1323078 - PHASE_SWEEP_MISC should be broken down more

Status: RESOLVED WORKSFORME

(Core :: JavaScript: GC, defect, P3)

Description

[Bill McCloskey [inactive unless it's an emergency] (:billm)]

This phase seems to cover a pretty broad variety of things. It's showing up a lot in telemetry, so we should break it down into smaller buckets.

Steve, one thing I don't understand is how the parallel accounting works. Do we account for time spent off the main thread? If we do, it seems like the numbers won't add up as one would expect. Or do we only add up the time spent waiting to join the worker threads?

Comment 1

[Steve Fink [:sfink] [:s:]]

(In reply to Bill McCloskey (:billm) from comment #0)
> This phase seems to cover a pretty broad variety of things. It's showing up
> a lot in telemetry, so we should break it down into smaller buckets.
> 
> Steve, one thing I don't understand is how the parallel accounting works. Do
> we account for time spent off the main thread? If we do, it seems like the
> numbers won't add up as one would expect. Or do we only add up the time
> spent waiting to join the worker threads?

You are correct, it's utterly bogus. I tried to fix it in bug 1306249, which ended up getting backed out because the numbers were intermittently not adding up on XP. (And I can't reproduce on try, despite 30 or so retriggers.) I'm intending to try landing that again with additional diagnostics.

Specifically, you are correct -- sometimes we're calculating elapsed mainthread time. Sometimes we add in the time spent in worker threads. Which means the units are different for different buckets. Which is crazy.