runtime: eliminate getfull barrier from concurrent mark
Currently dedicated mark workers participate in the getfull barrier
during concurrent mark. However, the getfull barrier wasn't designed
for concurrent work and this causes no end of headaches.
In the concurrent setting, participants come and go. This makes mark
completion susceptible to live-lock: since dedicated workers are only
periodically polling for completion, it's possible for the program to
be in some transient worker each time one of the dedicated workers
wakes up to check if it can exit the getfull barrier. It also
complicates reasoning about the system because dedicated workers
participate directly in the getfull barrier, but transient workers
must instead use trygetfull because they have exit conditions that
aren't captured by getfull (e.g., fractional workers exit when
preempted). The complexity of implementing these exit conditions
contributed to #11677. Furthermore, the getfull barrier is inefficient
because we could be running user code instead of spinning on a P. In
effect, we're dedicating 25% of the CPU to marking even if that means
we have to spin to make that 25%. It also causes issues on Windows
because we can't actually sleep for 100µs (#8687).
Fix this by making dedicated workers no longer participate in the
getfull barrier. Instead, dedicated workers simply return to the
scheduler when they fail to get more work, regardless of what others
workers are doing, and the scheduler only starts new dedicated workers
if there's work available. Everything that needs to be handled by this
barrier is already handled by detection of mark completion.
This makes the system much more symmetric because all workers and
assists now use trygetfull during concurrent mark. It also loosens the
25% CPU target so that we can give some of that 25% back to user code
if there isn't enough work to keep the mark worker busy. And it
eliminates the problematic 100µs sleep on Windows during concurrent
mark (though not during mark termination).
The downside of this is that if we hit a bottleneck in the heap graph
that then expands back out, the system may shut down dedicated workers
and take a while to start them back up. We'll address this in the next
commit.
Updates #12041 and #8687.
No effect on the go1 benchmarks. This slows down the garbage benchmark
by 9%, but we'll more than make it up in the next commit.
name old time/op new time/op delta
XBenchGarbage-12 5.80ms ± 2% 6.32ms ± 4% +9.03% (p=0.000 n=20+20)
Change-Id: I65100a9ba005a8b5cf97940798918672ea9dd09b
Reviewed-on: https://go-review.googlesource.com/16297 Reviewed-by: Rick Hudson <rlh@golang.org>