Per-thread cpu time #181
Per-thread cpu time #181
Conversation
github-actions
bot
added
port-to-v1.10
d-netto
force-pushed
the
dcn-user-cpu-time
branch
3 times, most recently
from
4e1bf4c
to
296eeb4
Compare
d-netto
force-pushed
the
dcn-user-cpu-time
branch
from
296eeb4
to
898fffb
Compare
| JL_DLLEXPORT jl_task_t *jl_task_get_next(jl_value_t *trypoptask, jl_value_t *q, jl_value_t *checkempty) | ||
| { | ||
| uint64_t t0 = jl_record_time_for_tls_metric(); | ||
| jl_task_t *task = _jl_task_get_next(trypoptask, q, checkempty); | ||
| jl_increment_timing_tls_metric(jl_current_task->ptls, scheduler_time, jl_record_time_for_tls_metric() - t0); | ||
| return task; | ||
| } |
There was a problem hiding this comment.
I think this got lost in the handoff, but the original plan was to not record the current time twice for the "fast-path" of task switching.
The concern is that adding two syscalls can meaningfully slow down task switching, and the idea was that we probably don't need to do this anyway since the time we'd be measuring in the fast-path case is trivial (we believe).
So ideally I think we'd want to only do this in the slow paths inside jl_task_get_next?
BUT to be honest, I find your approach much cleaner and clearer - and it's nice that we would get an accurate measure of that switch time which is supposed to be cheap......... AND we are going to need something like this anyway for the per-task CPU time in https://relationalai.atlassian.net/browse/RAI-30482...
So maybe if we could solve the vDSO / fast instruction problem, then we could do it this way after all?
Something to resolve CC @kpamnany.
d-netto
force-pushed
the
dcn-user-cpu-time
branch
5 times, most recently
from
abe0d2e
to
d7c0b4c
Compare
d-netto
force-pushed
the
dcn-user-cpu-time
branch
from
d7c0b4c
to
1cbe632
Compare
PR Description
Expands on #175 by:
Checklist
Requirements for merging:
port-to-*labels that don't apply.