Source file src/testing/benchmark.go

     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.
     4  
     5  package testing
     6  
     7  import (
     8  	"flag"
     9  	"fmt"
    10  	"internal/race"
    11  	"internal/sysinfo"
    12  	"io"
    13  	"math"
    14  	"os"
    15  	"runtime"
    16  	"sort"
    17  	"strconv"
    18  	"strings"
    19  	"sync"
    20  	"sync/atomic"
    21  	"time"
    22  	"unicode"
    23  )
    24  
    25  func initBenchmarkFlags() {
    26  	matchBenchmarks = flag.String("test.bench", "", "run only benchmarks matching `regexp`")
    27  	benchmarkMemory = flag.Bool("test.benchmem", false, "print memory allocations for benchmarks")
    28  	flag.Var(&benchTime, "test.benchtime", "run each benchmark for duration `d`")
    29  }
    30  
    31  var (
    32  	matchBenchmarks *string
    33  	benchmarkMemory *bool
    34  
    35  	benchTime = durationOrCountFlag{d: 1 * time.Second} // changed during test of testing package
    36  )
    37  
    38  type durationOrCountFlag struct {
    39  	d         time.Duration
    40  	n         int
    41  	allowZero bool
    42  }
    43  
    44  func (f *durationOrCountFlag) String() string {
    45  	if f.n > 0 {
    46  		return fmt.Sprintf("%dx", f.n)
    47  	}
    48  	return f.d.String()
    49  }
    50  
    51  func (f *durationOrCountFlag) Set(s string) error {
    52  	if strings.HasSuffix(s, "x") {
    53  		n, err := strconv.ParseInt(s[:len(s)-1], 10, 0)
    54  		if err != nil || n < 0 || (!f.allowZero && n == 0) {
    55  			return fmt.Errorf("invalid count")
    56  		}
    57  		*f = durationOrCountFlag{n: int(n)}
    58  		return nil
    59  	}
    60  	d, err := time.ParseDuration(s)
    61  	if err != nil || d < 0 || (!f.allowZero && d == 0) {
    62  		return fmt.Errorf("invalid duration")
    63  	}
    64  	*f = durationOrCountFlag{d: d}
    65  	return nil
    66  }
    67  
    68  // Global lock to ensure only one benchmark runs at a time.
    69  var benchmarkLock sync.Mutex
    70  
    71  // Used for every benchmark for measuring memory.
    72  var memStats runtime.MemStats
    73  
    74  // InternalBenchmark is an internal type but exported because it is cross-package;
    75  // it is part of the implementation of the "go test" command.
    76  type InternalBenchmark struct {
    77  	Name string
    78  	F    func(b *B)
    79  }
    80  
    81  // B is a type passed to Benchmark functions to manage benchmark
    82  // timing and to specify the number of iterations to run.
    83  //
    84  // A benchmark ends when its Benchmark function returns or calls any of the methods
    85  // FailNow, Fatal, Fatalf, SkipNow, Skip, or Skipf. Those methods must be called
    86  // only from the goroutine running the Benchmark function.
    87  // The other reporting methods, such as the variations of Log and Error,
    88  // may be called simultaneously from multiple goroutines.
    89  //
    90  // Like in tests, benchmark logs are accumulated during execution
    91  // and dumped to standard output when done. Unlike in tests, benchmark logs
    92  // are always printed, so as not to hide output whose existence may be
    93  // affecting benchmark results.
    94  type B struct {
    95  	common
    96  	importPath       string // import path of the package containing the benchmark
    97  	context          *benchContext
    98  	N                int
    99  	previousN        int           // number of iterations in the previous run
   100  	previousDuration time.Duration // total duration of the previous run
   101  	benchFunc        func(b *B)
   102  	benchTime        durationOrCountFlag
   103  	bytes            int64
   104  	missingBytes     bool // one of the subbenchmarks does not have bytes set.
   105  	timerOn          bool
   106  	showAllocResult  bool
   107  	result           BenchmarkResult
   108  	parallelism      int // RunParallel creates parallelism*GOMAXPROCS goroutines
   109  	// The initial states of memStats.Mallocs and memStats.TotalAlloc.
   110  	startAllocs uint64
   111  	startBytes  uint64
   112  	// The net total of this test after being run.
   113  	netAllocs uint64
   114  	netBytes  uint64
   115  	// Extra metrics collected by ReportMetric.
   116  	extra map[string]float64
   117  }
   118  
   119  // StartTimer starts timing a test. This function is called automatically
   120  // before a benchmark starts, but it can also be used to resume timing after
   121  // a call to StopTimer.
   122  func (b *B) StartTimer() {
   123  	if !b.timerOn {
   124  		runtime.ReadMemStats(&memStats)
   125  		b.startAllocs = memStats.Mallocs
   126  		b.startBytes = memStats.TotalAlloc
   127  		b.start = time.Now()
   128  		b.timerOn = true
   129  	}
   130  }
   131  
   132  // StopTimer stops timing a test. This can be used to pause the timer
   133  // while performing complex initialization that you don't
   134  // want to measure.
   135  func (b *B) StopTimer() {
   136  	if b.timerOn {
   137  		b.duration += time.Since(b.start)
   138  		runtime.ReadMemStats(&memStats)
   139  		b.netAllocs += memStats.Mallocs - b.startAllocs
   140  		b.netBytes += memStats.TotalAlloc - b.startBytes
   141  		b.timerOn = false
   142  	}
   143  }
   144  
   145  // ResetTimer zeroes the elapsed benchmark time and memory allocation counters
   146  // and deletes user-reported metrics.
   147  // It does not affect whether the timer is running.
   148  func (b *B) ResetTimer() {
   149  	if b.extra == nil {
   150  		// Allocate the extra map before reading memory stats.
   151  		// Pre-size it to make more allocation unlikely.
   152  		b.extra = make(map[string]float64, 16)
   153  	} else {
   154  		for k := range b.extra {
   155  			delete(b.extra, k)
   156  		}
   157  	}
   158  	if b.timerOn {
   159  		runtime.ReadMemStats(&memStats)
   160  		b.startAllocs = memStats.Mallocs
   161  		b.startBytes = memStats.TotalAlloc
   162  		b.start = time.Now()
   163  	}
   164  	b.duration = 0
   165  	b.netAllocs = 0
   166  	b.netBytes = 0
   167  }
   168  
   169  // SetBytes records the number of bytes processed in a single operation.
   170  // If this is called, the benchmark will report ns/op and MB/s.
   171  func (b *B) SetBytes(n int64) { b.bytes = n }
   172  
   173  // ReportAllocs enables malloc statistics for this benchmark.
   174  // It is equivalent to setting -test.benchmem, but it only affects the
   175  // benchmark function that calls ReportAllocs.
   176  func (b *B) ReportAllocs() {
   177  	b.showAllocResult = true
   178  }
   179  
   180  // runN runs a single benchmark for the specified number of iterations.
   181  func (b *B) runN(n int) {
   182  	benchmarkLock.Lock()
   183  	defer benchmarkLock.Unlock()
   184  	defer b.runCleanup(normalPanic)
   185  	// Try to get a comparable environment for each run
   186  	// by clearing garbage from previous runs.
   187  	runtime.GC()
   188  	b.raceErrors = -race.Errors()
   189  	b.N = n
   190  	b.parallelism = 1
   191  	b.ResetTimer()
   192  	b.StartTimer()
   193  	b.benchFunc(b)
   194  	b.StopTimer()
   195  	b.previousN = n
   196  	b.previousDuration = b.duration
   197  	b.raceErrors += race.Errors()
   198  	if b.raceErrors > 0 {
   199  		b.Errorf("race detected during execution of benchmark")
   200  	}
   201  }
   202  
   203  func min(x, y int64) int64 {
   204  	if x > y {
   205  		return y
   206  	}
   207  	return x
   208  }
   209  
   210  func max(x, y int64) int64 {
   211  	if x < y {
   212  		return y
   213  	}
   214  	return x
   215  }
   216  
   217  // run1 runs the first iteration of benchFunc. It reports whether more
   218  // iterations of this benchmarks should be run.
   219  func (b *B) run1() bool {
   220  	if ctx := b.context; ctx != nil {
   221  		// Extend maxLen, if needed.
   222  		if n := len(b.name) + ctx.extLen + 1; n > ctx.maxLen {
   223  			ctx.maxLen = n + 8 // Add additional slack to avoid too many jumps in size.
   224  		}
   225  	}
   226  	go func() {
   227  		// Signal that we're done whether we return normally
   228  		// or by FailNow's runtime.Goexit.
   229  		defer func() {
   230  			b.signal <- true
   231  		}()
   232  
   233  		b.runN(1)
   234  	}()
   235  	<-b.signal
   236  	if b.failed {
   237  		fmt.Fprintf(b.w, "--- FAIL: %s\n%s", b.name, b.output)
   238  		return false
   239  	}
   240  	// Only print the output if we know we are not going to proceed.
   241  	// Otherwise it is printed in processBench.
   242  	b.mu.RLock()
   243  	finished := b.finished
   244  	b.mu.RUnlock()
   245  	if atomic.LoadInt32(&b.hasSub) != 0 || finished {
   246  		tag := "BENCH"
   247  		if b.skipped {
   248  			tag = "SKIP"
   249  		}
   250  		if b.chatty != nil && (len(b.output) > 0 || finished) {
   251  			b.trimOutput()
   252  			fmt.Fprintf(b.w, "--- %s: %s\n%s", tag, b.name, b.output)
   253  		}
   254  		return false
   255  	}
   256  	return true
   257  }
   258  
   259  var labelsOnce sync.Once
   260  
   261  // run executes the benchmark in a separate goroutine, including all of its
   262  // subbenchmarks. b must not have subbenchmarks.
   263  func (b *B) run() {
   264  	labelsOnce.Do(func() {
   265  		fmt.Fprintf(b.w, "goos: %s\n", runtime.GOOS)
   266  		fmt.Fprintf(b.w, "goarch: %s\n", runtime.GOARCH)
   267  		if b.importPath != "" {
   268  			fmt.Fprintf(b.w, "pkg: %s\n", b.importPath)
   269  		}
   270  		if cpu := sysinfo.CPU.Name(); cpu != "" {
   271  			fmt.Fprintf(b.w, "cpu: %s\n", cpu)
   272  		}
   273  	})
   274  	if b.context != nil {
   275  		// Running go test --test.bench
   276  		b.context.processBench(b) // Must call doBench.
   277  	} else {
   278  		// Running func Benchmark.
   279  		b.doBench()
   280  	}
   281  }
   282  
   283  func (b *B) doBench() BenchmarkResult {
   284  	go b.launch()
   285  	<-b.signal
   286  	return b.result
   287  }
   288  
   289  // launch launches the benchmark function. It gradually increases the number
   290  // of benchmark iterations until the benchmark runs for the requested benchtime.
   291  // launch is run by the doBench function as a separate goroutine.
   292  // run1 must have been called on b.
   293  func (b *B) launch() {
   294  	// Signal that we're done whether we return normally
   295  	// or by FailNow's runtime.Goexit.
   296  	defer func() {
   297  		b.signal <- true
   298  	}()
   299  
   300  	// Run the benchmark for at least the specified amount of time.
   301  	if b.benchTime.n > 0 {
   302  		// We already ran a single iteration in run1.
   303  		// If -benchtime=1x was requested, use that result.
   304  		// See https://golang.org/issue/32051.
   305  		if b.benchTime.n > 1 {
   306  			b.runN(b.benchTime.n)
   307  		}
   308  	} else {
   309  		d := b.benchTime.d
   310  		for n := int64(1); !b.failed && b.duration < d && n < 1e9; {
   311  			last := n
   312  			// Predict required iterations.
   313  			goalns := d.Nanoseconds()
   314  			prevIters := int64(b.N)
   315  			prevns := b.duration.Nanoseconds()
   316  			if prevns <= 0 {
   317  				// Round up, to avoid div by zero.
   318  				prevns = 1
   319  			}
   320  			// Order of operations matters.
   321  			// For very fast benchmarks, prevIters ~= prevns.
   322  			// If you divide first, you get 0 or 1,
   323  			// which can hide an order of magnitude in execution time.
   324  			// So multiply first, then divide.
   325  			n = goalns * prevIters / prevns
   326  			// Run more iterations than we think we'll need (1.2x).
   327  			n += n / 5
   328  			// Don't grow too fast in case we had timing errors previously.
   329  			n = min(n, 100*last)
   330  			// Be sure to run at least one more than last time.
   331  			n = max(n, last+1)
   332  			// Don't run more than 1e9 times. (This also keeps n in int range on 32 bit platforms.)
   333  			n = min(n, 1e9)
   334  			b.runN(int(n))
   335  		}
   336  	}
   337  	b.result = BenchmarkResult{b.N, b.duration, b.bytes, b.netAllocs, b.netBytes, b.extra}
   338  }
   339  
   340  // ReportMetric adds "n unit" to the reported benchmark results.
   341  // If the metric is per-iteration, the caller should divide by b.N,
   342  // and by convention units should end in "/op".
   343  // ReportMetric overrides any previously reported value for the same unit.
   344  // ReportMetric panics if unit is the empty string or if unit contains
   345  // any whitespace.
   346  // If unit is a unit normally reported by the benchmark framework itself
   347  // (such as "allocs/op"), ReportMetric will override that metric.
   348  // Setting "ns/op" to 0 will suppress that built-in metric.
   349  func (b *B) ReportMetric(n float64, unit string) {
   350  	if unit == "" {
   351  		panic("metric unit must not be empty")
   352  	}
   353  	if strings.IndexFunc(unit, unicode.IsSpace) >= 0 {
   354  		panic("metric unit must not contain whitespace")
   355  	}
   356  	b.extra[unit] = n
   357  }
   358  
   359  // BenchmarkResult contains the results of a benchmark run.
   360  type BenchmarkResult struct {
   361  	N         int           // The number of iterations.
   362  	T         time.Duration // The total time taken.
   363  	Bytes     int64         // Bytes processed in one iteration.
   364  	MemAllocs uint64        // The total number of memory allocations.
   365  	MemBytes  uint64        // The total number of bytes allocated.
   366  
   367  	// Extra records additional metrics reported by ReportMetric.
   368  	Extra map[string]float64
   369  }
   370  
   371  // NsPerOp returns the "ns/op" metric.
   372  func (r BenchmarkResult) NsPerOp() int64 {
   373  	if v, ok := r.Extra["ns/op"]; ok {
   374  		return int64(v)
   375  	}
   376  	if r.N <= 0 {
   377  		return 0
   378  	}
   379  	return r.T.Nanoseconds() / int64(r.N)
   380  }
   381  
   382  // mbPerSec returns the "MB/s" metric.
   383  func (r BenchmarkResult) mbPerSec() float64 {
   384  	if v, ok := r.Extra["MB/s"]; ok {
   385  		return v
   386  	}
   387  	if r.Bytes <= 0 || r.T <= 0 || r.N <= 0 {
   388  		return 0
   389  	}
   390  	return (float64(r.Bytes) * float64(r.N) / 1e6) / r.T.Seconds()
   391  }
   392  
   393  // AllocsPerOp returns the "allocs/op" metric,
   394  // which is calculated as r.MemAllocs / r.N.
   395  func (r BenchmarkResult) AllocsPerOp() int64 {
   396  	if v, ok := r.Extra["allocs/op"]; ok {
   397  		return int64(v)
   398  	}
   399  	if r.N <= 0 {
   400  		return 0
   401  	}
   402  	return int64(r.MemAllocs) / int64(r.N)
   403  }
   404  
   405  // AllocedBytesPerOp returns the "B/op" metric,
   406  // which is calculated as r.MemBytes / r.N.
   407  func (r BenchmarkResult) AllocedBytesPerOp() int64 {
   408  	if v, ok := r.Extra["B/op"]; ok {
   409  		return int64(v)
   410  	}
   411  	if r.N <= 0 {
   412  		return 0
   413  	}
   414  	return int64(r.MemBytes) / int64(r.N)
   415  }
   416  
   417  // String returns a summary of the benchmark results.
   418  // It follows the benchmark result line format from
   419  // https://golang.org/design/14313-benchmark-format, not including the
   420  // benchmark name.
