CC @randall77

I was curious and bisected with the given test cases. I found the following:

• Setting GOGC=off eliminates the difference between 1.17 and 1.18.
• The beginning of the regression (about half of the perf loss) starts at CL 358674, when the pacer redesign was enabled. Setting GOEXPERIMENT=nopacerredesign recovers this loss.
• Keeping GOEXPERIMENT=nopacerredesign and re-bisecting shows the other half of the regression coming from CL 353975.

Thanks for bisecting. cc @mknyszek too, then.

I suspect this is just the new minimum heap size (512 KiB). Microbenchmarks suffer because they have a small heap, so you GC more often to save memory. Instead of GOEXPERIMENT=nopacerredesign, try to modify https://cs.opensource.google/go/go/+/master:src/runtime/mgcpacer.go;l=56?q=defaultHeapMinimum&ss=go%2Fgo and see if performance recovers.

We're thinking we're going to roll back just the new minimum heap size for 1.18, and try again later.

I can confirm that restoring defaultHeapMinimum to 4 << 20 eliminates the slowdown in my testing.

Cool. I think this is a non-issue for this release, but it's helped find a lot of various finalizer bugs and such in our tests (and elsewhere) so I'm leaving it in for just a little longer.

I'll leave this open until we roll back the heap minimum.

Random thought: If/when PGO (profile guided optimization) lands, perhaps it could be used to determine a reasonable minimum heap? This could be useful for short-lived utilities.

@mpx, that's an interesting idea. I think first we're going to try to see if we can just make the GC scale down better. My hypothesis is that the smaller minimum heap size is primarily a problem because of the high constant overheads of starting a GC (it used to be the "small heap amortization problem"—we didn't account properly for all sources of GC work and non-heap work could dominate for small heaps—but the new pacer fixes that). There's probably always going to be some amount of cost for things that genuinely run on very small heaps (after all, a smaller minimum heap trades CPU for RAM in that regime), but my instinct is that it shouldn't be nearly as high as what we're seeing. If we can fix it that way, it may simply obviate the issue.

Change https://golang.org/cl/368137 mentions this issue: runtime: switch back to a 4 MiB min heap for GOEXPERIMENT=pacerredesign