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.
18 var matchBenchmarks = flag.String("test.bench", "", "run only benchmarks matching `regexp`")
19 var benchTime = flag.Duration("test.benchtime", 1*time.Second, "run each benchmark for duration `d`")
20 var benchmarkMemory = flag.Bool("test.benchmem", false, "print memory allocations for benchmarks")
22 // Global lock to ensure only one benchmark runs at a time.
23 var benchmarkLock sync.Mutex
25 // Used for every benchmark for measuring memory.
26 var memStats runtime.MemStats
28 // An internal type but exported because it is cross-package; part of the implementation
29 // of the "go test" command.
30 type InternalBenchmark struct {
35 // B is a type passed to Benchmark functions to manage benchmark
36 // timing and to specify the number of iterations to run.
38 // A benchmark ends when its Benchmark function returns or calls any of the methods
39 // FailNow, Fatal, Fatalf, SkipNow, Skip, or Skipf. Those methods must be called
40 // only from the goroutine running the Benchmark function.
41 // The other reporting methods, such as the variations of Log and Error,
42 // may be called simultaneously from multiple goroutines.
44 // Like in tests, benchmark logs are accumulated during execution
45 // and dumped to standard error when done. Unlike in tests, benchmark logs
46 // are always printed, so as not to hide output whose existence may be
47 // affecting benchmark results.
50 importPath string // import path of the package containing the benchmark
53 previousN int // number of iterations in the previous run
54 previousDuration time.Duration // total duration of the previous run
56 benchTime time.Duration
58 missingBytes bool // one of the subbenchmarks does not have bytes set.
61 result BenchmarkResult
62 parallelism int // RunParallel creates parallelism*GOMAXPROCS goroutines
63 // The initial states of memStats.Mallocs and memStats.TotalAlloc.
66 // The net total of this test after being run.
71 // StartTimer starts timing a test. This function is called automatically
72 // before a benchmark starts, but it can also used to resume timing after
73 // a call to StopTimer.
74 func (b *B) StartTimer() {
76 runtime.ReadMemStats(&memStats)
77 b.startAllocs = memStats.Mallocs
78 b.startBytes = memStats.TotalAlloc
84 // StopTimer stops timing a test. This can be used to pause the timer
85 // while performing complex initialization that you don't
87 func (b *B) StopTimer() {
89 b.duration += time.Since(b.start)
90 runtime.ReadMemStats(&memStats)
91 b.netAllocs += memStats.Mallocs - b.startAllocs
92 b.netBytes += memStats.TotalAlloc - b.startBytes
97 // ResetTimer zeros the elapsed benchmark time and memory allocation counters.
98 // It does not affect whether the timer is running.
99 func (b *B) ResetTimer() {
101 runtime.ReadMemStats(&memStats)
102 b.startAllocs = memStats.Mallocs
103 b.startBytes = memStats.TotalAlloc
111 // SetBytes records the number of bytes processed in a single operation.
112 // If this is called, the benchmark will report ns/op and MB/s.
113 func (b *B) SetBytes(n int64) { b.bytes = n }
115 // ReportAllocs enables malloc statistics for this benchmark.
116 // It is equivalent to setting -test.benchmem, but it only affects the
117 // benchmark function that calls ReportAllocs.
118 func (b *B) ReportAllocs() {
119 b.showAllocResult = true
122 func (b *B) nsPerOp() int64 {
126 return b.duration.Nanoseconds() / int64(b.N)
129 // runN runs a single benchmark for the specified number of iterations.
130 func (b *B) runN(n int) {
132 defer benchmarkLock.Unlock()
133 // Try to get a comparable environment for each run
134 // by clearing garbage from previous runs.
136 b.raceErrors = -race.Errors()
144 b.previousDuration = b.duration
145 b.raceErrors += race.Errors()
146 if b.raceErrors > 0 {
147 b.Errorf("race detected during execution of benchmark")
151 func min(x, y int) int {
158 func max(x, y int) int {
165 // roundDown10 rounds a number down to the nearest power of 10.
166 func roundDown10(n int) int {
168 // tens = floor(log_10(n))
175 for i := 0; i < tens; i++ {
181 // roundUp rounds x up to a number of the form [1eX, 2eX, 3eX, 5eX].
182 func roundUp(n int) int {
183 base := roundDown10(n)
187 case n <= (2 * base):
189 case n <= (3 * base):
191 case n <= (5 * base):
198 // run1 runs the first iteration of benchFunc. It returns whether more
199 // iterations of this benchmarks should be run.
