1 // Copyright 2009 The Go Authors. All rights reserved.
2 // Use of this source code is governed by a BSD-style
3 // license that can be found in the LICENSE file.
24 func initBenchmarkFlags() {
25 matchBenchmarks = flag.String("test.bench", "", "run only benchmarks matching `regexp`")
26 benchmarkMemory = flag.Bool("test.benchmem", false, "print memory allocations for benchmarks")
27 flag.Var(&benchTime, "test.benchtime", "run each benchmark for duration `d` or N times if `d` is of the form Nx")
31 matchBenchmarks *string
34 benchTime = durationOrCountFlag{d: 1 * time.Second} // changed during test of testing package
37 type durationOrCountFlag struct {
43 func (f *durationOrCountFlag) String() string {
45 return fmt.Sprintf("%dx", f.n)
50 func (f *durationOrCountFlag) Set(s string) error {
51 if strings.HasSuffix(s, "x") {
52 n, err := strconv.ParseInt(s[:len(s)-1], 10, 0)
53 if err != nil || n < 0 || (!f.allowZero && n == 0) {
54 return fmt.Errorf("invalid count")
56 *f = durationOrCountFlag{n: int(n)}
59 d, err := time.ParseDuration(s)
60 if err != nil || d < 0 || (!f.allowZero && d == 0) {
61 return fmt.Errorf("invalid duration")
63 *f = durationOrCountFlag{d: d}
67 // Global lock to ensure only one benchmark runs at a time.
68 var benchmarkLock sync.Mutex
70 // Used for every benchmark for measuring memory.
71 var memStats runtime.MemStats
73 // InternalBenchmark is an internal type but exported because it is cross-package;
74 // it is part of the implementation of the "go test" command.
75 type InternalBenchmark struct {
80 // B is a type passed to Benchmark functions to manage benchmark
81 // timing and to specify the number of iterations to run.
83 // A benchmark ends when its Benchmark function returns or calls any of the methods
84 // FailNow, Fatal, Fatalf, SkipNow, Skip, or Skipf. Those methods must be called
85 // only from the goroutine running the Benchmark function.
86 // The other reporting methods, such as the variations of Log and Error,
87 // may be called simultaneously from multiple goroutines.
89 // Like in tests, benchmark logs are accumulated during execution
90 // and dumped to standard output when done. Unlike in tests, benchmark logs
91 // are always printed, so as not to hide output whose existence may be
92 // affecting benchmark results.
95 importPath string // import path of the package containing the benchmark
98 previousN int // number of iterations in the previous run
99 previousDuration time.Duration // total duration of the previous run
101 benchTime durationOrCountFlag
103 missingBytes bool // one of the subbenchmarks does not have bytes set.
106 result BenchmarkResult
107 parallelism int // RunParallel creates parallelism*GOMAXPROCS goroutines
108 // The initial states of memStats.Mallocs and memStats.TotalAlloc.
111 // The net total of this test after being run.
114 // Extra metrics collected by ReportMetric.
115 extra map[string]float64
118 // StartTimer starts timing a test. This function is called automatically
119 // before a benchmark starts, but it can also be used to resume timing after
120 // a call to StopTimer.
121 func (b *B) StartTimer() {
123 runtime.ReadMemStats(&memStats)
124 b.startAllocs = memStats.Mallocs
125 b.startBytes = memStats.TotalAlloc
131 // StopTimer stops timing a test. This can be used to pause the timer
132 // while performing complex initialization that you don't
134 func (b *B) StopTimer() {
136 b.duration += time.Since(b.start)
137 runtime.ReadMemStats(&memStats)
138 b.netAllocs += memStats.Mallocs - b.startAllocs
139 b.netBytes += memStats.TotalAlloc - b.startBytes
144 // ResetTimer zeroes the elapsed benchmark time and memory allocation counters
145 // and deletes user-reported metrics.
146 // It does not affect whether the timer is running.
147 func (b *B) ResetTimer() {
149 // Allocate the extra map before reading memory stats.
150 // Pre-size it to make more allocation unlikely.
151 b.extra = make(map[string]float64, 16)
156 runtime.ReadMemStats(&memStats)
157 b.startAllocs = memStats.Mallocs
158 b.startBytes = memStats.TotalAlloc
166 // SetBytes records the number of bytes processed in a single operation.
167 // If this is called, the benchmark will report ns/op and MB/s.
168 func (b *B) SetBytes(n int64) { b.bytes = n }
170 // ReportAllocs enables malloc statistics for this benchmark.
171 // It is equivalent to setting -test.benchmem, but it only affects the
172 // benchmark function that calls ReportAllocs.
