runtime: non-spinning Ms spin uselessly when work exists #43997
Comments
CC @dvyukov |
Change https://golang.org/cl/288232 mentions this issue: |
For reference, here's an application reproducer similar to the one with which I diagnosed this issue (though quite a bit noisier):
https://golang.org/cl/288232 is a basic prototype that takes aim at the first sub-optimal bullet above by skipping the recheck entirely for non-spinning Ms, allowing them to stop. Before and after that CL, I get:
This particular benchmark is unfortunately much noisier than my internal benchmark, but show promise (CpuTime samples are the sum of SystemTime and UserTime samples, so UserTime shouldn't be able to drop without both SystemTime and CpuTime flat). For reference, my internal workload for the same change:
|
Interesting find! It's indeed very old and was present in the scheduler since the rewrite. I guess I added this work conservation check during late fine-tuning stage and failed to properly integrate it with the overall logic. I agree the current spinning behavior is bad and https://golang.org/cl/288232 looks like reasonable to me.
I think it should work. The invariant scheduler maintains is that there should be at least 1 spinning M if there is any work, and that 1 spinning M will wake up more if necessary. Since spinning Ms already checked workqueues of all Ps during steal phase, it looks reasonable to not re-do it again for all but last spinning M. That "delicate dance" re-checking is there only to prevent bad edge cases of not realized parallelism, it's not intended to be part of the performance story. |
For what's it's worth, I also observed this behavior of the scheduler while working on the fix for #38860 and thought it was curious, but didn't worry about it since it was not the focus of my work at the time. That said, as the submitter of #28808 I welcome effort to improve the scheduler in this area. |
Background: The scheduler differentiates "spinning" and "non-spinning" Ms. We will allow up to
GOMAXPROCS/2
Ms to enter the spinning state (assuming all Ps are busy).Spinning Ms are eligible to steal runnable goroutines or timers from other Ps and will loop several times over all Ps attempting to steal before giving up and stopping.
For work conservation, the scheduler ensures that there is at least one spinning thread after readying a goroutine. To support this, spinning Ms adhere to a contract: any work submitted when sched.nmspinning > 0 is guaranteed to be discovered.
To maintain this invariant, spinning Ms must:
This is all good for spinning Ms, but there is a bad case for non-spinning Ms.
If there are already enough spinning Ms, the remaining Ms will skip the main steal loop and go “directly” to stop. However, after the stop label is the steal recheck primarily intended for spinning -> non-spinning transitions, but executed by all Ms.
A non-spinning M can detect work in this loop and jump back to the top to try to steal it (we don’t steal directly in this loop to avoid duplication of the somewhat complex stealing logic). Unfortunately this is a non-spinning M, so assuming other spinning Ms don’t go away we will remain non-spinning and skip right over the steal loop back to stop, at which point we’ll notice the same work again, jump to the top and start all over again.
Note that this can only occur while there is at least one spinning M, which means that this behavior is bounded: either the spinning M will take the work for itself or will take some other work and the non-spinning M can transition to spinning and take work for itself.
I argue that the current behavior is nearly always sub-optimal:
sched.nmspinning
total available units of work, then the spinning Ms will take all of the work and the non-spinning one will have to ultimately sleep anyways.sched.nmspinning
total available units of work, then the non-spinning M should eventually get some work. However, the current approach needlessly delays stealing until the existing spinning Ms have found work. I think it would be better if the non-spinning Ms could upgrade themselves to spinning (in excess of the normal GOMAXPROCS/2 limit) if they notice that there is extra work to do, as this would reduce scheduling latency without increasing work (remember that we are basically in a busy-loop anyways).I discovered this behavior on a Google-internal workload which is basically the worst case scenario for this: one M/P/G is running continuously generating very short-lived work that must be stolen by other Ms. The other Ms spend most of their time in the scheduler looking for work. A simple change to make non-spinning Ms skip the steal recheck reduces application wall time by 12% and CPU time by 23%. I hope to create a similar open source reproducer soon.
Note that this behavior has existed since the introduction of spinning Ms, so this is not a new regression.
Something that is not directly related, but perhaps worth investigating is whether we should even do all of the “delicate dance” checks on all spinning Ms, as it should not affect correctness to limit it to only the 1->0 nmspinning transition, which would further reduce usage.
cc @aclements @mknyszek
See also #18237, #28808
The text was updated successfully, but these errors were encountered: