1857828 - Use an RWLock for atom tables.

Status: RESOLVED FIXED

(Core :: XPCOM, task)

Description

[Emilio Cobos Álvarez (:emilio)]

Seems like a consistent progression in CSS-heavy speedometer tests (win64 example).

Comment 1

[Emilio Cobos Álvarez (:emilio)]

Attachment: Bug 1857828 - Use an RWLock for atom tables. r=smaug,#xpcom-reviewers

See comment 0. This should avoid some lock contention, specially when
parsing CSS and other off-main-thread uses of atoms.

Comment 2

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/70c980c054e9
Use an RWLock for atom tables. r=smaug

Comment 3

[Bobby Holley (:bholley)]

If we're seeing lock contention, we might also want to consider bumping the number of subtables. The idea is that we try to have basically no contention for any two non-identical atoms (though of course if both threads are using the same atom we can't solve that with a subtable).

Any sense of which case it is?

Comment 4

[Narcis Beleuzu [:NarcisB]]

https://hg.mozilla.org/mozilla-central/rev/70c980c054e9

Comment 5

[Emilio Cobos Álvarez (:emilio)]

I didn't dig too much, but my intuition was that it seems reasonable (and likely) that CSS being parsed off the main thread contains the same atoms that the main thread uses and also atomizes (think class names ids etc).

Comment 6

[Bobby Holley (:bholley)]

Maybe just briefly test if 2x-ing or 10x-ing the number of buckets has any impact beyond what we got with this change?

Comment 7

[Emilio Cobos Álvarez (:emilio)]

Twice as many: https://treeherder.mozilla.org/perfherder/compare?originalProject=try&originalRevision=120739cce5873b60be4c60065082bc0ffe1ebe3b&newProject=try&newRevision=09a9fee90d6c9fa4a3128a0e564884d256c026d5&page=1

Just out of curiosity, some less tables: https://treeherder.mozilla.org/perfherder/compare?originalProject=try&originalRevision=120739cce5873b60be4c60065082bc0ffe1ebe3b&newProject=try&newRevision=c77d1d98561e4e82d56212623272a8208b12ee64