   421  // Extra metrics override built-in metrics of the same name.
   422  // String does not include allocs/op or B/op, since those are reported
   423  // by MemString.
   424  func (r BenchmarkResult) String() string {
   425  	buf := new(strings.Builder)
   426  	fmt.Fprintf(buf, "%8d", r.N)
   427  
   428  	// Get ns/op as a float.
   429  	ns, ok := r.Extra["ns/op"]
   430  	if !ok {
   431  		ns = float64(r.T.Nanoseconds()) / float64(r.N)
   432  	}
   433  	if ns != 0 {
   434  		buf.WriteByte('\t')
   435  		prettyPrint(buf, ns, "ns/op")
   436  	}
   437  
   438  	if mbs := r.mbPerSec(); mbs != 0 {
   439  		fmt.Fprintf(buf, "\t%7.2f MB/s", mbs)
   440  	}
   441  
   442  	// Print extra metrics that aren't represented in the standard
   443  	// metrics.
   444  	var extraKeys []string
   445  	for k := range r.Extra {
   446  		switch k {
   447  		case "ns/op", "MB/s", "B/op", "allocs/op":
   448  			// Built-in metrics reported elsewhere.
   449  			continue
   450  		}
   451  		extraKeys = append(extraKeys, k)
   452  	}
   453  	sort.Strings(extraKeys)
   454  	for _, k := range extraKeys {
   455  		buf.WriteByte('\t')
   456  		prettyPrint(buf, r.Extra[k], k)
   457  	}
   458  	return buf.String()
   459  }
   460  
   461  func prettyPrint(w io.Writer, x float64, unit string) {
   462  	// Print all numbers with 10 places before the decimal point
   463  	// and small numbers with four sig figs. Field widths are
   464  	// chosen to fit the whole part in 10 places while aligning
   465  	// the decimal point of all fractional formats.
   466  	var format string
   467  	switch y := math.Abs(x); {
   468  	case y == 0 || y >= 999.95:
   469  		format = "%10.0f %s"
   470  	case y >= 99.995:
   471  		format = "%12.1f %s"
   472  	case y >= 9.9995:
   473  		format = "%13.2f %s"
   474  	case y >= 0.99995:
   475  		format = "%14.3f %s"
   476  	case y >= 0.099995:
   477  		format = "%15.4f %s"
   478  	case y >= 0.0099995:
   479  		format = "%16.5f %s"
   480  	case y >= 0.00099995:
   481  		format = "%17.6f %s"
   482  	default:
   483  		format = "%18.7f %s"
   484  	}
   485  	fmt.Fprintf(w, format, x, unit)
   486  }
   487  
   488  // MemString returns r.AllocedBytesPerOp and r.AllocsPerOp in the same format as 'go test'.
   489  func (r BenchmarkResult) MemString() string {
   490  	return fmt.Sprintf("%8d B/op\t%8d allocs/op",
   491  		r.AllocedBytesPerOp(), r.AllocsPerOp())
   492  }
   493  
   494  // benchmarkName returns full name of benchmark including procs suffix.
   495  func benchmarkName(name string, n int) string {
   496  	if n != 1 {
   497  		return fmt.Sprintf("%s-%d", name, n)
   498  	}
   499  	return name
   500  }
   501  
   502  type benchContext struct {
   503  	match *matcher
   504  
   505  	maxLen int // The largest recorded benchmark name.
   506  	extLen int // Maximum extension length.
   507  }
   508  
   509  // RunBenchmarks is an internal function but exported because it is cross-package;
   510  // it is part of the implementation of the "go test" command.
   511  func RunBenchmarks(matchString func(pat, str string) (bool, error), benchmarks []InternalBenchmark) {
   512  	runBenchmarks("", matchString, benchmarks)
   513  }
   514  
   515  func runBenchmarks(importPath string, matchString func(pat, str string) (bool, error), benchmarks []InternalBenchmark) bool {
   516  	// If no flag was specified, don't run benchmarks.
   517  	if len(*matchBenchmarks) == 0 {
   518  		return true
   519  	}
   520  	// Collect matching benchmarks and determine longest name.
   521  	maxprocs := 1
   522  	for _, procs := range cpuList {
   523  		if procs > maxprocs {
   524  			maxprocs = procs
   525  		}
   526  	}
   527  	ctx := &benchContext{
   528  		match:  newMatcher(matchString, *matchBenchmarks, "-test.bench"),
   529  		extLen: len(benchmarkName("", maxprocs)),
   530  	}
   531  	var bs []InternalBenchmark
   532  	for _, Benchmark := range benchmarks {
   533  		if _, matched, _ := ctx.match.fullName(nil, Benchmark.Name); matched {
   534  			bs = append(bs, Benchmark)
   535  			benchName := benchmarkName(Benchmark.Name, maxprocs)
   536  			if l := len(benchName) + ctx.extLen + 1; l > ctx.maxLen {
   537  				ctx.maxLen = l
   538  			}
   539  		}
   540  	}
   541  	main := &B{
   542  		common: common{
   543  			name:  "Main",
   544  			w:     os.Stdout,
   545  			bench: true,
   546  		},
   547  		importPath: importPath,
   548  		benchFunc: func(b *B) {
   549  			for _, Benchmark := range bs {
   550  				b.Run(Benchmark.Name, Benchmark.F)
   551  			}
   552  		},
   553  		benchTime: benchTime,
   554  		context:   ctx,
   555  	}
   556  	if Verbose() {
   557  		main.chatty = newChattyPrinter(main.w)
   558  	}
   559  	main.runN(1)
   560  	return !main.failed
   561  }
   562  
   563  // processBench runs bench b for the configured CPU counts and prints the results.
   564  func (ctx *benchContext) processBench(b *B) {
   565  	for i, procs := range cpuList {
   566  		for j := uint(0); j < *count; j++ {
   567  			runtime.GOMAXPROCS(procs)
   568  			benchName := benchmarkName(b.name, procs)
   569  
   570  			// If it's chatty, we've already printed this information.
   571  			if b.chatty == nil {
   572  				fmt.Fprintf(b.w, "%-*s\t", ctx.maxLen, benchName)
   573  			}
   574  			// Recompute the running time for all but the first iteration.
   575  			if i > 0 || j > 0 {
   576  				b = &B{
   577  					common: common{
   578  						signal: make(chan bool),
   579  						name:   b.name,
   580  						w:      b.w,
   581  						chatty: b.chatty,
   582  						bench:  true,
   583  					},
   584  					benchFunc: b.benchFunc,
   585  					benchTime: b.benchTime,
   586  				}
   587  				b.run1()
   588  			}
   589  			r := b.doBench()
   590  			if b.failed {
   591  				// The output could be very long here, but probably isn't.
   592  				// We print it all, regardless, because we don't want to trim the reason
   593  				// the benchmark failed.
   594  				fmt.Fprintf(b.w, "--- FAIL: %s\n%s", benchName, b.output)
   595  				continue
   596  			}
   597  			results := r.String()
   598  			if b.chatty != nil {
   599  				fmt.Fprintf(b.w, "%-*s\t", ctx.maxLen, benchName)
   600  			}
   601  			if *benchmarkMemory || b.showAllocResult {
   602  				results += "\t" + r.MemString()
   603  			}
   604  			fmt.Fprintln(b.w, results)
   605  			// Unlike with tests, we ignore the -chatty flag and always print output for
   606  			// benchmarks since the output generation time will skew the results.
   607  			if len(b.output) > 0 {
   608  				b.trimOutput()
   609  				fmt.Fprintf(b.w, "--- BENCH: %s\n%s", benchName, b.output)
   610  			}
   611  			if p := runtime.GOMAXPROCS(-1); p != procs {
   612  				fmt.Fprintf(os.Stderr, "testing: %s left GOMAXPROCS set to %d\n", benchName, p)
   613  			}
   614  		}
   615  	}
   616  }
   617  
   618  // Run benchmarks f as a subbenchmark with the given name. It reports
   619  // whether there were any failures.
   620  //
   621  // A subbenchmark is like any other benchmark. A benchmark that calls Run at
   622  // least once will not be measured itself and will be called once with N=1.
   623  func (b *B) Run(name string, f func(b *B)) bool {
   624  	// Since b has subbenchmarks, we will no longer run it as a benchmark itself.
