1495936 - Higher startup private bytes and VRAM usage under WebRender

Status: RESOLVED FIXED

(Core :: Graphics: WebRender, enhancement, P2)

Description

[Bobby Holley (:bholley)]

I did some testing today around memory usage in a clean startup of Firefox with and without WebRender. I loaded the browser and then opened a single page [1].

The private working sets, which are what the task manager reports, are about the same summed over all the processes. However, private bytes are around 150-200MB higher.

The working set size is more important, since that represents usage of physical RAM, and also corresponds to what users see in the task manager. However, private bytes count against the commit limit, which is RAM+swap. Eric Rahm says that swap on windows defaults to 1.5x ram size, so a machine with 2GB of ram would have a system-wide limit of 5GB. An extra 200MB there is not great.

The net value of "GPU committed bytes" is also 140MB higher, and "GPU dedicated bytes" is 100MB higher (I haven't worked out what the difference is). I suspect the private bytes and VRAM usage are related, so I'm filing them together. We can split out later if need be. 

[1] https://docs.microsoft.com/en-us/sysinternals/downloads/process-explorer

Comment 1

[Bobby Holley (:bholley)]

Attachment: memory usage screenshot

Attaching a screenshot with all the stats in one place.

PID 6764 is the WR GPU process. PID 5676 is the Fx GPU process.

Comment 2

[Bobby Holley (:bholley)]

Attachment: memory usage screenshot (intel reference laptop)

Here's a screenshot while performing a similar workload on the quantum reference laptop. On that device, the higher pbytes _do_ translate into higher working set. This may be a difference with the driver, or with overall system resources.

We don't need to worry about this one for the MVP (it's not NVIDIA), but we should keep it in mind, so I'm posting it here.

Comment 3

[Bobby Holley (:bholley)]

Attachment: GPU Process VirtualAlloc stacks, WebRender

Using WPR, I recorded all the large VirtualAlloc stacks in the GPU process during startup. This one is for WR.

Comment 4

[Bobby Holley (:bholley)]

Attachment: GPU Process VirtualAlloc stacks, Non-WebRender

Here are the non-WR stacks for comparison. It's not really a fair comparison because the GPU heavy lifting happens in the content process for non-WR.

Comment 5

[Bobby Holley (:bholley)]

I'll try to get back to this tomorrow, and investigate the largest stacks from comment 3.

Comment 6

[Bobby Holley (:bholley)]

Attachment: updated memory usage screenshot.png

Just re-did these measurements, and WR is now significantly better - 348.3 rather than 433.6, or 20% better. \o/