1898461 - Try tweaking our pretenuring parameters a little

Status: RESOLVED FIXED

(Core :: JavaScript: GC, enhancement)

Description

[Jon Coppeard (:jonco)]

A couple of things I noticed about our pretenuring parameters:

• our attention threshold is much higher at 500 than the equivalent in v8, which is 100
• we limit the number of allocation sites created per minor GC to 500 and speedometer3 hits this limit

Comment 1

[Jon Coppeard (:jonco)]

Attachment: Bug 1898461 - Tweak our pretenuring parameters a little r?jandem

This reduces the attention threshold to 200 and increases the limit on the
number of allocation sites per minor GC to 600.

This shows some large high confidence improvements on jetstream2 splay subtests
and some low confidence speedometer3 subtest improvements on linux.

Comment 2

[Jon Coppeard (:jonco)]

Benchmark comparisons are here: https://treeherder.mozilla.org/perfherder/compare?originalProject=try&originalRevision=244fc1c6647794d640b96bc9cddcb814bd41cb7b&newProject=try&newRevision=68a99e3bec6ebe977eb54b657e0551449ba3a6c5&framework=13&page=1

Comment 3

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/f6278961ec15
Tweak our pretenuring parameters a little r=jandem

Comment 4

[tszentpeteri]

https://hg.mozilla.org/mozilla-central/rev/f6278961ec15

Comment 5

[Mayank Bansal]

6% regression on AWFY-Jetstream2-acorn-wtb-Average
5.1% regression on babylon-wtb-Average
5% improvement on prepack-wtb-Average
incremental improvement on splay

Comment 6

[Jon Coppeard (:jonco)]

(In reply to Mayank Bansal from comment #5)

incremental improvement on splay

The graph linked reports 12.4% improvement on splay, but looking at recent history it could be greater than that if you take the average.