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 if *benchmarkMemory || b.showAllocResult {
77 runtime.ReadMemStats(&memStats)
78 b.startAllocs = memStats.Mallocs
79 b.startBytes = memStats.TotalAlloc
86 // StopTimer stops timing a test. This can be used to pause the timer
87 // while performing complex initialization that you don't
89 func (b *B) StopTimer() {
91 b.duration += time.Now().Sub(b.start)
92 if *benchmarkMemory || b.showAllocResult {
93 runtime.ReadMemStats(&memStats)
94 b.netAllocs += memStats.Mallocs - b.startAllocs
95 b.netBytes += memStats.TotalAlloc - b.startBytes
101 // ResetTimer zeros the elapsed benchmark time and memory allocation counters.
102 // It does not affect whether the timer is running.
103 func (b *B) ResetTimer() {
105 if *benchmarkMemory || b.showAllocResult {
106 runtime.ReadMemStats(&memStats)
107 b.startAllocs = memStats.Mallocs
108 b.startBytes = memStats.TotalAlloc
117 // SetBytes records the number of bytes processed in a single operation.
118 // If this is called, the benchmark will report ns/op and MB/s.
119 func (b *B) SetBytes(n int64) { b.bytes = n }
121 // ReportAllocs enables malloc statistics for this benchmark.
122 // It is equivalent to setting -test.benchmem, but it only affects the
123 // benchmark function that calls ReportAllocs.
124 func (b *B) ReportAllocs() {
125 b.showAllocResult = true
128 func (b *B) nsPerOp() int64 {
132 return b.duration.Nanoseconds() / int64(b.N)
135 // runN runs a single benchmark for the specified number of iterations.
136 func (b *B) runN(n int) {
138 defer benchmarkLock.Unlock()
139 // Try to get a comparable environment for each run
140 // by clearing garbage from previous runs.
142 b.raceErrors = -race.Errors()
150 b.previousDuration = b.duration
151 b.raceErrors += race.Errors()
152 if b.raceErrors > 0 {
153 b.Errorf("race detected during execution of benchmark")
157 func min(x, y int) int {
164 func max(x, y int) int {
171 // roundDown10 rounds a number down to the nearest power of 10.
172 func roundDown10(n int) int {
174 // tens = floor(log_10(n))
181 for i := 0; i < tens; i++ {
187 // roundUp rounds x up to a number of the form [1eX, 2eX, 3eX, 5eX].
188 func roundUp(n int) int {
189 base := roundDown10(n)
193 case n <= (2 * base):
195 case n <= (3 * base):
197 case n <= (5 * base):
204 // run1 runs the first iteration of benchFunc. It returns whether more
205 // iterations of this benchmarks should be run.
206 func (b *B) run1() bool {
207 if ctx := b.context; ctx != nil {
208 // Extend maxLen, if needed.
209 if n := len(b.name) + ctx.extLen + 1; n > ctx.maxLen {
210 ctx.maxLen = n + 8 // Add additional slack to avoid too many jumps in size.
214 // Signal that we're done whether we return normally
215 // or by FailNow's runtime.Goexit.
224 fmt.Fprintf(b.w, "--- FAIL: %s\n%s", b.name, b.output)
227 // Only print the output if we know we are not going to proceed.
228 // Otherwise it is printed in processBench.
229 if atomic.LoadInt32(&b.hasSub) != 0 || b.finished {
234 if b.chatty && (len(b.output) > 0 || b.finished) {
236 fmt.Fprintf(b.w, "--- %s: %s\n%s", tag, b.name, b.output)
243 var labelsOnce sync.Once
245 // run executes the benchmark in a separate goroutine, including all of its
246 // subbenchmarks. b must not have subbenchmarks.
247 func (b *B) run() BenchmarkResult {
248 labelsOnce.Do(func() {
249 fmt.Fprintf(b.w, "goos: %s\n", runtime.GOOS)
250 fmt.Fprintf(b.w, "goarch: %s\n", runtime.GOARCH)
251 if b.importPath != "" {
252 fmt.Fprintf(b.w, "pkg: %s\n", b.importPath)
255 if b.context != nil {
256 // Running go test --test.bench
257 b.context.processBench(b) // Must call doBench.
259 // 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.
284 for n := 1; !b.failed && b.duration < d && n < 1e9; {
286 // Predict required iterations.
287 n = int(d.Nanoseconds())
288 if nsop := b.nsPerOp(); nsop != 0 {
291 // Run more iterations than we think we'll need (1.2x).
292 // Don't grow too fast in case we had timing errors previously.
293 // Be sure to run at least one more than last time.
294 n = max(min(n+n/5, 100*last), last+1)
295 // Round up to something easy to read.
299 b.result = BenchmarkResult{b.N, b.duration, b.bytes, b.netAllocs, b.netBytes}
302 // The results of a benchmark run.
303 // MemAllocs and MemBytes may be zero if memory benchmarking is not requested
304 // using B.ReportAllocs or the -benchmem command line flag.
