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 var matchBenchmarks = flag.String("test.bench", "", "run only benchmarks matching `regexp`")
25 var benchTime = benchTimeFlag{d: 1 * time.Second}
26 var benchmarkMemory = flag.Bool("test.benchmem", false, "print memory allocations for benchmarks")
29 flag.Var(&benchTime, "test.benchtime", "run each benchmark for duration `d`")
32 type benchTimeFlag struct {
37 func (f *benchTimeFlag) String() string {
39 return fmt.Sprintf("%dx", f.n)
41 return time.Duration(f.d).String()
44 func (f *benchTimeFlag) Set(s string) error {
45 if strings.HasSuffix(s, "x") {
46 n, err := strconv.ParseInt(s[:len(s)-1], 10, 0)
47 if err != nil || n <= 0 {
48 return fmt.Errorf("invalid count")
50 *f = benchTimeFlag{n: int(n)}
53 d, err := time.ParseDuration(s)
54 if err != nil || d <= 0 {
55 return fmt.Errorf("invalid duration")
57 *f = benchTimeFlag{d: d}
61 // Global lock to ensure only one benchmark runs at a time.
62 var benchmarkLock sync.Mutex
64 // Used for every benchmark for measuring memory.
65 var memStats runtime.MemStats
67 // An internal type but exported because it is cross-package; part of the implementation
68 // of the "go test" command.
69 type InternalBenchmark struct {
74 // B is a type passed to Benchmark functions to manage benchmark
75 // timing and to specify the number of iterations to run.
77 // A benchmark ends when its Benchmark function returns or calls any of the methods
78 // FailNow, Fatal, Fatalf, SkipNow, Skip, or Skipf. Those methods must be called
79 // only from the goroutine running the Benchmark function.
80 // The other reporting methods, such as the variations of Log and Error,
81 // may be called simultaneously from multiple goroutines.
83 // Like in tests, benchmark logs are accumulated during execution
84 // and dumped to standard error when done. Unlike in tests, benchmark logs
85 // are always printed, so as not to hide output whose existence may be
86 // affecting benchmark results.
89 importPath string // import path of the package containing the benchmark
92 previousN int // number of iterations in the previous run
93 previousDuration time.Duration // total duration of the previous run
95 benchTime benchTimeFlag
97 missingBytes bool // one of the subbenchmarks does not have bytes set.
100 result BenchmarkResult
101 parallelism int // RunParallel creates parallelism*GOMAXPROCS goroutines
102 // The initial states of memStats.Mallocs and memStats.TotalAlloc.
105 // The net total of this test after being run.
108 // Extra metrics collected by ReportMetric.
109 extra map[string]float64
112 // StartTimer starts timing a test. This function is called automatically
113 // before a benchmark starts, but it can also be used to resume timing after
114 // a call to StopTimer.
115 func (b *B) StartTimer() {
117 runtime.ReadMemStats(&memStats)
118 b.startAllocs = memStats.Mallocs
119 b.startBytes = memStats.TotalAlloc
125 // StopTimer stops timing a test. This can be used to pause the timer
126 // while performing complex initialization that you don't
128 func (b *B) StopTimer() {
130 b.duration += time.Since(b.start)
131 runtime.ReadMemStats(&memStats)
132 b.netAllocs += memStats.Mallocs - b.startAllocs
133 b.netBytes += memStats.TotalAlloc - b.startBytes
138 // ResetTimer zeros the elapsed benchmark time and memory allocation counters
139 // and deletes user-reported metrics.
140 // It does not affect whether the timer is running.
141 func (b *B) ResetTimer() {
143 // Allocate the extra map before reading memory stats.
144 // Pre-size it to make more allocation unlikely.
145 b.extra = make(map[string]float64, 16)
147 for k := range b.extra {
152 runtime.ReadMemStats(&memStats)
153 b.startAllocs = memStats.Mallocs
154 b.startBytes = memStats.TotalAlloc
162 // SetBytes records the number of bytes processed in a single operation.
163 // If this is called, the benchmark will report ns/op and MB/s.
164 func (b *B) SetBytes(n int64) { b.bytes = n }
166 // ReportAllocs enables malloc statistics for this benchmark.
167 // It is equivalent to setting -test.benchmem, but it only affects the
168 // benchmark function that calls ReportAllocs.
169 func (b *B) ReportAllocs() {
170 b.showAllocResult = true
173 // runN runs a single benchmark for the specified number of iterations.
174 func (b *B) runN(n int) {
176 defer benchmarkLock.Unlock()
177 // Try to get a comparable environment for each run
178 // by clearing garbage from previous runs.
180 b.raceErrors = -race.Errors()
188 b.previousDuration = b.duration
189 b.raceErrors += race.Errors()
190 if b.raceErrors > 0 {
191 b.Errorf("race detected during execution of benchmark")
195 func min(x, y int64) int64 {
202 func max(x, y int64) int64 {
209 // run1 runs the first iteration of benchFunc. It reports whether more
210 // iterations of this benchmarks should be run.
211 func (b *B) run1() bool {
212 if ctx := b.context; ctx != nil {
213 // Extend maxLen, if needed.
214 if n := len(b.name) + ctx.extLen + 1; n > ctx.maxLen {
215 ctx.maxLen = n + 8 // Add additional slack to avoid too many jumps in size.
219 // Signal that we're done whether we return normally
220 // or by FailNow's runtime.Goexit.
229 fmt.Fprintf(b.w, "--- FAIL: %s\n%s", b.name, b.output)
232 // Only print the output if we know we are not going to proceed.
233 // Otherwise it is printed in processBench.
234 if atomic.LoadInt32(&b.hasSub) != 0 || b.finished {
239 if b.chatty && (len(b.output) > 0 || b.finished) {
241 fmt.Fprintf(b.w, "--- %s: %s\n%s", tag, b.name, b.output)
248 var labelsOnce sync.Once
250 // run executes the benchmark in a separate goroutine, including all of its
251 // subbenchmarks. b must not have subbenchmarks.
253 labelsOnce.Do(func() {
254 fmt.Fprintf(b.w, "goos: %s\n", runtime.GOOS)
255 fmt.Fprintf(b.w, "goarch: %s\n", runtime.GOARCH)
256 if b.importPath != "" {
257 fmt.Fprintf(b.w, "pkg: %s\n", b.importPath)
260 if b.context != nil {
261 // Running go test --test.bench
262 b.context.processBench(b) // Must call doBench.
264 // Running func Benchmark.
269 func (b *B) doBench() BenchmarkResult {
275 // launch launches the benchmark function. It gradually increases the number
276 // of benchmark iterations until the benchmark runs for the requested benchtime.
277 // launch is run by the doBench function as a separate goroutine.
278 // run1 must have been called on b.
279 func (b *B) launch() {
280 // Signal that we're done whether we return normally
281 // or by FailNow's runtime.Goexit.
286 // Run the benchmark for at least the specified amount of time.
287 if b.benchTime.n > 0 {
288 b.runN(b.benchTime.n)
291 for n := int64(1); !b.failed && b.duration < d && n < 1e9; {
293 // Predict required iterations.
294 goalns := d.Nanoseconds()
295 prevIters := int64(b.N)
296 prevns := b.duration.Nanoseconds()
298 // Round up, to avoid div by zero.
301 // Order of operations matters.
302 // For very fast benchmarks, prevIters ~= prevns.
303 // If you divide first, you get 0 or 1,
304 // which can hide an order of magnitude in execution time.
305 // So multiply first, then divide.
306 n = goalns * prevIters / prevns
307 // Run more iterations than we think we'll need (1.2x).
309 // Don't grow too fast in case we had timing errors previously.
311 // Be sure to run at least one more than last time.
313 // Don't run more than 1e9 times. (This also keeps n in int range on 32 bit platforms.)
318 b.result = BenchmarkResult{b.N, b.duration, b.bytes, b.netAllocs, b.netBytes, b.extra}
321 // ReportMetric adds "n unit" to the reported benchmark results.
322 // If the metric is per-iteration, the caller should divide by b.N,
323 // and by convention units should end in "/op".
324 // ReportMetric overrides any previously reported value for the same unit.
325 // ReportMetric panics if unit is the empty string or if unit contains
327 // If unit is a unit normally reported by the benchmark framework itself
328 // (such as "allocs/op"), ReportMetric will override that metric.
329 // Setting "ns/op" to 0 will suppress that built-in metric.
330 func (b *B) ReportMetric(n float64, unit string) {
332 panic("metric unit must not be empty")
334 if strings.IndexFunc(unit, unicode.IsSpace) >= 0 {
335 panic("metric unit must not contain whitespace")
340 // The results of a benchmark run.
341 type BenchmarkResult struct {
342 N int // The number of iterations.
343 T time.Duration // The total time taken.
344 Bytes int64 // Bytes processed in one iteration.
345 MemAllocs uint64 // The total number of memory allocations.
346 MemBytes uint64 // The total number of bytes allocated.
348 // Extra records additional metrics reported by ReportMetric.
349 Extra map[string]float64
352 // NsPerOp returns the "ns/op" metric.
353 func (r BenchmarkResult) NsPerOp() int64 {
354 if v, ok := r.Extra["ns/op"]; ok {
360 return r.T.Nanoseconds() / int64(r.N)
363 // mbPerSec returns the "MB/s" metric.
364 func (r BenchmarkResult) mbPerSec() float64 {
365 if v, ok := r.Extra["MB/s"]; ok {
368 if r.Bytes <= 0 || r.T <= 0 || r.N <= 0 {
371 return (float64(r.Bytes) * float64(r.N) / 1e6) / r.T.Seconds()
374 // AllocsPerOp returns the "allocs/op" metric,
375 // which is calculated as r.MemAllocs / r.N.
376 func (r BenchmarkResult) AllocsPerOp() int64 {
377 if v, ok := r.Extra["allocs/op"]; ok {
383 return int64(r.MemAllocs) / int64(r.N)
386 // AllocedBytesPerOp returns the "B/op" metric,
387 // which is calculated as r.MemBytes / r.N.
388 func (r BenchmarkResult) AllocedBytesPerOp() int64 {
389 if v, ok := r.Extra["B/op"]; ok {
395 return int64(r.MemBytes) / int64(r.N)
398 // String returns a summary of the benchmark results.
399 // It follows the benchmark result line format from
400 // https://golang.org/design/14313-benchmark-format, not including the
402 // Extra metrics override built-in metrics of the same name.
403 // String does not include allocs/op or B/op, since those are reported
405 func (r BenchmarkResult) String() string {
406 buf := new(strings.Builder)
407 fmt.Fprintf(buf, "%8d", r.N)
409 // Get ns/op as a float.
410 ns, ok := r.Extra["ns/op"]
412 ns = float64(r.T.Nanoseconds()) / float64(r.N)
416 prettyPrint(buf, ns, "ns/op")
419 if mbs := r.mbPerSec(); mbs != 0 {
420 fmt.Fprintf(buf, "\t%7.2f MB/s", mbs)
423 // Print extra metrics that aren't represented in the standard
425 var extraKeys []string
426 for k := range r.Extra {
428 case "ns/op", "MB/s", "B/op", "allocs/op":
429 // Built-in metrics reported elsewhere.
432 extraKeys = append(extraKeys, k)
434 sort.Strings(extraKeys)
435 for _, k := range extraKeys {
437 prettyPrint(buf, r.Extra[k], k)
442 func prettyPrint(w io.Writer, x float64, unit string) {
443 // Print all numbers with 10 places before the decimal point
444 // and small numbers with three sig figs.
446 switch y := math.Abs(x); {
447 case y == 0 || y >= 99.95:
462 fmt.Fprintf(w, format, x, unit)
465 // MemString returns r.AllocedBytesPerOp and r.AllocsPerOp in the same format as 'go test'.
466 func (r BenchmarkResult) MemString() string {
467 return fmt.Sprintf("%8d B/op\t%8d allocs/op",
468 r.AllocedBytesPerOp(), r.AllocsPerOp())
471 // benchmarkName returns full name of benchmark including procs suffix.
472 func benchmarkName(name string, n int) string {
474 return fmt.Sprintf("%s-%d", name, n)
479 type benchContext struct {
482 maxLen int // The largest recorded benchmark name.
483 extLen int // Maximum extension length.
486 // An internal function but exported because it is cross-package; part of the implementation
487 // of the "go test" command.
488 func RunBenchmarks(matchString func(pat, str string) (bool, error), benchmarks []InternalBenchmark) {
489 runBenchmarks("", matchString, benchmarks)
492 func runBenchmarks(importPath string, matchString func(pat, str string) (bool, error), benchmarks []InternalBenchmark) bool {
493 // If no flag was specified, don't run benchmarks.
494 if len(*matchBenchmarks) == 0 {
497 // Collect matching benchmarks and determine longest name.
499 for _, procs := range cpuList {
500 if procs > maxprocs {
504 ctx := &benchContext{
505 match: newMatcher(matchString, *matchBenchmarks, "-test.bench"),
506 extLen: len(benchmarkName("", maxprocs)),
508 var bs []InternalBenchmark
509 for _, Benchmark := range benchmarks {
510 if _, matched, _ := ctx.match.fullName(nil, Benchmark.Name); matched {
511 bs = append(bs, Benchmark)
512 benchName := benchmarkName(Benchmark.Name, maxprocs)
513 if l := len(benchName) + ctx.extLen + 1; l > ctx.maxLen {
524 importPath: importPath,
525 benchFunc: func(b *B) {
526 for _, Benchmark := range bs {
527 b.Run(Benchmark.Name, Benchmark.F)
530 benchTime: benchTime,
537 // processBench runs bench b for the configured CPU counts and prints the results.
538 func (ctx *benchContext) processBench(b *B) {
539 for i, procs := range cpuList {
540 for j := uint(0); j < *count; j++ {
541 runtime.GOMAXPROCS(procs)
542 benchName := benchmarkName(b.name, procs)
543 fmt.Fprintf(b.w, "%-*s\t", ctx.maxLen, benchName)
544 // Recompute the running time for all but the first iteration.
548 signal: make(chan bool),
553 benchFunc: b.benchFunc,
554 benchTime: b.benchTime,
560 // The output could be very long here, but probably isn't.
561 // We print it all, regardless, because we don't want to trim the reason
562 // the benchmark failed.
563 fmt.Fprintf(b.w, "--- FAIL: %s\n%s", benchName, b.output)
566 results := r.String()
567 if *benchmarkMemory || b.showAllocResult {
568 results += "\t" + r.MemString()
570 fmt.Fprintln(b.w, results)
571 // Unlike with tests, we ignore the -chatty flag and always print output for
572 // benchmarks since the output generation time will skew the results.
573 if len(b.output) > 0 {
575 fmt.Fprintf(b.w, "--- BENCH: %s\n%s", benchName, b.output)
577 if p := runtime.GOMAXPROCS(-1); p != procs {
578 fmt.Fprintf(os.Stderr, "testing: %s left GOMAXPROCS set to %d\n", benchName, p)
584 // Run benchmarks f as a subbenchmark with the given name. It reports
585 // whether there were any failures.
587 // A subbenchmark is like any other benchmark. A benchmark that calls Run at
588 // least once will not be measured itself and will be called once with N=1.
589 func (b *B) Run(name string, f func(b *B)) bool {
590 // Since b has subbenchmarks, we will no longer run it as a benchmark itself.
591 // Release the lock and acquire it on exit to ensure locks stay paired.
592 atomic.StoreInt32(&b.hasSub, 1)
593 benchmarkLock.Unlock()
594 defer benchmarkLock.Lock()
596 benchName, ok, partial := b.name, true, false
597 if b.context != nil {
598 benchName, ok, partial = b.context.match.fullName(&b.common, name)
603 var pc [maxStackLen]uintptr
604 n := runtime.Callers(2, pc[:])
607 signal: make(chan bool),
615 importPath: b.importPath,
617 benchTime: b.benchTime,
621 // Partial name match, like -bench=X/Y matching BenchmarkX.
622 // Only process sub-benchmarks, if any.
623 atomic.StoreInt32(&sub.hasSub, 1)
632 // add simulates running benchmarks in sequence in a single iteration. It is
633 // used to give some meaningful results in case func Benchmark is used in
634 // combination with Run.
635 func (b *B) add(other BenchmarkResult) {
637 // The aggregated BenchmarkResults resemble running all subbenchmarks as
638 // in sequence in a single benchmark.
640 r.T += time.Duration(other.NsPerOp())
641 if other.Bytes == 0 {
642 // Summing Bytes is meaningless in aggregate if not all subbenchmarks
644 b.missingBytes = true
648 r.Bytes += other.Bytes
650 r.MemAllocs += uint64(other.AllocsPerOp())
651 r.MemBytes += uint64(other.AllocedBytesPerOp())
654 // trimOutput shortens the output from a benchmark, which can be very long.
655 func (b *B) trimOutput() {
656 // The output is likely to appear multiple times because the benchmark
657 // is run multiple times, but at least it will be seen. This is not a big deal
658 // because benchmarks rarely print, but just in case, we trim it if it's too long.
659 const maxNewlines = 10
660 for nlCount, j := 0, 0; j < len(b.output); j++ {
661 if b.output[j] == '\n' {
663 if nlCount >= maxNewlines {
664 b.output = append(b.output[:j], "\n\t... [output truncated]\n"...)
671 // A PB is used by RunParallel for running parallel benchmarks.
673 globalN *uint64 // shared between all worker goroutines iteration counter
674 grain uint64 // acquire that many iterations from globalN at once
675 cache uint64 // local cache of acquired iterations
676 bN uint64 // total number of iterations to execute (b.N)
679 // Next reports whether there are more iterations to execute.
680 func (pb *PB) Next() bool {
682 n := atomic.AddUint64(pb.globalN, pb.grain)
685 } else if n < pb.bN+pb.grain {
686 pb.cache = pb.bN + pb.grain - n
695 // RunParallel runs a benchmark in parallel.
696 // It creates multiple goroutines and distributes b.N iterations among them.
697 // The number of goroutines defaults to GOMAXPROCS. To increase parallelism for
698 // non-CPU-bound benchmarks, call SetParallelism before RunParallel.
699 // RunParallel is usually used with the go test -cpu flag.
701 // The body function will be run in each goroutine. It should set up any
702 // goroutine-local state and then iterate until pb.Next returns false.
703 // It should not use the StartTimer, StopTimer, or ResetTimer functions,
704 // because they have global effect. It should also not call Run.
705 func (b *B) RunParallel(body func(*PB)) {
707 return // Nothing to do when probing.
709 // Calculate grain size as number of iterations that take ~100µs.
710 // 100µs is enough to amortize the overhead and provide sufficient
711 // dynamic load balancing.
713 if b.previousN > 0 && b.previousDuration > 0 {
714 grain = 1e5 * uint64(b.previousN) / uint64(b.previousDuration)
719 // We expect the inner loop and function call to take at least 10ns,
720 // so do not do more than 100µs/10ns=1e4 iterations.
726 numProcs := b.parallelism * runtime.GOMAXPROCS(0)
727 var wg sync.WaitGroup
729 for p := 0; p < numProcs; p++ {
741 if n <= uint64(b.N) && !b.Failed() {
742 b.Fatal("RunParallel: body exited without pb.Next() == false")
746 // SetParallelism sets the number of goroutines used by RunParallel to p*GOMAXPROCS.
747 // There is usually no need to call SetParallelism for CPU-bound benchmarks.
748 // If p is less than 1, this call will have no effect.
749 func (b *B) SetParallelism(p int) {
755 // Benchmark benchmarks a single function. Useful for creating
756 // custom benchmarks that do not use the "go test" command.
758 // If f calls Run, the result will be an estimate of running all its
759 // subbenchmarks that don't call Run in sequence in a single benchmark.
760 func Benchmark(f func(b *B)) BenchmarkResult {
763 signal: make(chan bool),
767 benchTime: benchTime,
775 type discard struct{}
777 func (discard) Write(b []byte) (n int, err error) { return len(b), nil }