unistack-org/metrics
4
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases 1 Wiki Activity
447d235cbb
metrics/vendor/github.com/valyala/fastrand
History
Aliaksandr Valialkin 447d235cbb
Do not panic on unsupported Go runtime metrics 
Log the unsupported Go runtime metrics on startup instead, so the user is aware of unsupported metrics.
The solution for removing the log lines is to upgrade Go builder.

Reduce the minimum supported Go version at go.mod from Go1.20 to Go1.16, where the runtime/metrics package has been added.
See https://tip.golang.org/doc/go1.16#runtime

Updates https://github.com/VictoriaMetrics/metrics/issues/59
Updates https://github.com/VictoriaMetrics/metrics/pull/60
2023-12-17 16:30:30 +02:00
..
.travis.yml vendor: update github.com/valyala/histogram from v1.0.1 to v1.1.1 2020-07-20 16:50:58 +03:00
fastrand.go vendor: update github.com/valyala/histogram from v1.1.2 to v1.2.0 in order to fix https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1612 2021-09-15 09:20:11 +03:00
go.mod Do not panic on unsupported Go runtime metrics 2023-12-17 16:30:30 +02:00
LICENSE vendor: update github.com/valyala/histogram from v1.0.1 to v1.1.1 2020-07-20 16:50:58 +03:00
README.md vendor: update github.com/valyala/histogram from v1.0.1 to v1.1.1 2020-07-20 16:50:58 +03:00

README.md

Build Status GoDoc Go Report

fastrand

Fast pseudorandom number generator.

Features

  • Optimized for speed.
  • Performance scales on multiple CPUs.

How does it work?

It abuses sync.Pool for maintaining "per-CPU" pseudorandom number generators.

TODO: firgure out how to use real per-CPU pseudorandom number generators.

Benchmark results

$ GOMAXPROCS=1 go test -bench=. github.com/valyala/fastrand
goos: linux
goarch: amd64
pkg: github.com/valyala/fastrand
BenchmarkUint32n                   	50000000	        29.7 ns/op
BenchmarkRNGUint32n                	200000000	         6.50 ns/op
BenchmarkRNGUint32nWithLock        	100000000	        21.5 ns/op
BenchmarkMathRandInt31n            	50000000	        31.8 ns/op
BenchmarkMathRandRNGInt31n         	100000000	        17.9 ns/op
BenchmarkMathRandRNGInt31nWithLock 	50000000	        30.2 ns/op
PASS
ok  	github.com/valyala/fastrand	10.634s

$ GOMAXPROCS=2 go test -bench=. github.com/valyala/fastrand
goos: linux
goarch: amd64
pkg: github.com/valyala/fastrand
BenchmarkUint32n-2                     	100000000	        17.6 ns/op
BenchmarkRNGUint32n-2                  	500000000	         3.36 ns/op
BenchmarkRNGUint32nWithLock-2          	50000000	        32.0 ns/op
BenchmarkMathRandInt31n-2              	20000000	        51.2 ns/op
BenchmarkMathRandRNGInt31n-2           	100000000	        11.0 ns/op
BenchmarkMathRandRNGInt31nWithLock-2   	20000000	        91.0 ns/op
PASS
ok  	github.com/valyala/fastrand	9.543s

$ GOMAXPROCS=4 go test -bench=. github.com/valyala/fastrand
goos: linux
goarch: amd64
pkg: github.com/valyala/fastrand
BenchmarkUint32n-4                     	100000000	        14.2 ns/op
BenchmarkRNGUint32n-4                  	500000000	         3.30 ns/op
BenchmarkRNGUint32nWithLock-4          	20000000	        88.7 ns/op
BenchmarkMathRandInt31n-4              	10000000	       145 ns/op
BenchmarkMathRandRNGInt31n-4           	200000000	         8.35 ns/op
BenchmarkMathRandRNGInt31nWithLock-4   	20000000	       102 ns/op
PASS
ok  	github.com/valyala/fastrand	11.534s

As you can see, fastrand.Uint32n scales on multiple CPUs, while rand.Int31n doesn't scale. Their performance is comparable on GOMAXPROCS=1, but fastrand.Uint32n runs 3x faster than rand.Int31n on GOMAXPROCS=2 and 10x faster than rand.Int31n on GOMAXPROCS=4.