Benchmarks and profiling
Write a BenchmarkX
A benchmark lives in a test file and its name starts with Benchmark. The testing framework calls it with a B value, and the loop over b.N repeats the measured code until timing stabilizes.
Good benchmarks compare two approaches to one task. The file below joins words with plain concatenation and with a strings.Builder, so one run shows which method wins.
package main
import (
"strings"
"testing"
)
func JoinPlus(parts []string) string {
out := ""
for _, p := range parts {
out += p
}
return out
}
func BenchmarkJoinPlus(b *testing.B) {
parts := []string{"go", "is", "fast"}
for i := 0; i < b.N; i++ {
JoinPlus(parts)
}
}
func BenchmarkJoinBuilder(b *testing.B) {
parts := []string{"go", "is", "fast"}
for i := 0; i < b.N; i++ {
var sb strings.Builder
for _, p := range parts {
sb.WriteString(p)
}
_ = sb.String()
}
}
go test -bench
The go test tool runs benchmarks when the bench flag matches their names. A dot matches every benchmark, run with NONE skips unit tests, and benchtime controls the workload per benchmark.
Each result line shows the name, the iteration count, and the average time per operation. Smaller time per operation wins, so the Builder line below beats plain concatenation.
go test -bench=. -benchtime=100x -run=NONE
BenchmarkJoinPlus-8 100 420 ns/op
BenchmarkJoinBuilder-8 100 310 ns/op
pprof basics
Timing tells which function is slow, while a profile tells which line inside it burns the time. The test tool writes a CPU profile to a file when the cpuprofile flag points at one.
The go tool pprof command opens that file for questions. The top command lists the hottest functions first, and the list command with a function name shows time per source line.
go test -bench=Builder -cpuprofile=cpu.out -run=NONE
BenchmarkJoinBuilder-8 100 310 ns/op
ok wordfreq 0.412s
go tool pprof -top cpu.out
Type: cpu
Time: Oct 6, 2026
main.JoinBuilder
Watch allocations
Speed is only half the story. Every heap allocation costs the garbage collector future work, so benchmarks should track bytes and allocation counts next to time.
The benchmem flag adds two columns to each result line: bytes per operation and allocations per operation. Fewer allocations often predict calmer production behavior under load.
go test -bench=. -benchmem -run=NONE
BenchmarkJoinPlus-8 100 420 ns/op 64 B/op 3 allocs/op
BenchmarkJoinBuilder-8 100 310 ns/op 24 B/op 1 allocs/op
For a focused check, AllocsPerRun measures one function in isolation. It runs the closure many times and returns the average allocation count as a float.
package main
import (
"testing"
)
func BenchmarkConcatAllocs(b *testing.B) {
parts := []string{"go", "is", "fast"}
n := testing.AllocsPerRun(100, func() {
out := ""
for _, p := range parts {
out += p
}
})
b.ReportMetric(n, "allocs")
}
Next: Toolchain and CI
Article author: Arthur Isaev
