runtime: TestPingPongHog fail randomly on linux/arm64

[williamweixiao]

What version of Go are you using (go version)?

go version devel +2d20ded Sat May 20 00:45:56 2017 +0000 linux/arm64

What operating system and processor architecture are you using (go env)?

GOARCH="arm64"
GOBIN=""
GOEXE=""
GOHOSTARCH="arm64"
GOHOSTOS="linux"
GOOS="linux"
GOPATH="/home/weixia01/go"
GORACE=""
GOROOT="/home/weixia01/workspace/llt/golang/golang"
GOTOOLDIR="/home/weixia01/workspace/llt/golang/golang/pkg/tool/linux_arm64"
GCCGO="gccgo"
CC="gcc"
GOGCCFLAGS="-fPIC -pthread -fmessage-length=0 -fdebug-prefix-map=/tmp/go-build257642744=/tmp/go-build -gno-record-gcc-switches"
CXX="g++"
CGO_ENABLED="1"
CGO_CFLAGS="-g -O2"
CGO_CPPFLAGS=""
CGO_CXXFLAGS="-g -O2"
CGO_FFLAGS="-g -O2"
CGO_LDFLAGS="-g -O2"
PKG_CONFIG="pkg-config"

What did you do?

$cd src/runtime
$go test -run="TestPingPongHog"

What did you expect to see?

PASS

What did you see instead?

--- FAIL: TestPingPongHog (0.15s)
proc_test.go:434: want hogCount/lightCount in [0.5, 2]; got 23000000/11097000 = 2.072632242948545
FAIL
exit status 1

[williamweixiao]

I'm not sure whether the required range [0.5, 2] is reasonable.
I would appreciate it if someone can shed some light on the range.

[mvdan]

How often does it fail? Does -race affect the frequency? I guess you used something like go test -run=TestPingPongHog -count 1000 to try many times and get a few failures?

[williamweixiao]

The failure rate is about: 458/1000 (-count 1000)
The "-race" is only supported on amd64 and I can't use it on arm64.

[williamweixiao]

I find there are failures even for amd64 but the rate is very small: 24/1000.
The "-race" will improve the failure rate to: 100%.
So it seems that arm64 just amplify the failure probability.

[williamweixiao]

/cc @ianlancetaylor @bradfitz
any comments on the failure rate?

[ianlancetaylor]

I just ran the test 1000 times on amd64 and saw 39 failures. The smallest ratio I saw was 0.4875967574815627 and the largest was 2.5314388371713212.

@aclements wrote the test so I will let him decide what to do.

[aclements]

I think we can just expand the range a bit. There's was nothing special about the 2X. If this test weren't working, I would expect the gap to be closer to 1000X.

[gopherbot]

CL https://golang.org/cl/44859 mentions this issue.