   625  	// Release the lock and acquire it on exit to ensure locks stay paired.
   626  	atomic.StoreInt32(&b.hasSub, 1)
   627  	benchmarkLock.Unlock()
   628  	defer benchmarkLock.Lock()
   629  
   630  	benchName, ok, partial := b.name, true, false
   631  	if b.context != nil {
   632  		benchName, ok, partial = b.context.match.fullName(&b.common, name)
   633  	}
   634  	if !ok {
   635  		return true
   636  	}
   637  	var pc [maxStackLen]uintptr
   638  	n := runtime.Callers(2, pc[:])
   639  	sub := &B{
   640  		common: common{
   641  			signal:  make(chan bool),
   642  			name:    benchName,
   643  			parent:  &b.common,
   644  			level:   b.level + 1,
   645  			creator: pc[:n],
   646  			w:       b.w,
   647  			chatty:  b.chatty,
   648  			bench:   true,
   649  		},
   650  		importPath: b.importPath,
   651  		benchFunc:  f,
   652  		benchTime:  b.benchTime,
   653  		context:    b.context,
   654  	}
   655  	if partial {
   656  		// Partial name match, like -bench=X/Y matching BenchmarkX.
   657  		// Only process sub-benchmarks, if any.
   658  		atomic.StoreInt32(&sub.hasSub, 1)
   659  	}
   660  
   661  	if b.chatty != nil {
   662  		labelsOnce.Do(func() {
   663  			fmt.Printf("goos: %s\n", runtime.GOOS)
   664  			fmt.Printf("goarch: %s\n", runtime.GOARCH)
   665  			if b.importPath != "" {
   666  				fmt.Printf("pkg: %s\n", b.importPath)
   667  			}
   668  			if cpu := sysinfo.CPU.Name(); cpu != "" {
   669  				fmt.Printf("cpu: %s\n", cpu)
   670  			}
   671  		})
   672  
   673  		fmt.Println(benchName)
   674  	}
   675  
   676  	if sub.run1() {
   677  		sub.run()
   678  	}
   679  	b.add(sub.result)
   680  	return !sub.failed
   681  }
   682  
   683  // add simulates running benchmarks in sequence in a single iteration. It is
   684  // used to give some meaningful results in case func Benchmark is used in
   685  // combination with Run.
   686  func (b *B) add(other BenchmarkResult) {
   687  	r := &b.result
   688  	// The aggregated BenchmarkResults resemble running all subbenchmarks as
   689  	// in sequence in a single benchmark.
   690  	r.N = 1
   691  	r.T += time.Duration(other.NsPerOp())
   692  	if other.Bytes == 0 {
   693  		// Summing Bytes is meaningless in aggregate if not all subbenchmarks
   694  		// set it.
   695  		b.missingBytes = true
   696  		r.Bytes = 0
   697  	}
   698  	if !b.missingBytes {
   699  		r.Bytes += other.Bytes
   700  	}
   701  	r.MemAllocs += uint64(other.AllocsPerOp())
   702  	r.MemBytes += uint64(other.AllocedBytesPerOp())
   703  }
   704  
   705  // trimOutput shortens the output from a benchmark, which can be very long.
   706  func (b *B) trimOutput() {
   707  	// The output is likely to appear multiple times because the benchmark
   708  	// is run multiple times, but at least it will be seen. This is not a big deal
   709  	// because benchmarks rarely print, but just in case, we trim it if it's too long.
   710  	const maxNewlines = 10
   711  	for nlCount, j := 0, 0; j < len(b.output); j++ {
   712  		if b.output[j] == '\n' {
   713  			nlCount++
   714  			if nlCount >= maxNewlines {
   715  				b.output = append(b.output[:j], "\n\t... [output truncated]\n"...)
   716  				break
   717  			}
   718  		}
   719  	}
   720  }
   721  
   722  // A PB is used by RunParallel for running parallel benchmarks.
   723  type PB struct {
   724  	globalN *uint64 // shared between all worker goroutines iteration counter
   725  	grain   uint64  // acquire that many iterations from globalN at once
   726  	cache   uint64  // local cache of acquired iterations
   727  	bN      uint64  // total number of iterations to execute (b.N)
   728  }
   729  
   730  // Next reports whether there are more iterations to execute.
