\cc @aclements

I note that the sysmon goroutine polls the network every 10ms if nothing else has polled. Normally findrunnable will poll the network whenever it doesn't have another goroutine to run. Perhaps during the GC there is always another goroutine to run (to do GC work) and therefore we fall back to the sysmon polling.

I forgot to add that I instrumented the Go runtime and can see lots of calls to netpoll via gcDrain -> pollWork. I also tried lowering the max sleep time in sysmon to 1ms, but that had no effect. Perhaps the 10ms is some OS level time slice on Darwin. I'll try and verify this same behavior on Linux.

Linux shows similar behavior:

trace.linux.out.zip

I'm not sure what to make of the 3 idle procs (5, 6 and 7).

Probably unrelated, but #20307 also observed that GC had a detrimental effect on goroutine scheduling.

@petermattis if you can reproduce easily, would you mind trying out tip (and maybe 1.7.5) to see whether anything has changed?

@josharian Tip as of go version devel +5e79787 Mon May 22 13:44:12 2017 +0000 darwin/amd64:

trace.tip.out.zip

The networking blip is still there during GC, but it is no longer precisely 10ms. The network blips during sweep seem to be much reduced:

I'm reasonably convinced that what I'm seeing is CPU starvation of my load generator. I tweaked the output of GODEBUG=gctrace=1 to compute the amount of CPU used by GC during the concurrent mark phase:

gc 203 @28.154s 0/98%: 0.007+9.5+0.039 ms clock, 0.060+0.21/18/56+0.31 ms cpu, 82->82->42 MB, 84 MB goal, 8 P

The 0/98% indicates that GC has used 0% of the CPU over the lifetime of the process, but used 98% of the CPU during the 9.5ms of concurrent marking. The use of the idle CPU for GC is reasonable on the one hand, but makes the Go process a poor neighbor.

I tried disabling idle GC workers which significantly reduced the CPU spikes during GC, but made latencies much worse:

trace.noidlegc.out.zip

In the above trace, G377 is running on Proc 5 when GC starts. G377 then blocks on network IO and the 2 dedicated GC workers get started. But there are only 2 Procs running and there does not seem to be a mechanism to kick off another one (or perhaps I disabled that mechanism in my hack to disable idle GC workers). Adding a wakep call after a dedicated GC worker is started fixes this issue. This gets latencies to a better place:

trace.noidlegc2.out.zip

CPU usage during marking is much lower than stock go1.8.1:

gc 201 @30.960s 1/29%: 0.025+22+0.080 ms clock, 0.20+11/41/0+0.64 ms cpu, 87->88->45 MB, 89 MB goal, 8 P

The downside to removing the idle GC workers is that the GC periods are about twice as long.

Seems like there is a high-level question to be answered about how much CPU GC should use when the process is not CPU limited.

Dup of #17969 ("aggressive GC completion is disruptive to co-tenants") then?

Yep, that seems like the same issue I'm seeing.

But there are only 2 Procs running and there does not seem to be a mechanism to kick off another one (or perhaps I disabled that mechanism in my hack to disable idle GC workers).

You're right that at the moment it's possible for not all idle Ps to help with GC. This is #14179 (which we fixed once, but the fix was broken, so then we unfixed it).

Seems like there is a high-level question to be answered about how much CPU GC should use when the process is not CPU limited.

Yes. :) But I don't know what the answer is. Ideally we'd get the OS scheduler involved, since it's the only thing that knows what CPU resources are really idle. But (right now) we need global coordination between the GC workers, so we can't just run them at OS idle priority.

Thanks for all the traces and debugging! Closing as a dup of #17969.