runtime: CPU bound goroutines cause unnecessary timer latency [1.14 backport]

[gopherbot]

@networkimprov requested issue #38860 to be considered for backport to the next 1.14 minor release.

@gopherbot please backport to 1.14. This is a regression with a simple fix. (It may have a more complex fix, too, but that won't land until 1.16.)

[ianlancetaylor]

We don't backport performance improvements, simple or not. Although 1.14 showed a regression in timer latency for programs with soft real time constraints, the timers in 1.14 improved in other ways. This doesn't rise to the level of something that merits a backport. Sorry.

[networkimprov]

Sorry, I had the impression that regressions without workarounds would be fixed... Is that not so?

@dmitshur @andybons @rsc

[ianlancetaylor]

Behavior regressions, yes. Not performance regressions.

This is a performance regression. All programs continue to run as expected. Certain specific kinds of program take a small amount more time before they invoke timers.

We can't promise that for each new release all programs will continue to run as fast as they did in previous releases. Of course we hope that they do. And we certainly try hard to ensure that on average programs run faster. But the fact that it is possible to write a program that runs slightly slower in a new release does not in itself constitute a regression that should be fixed.

[networkimprov]

@ianlancetaylor did you mean to reopen this?

this also needs a backport to 1.15, as it is milestoned for 1.16 -- sorry, got confused by the milestone change after fix landed.

[ianlancetaylor]

@networkimprov No, I raised the possibility of reopening it. I don't think we've made a decision.

[ianlancetaylor]

And the only fix we would be backporting would be the one that is already in 1.15.