1805255 - Firefox has 50% more soft page faults than Chrome running Speedometer 2

Status: NEW

(Core :: Performance: General, defect)

Description

[Alex Thayer [:alexical] (she/her)]

When running Speedometer 2 on Windows, I see 30k page faults / sec for a release build of Firefox, and only 20k page faults from Chrome. We should diagnose what the underlying difference is here and whether there is an opportunity to improve this situation. (To be clear, almost none of these faults were hard page faults.)

Comment 1

[Jeff Muizelaar [:jrmuizel]]

Running xperf -on ALL_FAULTS -stackwalk PagefaultDemandZero shows that most of the faults are DemandZeroFaults. This is what I expected.

Comment 2

[Jeff Muizelaar [:jrmuizel]]

Here's a quick profile using xperf -on latency+ALL_FAULTS -stackwalk PagefaultDemandZero I threw together using that should roughly show where the faults are coming from:
https://share.firefox.dev/3FfEEIs

Comment 3

[Jeff Muizelaar [:jrmuizel]]

It would be good if this profile also included markers for VirtualAlloc/VirtualFree

Comment 4

[Jeff Muizelaar [:jrmuizel]]

xperf -on latency+ALL_FAULTS+VIRT_ALLOC -stackwalk PagefaultDemandZero gives us events for VirtualAlloc/VirtualFree.

When comparing Firefox vs Chrome in WPA we definitely seem to have a higher rate of VirtualAlloc/VirtualFree when running React-TodoMVC. It also seems like Chrome hits a stable roughly stable "Outstanding commit" better than we do.

Comment 5

[Greg Mierzwinski [:sparky]]

:dthayer, we set the Performance Impact to ?, could you set it more appropriately please (if needed)?

Comment 6

[Olli Pettay [:smaug][[e-mail]]]

One possible reason is that our page caches are tiny
https://bugzilla.mozilla.org/show_bug.cgi?id=1805644#c16