runtime: account for idle mark time in the GC CPU limiter
Currently the GC CPU limiter doesn't account for idle application time
at all. This means that the GC could start thrashing, for example if the
live heap exceeds the max heap set by the memory limit, but the limiter
will fail to kick in when there's a lot of available idle time. User
goroutines will still be assisting at a really high rate because of
assist pacing rules, but the GC CPU limiter will fail to kick in because
the actual fraction of GC CPU time will be low if there's a lot of
otherwise idle time (for example, on an overprovisioned system).
Luckily, that idle time is usually eaten up entirely by idle mark
workers, at least during the GC cycle. And in these cases where we're
GCing continuously, that's all of our idle time. So we can take idle
mark work time and subtract it from the mutator time accumulated in the
GC CPU limiter, and that will give us a more accurate picture of how
much CPU is being spent by user goroutines on GC. This will allow the GC
CPU limiter to kick in, and reduce the impact of the thrashing.
There is a corner case here if the idle mark workers are disabled, for
example for the periodic GC, but in the case of the periodic GC, I don't
think it's possible for us to be thrashing at all, so it doesn't really
matter.
Fixes #52890.
Change-Id: Ie133a7d1f89b603434b415d51eb8733c2708a858
Reviewed-on: https://go-review.googlesource.com/c/go/+/405898 Reviewed-by: David Chase <drchase@google.com>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>