testing: tiny benchmark with StopTimer runs forever

[peterGo]

From at least Go 1,4 onward,

$ go version
go version devel +e03220a594 Sat Aug 25 02:39:49 2018 +0000 linux/amd64
$

If we run a tiny benchmark (count++) and use StopTimer() for initialization (count = 0) then the benchmark runs for a very long time, perhaps forever.

$ cat tiny_test.go
package main

import (
	"testing"
)

var count int64

func BenchmarkTiny(b *testing.B) {
	for n := 0; n < b.N; n++ {
		b.StopTimer()
		count = 0
		b.StartTimer()
		count++
	}
}
$

$ go test tiny_test.go -bench=.
goos: linux
goarch: amd64
BenchmarkTiny-4   	^Csignal: interrupt
FAIL	command-line-arguments	143.633s
$

If we comment out StopTimer() (//b.StopTimer()) then the benchmark quickly runs to completion.

$ cat tiny_test.go
package main

import (
	"testing"
)

var count int64

func BenchmarkTiny(b *testing.B) {
	for n := 0; n < b.N; n++ {
		// b.StopTimer()
		count = 0
		b.StartTimer()
		count++
	}
}
$

$ go test tiny_test.go -bench=.
goos: linux
goarch: amd64
BenchmarkTiny-4   	1000000000	         2.15 ns/op
PASS
ok  	command-line-arguments	2.374s
$ 

[meirf]

runtime.ReadMemStats which is called by both StartTimer and StopTimer is known to be slow, though I'm not sure it's expected to be as slow as shown here.

Calling just StartTimer without StopTimer is a noop because b.timerOn is true.

The uses of StartTimer/StopTimer are usually around one time setup/teardown code that happens before or after the b.N loop.

[josharian]

I suspect that this is (effectively) a duplicate of #20875.

[mvdan]

I encountered the same while benchmarking fairly small functions. Via pprof, I could see that ReadMemStats (called via both StartTimer and StopTimer) was taking up ~90% of the CPU time, while the func I was benchmarking was only taking about 3% of CPU time.

Initially, I too thought that go test -bench=. would never end. But that wasn't the problem.

For example, go test -bench=. -benchtime=100x runs fine. The issue is that when given a benchtime in a duration instead of number of iterations, the testing package has to estimate the number of times it will need to execute the benchmark loop.

And this is where it gets whacky. For example, on go test -bench=. -benchtime=0.1s, the total run-time ends up being over 6s on my machine. This seems to indicate that the b.N estimate that the testing package ends up with is way too high, as the time it spends running the benchmark is over 50 times larger than what the user wanted.

So I presume this is what you're seeing. By default the benchtime is 1s, so it's reasonable to think that your tiny benchmark could be slowed down enough to run for a few minutes. I'd suggest trying to run it with -benchtime=100x or -benchtime=0.01s to confirm that it's only the estimate that is wrong.

In which case I agree with @josharian that the root of the issue is #20875. If ReadMemStats wasn't called that often, I presume that the b.N estimates wouldn't be nearly as bad.

[mvdan]

Looks like other people ran into very similar puzzling cases before: https://stackoverflow.com/questions/37620251/golang-benchmarking-b-stoptimer-hangs-is-it-me

[josharian]

This seems to indicate that the b.N estimate that the testing package ends up with is way too high

Note that if the goal is for the benchmark code to execute for 1s, and there's 97% overhead due to StopTimer/StartTimer, then the correct overall execution time is 33s. So maybe the b.N estimate is correct.

Note that we should make the b.N estimate better. CL 112155 is an attempt at this. It needs to be revisited. There's also discussion of this at #24735. See also
#27168 (comment).

I've also thought for some time that we should set an absolute limit on b.N. To mangle Tolstoy: How many iterations does one benchmark need? If we capped b.N at something reasonably sized, like 100k, that'd mitigate these disaster scenarios. And also general speed up running microbenchmarks. (I suppose I should file a new issue for this?)