305 type BenchmarkResult struct {
306 N int // The number of iterations.
307 T time.Duration // The total time taken.
308 Bytes int64 // Bytes processed in one iteration.
309 MemAllocs uint64 // The total number of memory allocations.
310 MemBytes uint64 // The total number of bytes allocated.
313 func (r BenchmarkResult) NsPerOp() int64 {
317 return r.T.Nanoseconds() / int64(r.N)
320 func (r BenchmarkResult) mbPerSec() float64 {
321 if r.Bytes <= 0 || r.T <= 0 || r.N <= 0 {
324 return (float64(r.Bytes) * float64(r.N) / 1e6) / r.T.Seconds()
327 // AllocsPerOp returns r.MemAllocs / r.N.
328 func (r BenchmarkResult) AllocsPerOp() int64 {
332 return int64(r.MemAllocs) / int64(r.N)
335 // AllocedBytesPerOp returns r.MemBytes / r.N.
336 func (r BenchmarkResult) AllocedBytesPerOp() int64 {
340 return int64(r.MemBytes) / int64(r.N)
343 func (r BenchmarkResult) String() string {
347 mb = fmt.Sprintf("\t%7.2f MB/s", mbs)
350 ns := fmt.Sprintf("%10d ns/op", nsop)
351 if r.N > 0 && nsop < 100 {
352 // The format specifiers here make sure that
353 // the ones digits line up for all three possible formats.
355 ns = fmt.Sprintf("%13.2f ns/op", float64(r.T.Nanoseconds())/float64(r.N))
357 ns = fmt.Sprintf("%12.1f ns/op", float64(r.T.Nanoseconds())/float64(r.N))
360 return fmt.Sprintf("%8d\t%s%s", r.N, ns, mb)
363 // MemString returns r.AllocedBytesPerOp and r.AllocsPerOp in the same format as 'go test'.
364 func (r BenchmarkResult) MemString() string {
365 return fmt.Sprintf("%8d B/op\t%8d allocs/op",
366 r.AllocedBytesPerOp(), r.AllocsPerOp())
369 // benchmarkName returns full name of benchmark including procs suffix.
370 func benchmarkName(name string, n int) string {
372 return fmt.Sprintf("%s-%d", name, n)
377 type benchContext struct {
380 maxLen int // The largest recorded benchmark name.
381 extLen int // Maximum extension length.
384 // An internal function but exported because it is cross-package; part of the implementation
385 // of the "go test" command.
386 func RunBenchmarks(matchString func(pat, str string) (bool, error), benchmarks []InternalBenchmark) {
387 runBenchmarks("", matchString, benchmarks)
390 func runBenchmarks(importPath string, matchString func(pat, str string) (bool, error), benchmarks []InternalBenchmark) bool {
391 // If no flag was specified, don't run benchmarks.
392 if len(*matchBenchmarks) == 0 {
395 // Collect matching benchmarks and determine longest name.
397 for _, procs := range cpuList {
398 if procs > maxprocs {
402 ctx := &benchContext{
403 match: newMatcher(matchString, *matchBenchmarks, "-test.bench"),
404 extLen: len(benchmarkName("", maxprocs)),
406 var bs []InternalBenchmark
407 for _, Benchmark := range benchmarks {
408 if _, matched := ctx.match.fullName(nil, Benchmark.Name); matched {
409 bs = append(bs, Benchmark)
410 benchName := benchmarkName(Benchmark.Name, maxprocs)
411 if l := len(benchName) + ctx.extLen + 1; l > ctx.maxLen {
422 importPath: importPath,
423 benchFunc: func(b *B) {
424 for _, Benchmark := range bs {
425 b.Run(Benchmark.Name, Benchmark.F)
428 benchTime: *benchTime,
435 // processBench runs bench b for the configured CPU counts and prints the results.
436 func (ctx *benchContext) processBench(b *B) {
437 for i, procs := range cpuList {
438 runtime.GOMAXPROCS(procs)
439 benchName := benchmarkName(b.name, procs)
440 fmt.Fprintf(b.w, "%-*s\t", ctx.maxLen, benchName)
441 // Recompute the running time for all but the first iteration.
445 signal: make(chan bool),
450 benchFunc: b.benchFunc,
451 benchTime: b.benchTime,
457 // The output could be very long here, but probably isn't.
458 // We print it all, regardless, because we don't want to trim the reason
459 // the benchmark failed.
460 fmt.Fprintf(b.w, "--- FAIL: %s\n%s", benchName, b.output)
463 results := r.String()
464 if *benchmarkMemory || b.showAllocResult {
465 results += "\t" + r.MemString()
467 fmt.Fprintln(b.w, results)
468 // Unlike with tests, we ignore the -chatty flag and always print output for
469 // benchmarks since the output generation time will skew the results.
470 if len(b.output) > 0 {
472 fmt.Fprintf(b.w, "--- BENCH: %s\n%s", benchName, b.output)
474 if p := runtime.GOMAXPROCS(-1); p != procs {
475 fmt.Fprintf(os.Stderr, "testing: %s left GOMAXPROCS set to %d\n", benchName, p)
480 // Run benchmarks f as a subbenchmark with the given name. It reports
481 // whether there were any failures.
