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.
17 var matchBenchmarks = flag.String("test.bench", "", "run only benchmarks matching `regexp`")
18 var benchTime = flag.Duration("test.benchtime", 1*time.Second, "run each benchmark for duration `d`")
19 var benchmarkMemory = flag.Bool("test.benchmem", false, "print memory allocations for benchmarks")
21 // Global lock to ensure only one benchmark runs at a time.
22 var benchmarkLock sync.Mutex
24 // Used for every benchmark for measuring memory.
25 var memStats runtime.MemStats
27 // An internal type but exported because it is cross-package; part of the implementation
28 // of the "go test" command.
29 type InternalBenchmark struct {
34 // B is a type passed to Benchmark functions to manage benchmark
35 // timing and to specify the number of iterations to run.
37 // A benchmark ends when its Benchmark function returns or calls any of the methods
38 // FailNow, Fatal, Fatalf, SkipNow, Skip, or Skipf. Those methods must be called
39 // only from the goroutine running the Benchmark function.
40 // The other reporting methods, such as the variations of Log and Error,
41 // may be called simultaneously from multiple goroutines.
43 // Like in tests, benchmark logs are accumulated during execution
44 // and dumped to standard error when done. Unlike in tests, benchmark logs
45 // are always printed, so as not to hide output whose existence may be
46 // affecting benchmark results.
51 previousN int // number of iterations in the previous run
52 previousDuration time.Duration // total duration of the previous run
54 benchTime time.Duration
56 missingBytes bool // one of the subbenchmarks does not have bytes set.
59 result BenchmarkResult
60 parallelism int // RunParallel creates parallelism*GOMAXPROCS goroutines
61 // The initial states of memStats.Mallocs and memStats.TotalAlloc.
64 // The net total of this test after being run.
69 // StartTimer starts timing a test. This function is called automatically
70 // before a benchmark starts, but it can also used to resume timing after
71 // a call to StopTimer.
72 func (b *B) StartTimer() {
74 runtime.ReadMemStats(&memStats)
75 b.startAllocs = memStats.Mallocs
76 b.startBytes = memStats.TotalAlloc
82 // StopTimer stops timing a test. This can be used to pause the timer
83 // while performing complex initialization that you don't
85 func (b *B) StopTimer() {
87 b.duration += time.Now().Sub(b.start)
88 runtime.ReadMemStats(&memStats)
89 b.netAllocs += memStats.Mallocs - b.startAllocs
90 b.netBytes += memStats.TotalAlloc - b.startBytes
95 // ResetTimer zeros the elapsed benchmark time and memory allocation counters.
96 // It does not affect whether the timer is running.
97 func (b *B) ResetTimer() {
99 runtime.ReadMemStats(&memStats)
100 b.startAllocs = memStats.Mallocs
101 b.startBytes = memStats.TotalAlloc
109 // SetBytes records the number of bytes processed in a single operation.
110 // If this is called, the benchmark will report ns/op and MB/s.
111 func (b *B) SetBytes(n int64) { b.bytes = n }
113 // ReportAllocs enables malloc statistics for this benchmark.
114 // It is equivalent to setting -test.benchmem, but it only affects the
115 // benchmark function that calls ReportAllocs.
116 func (b *B) ReportAllocs() {
117 b.showAllocResult = true
120 func (b *B) nsPerOp() int64 {
124 return b.duration.Nanoseconds() / int64(b.N)
127 // runN runs a single benchmark for the specified number of iterations.
128 func (b *B) runN(n int) {
130 defer benchmarkLock.Unlock()
131 // Try to get a comparable environment for each run
132 // by clearing garbage from previous runs.
141 b.previousDuration = b.duration
144 func min(x, y int) int {
151 func max(x, y int) int {
158 // roundDown10 rounds a number down to the nearest power of 10.
159 func roundDown10(n int) int {
161 // tens = floor(log_10(n))
168 for i := 0; i < tens; i++ {
174 // roundUp rounds x up to a number of the form [1eX, 2eX, 3eX, 5eX].
175 func roundUp(n int) int {
176 base := roundDown10(n)
180 case n <= (2 * base):
182 case n <= (3 * base):
184 case n <= (5 * base):
191 // run1 runs the first iteration of benchFunc. It returns whether more
192 // iterations of this benchmarks should be run.
193 func (b *B) run1() bool {
194 if ctx := b.context; ctx != nil {
195 // Extend maxLen, if needed.
196 if n := len(b.name) + ctx.extLen + 1; n > ctx.maxLen {
197 ctx.maxLen = n + 8 // Add additional slack to avoid too many jumps in size.
201 // Signal that we're done whether we return normally
202 // or by FailNow's runtime.Goexit.
211 fmt.Fprintf(b.w, "--- FAIL: %s\n%s", b.name, b.output)
214 // Only print the output if we know we are not going to proceed.
215 // Otherwise it is printed in processBench.
216 if b.hasSub || b.finished {
221 if b.chatty && (len(b.output) > 0 || b.finished) {
223 fmt.Fprintf(b.w, "--- %s: %s\n%s", tag, b.name, b.output)
230 // run executes the benchmark in a separate goroutine, including all of its
231 // subbenchmarks. b must not have subbenchmarks.
232 func (b *B) run() BenchmarkResult {
233 if b.context != nil {
234 // Running go test --test.bench
235 b.context.processBench(b) // Must call doBench.
237 // Running func Benchmark.
243 func (b *B) doBench() BenchmarkResult {
249 // launch launches the benchmark function. It gradually increases the number
250 // of benchmark iterations until the benchmark runs for the requested benchtime.
251 // launch is run by the doBench function as a separate goroutine.
252 // run1 must have been called on b.
253 func (b *B) launch() {
254 // Signal that we're done whether we return normally
255 // or by FailNow's runtime.Goexit.
260 // Run the benchmark for at least the specified amount of time.
262 for n := 1; !b.failed && b.duration < d && n < 1e9; {
264 // Predict required iterations.
265 n = int(d.Nanoseconds())
266 if nsop := b.nsPerOp(); nsop != 0 {
269 // Run more iterations than we think we'll need (1.2x).
270 // Don't grow too fast in case we had timing errors previously.
271 // Be sure to run at least one more than last time.
272 n = max(min(n+n/5, 100*last), last+1)
273 // Round up to something easy to read.
277 b.result = BenchmarkResult{b.N, b.duration, b.bytes, b.netAllocs, b.netBytes}
280 // The results of a benchmark run.
281 type BenchmarkResult struct {
282 N int // The number of iterations.
283 T time.Duration // The total time taken.
284 Bytes int64 // Bytes processed in one iteration.
285 MemAllocs uint64 // The total number of memory allocations.
286 MemBytes uint64 // The total number of bytes allocated.
289 func (r BenchmarkResult) NsPerOp() int64 {
293 return r.T.Nanoseconds() / int64(r.N)
296 func (r BenchmarkResult) mbPerSec() float64 {
297 if r.Bytes <= 0 || r.T <= 0 || r.N <= 0 {
300 return (float64(r.Bytes) * float64(r.N) / 1e6) / r.T.Seconds()
303 func (r BenchmarkResult) AllocsPerOp() int64 {
307 return int64(r.MemAllocs) / int64(r.N)
310 func (r BenchmarkResult) AllocedBytesPerOp() int64 {
314 return int64(r.MemBytes) / int64(r.N)
317 func (r BenchmarkResult) String() string {
321 mb = fmt.Sprintf("\t%7.2f MB/s", mbs)
324 ns := fmt.Sprintf("%10d ns/op", nsop)
325 if r.N > 0 && nsop < 100 {
326 // The format specifiers here make sure that
327 // the ones digits line up for all three possible formats.
329 ns = fmt.Sprintf("%13.2f ns/op", float64(r.T.Nanoseconds())/float64(r.N))
331 ns = fmt.Sprintf("%12.1f ns/op", float64(r.T.Nanoseconds())/float64(r.N))
334 return fmt.Sprintf("%8d\t%s%s", r.N, ns, mb)
337 func (r BenchmarkResult) MemString() string {
338 return fmt.Sprintf("%8d B/op\t%8d allocs/op",
339 r.AllocedBytesPerOp(), r.AllocsPerOp())
342 // benchmarkName returns full name of benchmark including procs suffix.
343 func benchmarkName(name string, n int) string {
345 return fmt.Sprintf("%s-%d", name, n)
350 type benchContext struct {
353 maxLen int // The largest recorded benchmark name.
354 extLen int // Maximum extension length.
357 // An internal function but exported because it is cross-package; part of the implementation
358 // of the "go test" command.
359 func RunBenchmarks(matchString func(pat, str string) (bool, error), benchmarks []InternalBenchmark) {
360 runBenchmarks(matchString, benchmarks)
363 func runBenchmarks(matchString func(pat, str string) (bool, error), benchmarks []InternalBenchmark) bool {
364 // If no flag was specified, don't run benchmarks.
365 if len(*matchBenchmarks) == 0 {
368 // Collect matching benchmarks and determine longest name.
370 for _, procs := range cpuList {
371 if procs > maxprocs {
375 ctx := &benchContext{
376 match: newMatcher(matchString, *matchBenchmarks, "-test.bench"),
377 extLen: len(benchmarkName("", maxprocs)),
379 var bs []InternalBenchmark
380 for _, Benchmark := range benchmarks {
381 if _, matched := ctx.match.fullName(nil, Benchmark.Name); matched {
382 bs = append(bs, Benchmark)
383 benchName := benchmarkName(Benchmark.Name, maxprocs)
384 if l := len(benchName) + ctx.extLen + 1; l > ctx.maxLen {
395 benchFunc: func(b *B) {
396 for _, Benchmark := range bs {
397 b.Run(Benchmark.Name, Benchmark.F)
400 benchTime: *benchTime,
407 // processBench runs bench b for the configured CPU counts and prints the results.
408 func (ctx *benchContext) processBench(b *B) {
409 for i, procs := range cpuList {
410 runtime.GOMAXPROCS(procs)
411 benchName := benchmarkName(b.name, procs)
412 fmt.Fprintf(b.w, "%-*s\t", ctx.maxLen, benchName)
413 // Recompute the running time for all but the first iteration.
417 signal: make(chan bool),
422 benchFunc: b.benchFunc,
423 benchTime: b.benchTime,
429 // The output could be very long here, but probably isn't.
430 // We print it all, regardless, because we don't want to trim the reason
431 // the benchmark failed.
432 fmt.Fprintf(b.w, "--- FAIL: %s\n%s", benchName, b.output)
435 results := r.String()
436 if *benchmarkMemory || b.showAllocResult {
437 results += "\t" + r.MemString()
439 fmt.Fprintln(b.w, results)
440 // Unlike with tests, we ignore the -chatty flag and always print output for
441 // benchmarks since the output generation time will skew the results.
442 if len(b.output) > 0 {
444 fmt.Fprintf(b.w, "--- BENCH: %s\n%s", benchName, b.output)
446 if p := runtime.GOMAXPROCS(-1); p != procs {
447 fmt.Fprintf(os.Stderr, "testing: %s left GOMAXPROCS set to %d\n", benchName, p)
452 // Run benchmarks f as a subbenchmark with the given name. It reports
453 // whether there were any failures.
455 // A subbenchmark is like any other benchmark. A benchmark that calls Run at
456 // least once will not be measured itself and will be called once with N=1.
457 func (b *B) Run(name string, f func(b *B)) bool {
458 // Since b has subbenchmarks, we will no longer run it as a benchmark itself.
459 // Release the lock and acquire it on exit to ensure locks stay paired.
461 benchmarkLock.Unlock()
462 defer benchmarkLock.Lock()
464 benchName, ok := b.name, true
465 if b.context != nil {
466 benchName, ok = b.context.match.fullName(&b.common, name)
473 signal: make(chan bool),
481 benchTime: b.benchTime,
491 // add simulates running benchmarks in sequence in a single iteration. It is
492 // used to give some meaningful results in case func Benchmark is used in
493 // combination with Run.
494 func (b *B) add(other BenchmarkResult) {
496 // The aggregated BenchmarkResults resemble running all subbenchmarks as
497 // in sequence in a single benchmark.
499 r.T += time.Duration(other.NsPerOp())
500 if other.Bytes == 0 {
501 // Summing Bytes is meaningless in aggregate if not all subbenchmarks
503 b.missingBytes = true
507 r.Bytes += other.Bytes
509 r.MemAllocs += uint64(other.AllocsPerOp())
510 r.MemBytes += uint64(other.AllocedBytesPerOp())
513 // trimOutput shortens the output from a benchmark, which can be very long.
514 func (b *B) trimOutput() {
515 // The output is likely to appear multiple times because the benchmark
516 // is run multiple times, but at least it will be seen. This is not a big deal
517 // because benchmarks rarely print, but just in case, we trim it if it's too long.
518 const maxNewlines = 10
519 for nlCount, j := 0, 0; j < len(b.output); j++ {
520 if b.output[j] == '\n' {
522 if nlCount >= maxNewlines {
523 b.output = append(b.output[:j], "\n\t... [output truncated]\n"...)
530 // A PB is used by RunParallel for running parallel benchmarks.
532 globalN *uint64 // shared between all worker goroutines iteration counter
533 grain uint64 // acquire that many iterations from globalN at once
534 cache uint64 // local cache of acquired iterations
535 bN uint64 // total number of iterations to execute (b.N)
538 // Next reports whether there are more iterations to execute.
539 func (pb *PB) Next() bool {
541 n := atomic.AddUint64(pb.globalN, pb.grain)
544 } else if n < pb.bN+pb.grain {
545 pb.cache = pb.bN + pb.grain - n
554 // RunParallel runs a benchmark in parallel.
555 // It creates multiple goroutines and distributes b.N iterations among them.
556 // The number of goroutines defaults to GOMAXPROCS. To increase parallelism for
557 // non-CPU-bound benchmarks, call SetParallelism before RunParallel.
558 // RunParallel is usually used with the go test -cpu flag.
560 // The body function will be run in each goroutine. It should set up any
561 // goroutine-local state and then iterate until pb.Next returns false.
562 // It should not use the StartTimer, StopTimer, or ResetTimer functions,
563 // because they have global effect. It should also not call Run.
564 func (b *B) RunParallel(body func(*PB)) {
566 return // Nothing to do when probing.
568 // Calculate grain size as number of iterations that take ~100µs.
569 // 100µs is enough to amortize the overhead and provide sufficient
570 // dynamic load balancing.
572 if b.previousN > 0 && b.previousDuration > 0 {
573 grain = 1e5 * uint64(b.previousN) / uint64(b.previousDuration)
578 // We expect the inner loop and function call to take at least 10ns,
579 // so do not do more than 100µs/10ns=1e4 iterations.
585 numProcs := b.parallelism * runtime.GOMAXPROCS(0)
586 var wg sync.WaitGroup
588 for p := 0; p < numProcs; p++ {
600 if n <= uint64(b.N) && !b.Failed() {
601 b.Fatal("RunParallel: body exited without pb.Next() == false")
605 // SetParallelism sets the number of goroutines used by RunParallel to p*GOMAXPROCS.
606 // There is usually no need to call SetParallelism for CPU-bound benchmarks.
607 // If p is less than 1, this call will have no effect.
608 func (b *B) SetParallelism(p int) {
614 // Benchmark benchmarks a single function. Useful for creating
615 // custom benchmarks that do not use the "go test" command.
617 // If f calls Run, the result will be an estimate of running all its
618 // subbenchmarks that don't call Run in sequence in a single benchmark.
619 func Benchmark(f func(b *B)) BenchmarkResult {
622 signal: make(chan bool),
626 benchTime: *benchTime,
629 return BenchmarkResult{}
634 type discard struct{}
636 func (discard) Write(b []byte) (n int, err error) { return len(b), nil }