1649657 - Unbounded GPU process memory growth over long-running sessions

Status: RESOLVED WORKSFORME

(Core :: Graphics, defect, P3)

Description

[Ryan VanderMeulen [:RyanVM]]

Attachment: about:memory report

I've been noticing lately that by the end of the day, my GPU process is using large amounts of memory (currently over 4GB) and my browser process is bouncing between 20-30% CPU usage constantly.

When looking about about:memory, I can see that the resident number reported is in line with what I'm seeing elsewhere, but nothing in the list above it comes close to adding up to such a large number.

Comment 1

[Ryan VanderMeulen [:RyanVM]]

Attachment: about:support info

Comment 2

[Ryan VanderMeulen [:RyanVM]]

I captured a profile and it appears to be showing a steady uptick in GPU process memory usage every 2 seconds or so.
https://share.firefox.dev/2VAO8bl

Comment 3

[Ryan VanderMeulen [:RyanVM]]

On a fresh browser launch, my GPU process is using around 150MB. I'll keep an eye on it to see how it grows throughout the day.

Comment 4

[Ryan VanderMeulen [:RyanVM]]

Attachment: DMD report

I'm over 3GB again today at the moment. I grabbed a DMD report but nothing's obviously standing out there that I can see.

Comment 5

[Jeff Muizelaar [:jrmuizel]]

I think we'll need to switch to ETW to try get lower level info. Let's try getting VirtualAlloc info https://docs.microsoft.com/en-us/windows-hardware/test/wpt/memory-footprint-optimization-exercise-2. You can likely ignore the heap allocation stuff.

Comment 6

[Ryan VanderMeulen [:RyanVM]]

Memory usage appears to be much more stable with gfx.webrender.compositor set to false.

Comment 7

[Ryan VanderMeulen [:RyanVM]]

After a few days of running with gfx.webrender.compositor set to false, I can say it's definitely improving things (albeit with the occasional rendering glitch). Also, I'm updating the summary because the CPU usage part was a red herring caused by about:processes itself using tons of CPU.

Comment 8

[Ryan VanderMeulen [:RyanVM]]

Haven't seen this in a long time. Probably fixed by bug 1640564.