runtime: skip work recheck for non-spinning Ms
When an M transitions from spinning to non-spinning state, it must
recheck most sources of work to avoid missing work submitted between its
initial check and decrementing sched.nmspinning (see "delicate dance"
comment).
Ever since the scheduler rewrite in Go 1.1 (golang.org/cl/
7314062), we
have performed this recheck on all Ms before stopping, regardless of
whether or not they were spinning.
Unfortunately, there is a problem with this approach: non-spinning Ms
are not eligible to steal work (note the skip over the stealWork block),
but can detect work during the recheck. If there is work available, this
non-spinning M will jump to top, skip stealing, land in recheck again,
and repeat. i.e., it will spin uselessly.
The spin is bounded. This can only occur if there is another spinning M,
which will either take the work, allowing this M to stop, or take some
other work, allowing this M to upgrade to spinning. But the spinning is
ultimately just a fancy spin-wait.
golang.org/issue/43997 discusses several ways to address this. This CL
takes the simplest approach: skipping the recheck on non-spinning Ms and
allowing them to go to stop.
Results for scheduler-relevant runtime and time benchmarks can be found
at https://perf.golang.org/search?q=upload:
20210420.5.
The new BenchmarkCreateGoroutinesSingle is a characteristic example
workload that hits this issue hard. A single M readies lots of work
without itself parking. Other Ms must spin to steal work, which is very
short-lived, forcing those Ms to spin again. Some of the Ms will be
non-spinning and hit the above bug.
With this fixed, that benchmark drops in CPU usage by a massive 68%, and
wall time 24%. BenchmarkNetpollBreak shows similar drops because it is
unintentionally almost the same benchmark (create short-living Gs in a
loop). Typical well-behaved programs show little change.
We also measure scheduling latency (time from goready to execute). Note
that many of these benchmarks are very noisy because they don't involve
much scheduling. Those that do, like CreateGoroutinesSingle, are
expected to increase as we are replacing unintentional spin waiting with
a real park.
Fixes #43997
Change-Id: Ie1d1e1800f393cee1792455412caaa5865d13562
Reviewed-on: https://go-review.googlesource.com/c/go/+/310850
Trust: Michael Pratt <mpratt@google.com>
Run-TryBot: Michael Pratt <mpratt@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Michael Knyszek <mknyszek@google.com>