   731  func (pb *PB) Next() bool {
   732  	if pb.cache == 0 {
   733  		n := atomic.AddUint64(pb.globalN, pb.grain)
   734  		if n <= pb.bN {
   735  			pb.cache = pb.grain
   736  		} else if n < pb.bN+pb.grain {
   737  			pb.cache = pb.bN + pb.grain - n
   738  		} else {
   739  			return false
   740  		}
   741  	}
   742  	pb.cache--
   743  	return true
   744  }
   745  
   746  // RunParallel runs a benchmark in parallel.
   747  // It creates multiple goroutines and distributes b.N iterations among them.
   748  // The number of goroutines defaults to GOMAXPROCS. To increase parallelism for
   749  // non-CPU-bound benchmarks, call SetParallelism before RunParallel.
   750  // RunParallel is usually used with the go test -cpu flag.
   751  //
   752  // The body function will be run in each goroutine. It should set up any
   753  // goroutine-local state and then iterate until pb.Next returns false.
   754  // It should not use the StartTimer, StopTimer, or ResetTimer functions,
   755  // because they have global effect. It should also not call Run.
   756  func (b *B) RunParallel(body func(*PB)) {
   757  	if b.N == 0 {
   758  		return // Nothing to do when probing.
   759  	}
   760  	// Calculate grain size as number of iterations that take ~100µs.
   761  	// 100µs is enough to amortize the overhead and provide sufficient
   762  	// dynamic load balancing.
   763  	grain := uint64(0)
   764  	if b.previousN > 0 && b.previousDuration > 0 {
   765  		grain = 1e5 * uint64(b.previousN) / uint64(b.previousDuration)
   766  	}
   767  	if grain < 1 {
   768  		grain = 1
   769  	}
   770  	// We expect the inner loop and function call to take at least 10ns,
   771  	// so do not do more than 100µs/10ns=1e4 iterations.
   772  	if grain > 1e4 {
   773  		grain = 1e4
   774  	}
   775  
   776  	n := uint64(0)
   777  	numProcs := b.parallelism * runtime.GOMAXPROCS(0)
   778  	var wg sync.WaitGroup
   779  	wg.Add(numProcs)
   780  	for p := 0; p < numProcs; p++ {
   781  		go func() {
   782  			defer wg.Done()
   783  			pb := &PB{
   784  				globalN: &n,
   785  				grain:   grain,
   786  				bN:      uint64(b.N),
   787  			}
   788  			body(pb)
   789  		}()
   790  	}
   791  	wg.Wait()
   792  	if n <= uint64(b.N) && !b.Failed() {
   793  		b.Fatal("RunParallel: body exited without pb.Next() == false")
   794  	}
   795  }
   796  
   797  // SetParallelism sets the number of goroutines used by RunParallel to p*GOMAXPROCS.
   798  // There is usually no need to call SetParallelism for CPU-bound benchmarks.
   799  // If p is less than 1, this call will have no effect.
   800  func (b *B) SetParallelism(p int) {
   801  	if p >= 1 {
   802  		b.parallelism = p
   803  	}
   804  }
   805  
   806  // Benchmark benchmarks a single function. It is useful for creating
   807  // custom benchmarks that do not use the "go test" command.
   808  //
   809  // If f depends on testing flags, then Init must be used to register
   810  // those flags before calling Benchmark and before calling flag.Parse.
   811  //
   812  // If f calls Run, the result will be an estimate of running all its
   813  // subbenchmarks that don't call Run in sequence in a single benchmark.
   814  func Benchmark(f func(b *B)) BenchmarkResult {
   815  	b := &B{
   816  		common: common{
   817  			signal: make(chan bool),
   818  			w:      discard{},
   819  		},
   820  		benchFunc: f,
   821  		benchTime: benchTime,
   822  	}
   823  	if b.run1() {
   824  		b.run()
   825  	}
   826  	return b.result
   827  }
   828  
   829  type discard struct{}
   830  
   831  func (discard) Write(b []byte) (n int, err error) { return len(b), nil }
   832  

View as plain text