483 // A subbenchmark is like any other benchmark. A benchmark that calls Run at
484 // least once will not be measured itself and will be called once with N=1.
486 // Run may be called simultaneously from multiple goroutines, but all such
487 // calls must happen before the outer benchmark function for b returns.
488 func (b *B) Run(name string, f func(b *B)) bool {
489 // Since b has subbenchmarks, we will no longer run it as a benchmark itself.
490 // Release the lock and acquire it on exit to ensure locks stay paired.
491 atomic.StoreInt32(&b.hasSub, 1)
492 benchmarkLock.Unlock()
493 defer benchmarkLock.Lock()
495 benchName, ok := b.name, true
496 if b.context != nil {
497 benchName, ok = b.context.match.fullName(&b.common, name)
504 signal: make(chan bool),
511 importPath: b.importPath,
513 benchTime: b.benchTime,
523 // add simulates running benchmarks in sequence in a single iteration. It is
524 // used to give some meaningful results in case func Benchmark is used in
525 // combination with Run.
526 func (b *B) add(other BenchmarkResult) {
528 // The aggregated BenchmarkResults resemble running all subbenchmarks as
529 // in sequence in a single benchmark.
531 r.T += time.Duration(other.NsPerOp())
532 if other.Bytes == 0 {
533 // Summing Bytes is meaningless in aggregate if not all subbenchmarks
535 b.missingBytes = true
539 r.Bytes += other.Bytes
541 r.MemAllocs += uint64(other.AllocsPerOp())
542 r.MemBytes += uint64(other.AllocedBytesPerOp())
545 // trimOutput shortens the output from a benchmark, which can be very long.
546 func (b *B) trimOutput() {
547 // The output is likely to appear multiple times because the benchmark
548 // is run multiple times, but at least it will be seen. This is not a big deal
549 // because benchmarks rarely print, but just in case, we trim it if it's too long.
550 const maxNewlines = 10
551 for nlCount, j := 0, 0; j < len(b.output); j++ {
552 if b.output[j] == '\n' {
554 if nlCount >= maxNewlines {
555 b.output = append(b.output[:j], "\n\t... [output truncated]\n"...)
562 // A PB is used by RunParallel for running parallel benchmarks.
564 globalN *uint64 // shared between all worker goroutines iteration counter
565 grain uint64 // acquire that many iterations from globalN at once
566 cache uint64 // local cache of acquired iterations
567 bN uint64 // total number of iterations to execute (b.N)
570 // Next reports whether there are more iterations to execute.
571 func (pb *PB) Next() bool {
573 n := atomic.AddUint64(pb.globalN, pb.grain)
576 } else if n < pb.bN+pb.grain {
577 pb.cache = pb.bN + pb.grain - n
586 // RunParallel runs a benchmark in parallel.
587 // It creates multiple goroutines and distributes b.N iterations among them.
588 // The number of goroutines defaults to GOMAXPROCS. To increase parallelism for
589 // non-CPU-bound benchmarks, call SetParallelism before RunParallel.
590 // RunParallel is usually used with the go test -cpu flag.
592 // The body function will be run in each goroutine. It should set up any
593 // goroutine-local state and then iterate until pb.Next returns false.
594 // It should not use the StartTimer, StopTimer, or ResetTimer functions,
595 // because they have global effect. It should also not call Run.
596 func (b *B) RunParallel(body func(*PB)) {
598 return // Nothing to do when probing.
600 // Calculate grain size as number of iterations that take ~100µs.
601 // 100µs is enough to amortize the overhead and provide sufficient
602 // dynamic load balancing.
604 if b.previousN > 0 && b.previousDuration > 0 {
605 grain = 1e5 * uint64(b.previousN) / uint64(b.previousDuration)
610 // We expect the inner loop and function call to take at least 10ns,
611 // so do not do more than 100µs/10ns=1e4 iterations.
617 numProcs := b.parallelism * runtime.GOMAXPROCS(0)
618 var wg sync.WaitGroup
620 for p := 0; p < numProcs; p++ {
632 if n <= uint64(b.N) && !b.Failed() {
633 b.Fatal("RunParallel: body exited without pb.Next() == false")
637 // SetParallelism sets the number of goroutines used by RunParallel to p*GOMAXPROCS.
638 // There is usually no need to call SetParallelism for CPU-bound benchmarks.
639 // If p is less than 1, this call will have no effect.
640 func (b *B) SetParallelism(p int) {
646 // Benchmark benchmarks a single function. Useful for creating
647 // custom benchmarks that do not use the "go test" command.
649 // If f calls Run, the result will be an estimate of running all its
650 // subbenchmarks that don't call Run in sequence in a single benchmark.
651 func Benchmark(f func(b *B)) BenchmarkResult {
654 signal: make(chan bool),
658 benchTime: *benchTime,
661 return BenchmarkResult{}
666 type discard struct{}
668 func (discard) Write(b []byte) (n int, err error) { return len(b), nil }