173 func (b *B) ReportAllocs() {
174 b.showAllocResult = true
177 // runN runs a single benchmark for the specified number of iterations.
178 func (b *B) runN(n int) {
180 defer benchmarkLock.Unlock()
182 b.runCleanup(normalPanic)
185 // Try to get a comparable environment for each run
186 // by clearing garbage from previous runs.
196 b.previousDuration = b.duration
199 // run1 runs the first iteration of benchFunc. It reports whether more
200 // iterations of this benchmarks should be run.
201 func (b *B) run1() bool {
202 if ctx := b.context; ctx != nil {
203 // Extend maxLen, if needed.
204 if n := len(b.name) + ctx.extLen + 1; n > ctx.maxLen {
205 ctx.maxLen = n + 8 // Add additional slack to avoid too many jumps in size.
209 // Signal that we're done whether we return normally
210 // or by FailNow's runtime.Goexit.
219 fmt.Fprintf(b.w, "%s--- FAIL: %s\n%s", b.chatty.prefix(), b.name, b.output)
222 // Only print the output if we know we are not going to proceed.
223 // Otherwise it is printed in processBench.
225 finished := b.finished
227 if b.hasSub.Load() || finished {
232 if b.chatty != nil && (len(b.output) > 0 || finished) {
234 fmt.Fprintf(b.w, "%s--- %s: %s\n%s", b.chatty.prefix(), tag, b.name, b.output)
241 var labelsOnce sync.Once
243 // run executes the benchmark in a separate goroutine, including all of its
244 // subbenchmarks. b must not have subbenchmarks.
246 labelsOnce.Do(func() {
247 fmt.Fprintf(b.w, "goos: %s\n", runtime.GOOS)
248 fmt.Fprintf(b.w, "goarch: %s\n", runtime.GOARCH)
249 if b.importPath != "" {
250 fmt.Fprintf(b.w, "pkg: %s\n", b.importPath)
252 if cpu := sysinfo.CPUName(); cpu != "" {
253 fmt.Fprintf(b.w, "cpu: %s\n", cpu)
256 if b.context != nil {
257 // Running go test --test.bench
258 b.context.processBench(b) // Must call doBench.
260 // Running func Benchmark.
265 func (b *B) doBench() BenchmarkResult {
271 // launch launches the benchmark function. It gradually increases the number
272 // of benchmark iterations until the benchmark runs for the requested benchtime.
273 // launch is run by the doBench function as a separate goroutine.
274 // run1 must have been called on b.
275 func (b *B) launch() {
276 // Signal that we're done whether we return normally
277 // or by FailNow's runtime.Goexit.
282 // Run the benchmark for at least the specified amount of time.
283 if b.benchTime.n > 0 {
284 // We already ran a single iteration in run1.
285 // If -benchtime=1x was requested, use that result.
286 // See https://golang.org/issue/32051.
287 if b.benchTime.n > 1 {
288 b.runN(b.benchTime.n)
292 for n := int64(1); !b.failed && b.duration < d && n < 1e9; {
294 // Predict required iterations.
295 goalns := d.Nanoseconds()
296 prevIters := int64(b.N)
297 prevns := b.duration.Nanoseconds()
299 // Round up, to avoid div by zero.
302 // Order of operations matters.
303 // For very fast benchmarks, prevIters ~= prevns.
304 // If you divide first, you get 0 or 1,
305 // which can hide an order of magnitude in execution time.
306 // So multiply first, then divide.
307 n = goalns * prevIters / prevns
308 // Run more iterations than we think we'll need (1.2x).
310 // Don't grow too fast in case we had timing errors previously.
312 // Be sure to run at least one more than last time.
314 // Don't run more than 1e9 times. (This also keeps n in int range on 32 bit platforms.)
319 b.result = BenchmarkResult{b.N, b.duration, b.bytes, b.netAllocs, b.netBytes, b.extra}
322 // Elapsed returns the measured elapsed time of the benchmark.
323 // The duration reported by Elapsed matches the one measured by
324 // StartTimer, StopTimer, and ResetTimer.
325 func (b *B) Elapsed() time.Duration {
328 d += time.Since(b.start)
333 // ReportMetric adds "n unit" to the reported benchmark results.
334 // If the metric is per-iteration, the caller should divide by b.N,
335 // and by convention units should end in "/op".
336 // ReportMetric overrides any previously reported value for the same unit.
337 // ReportMetric panics if unit is the empty string or if unit contains
339 // If unit is a unit normally reported by the benchmark framework itself
340 // (such as "allocs/op"), ReportMetric will override that metric.