200 func (b *B) run1() bool {
201 if ctx := b.context; ctx != nil {
202 // Extend maxLen, if needed.
203 if n := len(b.name) + ctx.extLen + 1; n > ctx.maxLen {
204 ctx.maxLen = n + 8 // Add additional slack to avoid too many jumps in size.
208 // Signal that we're done whether we return normally
209 // or by FailNow's runtime.Goexit.
218 fmt.Fprintf(b.w, "--- FAIL: %s\n%s", b.name, b.output)
221 // Only print the output if we know we are not going to proceed.
222 // Otherwise it is printed in processBench.
223 if atomic.LoadInt32(&b.hasSub) != 0 || b.finished {
228 if b.chatty && (len(b.output) > 0 || b.finished) {
230 fmt.Fprintf(b.w, "--- %s: %s\n%s", tag, b.name, b.output)
237 var labelsOnce sync.Once
239 // run executes the benchmark in a separate goroutine, including all of its
240 // subbenchmarks. b must not have subbenchmarks.
242 labelsOnce.Do(func() {
243 fmt.Fprintf(b.w, "goos: %s\n", runtime.GOOS)
244 fmt.Fprintf(b.w, "goarch: %s\n", runtime.GOARCH)
245 if b.importPath != "" {
246 fmt.Fprintf(b.w, "pkg: %s\n", b.importPath)
249 if b.context != nil {
250 // Running go test --test.bench
251 b.context.processBench(b) // Must call doBench.
253 // Running func Benchmark.
258 func (b *B) doBench(hint int) BenchmarkResult {
264 // autodetectN runs the benchmark function, gradually increasing the
265 // number of iterations until the benchmark runs for the requested
267 func (b *B) autodetectN() {
268 // Run the benchmark for at least the specified amount of time.
270 for n := 1; !b.failed && b.duration < d && n < 1e9; {
272 // Predict required iterations.
273 n = int(d.Nanoseconds())
274 if nsop := b.nsPerOp(); nsop != 0 {
277 // Run more iterations than we think we'll need (1.2x).
278 // Don't grow too fast in case we had timing errors previously.
279 // Be sure to run at least one more than last time.
280 n = max(min(n+n/5, 100*last), last+1)
281 // Round up to something easy to read.
287 // launch launches the benchmark function for hintN iterations. If
288 // hintN == 0, it autodetects the number of benchmark iterations based
289 // on the requested benchtime.
290 // launch is run by the doBench function as a separate goroutine.
291 // run1 must have been called on b.
292 func (b *B) launch(hintN int) {
293 // Signal that we're done whether we return normally
294 // or by FailNow's runtime.Goexit.
305 b.result = BenchmarkResult{b.N, b.duration, b.bytes, b.netAllocs, b.netBytes}
308 // The results of a benchmark run.
309 type BenchmarkResult struct {
310 N int // The number of iterations.
311 T time.Duration // The total time taken.
312 Bytes int64 // Bytes processed in one iteration.
313 MemAllocs uint64 // The total number of memory allocations.
314 MemBytes uint64 // The total number of bytes allocated.
317 func (r BenchmarkResult) NsPerOp() int64 {
321 return r.T.Nanoseconds() / int64(r.N)
324 func (r BenchmarkResult) mbPerSec() float64 {
325 if r.Bytes <= 0 || r.T <= 0 || r.N <= 0 {
328 return (float64(r.Bytes) * float64(r.N) / 1e6) / r.T.Seconds()
331 // AllocsPerOp returns r.MemAllocs / r.N.
332 func (r BenchmarkResult) AllocsPerOp() int64 {
336 return int64(r.MemAllocs) / int64(r.N)
339 // AllocedBytesPerOp returns r.MemBytes / r.N.
340 func (r BenchmarkResult) AllocedBytesPerOp() int64 {
344 return int64(r.MemBytes) / int64(r.N)
347 func (r BenchmarkResult) String() string {
351 mb = fmt.Sprintf("\t%7.2f MB/s", mbs)
354 ns := fmt.Sprintf("%10d ns/op", nsop)
355 if r.N > 0 && nsop < 100 {
356 // The format specifiers here make sure that
357 // the ones digits line up for all three possible formats.
