PA-E for macOS status update

14 views

Skip to first unread message

Yuki Shiino

unread,

Nov 30, 2021, 7:17:16 AM11/30/21

to memor...@chromium.org, li...@chromium.org, bar...@chromium.org, Kentaro Hara

Hi team,

We're glad to report that we enabled PartitionAlloc-Everywhere for macOS on Nov 19 (CL; 2nd time) and it has stuck this time. (A little more details: We enabled PA-E on Nov 10 for the first time (CL) but it got reverted due to several test failures.)

The performance of PA-E looks very good so far.

On bots, total private memory footprint improves by up to ~20% for single-tab stories (link):

Performance regressed at first (by up to 14% on Jetstream 2.0), but recovered (and is perhaps even a bit better than before) when the spinning mutex relanded (vs spinlock: this also shows that spin locks are really bad). See below on M1 (link):

In the field, Chrome Canary shows private memory footprint improvements of ~20% for the browser process, 6-8% (noisy) for the GPU process, and ~15% for renderer processes, across quantiles (source: UMA, gradual transition due to 7 day aggregation):

We have found a few issues that we are addressing:

Initialization is racy, causing very early crashes: these are invisible to users except for the browser process. There is a fix in review.
Some WebGPU ANGLE tests are failing with an AMD GPU, with an invalid memory access inside the GPU driver: we are investigating, but it does not seem to impact users in the wild.