cmd/compile: unable to fully utilize CPU after upgrading from Go 1.20 to 1.21

[necheffa]

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

$ go version
go version go1.21.1 linux/amd64

Does this issue reproduce with the latest release?

Yes.

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

Debian 12.1 with an i7-1165g7.

go env Output

$ go env
GO111MODULE=''
GOARCH='amd64'
GOBIN=''
GOCACHE='/home/alex/.cache/go-build'
GOENV='/home/alex/.config/go/env'
GOEXE=''
GOEXPERIMENT=''
GOFLAGS=''
GOHOSTARCH='amd64'
GOHOSTOS='linux'
GOINSECURE=''
GOMODCACHE='/home/alex/go/pkg/mod'
GONOPROXY=''
GONOSUMDB=''
GOOS='linux'
GOPATH='/home/alex/go'
GOPRIVATE=''
GOPROXY='https://proxy.golang.org,direct'
GOROOT='/opt/golang/go-1.21.1'
GOSUMDB='sum.golang.org'
GOTMPDIR=''
GOTOOLCHAIN='auto'
GOTOOLDIR='/opt/golang/go-1.21.1/pkg/tool/linux_amd64'
GOVCS=''
GOVERSION='go1.21.1'
GCCGO='gccgo'
GOAMD64='v1'
AR='ar'
CC='gcc'
CXX='g++'
CGO_ENABLED='1'
GOMOD='/home/alex/repaste_mini/go.mod'
GOWORK=''
CGO_CFLAGS='-O2 -g'
CGO_CPPFLAGS=''
CGO_CXXFLAGS='-O2 -g'
CGO_FFLAGS='-O2 -g'
CGO_LDFLAGS='-O2 -g'
PKG_CONFIG='pkg-config'
GOGCCFLAGS='-fPIC -m64 -pthread -Wl,--no-gc-sections -fmessage-length=0 -ffile-prefix-map=/tmp/go-build2302350423=/tmp/go-build -gno-record-gcc-switches'

What did you do?

I compiled and executed the snippet of code below.

package main

import (
	"math"
	"math/rand"
	"runtime"
	"time"
)

const (
	timeLimit   = 1  // timeout in minutes
	timeSegment = 10 // sample rate in seconds
)

func Repaste(upTo int, done chan any) {
	for i := 0; i < upTo; i++ {
		select {
		case <-done:
			break
		default:
		}
		_ = math.Sqrt(rand.Float64())
	}
}

func main() {
	lim := runtime.NumCPU()
	done := make(chan any, lim)
	for i := 0; i < lim; i++ {
		go Repaste(math.MaxInt, done)
	}

	runTime := timeLimit * 60 // convert to seconds
	for i := 0; i < runTime; i += timeSegment {
		time.Sleep(timeSegment * time.Second)
	}

	for i := 0; i < lim; i++ {
		done <- i
	}

}

What did you expect to see?

I expected to see all 4 cores / 8 threads hit 4.1GHz during turbo boost.
NOTE: I am able to get this all-core clock speed with the exact same code snippet but built with Go 1.20.8 on the same machine, so I know it isn't a hardware issue like hitting thermal/power limits.

What did you see instead?

All cores / threads were at 100% utilization but only peaked at 3.6GHz.

It is also worth mentioning that I have tried other variations of "load" besides math.Sqrt(), such as just primitive arithmetic expressions, so this would not reflect a regression in the math package, the issue lies elsewhere.

[necheffa]

Removing channels entirely and letting the goroutines spin for a period of time before mashing Ctrl+C seems to allow full CPU utilization when building with Go 1.21.1, so it would seem to me that the perf regression likely lies somewhere in the implementation of channels.

[mdempsky]

I expected to see all 4 cores / 8 threads hit 4.1GHz during turbo boost.

All cores / threads were at 100% utilization but only peaked at 3.6GHz.

Are you able to quantify the impact of this further? E.g., do you have some benchmarks that show latency or throughput worsening?

The Go compiler and runtime don't manage CPU frequency. I also don't think we care about CPU frequency in isolation, unless it's affecting some other performance benchmarks (e.g., latency or throughput).

[gopherbot]

Timed out in state WaitingForInfo. Closing.

(I am just a bot, though. Please speak up if this is a mistake or you have the requested information.)