359 ns = fmt.Sprintf("%13.2f ns/op", float64(r.T.Nanoseconds())/float64(r.N))
361 ns = fmt.Sprintf("%12.1f ns/op", float64(r.T.Nanoseconds())/float64(r.N))
364 return fmt.Sprintf("%8d\t%s%s", r.N, ns, mb)
367 // MemString returns r.AllocedBytesPerOp and r.AllocsPerOp in the same format as 'go test'.
368 func (r BenchmarkResult) MemString() string {
369 return fmt.Sprintf("%8d B/op\t%8d allocs/op",
370 r.AllocedBytesPerOp(), r.AllocsPerOp())
373 // benchmarkName returns full name of benchmark including procs suffix.
374 func benchmarkName(name string, n int) string {
376 return fmt.Sprintf("%s-%d", name, n)
381 type benchContext struct {
384 maxLen int // The largest recorded benchmark name.
385 extLen int // Maximum extension length.
388 // An internal function but exported because it is cross-package; part of the implementation
389 // of the "go test" command.
390 func RunBenchmarks(matchString func(pat, str string) (bool, error), benchmarks []InternalBenchmark) {
391 runBenchmarks("", matchString, benchmarks)
394 func runBenchmarks(importPath string, matchString func(pat, str string) (bool, error), benchmarks []InternalBenchmark) bool {
395 // If no flag was specified, don't run benchmarks.
396 if len(*matchBenchmarks) == 0 {
399 // Collect matching benchmarks and determine longest name.
401 for _, procs := range cpuList {
402 if procs > maxprocs {
406 ctx := &benchContext{
407 match: newMatcher(matchString, *matchBenchmarks, "-test.bench"),
408 extLen: len(benchmarkName("", maxprocs)),
410 var bs []InternalBenchmark
411 for _, Benchmark := range benchmarks {
412 if _, matched, _ := ctx.match.fullName(nil, Benchmark.Name); matched {
413 bs = append(bs, Benchmark)
414 benchName := benchmarkName(Benchmark.Name, maxprocs)
415 if l := len(benchName) + ctx.extLen + 1; l > ctx.maxLen {
426 importPath: importPath,
427 benchFunc: func(b *B) {
428 for _, Benchmark := range bs {
429 b.Run(Benchmark.Name, Benchmark.F)
432 benchTime: *benchTime,
439 // processBench runs bench b for the configured CPU counts and prints the results.
440 func (ctx *benchContext) processBench(b *B) {
441 for i, procs := range cpuList {
443 for j := uint(0); j < *count; j++ {
444 runtime.GOMAXPROCS(procs)
445 benchName := benchmarkName(b.name, procs)
446 fmt.Fprintf(b.w, "%-*s\t", ctx.maxLen, benchName)
447 // Recompute the running time for all but the first iteration.
451 signal: make(chan bool),
456 benchFunc: b.benchFunc,
457 benchTime: b.benchTime,
461 r := b.doBench(nHint)
466 // The output could be very long here, but probably isn't.
467 // We print it all, regardless, because we don't want to trim the reason
468 // the benchmark failed.
469 fmt.Fprintf(b.w, "--- FAIL: %s\n%s", benchName, b.output)
472 results := r.String()
473 if *benchmarkMemory || b.showAllocResult {
474 results += "\t" + r.MemString()
476 fmt.Fprintln(b.w, results)
477 // Unlike with tests, we ignore the -chatty flag and always print output for
478 // benchmarks since the output generation time will skew the results.
479 if len(b.output) > 0 {
481 fmt.Fprintf(b.w, "--- BENCH: %s\n%s", benchName, b.output)
483 if p := runtime.GOMAXPROCS(-1); p != procs {
484 fmt.Fprintf(os.Stderr, "testing: %s left GOMAXPROCS set to %d\n", benchName, p)
490 // Run benchmarks f as a subbenchmark with the given name. It reports
491 // whether there were any failures.
493 // A subbenchmark is like any other benchmark. A benchmark that calls Run at
494 // least once will not be measured itself and will be called once with N=1.
495 func (b *B) Run(name string, f func(b *B)) bool {
496 // Since b has subbenchmarks, we will no longer run it as a benchmark itself.
497 // Release the lock and acquire it on exit to ensure locks stay paired.
498 atomic.StoreInt32(&b.hasSub, 1)
499 benchmarkLock.Unlock()
500 defer benchmarkLock.Lock()
502 benchName, ok, partial := b.name, true, false
503 if b.context != nil {
504 benchName, ok, partial = b.context.match.fullName(&b.common, name)
511 signal: make(chan bool),
518 importPath: b.importPath,
520 benchTime: b.benchTime,
524 // Partial name match, like -bench=X/Y matching BenchmarkX.