341 // Setting "ns/op" to 0 will suppress that built-in metric.
342 func (b *B) ReportMetric(n float64, unit string) {
344 panic("metric unit must not be empty")
346 if strings.IndexFunc(unit, unicode.IsSpace) >= 0 {
347 panic("metric unit must not contain whitespace")
352 // BenchmarkResult contains the results of a benchmark run.
353 type BenchmarkResult struct {
354 N int // The number of iterations.
355 T time.Duration // The total time taken.
356 Bytes int64 // Bytes processed in one iteration.
357 MemAllocs uint64 // The total number of memory allocations.
358 MemBytes uint64 // The total number of bytes allocated.
360 // Extra records additional metrics reported by ReportMetric.
361 Extra map[string]float64
364 // NsPerOp returns the "ns/op" metric.
365 func (r BenchmarkResult) NsPerOp() int64 {
366 if v, ok := r.Extra["ns/op"]; ok {
372 return r.T.Nanoseconds() / int64(r.N)
375 // mbPerSec returns the "MB/s" metric.
376 func (r BenchmarkResult) mbPerSec() float64 {
377 if v, ok := r.Extra["MB/s"]; ok {
380 if r.Bytes <= 0 || r.T <= 0 || r.N <= 0 {
383 return (float64(r.Bytes) * float64(r.N) / 1e6) / r.T.Seconds()
386 // AllocsPerOp returns the "allocs/op" metric,
387 // which is calculated as r.MemAllocs / r.N.
388 func (r BenchmarkResult) AllocsPerOp() int64 {
389 if v, ok := r.Extra["allocs/op"]; ok {
395 return int64(r.MemAllocs) / int64(r.N)
398 // AllocedBytesPerOp returns the "B/op" metric,
399 // which is calculated as r.MemBytes / r.N.
400 func (r BenchmarkResult) AllocedBytesPerOp() int64 {
401 if v, ok := r.Extra["B/op"]; ok {
407 return int64(r.MemBytes) / int64(r.N)
410 // String returns a summary of the benchmark results.
411 // It follows the benchmark result line format from
412 // https://golang.org/design/14313-benchmark-format, not including the
414 // Extra metrics override built-in metrics of the same name.
415 // String does not include allocs/op or B/op, since those are reported
417 func (r BenchmarkResult) String() string {
418 buf := new(strings.Builder)
419 fmt.Fprintf(buf, "%8d", r.N)
421 // Get ns/op as a float.
422 ns, ok := r.Extra["ns/op"]
424 ns = float64(r.T.Nanoseconds()) / float64(r.N)
428 prettyPrint(buf, ns, "ns/op")
431 if mbs := r.mbPerSec(); mbs != 0 {
432 fmt.Fprintf(buf, "\t%7.2f MB/s", mbs)
435 // Print extra metrics that aren't represented in the standard
437 var extraKeys []string
438 for k := range r.Extra {
440 case "ns/op", "MB/s", "B/op", "allocs/op":
441 // Built-in metrics reported elsewhere.
444 extraKeys = append(extraKeys, k)
446 sort.Strings(extraKeys)
447 for _, k := range extraKeys {
449 prettyPrint(buf, r.Extra[k], k)
454 func prettyPrint(w io.Writer, x float64, unit string) {
455 // Print all numbers with 10 places before the decimal point
456 // and small numbers with four sig figs. Field widths are
457 // chosen to fit the whole part in 10 places while aligning
458 // the decimal point of all fractional formats.
460 switch y := math.Abs(x); {
461 case y == 0 || y >= 999.95:
473 case y >= 0.00099995:
478 fmt.Fprintf(w, format, x, unit)
481 // MemString returns r.AllocedBytesPerOp and r.AllocsPerOp in the same format as 'go test'.
482 func (r BenchmarkResult) MemString() string {
483 return fmt.Sprintf("%8d B/op\t%8d allocs/op",
484 r.AllocedBytesPerOp(), r.AllocsPerOp())
487 // benchmarkName returns full name of benchmark including procs suffix.
488 func benchmarkName(name string, n int) string {
490 return fmt.Sprintf("%s-%d", name, n)
495 type benchContext struct {
498 maxLen int // The largest recorded benchmark name.
499 extLen int // Maximum extension length.
502 // RunBenchmarks is an internal function but exported because it is cross-package;
503 // it is part of the implementation of the "go test" command.
