1151836 - WholeCellEdges store buffer can slow down minor GCs a lot

Status: NEW

(Core :: JavaScript: GC, defect)

Description

[Jan de Mooij [:jandem]]

Attachment: Shell testcase

Luke told me about a webgl demo where we are a lot slower than Chrome and IE. A big part of the problem is that it hits a store buffer worst case scenario.

The attached shell micro-benchmark demonstrates the problem:

d8:           260 ms
js --no-ggc:  480 ms
js:          1630 ms

We create a huge number of WholeCellEdges store buffer entries, and have to mark a ton of objects on every minor GC.

Comment 1

[Terrence Cole [:terrence]]

I'm currently working on speeding up the marking we do for store buffers. Unfortunately, depending on the object sizes we may not be able to get down to d8 levels until we can do page marking.

Comment 2

[Mayank Bansal]

We are 4x-5x slower on this testcase.

Nightly : https://share.firefox.dev/4fOkwib (420ms-450ms without the profiler)
Chrome: https://share.firefox.dev/3WRNncT (80ms-100ms without the profiler)

Edit: and a profile with 10x the loop count : https://share.firefox.dev/3Xhmh0l