525 // Only process sub-benchmarks, if any.
526 atomic.StoreInt32(&sub.hasSub, 1)
535 // add simulates running benchmarks in sequence in a single iteration. It is
536 // used to give some meaningful results in case func Benchmark is used in
537 // combination with Run.
538 func (b *B) add(other BenchmarkResult) {
540 // The aggregated BenchmarkResults resemble running all subbenchmarks as
541 // in sequence in a single benchmark.
543 r.T += time.Duration(other.NsPerOp())
544 if other.Bytes == 0 {
545 // Summing Bytes is meaningless in aggregate if not all subbenchmarks
547 b.missingBytes = true
551 r.Bytes += other.Bytes
553 r.MemAllocs += uint64(other.AllocsPerOp())
554 r.MemBytes += uint64(other.AllocedBytesPerOp())
557 // trimOutput shortens the output from a benchmark, which can be very long.
558 func (b *B) trimOutput() {
559 // The output is likely to appear multiple times because the benchmark
560 // is run multiple times, but at least it will be seen. This is not a big deal
561 // because benchmarks rarely print, but just in case, we trim it if it's too long.
562 const maxNewlines = 10
563 for nlCount, j := 0, 0; j < len(b.output); j++ {
564 if b.output[j] == '\n' {
566 if nlCount >= maxNewlines {
567 b.output = append(b.output[:j], "\n\t... [output truncated]\n"...)
574 // A PB is used by RunParallel for running parallel benchmarks.
576 globalN *uint64 // shared between all worker goroutines iteration counter
577 grain uint64 // acquire that many iterations from globalN at once
578 cache uint64 // local cache of acquired iterations
579 bN uint64 // total number of iterations to execute (b.N)
582 // Next reports whether there are more iterations to execute.
583 func (pb *PB) Next() bool {
585 n := atomic.AddUint64(pb.globalN, pb.grain)
588 } else if n < pb.bN+pb.grain {
589 pb.cache = pb.bN + pb.grain - n
598 // RunParallel runs a benchmark in parallel.
599 // It creates multiple goroutines and distributes b.N iterations among them.
600 // The number of goroutines defaults to GOMAXPROCS. To increase parallelism for
601 // non-CPU-bound benchmarks, call SetParallelism before RunParallel.
602 // RunParallel is usually used with the go test -cpu flag.
604 // The body function will be run in each goroutine. It should set up any
605 // goroutine-local state and then iterate until pb.Next returns false.
606 // It should not use the StartTimer, StopTimer, or ResetTimer functions,
607 // because they have global effect. It should also not call Run.
608 func (b *B) RunParallel(body func(*PB)) {
610 return // Nothing to do when probing.
612 // Calculate grain size as number of iterations that take ~100µs.
613 // 100µs is enough to amortize the overhead and provide sufficient
614 // dynamic load balancing.
616 if b.previousN > 0 && b.previousDuration > 0 {
617 grain = 1e5 * uint64(b.previousN) / uint64(b.previousDuration)
622 // We expect the inner loop and function call to take at least 10ns,
623 // so do not do more than 100µs/10ns=1e4 iterations.
629 numProcs := b.parallelism * runtime.GOMAXPROCS(0)
630 var wg sync.WaitGroup
632 for p := 0; p < numProcs; p++ {
644 if n <= uint64(b.N) && !b.Failed() {
645 b.Fatal("RunParallel: body exited without pb.Next() == false")
649 // SetParallelism sets the number of goroutines used by RunParallel to p*GOMAXPROCS.
650 // There is usually no need to call SetParallelism for CPU-bound benchmarks.
651 // If p is less than 1, this call will have no effect.
652 func (b *B) SetParallelism(p int) {
658 // Benchmark benchmarks a single function. Useful for creating
659 // custom benchmarks that do not use the "go test" command.
661 // If f calls Run, the result will be an estimate of running all its
662 // subbenchmarks that don't call Run in sequence in a single benchmark.
663 func Benchmark(f func(b *B)) BenchmarkResult {
666 signal: make(chan bool),
670 benchTime: *benchTime,
678 type discard struct{}
680 func (discard) Write(b []byte) (n int, err error) { return len(b), nil }