504 func RunBenchmarks(matchString func(pat, str string) (bool, error), benchmarks []InternalBenchmark) {
505 runBenchmarks("", matchString, benchmarks)
508 func runBenchmarks(importPath string, matchString func(pat, str string) (bool, error), benchmarks []InternalBenchmark) bool {
509 // If no flag was specified, don't run benchmarks.
510 if len(*matchBenchmarks) == 0 {
513 // Collect matching benchmarks and determine longest name.
515 for _, procs := range cpuList {
516 if procs > maxprocs {
520 ctx := &benchContext{
521 match: newMatcher(matchString, *matchBenchmarks, "-test.bench", *skip),
522 extLen: len(benchmarkName("", maxprocs)),
524 var bs []InternalBenchmark
525 for _, Benchmark := range benchmarks {
526 if _, matched, _ := ctx.match.fullName(nil, Benchmark.Name); matched {
527 bs = append(bs, Benchmark)
528 benchName := benchmarkName(Benchmark.Name, maxprocs)
529 if l := len(benchName) + ctx.extLen + 1; l > ctx.maxLen {
540 importPath: importPath,
541 benchFunc: func(b *B) {
542 for _, Benchmark := range bs {
543 b.Run(Benchmark.Name, Benchmark.F)
546 benchTime: benchTime,
550 main.chatty = newChattyPrinter(main.w)
556 // processBench runs bench b for the configured CPU counts and prints the results.
557 func (ctx *benchContext) processBench(b *B) {
558 for i, procs := range cpuList {
559 for j := uint(0); j < *count; j++ {
560 runtime.GOMAXPROCS(procs)
561 benchName := benchmarkName(b.name, procs)
563 // If it's chatty, we've already printed this information.
565 fmt.Fprintf(b.w, "%-*s\t", ctx.maxLen, benchName)
567 // Recompute the running time for all but the first iteration.
571 signal: make(chan bool),
577 benchFunc: b.benchFunc,
578 benchTime: b.benchTime,
584 // The output could be very long here, but probably isn't.
585 // We print it all, regardless, because we don't want to trim the reason
586 // the benchmark failed.
587 fmt.Fprintf(b.w, "%s--- FAIL: %s\n%s", b.chatty.prefix(), benchName, b.output)
590 results := r.String()
592 fmt.Fprintf(b.w, "%-*s\t", ctx.maxLen, benchName)
594 if *benchmarkMemory || b.showAllocResult {
595 results += "\t" + r.MemString()
597 fmt.Fprintln(b.w, results)
598 // Unlike with tests, we ignore the -chatty flag and always print output for
599 // benchmarks since the output generation time will skew the results.
600 if len(b.output) > 0 {
602 fmt.Fprintf(b.w, "%s--- BENCH: %s\n%s", b.chatty.prefix(), benchName, b.output)
604 if p := runtime.GOMAXPROCS(-1); p != procs {
605 fmt.Fprintf(os.Stderr, "testing: %s left GOMAXPROCS set to %d\n", benchName, p)
607 if b.chatty != nil && b.chatty.json {
608 b.chatty.Updatef("", "=== NAME %s\n", "")
614 // If hideStdoutForTesting is true, Run does not print the benchName.
615 // This avoids a spurious print during 'go test' on package testing itself,
616 // which invokes b.Run in its own tests (see sub_test.go).
617 var hideStdoutForTesting = false
619 // Run benchmarks f as a subbenchmark with the given name. It reports
620 // whether there were any failures.
622 // A subbenchmark is like any other benchmark. A benchmark that calls Run at
623 // least once will not be measured itself and will be called once with N=1.
624 func (b *B) Run(name string, f func(b *B)) bool {
625 // Since b has subbenchmarks, we will no longer run it as a benchmark itself.
626 // Release the lock and acquire it on exit to ensure locks stay paired.
628 benchmarkLock.Unlock()
629 defer benchmarkLock.Lock()
631 benchName, ok, partial := b.name, true, false
632 if b.context != nil {
633 benchName, ok, partial = b.context.match.fullName(&b.common, name)
638 var pc [maxStackLen]uintptr
639 n := runtime.Callers(2, pc[:])
642 signal: make(chan bool),
651 importPath: b.importPath,
653 benchTime: b.benchTime,
657 // Partial name match, like -bench=X/Y matching BenchmarkX.
658 // Only process sub-benchmarks, if any.
659 sub.hasSub.Store(true)
663 labelsOnce.Do(func() {
664 fmt.Printf("goos: %s\n", runtime.GOOS)
665 fmt.Printf("goarch: %s\n", runtime.GOARCH)
666 if b.importPath != "" {
667 fmt.Printf("pkg: %s\n", b.importPath)
669 if cpu := sysinfo.CPUName(); cpu != "" {
670 fmt.Printf("cpu: %s\n", cpu)
674 if !hideStdoutForTesting {
676 b.chatty.Updatef(benchName, "=== RUN %s\n", benchName)
678 fmt.Println(benchName)
689 // add simulates running benchmarks in sequence in a single iteration. It is
690 // used to give some meaningful results in case func Benchmark is used in
691 // combination with Run.
692 func (b *B) add(other BenchmarkResult) {
694 // The aggregated BenchmarkResults resemble running all subbenchmarks as
695 // in sequence in a single benchmark.
697 r.T += time.Duration(other.NsPerOp())
698 if other.Bytes == 0 {
699 // Summing Bytes is meaningless in aggregate if not all subbenchmarks
701 b.missingBytes = true
705 r.Bytes += other.Bytes
707 r.MemAllocs += uint64(other.AllocsPerOp())
708 r.MemBytes += uint64(other.AllocedBytesPerOp())
711 // trimOutput shortens the output from a benchmark, which can be very long.
712 func (b *B) trimOutput() {
713 // The output is likely to appear multiple times because the benchmark
714 // is run multiple times, but at least it will be seen. This is not a big deal
715 // because benchmarks rarely print, but just in case, we trim it if it's too long.
716 const maxNewlines = 10
717 for nlCount, j := 0, 0; j < len(b.output); j++ {
718 if b.output[j] == '\n' {
720 if nlCount >= maxNewlines {
721 b.output = append(b.output[:j], "\n\t... [output truncated]\n"...)
728 // A PB is used by RunParallel for running parallel benchmarks.
730 globalN *uint64 // shared between all worker goroutines iteration counter
731 grain uint64 // acquire that many iterations from globalN at once
732 cache uint64 // local cache of acquired iterations
733 bN uint64 // total number of iterations to execute (b.N)
736 // Next reports whether there are more iterations to execute.
737 func (pb *PB) Next() bool {
739 n := atomic.AddUint64(pb.globalN, pb.grain)
742 } else if n < pb.bN+pb.grain {
743 pb.cache = pb.bN + pb.grain - n
752 // RunParallel runs a benchmark in parallel.
753 // It creates multiple goroutines and distributes b.N iterations among them.
754 // The number of goroutines defaults to GOMAXPROCS. To increase parallelism for
755 // non-CPU-bound benchmarks, call SetParallelism before RunParallel.
756 // RunParallel is usually used with the go test -cpu flag.
758 // The body function will be run in each goroutine. It should set up any
759 // goroutine-local state and then iterate until pb.Next returns false.
760 // It should not use the StartTimer, StopTimer, or ResetTimer functions,
761 // because they have global effect. It should also not call Run.
763 // RunParallel reports ns/op values as wall time for the benchmark as a whole,
764 // not the sum of wall time or CPU time over each parallel goroutine.
765 func (b *B) RunParallel(body func(*PB)) {
767 return // Nothing to do when probing.
769 // Calculate grain size as number of iterations that take ~100µs.
770 // 100µs is enough to amortize the overhead and provide sufficient
771 // dynamic load balancing.
773 if b.previousN > 0 && b.previousDuration > 0 {
774 grain = 1e5 * uint64(b.previousN) / uint64(b.previousDuration)
779 // We expect the inner loop and function call to take at least 10ns,
780 // so do not do more than 100µs/10ns=1e4 iterations.
786 numProcs := b.parallelism * runtime.GOMAXPROCS(0)
787 var wg sync.WaitGroup
789 for p := 0; p < numProcs; p++ {
801 if n <= uint64(b.N) && !b.Failed() {
802 b.Fatal("RunParallel: body exited without pb.Next() == false")
806 // SetParallelism sets the number of goroutines used by RunParallel to p*GOMAXPROCS.
807 // There is usually no need to call SetParallelism for CPU-bound benchmarks.
808 // If p is less than 1, this call will have no effect.
809 func (b *B) SetParallelism(p int) {
815 // Benchmark benchmarks a single function. It is useful for creating
816 // custom benchmarks that do not use the "go test" command.
818 // If f depends on testing flags, then Init must be used to register
819 // those flags before calling Benchmark and before calling flag.Parse.
821 // If f calls Run, the result will be an estimate of running all its
822 // subbenchmarks that don't call Run in sequence in a single benchmark.
823 func Benchmark(f func(b *B)) BenchmarkResult {
826 signal: make(chan bool),
830 benchTime: benchTime,
838 type discard struct{}
840 func (discard) Write(b []byte) (n int, err error) { return len(b), nil }