1643805 - raptor-tp6-netflix-firefox load time is slower with Fission than non-Fission (10% on Linux and 60% on Windows)

Status: VERIFIED WORKSFORME

(Core :: DOM: Content Processes, defect, P2)

Description

[Chris Peterson [:cpeterson]]

Attachment: Screenshot_2020-06-05 Treeherder.png

Link to Perfherder results:

https://treeherder.mozilla.org/perf.html#/graphs?highlightAlerts=1&series=mozilla-central,2072626,1,10&series=mozilla-central,2237604,1,10&series=mozilla-central,2072660,1,10&series=mozilla-central,2237458,1,10&timerange=7776000

In contrast to bug 1643802, why is Windows so much slower than Linux?

See the attached Perfherder screenshot for posterity.

Comment 1

[Chris Peterson [:cpeterson]]

@ Jesup, looks like bug 1654201 ("Record netflix with mitm5") on August 8 that greatly improved netflix.com page load perf for both Fission and non-Fission.

Can we close this bug as WORKSFORME?

The Treeherder URL changed to:

https://treeherder.mozilla.org/perf.html#/graphs?highlightAlerts=1&series=mozilla-central,2136344,1,10&series=mozilla-central,2529259,1,10&series=mozilla-central,2565348,1,10&timerange=2592000

Comment 2

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

(In reply to Chris Peterson [:cpeterson] from comment #1)

@ Jesup, looks like bug 1654201 ("Record netflix with mitm5") on August 8 that greatly improved netflix.com page load perf for both Fission and non-Fission.

Can we close this bug as WORKSFORME?

And then it got bad again, and stayed that way; around a 40-50% difference again.

This is likely a timing issue, since we don't see large differences in visualmetrics. Something is being preloaded (but not visible) and depending on luck it either ends up before onload (and delays it a lot), or after onload.

Being a timing issue, it can vary by platform, and even by randomness (like where we were tied for a few days, then regressed again)

Comment 3

[Chris Peterson [:cpeterson]]

M7

Looks like Fission improved again, though it's still about 20% slower than non-Fission.

Comment 4

[Neha Kochar [:neha]]

We'll monitor visual metrics and not pageload so closing this. We'll file separate bugs for visual metrics Fission regressions.

Comment 5

[Chris Peterson [:cpeterson]]

Changing resolution to WORKSFORME because it looks like the regression went away in early December, possibly related to infra changes:

https://treeherder.mozilla.org/perfherder/graphs?highlightAlerts=1&highlightChangelogData=1&series=mozilla-central,2136344,1,10&series=mozilla-central,2529259,1,10&series=mozilla-central,2565348,1,10&timerange=7776000