Add Runtime_events.EV_FUTEX_WAIT and EV_EMPTY_MINOR #13407
Conversation
| | EV_OS_WAIT | ||
| (** | ||
| Event spanning calls to the OS to wait for another domain. | ||
| @since 5.3 |
There was a problem hiding this comment.
This should be 5.4, as 5.3 has already branched.
runtime/platform.c
Outdated
| @@ -285,7 286,9 @@ void caml_plat_futex_free(caml_plat_futex* ftx) { | |||
| void caml_plat_futex_wait(caml_plat_futex* ftx, | |||
| caml_plat_futex_value undesired) { | |||
| while (atomic_load_acquire(&ftx->value) == undesired) { | |||
| CAML_EV_BEGIN(EV_OS_WAIT); | |||
There was a problem hiding this comment.
OS_WAIT seems too general for what's happening here. For example, waiting on select or accept call would also be "waiting on the OS". Would FUTEX_WAIT be more appropriate?
EV_FUTEX_WAIT is useful to see in traces when OCaml asked the OS to suspend a domain. If the domain is waiting for another domain on the same CPU, this is the point where it is likely to be scheduled. EV_EMPTY_MINOR shows when a domain is trying to empty its minor heap. It may be a long time between starting this process and actually performing a minor GC if, for example, another domain is holding the platform lock. Without this event, profiling tools tend to under-report the amount of time spent on GC.
talex5
changed the title
|
The "MSVC 32 bits" test |
|
Thanks @talex5 , these look good to have. The only thing I'm worried about is the FUTEX_WAIT event being inside the loop - if for some reason it spins frequently it will cause lots of other events in the ring to be dropped. Does it make sense to move it outside of the while loop? |
I don't mind either way. My thinking was that if it keeps being woken up unnecessarily then that's interesting to see. |
|
Also, I guess putting it outside the loop would need some extra logic to avoid tracing the fast path where we never suspend. |
EV_OS_WAITis useful to see in traces when OCaml asked the OS to suspend a domain. If the domain is waiting for another domain on the same CPU, this is the point where it is likely to be scheduled.EV_EMPTY_MINORshows when a domain is trying to empty its minor heap. It may be a long time between starting this process and actually performing a minor GC if, for example, another domain is holding the platform lock. Without this event, profiling tools tend to under-report the amount of time spent on GC./cc @sadiqj @kayceesrk (https://discuss.ocaml.org/t/ocaml-5-performance/15014/19?u=talex5)
Example trace showing the new events: