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`")
31 matchBenchmarks *string
34 benchTime = benchTimeFlag{d: 1 * time.Second} // changed during test of testing package
37 type benchTimeFlag struct {
42 func (f *benchTimeFlag) String() string {
44 return fmt.Sprintf("%dx", f.n)
46 return time.Duration(f.d).String()
49 func (f *benchTimeFlag) Set(s string) error {
50 if strings.HasSuffix(s, "x") {
51 n, err := strconv.ParseInt(s[:len(s)-1], 10, 0)
52 if err != nil || n <= 0 {
53 return fmt.Errorf("invalid count")
55 *f = benchTimeFlag{n: int(n)}
58 d, err := time.ParseDuration(s)
59 if err != nil || d <= 0 {
60 return fmt.Errorf("invalid duration")
62 *f = benchTimeFlag{d: d}
66 // Global lock to ensure only one benchmark runs at a time.
67 var benchmarkLock sync.Mutex
69 // Used for every benchmark for measuring memory.
70 var memStats runtime.MemStats
72 // InternalBenchmark is an internal type but exported because it is cross-package;
73 // it is part of the implementation of the "go test" command.
74 type InternalBenchmark struct {
79 // B is a type passed to Benchmark functions to manage benchmark
80 // timing and to specify the number of iterations to run.
82 // A benchmark ends when its Benchmark function returns or calls any of the methods
83 // FailNow, Fatal, Fatalf, SkipNow, Skip, or Skipf. Those methods must be called
84 // only from the goroutine running the Benchmark function.
85 // The other reporting methods, such as the variations of Log and Error,
86 // may be called simultaneously from multiple goroutines.
88 // Like in tests, benchmark logs are accumulated during execution
89 // and dumped to standard error when done. Unlike in tests, benchmark logs
90 // are always printed, so as not to hide output whose existence may be
91 // affecting benchmark results.
94 importPath string // import path of the package containing the benchmark
97 previousN int // number of iterations in the previous run
98 previousDuration time.Duration // total duration of the previous run
100 benchTime benchTimeFlag
102 missingBytes bool // one of the subbenchmarks does not have bytes set.
105 result BenchmarkResult
106 parallelism int // RunParallel creates parallelism*GOMAXPROCS goroutines
107 // The initial states of memStats.Mallocs and memStats.TotalAlloc.
110 // The net total of this test after being run.
113 // Extra metrics collected by ReportMetric.
114 extra map[string]float64
117 // StartTimer starts timing a test. This function is called automatically
118 // before a benchmark starts, but it can also be used to resume timing after
119 // a call to StopTimer.
120 func (b *B) StartTimer() {
122 runtime.ReadMemStats(&memStats)
123 b.startAllocs = memStats.Mallocs
124 b.startBytes = memStats.TotalAlloc
130 // StopTimer stops timing a test. This can be used to pause the timer
131 // while performing complex initialization that you don't
133 func (b *B) StopTimer() {
135 b.duration += time.Since(b.start)
136 runtime.ReadMemStats(&memStats)
137 b.netAllocs += memStats.Mallocs - b.startAllocs
138 b.netBytes += memStats.TotalAlloc - b.startBytes
143 // ResetTimer zeroes the elapsed benchmark time and memory allocation counters
144 // and deletes user-reported metrics.
145 // It does not affect whether the timer is running.
146 func (b *B) ResetTimer() {
148 // Allocate the extra map before reading memory stats.
149 // Pre-size it to make more allocation unlikely.
150 b.extra = make(map[string]float64, 16)
152 for k := range b.extra {
157 runtime.ReadMemStats(&memStats)
158 b.startAllocs = memStats.Mallocs
159 b.startBytes = memStats.TotalAlloc
167 // SetBytes records the number of bytes processed in a single operation.
168 // If this is called, the benchmark will report ns/op and MB/s.
169 func (b *B) SetBytes(n int64) { b.bytes = n }
171 // ReportAllocs enables malloc statistics for this benchmark.
172 // It is equivalent to setting -test.benchmem, but it only affects the
173 // benchmark function that calls ReportAllocs.
174 func (b *B) ReportAllocs() {
175 b.showAllocResult = true
178 // runN runs a single benchmark for the specified number of iterations.
179 func (b *B) runN(n int) {
181 defer benchmarkLock.Unlock()
182 // Try to get a comparable environment for each run
183 // by clearing garbage from previous runs.
185 b.raceErrors = -race.Errors()
193 b.previousDuration = b.duration
194 b.raceErrors += race.Errors()
195 if b.raceErrors > 0 {
196 b.Errorf("race detected during execution of benchmark")
200 func min(x, y int64) int64 {
207 func max(x, y int64) int64 {
214 // run1 runs the first iteration of benchFunc. It reports whether more
215 // iterations of this benchmarks should be run.
216 func (b *B) run1() bool {
217 if ctx := b.context; ctx != nil {
218 // Extend maxLen, if needed.
219 if n := len(b.name) + ctx.extLen + 1; n > ctx.maxLen {
220 ctx.maxLen = n + 8 // Add additional slack to avoid too many jumps in size.
224 // Signal that we're done whether we return normally
225 // or by FailNow's runtime.Goexit.
234 fmt.Fprintf(b.w, "--- FAIL: %s\n%s", b.name, b.output)
237 // Only print the output if we know we are not going to proceed.
238 // Otherwise it is printed in processBench.
239 if atomic.LoadInt32(&b.hasSub) != 0 || b.finished {
244 if b.chatty && (len(b.output) > 0 || b.finished) {
246 fmt.Fprintf(b.w, "--- %s: %s\n%s", tag, b.name, b.output)
253 var labelsOnce sync.Once
255 // run executes the benchmark in a separate goroutine, including all of its
256 // subbenchmarks. b must not have subbenchmarks.
258 labelsOnce.Do(func() {
259 fmt.Fprintf(b.w, "goos: %s\n", runtime.GOOS)
260 fmt.Fprintf(b.w, "goarch: %s\n", runtime.GOARCH)
261 if b.importPath != "" {
262 fmt.Fprintf(b.w, "pkg: %s\n", b.importPath)
265 if b.context != nil {
266 // Running go test --test.bench
267 b.context.processBench(b) // Must call doBench.
269 // Running func Benchmark.
274 func (b *B) doBench() BenchmarkResult {
280 // launch launches the benchmark function. It gradually increases the number
281 // of benchmark iterations until the benchmark runs for the requested benchtime.
282 // launch is run by the doBench function as a separate goroutine.
283 // run1 must have been called on b.
284 func (b *B) launch() {
285 // Signal that we're done whether we return normally
286 // or by FailNow's runtime.Goexit.
291 // Run the benchmark for at least the specified amount of time.
292 if b.benchTime.n > 0 {
293 b.runN(b.benchTime.n)
296 for n := int64(1); !b.failed && b.duration < d && n < 1e9; {
298 // Predict required iterations.
299 goalns := d.Nanoseconds()
300 prevIters := int64(b.N)
301 prevns := b.duration.Nanoseconds()
303 // Round up, to avoid div by zero.
306 // Order of operations matters.
307 // For very fast benchmarks, prevIters ~= prevns.
308 // If you divide first, you get 0 or 1,
309 // which can hide an order of magnitude in execution time.
310 // So multiply first, then divide.
311 n = goalns * prevIters / prevns
312 // Run more iterations than we think we'll need (1.2x).
314 // Don't grow too fast in case we had timing errors previously.
316 // Be sure to run at least one more than last time.
318 // Don't run more than 1e9 times. (This also keeps n in int range on 32 bit platforms.)
323 b.result = BenchmarkResult{b.N, b.duration, b.bytes, b.netAllocs, b.netBytes, b.extra}
326 // ReportMetric adds "n unit" to the reported benchmark results.
327 // If the metric is per-iteration, the caller should divide by b.N,
328 // and by convention units should end in "/op".
329 // ReportMetric overrides any previously reported value for the same unit.
330 // ReportMetric panics if unit is the empty string or if unit contains
332 // If unit is a unit normally reported by the benchmark framework itself
333 // (such as "allocs/op"), ReportMetric will override that metric.
334 // Setting "ns/op" to 0 will suppress that built-in metric.
335 func (b *B) ReportMetric(n float64, unit string) {
337 panic("metric unit must not be empty")
339 if strings.IndexFunc(unit, unicode.IsSpace) >= 0 {
340 panic("metric unit must not contain whitespace")
345 // BenchmarkResult contains the results of a benchmark run.
346 type BenchmarkResult struct {
347 N int // The number of iterations.
348 T time.Duration // The total time taken.
349 Bytes int64 // Bytes processed in one iteration.
350 MemAllocs uint64 // The total number of memory allocations.
351 MemBytes uint64 // The total number of bytes allocated.
353 // Extra records additional metrics reported by ReportMetric.
354 Extra map[string]float64
357 // NsPerOp returns the "ns/op" metric.
358 func (r BenchmarkResult) NsPerOp() int64 {
359 if v, ok := r.Extra["ns/op"]; ok {
365 return r.T.Nanoseconds() / int64(r.N)
368 // mbPerSec returns the "MB/s" metric.
369 func (r BenchmarkResult) mbPerSec() float64 {
370 if v, ok := r.Extra["MB/s"]; ok {
373 if r.Bytes <= 0 || r.T <= 0 || r.N <= 0 {
376 return (float64(r.Bytes) * float64(r.N) / 1e6) / r.T.Seconds()
379 // AllocsPerOp returns the "allocs/op" metric,
380 // which is calculated as r.MemAllocs / r.N.
381 func (r BenchmarkResult) AllocsPerOp() int64 {
382 if v, ok := r.Extra["allocs/op"]; ok {
388 return int64(r.MemAllocs) / int64(r.N)
391 // AllocedBytesPerOp returns the "B/op" metric,
392 // which is calculated as r.MemBytes / r.N.
393 func (r BenchmarkResult) AllocedBytesPerOp() int64 {
394 if v, ok := r.Extra["B/op"]; ok {
400 return int64(r.MemBytes) / int64(r.N)
403 // String returns a summary of the benchmark results.
404 // It follows the benchmark result line format from
405 // https://golang.org/design/14313-benchmark-format, not including the
407 // Extra metrics override built-in metrics of the same name.
408 // String does not include allocs/op or B/op, since those are reported
410 func (r BenchmarkResult) String() string {
411 buf := new(strings.Builder)
412 fmt.Fprintf(buf, "%8d", r.N)
414 // Get ns/op as a float.
415 ns, ok := r.Extra["ns/op"]
417 ns = float64(r.T.Nanoseconds()) / float64(r.N)
421 prettyPrint(buf, ns, "ns/op")
424 if mbs := r.mbPerSec(); mbs != 0 {
425 fmt.Fprintf(buf, "\t%7.2f MB/s", mbs)
428 // Print extra metrics that aren't represented in the standard
430 var extraKeys []string
431 for k := range r.Extra {
433 case "ns/op", "MB/s", "B/op", "allocs/op":
434 // Built-in metrics reported elsewhere.
437 extraKeys = append(extraKeys, k)
439 sort.Strings(extraKeys)
440 for _, k := range extraKeys {
442 prettyPrint(buf, r.Extra[k], k)
447 func prettyPrint(w io.Writer, x float64, unit string) {
448 // Print all numbers with 10 places before the decimal point
449 // and small numbers with three sig figs.
451 switch y := math.Abs(x); {
452 case y == 0 || y >= 99.95:
467 fmt.Fprintf(w, format, x, unit)
470 // MemString returns r.AllocedBytesPerOp and r.AllocsPerOp in the same format as 'go test'.
471 func (r BenchmarkResult) MemString() string {
472 return fmt.Sprintf("%8d B/op\t%8d allocs/op",
473 r.AllocedBytesPerOp(), r.AllocsPerOp())
476 // benchmarkName returns full name of benchmark including procs suffix.
477 func benchmarkName(name string, n int) string {
479 return fmt.Sprintf("%s-%d", name, n)
484 type benchContext struct {
487 maxLen int // The largest recorded benchmark name.
488 extLen int // Maximum extension length.
491 // RunBenchmarks is an internal function but exported because it is cross-package;
492 // it is part of the implementation of the "go test" command.
493 func RunBenchmarks(matchString func(pat, str string) (bool, error), benchmarks []InternalBenchmark) {
494 runBenchmarks("", matchString, benchmarks)
497 func runBenchmarks(importPath string, matchString func(pat, str string) (bool, error), benchmarks []InternalBenchmark) bool {
498 // If no flag was specified, don't run benchmarks.
499 if len(*matchBenchmarks) == 0 {
502 // Collect matching benchmarks and determine longest name.
504 for _, procs := range cpuList {
505 if procs > maxprocs {
509 ctx := &benchContext{
510 match: newMatcher(matchString, *matchBenchmarks, "-test.bench"),
511 extLen: len(benchmarkName("", maxprocs)),
513 var bs []InternalBenchmark
514 for _, Benchmark := range benchmarks {
515 if _, matched, _ := ctx.match.fullName(nil, Benchmark.Name); matched {
516 bs = append(bs, Benchmark)
517 benchName := benchmarkName(Benchmark.Name, maxprocs)
518 if l := len(benchName) + ctx.extLen + 1; l > ctx.maxLen {
529 importPath: importPath,
530 benchFunc: func(b *B) {
531 for _, Benchmark := range bs {
532 b.Run(Benchmark.Name, Benchmark.F)
535 benchTime: benchTime,
542 // processBench runs bench b for the configured CPU counts and prints the results.
543 func (ctx *benchContext) processBench(b *B) {
544 for i, procs := range cpuList {
545 for j := uint(0); j < *count; j++ {
546 runtime.GOMAXPROCS(procs)
547 benchName := benchmarkName(b.name, procs)
549 // If it's chatty, we've already printed this information.
551 fmt.Fprintf(b.w, "%-*s\t", ctx.maxLen, benchName)
553 // Recompute the running time for all but the first iteration.
557 signal: make(chan bool),
562 benchFunc: b.benchFunc,
563 benchTime: b.benchTime,
569 // The output could be very long here, but probably isn't.
570 // We print it all, regardless, because we don't want to trim the reason
571 // the benchmark failed.
572 fmt.Fprintf(b.w, "--- FAIL: %s\n%s", benchName, b.output)
575 results := r.String()
577 fmt.Fprintf(b.w, "%-*s\t", ctx.maxLen, benchName)
579 if *benchmarkMemory || b.showAllocResult {
580 results += "\t" + r.MemString()
582 fmt.Fprintln(b.w, results)
583 // Unlike with tests, we ignore the -chatty flag and always print output for
584 // benchmarks since the output generation time will skew the results.
585 if len(b.output) > 0 {
587 fmt.Fprintf(b.w, "--- BENCH: %s\n%s", benchName, b.output)
589 if p := runtime.GOMAXPROCS(-1); p != procs {
590 fmt.Fprintf(os.Stderr, "testing: %s left GOMAXPROCS set to %d\n", benchName, p)
596 // Run benchmarks f as a subbenchmark with the given name. It reports
597 // whether there were any failures.
599 // A subbenchmark is like any other benchmark. A benchmark that calls Run at
600 // least once will not be measured itself and will be called once with N=1.
601 func (b *B) Run(name string, f func(b *B)) bool {
602 // Since b has subbenchmarks, we will no longer run it as a benchmark itself.
603 // Release the lock and acquire it on exit to ensure locks stay paired.
604 atomic.StoreInt32(&b.hasSub, 1)
605 benchmarkLock.Unlock()
606 defer benchmarkLock.Lock()
608 benchName, ok, partial := b.name, true, false
609 if b.context != nil {
610 benchName, ok, partial = b.context.match.fullName(&b.common, name)
615 var pc [maxStackLen]uintptr
616 n := runtime.Callers(2, pc[:])
619 signal: make(chan bool),
627 importPath: b.importPath,
629 benchTime: b.benchTime,
633 // Partial name match, like -bench=X/Y matching BenchmarkX.
634 // Only process sub-benchmarks, if any.
635 atomic.StoreInt32(&sub.hasSub, 1)
639 labelsOnce.Do(func() {
640 fmt.Printf("goos: %s\n", runtime.GOOS)
641 fmt.Printf("goarch: %s\n", runtime.GOARCH)
642 if b.importPath != "" {
643 fmt.Printf("pkg: %s\n", b.importPath)
647 fmt.Println(benchName)
657 // add simulates running benchmarks in sequence in a single iteration. It is
658 // used to give some meaningful results in case func Benchmark is used in
659 // combination with Run.
660 func (b *B) add(other BenchmarkResult) {
662 // The aggregated BenchmarkResults resemble running all subbenchmarks as
663 // in sequence in a single benchmark.
665 r.T += time.Duration(other.NsPerOp())
666 if other.Bytes == 0 {
667 // Summing Bytes is meaningless in aggregate if not all subbenchmarks
669 b.missingBytes = true
673 r.Bytes += other.Bytes
675 r.MemAllocs += uint64(other.AllocsPerOp())
676 r.MemBytes += uint64(other.AllocedBytesPerOp())
679 // trimOutput shortens the output from a benchmark, which can be very long.
680 func (b *B) trimOutput() {
681 // The output is likely to appear multiple times because the benchmark
682 // is run multiple times, but at least it will be seen. This is not a big deal
683 // because benchmarks rarely print, but just in case, we trim it if it's too long.
684 const maxNewlines = 10
685 for nlCount, j := 0, 0; j < len(b.output); j++ {
686 if b.output[j] == '\n' {
688 if nlCount >= maxNewlines {
689 b.output = append(b.output[:j], "\n\t... [output truncated]\n"...)
696 // A PB is used by RunParallel for running parallel benchmarks.
698 globalN *uint64 // shared between all worker goroutines iteration counter
699 grain uint64 // acquire that many iterations from globalN at once
700 cache uint64 // local cache of acquired iterations
701 bN uint64 // total number of iterations to execute (b.N)
704 // Next reports whether there are more iterations to execute.
705 func (pb *PB) Next() bool {
707 n := atomic.AddUint64(pb.globalN, pb.grain)
710 } else if n < pb.bN+pb.grain {
711 pb.cache = pb.bN + pb.grain - n
720 // RunParallel runs a benchmark in parallel.
721 // It creates multiple goroutines and distributes b.N iterations among them.
722 // The number of goroutines defaults to GOMAXPROCS. To increase parallelism for
723 // non-CPU-bound benchmarks, call SetParallelism before RunParallel.
724 // RunParallel is usually used with the go test -cpu flag.
726 // The body function will be run in each goroutine. It should set up any
727 // goroutine-local state and then iterate until pb.Next returns false.
728 // It should not use the StartTimer, StopTimer, or ResetTimer functions,
729 // because they have global effect. It should also not call Run.
730 func (b *B) RunParallel(body func(*PB)) {
732 return // Nothing to do when probing.
734 // Calculate grain size as number of iterations that take ~100µs.
735 // 100µs is enough to amortize the overhead and provide sufficient
736 // dynamic load balancing.
738 if b.previousN > 0 && b.previousDuration > 0 {
739 grain = 1e5 * uint64(b.previousN) / uint64(b.previousDuration)
744 // We expect the inner loop and function call to take at least 10ns,
745 // so do not do more than 100µs/10ns=1e4 iterations.
751 numProcs := b.parallelism * runtime.GOMAXPROCS(0)
752 var wg sync.WaitGroup
754 for p := 0; p < numProcs; p++ {
766 if n <= uint64(b.N) && !b.Failed() {
767 b.Fatal("RunParallel: body exited without pb.Next() == false")
771 // SetParallelism sets the number of goroutines used by RunParallel to p*GOMAXPROCS.
772 // There is usually no need to call SetParallelism for CPU-bound benchmarks.
773 // If p is less than 1, this call will have no effect.
774 func (b *B) SetParallelism(p int) {
780 // Benchmark benchmarks a single function. It is useful for creating
781 // custom benchmarks that do not use the "go test" command.
783 // If f depends on testing flags, then Init must be used to register
784 // those flags before calling Benchmark and before calling flag.Parse.
786 // If f calls Run, the result will be an estimate of running all its
787 // subbenchmarks that don't call Run in sequence in a single benchmark.
788 func Benchmark(f func(b *B)) BenchmarkResult {
791 signal: make(chan bool),
795 benchTime: benchTime,
803 type discard struct{}
805 func (discard) Write(b []byte) (n int, err error) { return len(b), nil }