In both cases, it was loaded at roughly the same point; it didn't come in late

Comment 8

[Olli Pettay [:smaug][[e-mail]]]

This is tracked as part of Fission, so I don't think tracking this also as a qf bug is particularly useful.

Comment 9

[Randell Jesup [:jesup] (needinfo me)]

Regression is down to 1.85% with NNT and preallocation delay(1000ms).