1997702 - Testcase creating long chains of microtasks is 8x slower with the new microtaskQueue. Spends all the time in MinorGC tracing js::MicroTaskQueueSet::trace and TraceableFifo::trace

Status: RESOLVED FIXED

(Core :: JavaScript Engine, task, P3)

Description

[Mayank Bansal]

Attachment: Microtask stress.HTML

enable the new js-managed-microtask
Open attached testcase
click run

default microtask: https://share.firefox.dev/4hLOHIo (9s)
new microtask: https://share.firefox.dev/4oMt1xS (72s)

Testcase created using chatgpt. 100% artificial.
Deliberately not making this bug bug 1990841 as this is a stress-test.

Comment 1

[Mayank Bansal]

Attachment: chained promises.HTML

Perhaps an equivalent and simpler testcase.
N=5000000
Profile: https://share.firefox.dev/4ol5PXT (110s)
Chrome: https://share.firefox.dev/48RAlUm (1.5s)

Comment 2

[Jon Coppeard (:jonco)]

This is probably because of the PersistentRooted collections in JSContext used for the microtask queue. These are always a risk for causing quadratic behaviour. To fix this we could change the microtask queues to use TracebleFifo<HeapPtr<Value>> (to put nursery values into the store buffer) and trace the queues in JSContext::trace, but only if we're not doing a minor GC.

Comment 3

[Mayank Bansal]

this testcase is kinda silly. Feel free to wontfix.

Comment 4

[Matthew Gaudet (he/him) [:mgaudet]]

Attachment: Bug 1997702 - Remove use of PersistentRooted on JSContext for MicroTaskQueues r?jonco

Comment 5

[Pulsebot]

Pushed by [e-mail]:
https://github.com/mozilla-firefox/firefox/commit/c858cd0fdd61
https://hg.mozilla.org/integration/autoland/rev/2c4ef91d4c59
Remove use of PersistentRooted on JSContext for MicroTaskQueues r=jonco

Comment 6

[Cosmin Sabou [:CosminS]]

https://hg.mozilla.org/mozilla-central/rev/2c4ef91d4c59

Comment 7

[Mayank Bansal]

Testcase1: https://share.firefox.dev/491RV8e (9.5s + 0.5s cleanup)
Testcase2: https://share.firefox.dev/4oW0ead (13s)

Comment 8

[Matthew Gaudet (he/him) [:mgaudet]]

We're going to back the solution to this out (See Bug 2002849), and re-open this bug. We'd like to get to a better solution, but for now the performance impact of this on JS3 was too high.

Oh, right, and I'll clone this bug rather than re-open it.