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, "%s--- FAIL: %s\n%s", b.chatty.prefix(), 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 b.hasSub.Load() || 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: %s\n%s", b.chatty.prefix(), 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  // Elapsed returns the measured elapsed time of the benchmark.
   341  // The duration reported by Elapsed matches the one measured by
   342  // StartTimer, StopTimer, and ResetTimer.
   343  func (b *B) Elapsed() time.Duration {
   344  	d := b.duration
   345  	if b.timerOn {
   346  		d += time.Since(b.start)
   347  	}
   348  	return d
   349  }
   350  
   351  // ReportMetric adds "n unit" to the reported benchmark results.
   352  // If the metric is per-iteration, the caller should divide by b.N,
   353  // and by convention units should end in "/op".
   354  // ReportMetric overrides any previously reported value for the same unit.
   355  // ReportMetric panics if unit is the empty string or if unit contains
   356  // any whitespace.
   357  // If unit is a unit normally reported by the benchmark framework itself
   358  // (such as "allocs/op"), ReportMetric will override that metric.
   359  // Setting "ns/op" to 0 will suppress that built-in metric.
   360  func (b *B) ReportMetric(n float64, unit string) {
   361  	if unit == "" {
   362  		panic("metric unit must not be empty")
   363  	}
   364  	if strings.IndexFunc(unit, unicode.IsSpace) >= 0 {
   365  		panic("metric unit must not contain whitespace")
   366  	}
   367  	b.extra[unit] = n
   368  }
   369  
   370  // BenchmarkResult contains the results of a benchmark run.
   371  type BenchmarkResult struct {
   372  	N         int           // The number of iterations.
   373  	T         time.Duration // The total time taken.
   374  	Bytes     int64         // Bytes processed in one iteration.
   375  	MemAllocs uint64        // The total number of memory allocations.
   376  	MemBytes  uint64        // The total number of bytes allocated.
   377  
   378  	// Extra records additional metrics reported by ReportMetric.
   379  	Extra map[string]float64
   380  }
   381  
   382  // NsPerOp returns the "ns/op" metric.
   383  func (r BenchmarkResult) NsPerOp() int64 {
   384  	if v, ok := r.Extra["ns/op"]; ok {
   385  		return int64(v)
   386  	}
   387  	if r.N <= 0 {
   388  		return 0
   389  	}
   390  	return r.T.Nanoseconds() / int64(r.N)
   391  }
   392  
   393  // mbPerSec returns the "MB/s" metric.
   394  func (r BenchmarkResult) mbPerSec() float64 {
   395  	if v, ok := r.Extra["MB/s"]; ok {
   396  		return v
   397  	}
   398  	if r.Bytes <= 0 || r.T <= 0 || r.N <= 0 {
   399  		return 0
   400  	}
   401  	return (float64(r.Bytes) * float64(r.N) / 1e6) / r.T.Seconds()
   402  }
   403  
   404  // AllocsPerOp returns the "allocs/op" metric,
   405  // which is calculated as r.MemAllocs / r.N.
   406  func (r BenchmarkResult) AllocsPerOp() int64 {
   407  	if v, ok := r.Extra["allocs/op"]; ok {
   408  		return int64(v)
   409  	}
   410  	if r.N <= 0 {
   411  		return 0
   412  	}
   413  	return int64(r.MemAllocs) / int64(r.N)
   414  }
   415  
   416  // AllocedBytesPerOp returns the "B/op" metric,
   417  // which is calculated as r.MemBytes / r.N.
   418  func (r BenchmarkResult) AllocedBytesPerOp() int64 {
   419  	if v, ok := r.Extra["B/op"]; ok {
   420  		return int64(v)
   421  	}
   422  	if r.N <= 0 {
   423  		return 0
   424  	}
   425  	return int64(r.MemBytes) / int64(r.N)
   426  }
   427  
   428  // String returns a summary of the benchmark results.
   429  // It follows the benchmark result line format from
   430  // https://golang.org/design/14313-benchmark-format, not including the
   431  // benchmark name.
   432  // Extra metrics override built-in metrics of the same name.
   433  // String does not include allocs/op or B/op, since those are reported
   434  // by MemString.
   435  func (r BenchmarkResult) String() string {
   436  	buf := new(strings.Builder)
   437  	fmt.Fprintf(buf, "%8d", r.N)
   438  
   439  	// Get ns/op as a float.
   440  	ns, ok := r.Extra["ns/op"]
   441  	if !ok {
   442  		ns = float64(r.T.Nanoseconds()) / float64(r.N)
   443  	}
   444  	if ns != 0 {
   445  		buf.WriteByte('\t')
   446  		prettyPrint(buf, ns, "ns/op")
   447  	}
   448  
   449  	if mbs := r.mbPerSec(); mbs != 0 {
   450  		fmt.Fprintf(buf, "\t%7.2f MB/s", mbs)
   451  	}
   452  
   453  	// Print extra metrics that aren't represented in the standard
   454  	// metrics.
   455  	var extraKeys []string
   456  	for k := range r.Extra {
   457  		switch k {
   458  		case "ns/op", "MB/s", "B/op", "allocs/op":
   459  			// Built-in metrics reported elsewhere.
   460  			continue
   461  		}
   462  		extraKeys = append(extraKeys, k)
   463  	}
   464  	sort.Strings(extraKeys)
   465  	for _, k := range extraKeys {
   466  		buf.WriteByte('\t')
   467  		prettyPrint(buf, r.Extra[k], k)
   468  	}
   469  	return buf.String()
   470  }
   471  
   472  func prettyPrint(w io.Writer, x float64, unit string) {
   473  	// Print all numbers with 10 places before the decimal point
   474  	// and small numbers with four sig figs. Field widths are
   475  	// chosen to fit the whole part in 10 places while aligning
   476  	// the decimal point of all fractional formats.
   477  	var format string
   478  	switch y := math.Abs(x); {
   479  	case y == 0 || y >= 999.95:
   480  		format = "%10.0f %s"
   481  	case y >= 99.995:
   482  		format = "%12.1f %s"
   483  	case y >= 9.9995:
   484  		format = "%13.2f %s"
   485  	case y >= 0.99995:
   486  		format = "%14.3f %s"
   487  	case y >= 0.099995:
   488  		format = "%15.4f %s"
   489  	case y >= 0.0099995:
   490  		format = "%16.5f %s"
   491  	case y >= 0.00099995:
   492  		format = "%17.6f %s"
   493  	default:
   494  		format = "%18.7f %s"
   495  	}
   496  	fmt.Fprintf(w, format, x, unit)
   497  }
   498  
   499  // MemString returns r.AllocedBytesPerOp and r.AllocsPerOp in the same format as 'go test'.
   500  func (r BenchmarkResult) MemString() string {
   501  	return fmt.Sprintf("%8d B/op\t%8d allocs/op",
   502  		r.AllocedBytesPerOp(), r.AllocsPerOp())
   503  }
   504  
   505  // benchmarkName returns full name of benchmark including procs suffix.
   506  func benchmarkName(name string, n int) string {
   507  	if n != 1 {
   508  		return fmt.Sprintf("%s-%d", name, n)
   509  	}
   510  	return name
   511  }
   512  
   513  type benchContext struct {
   514  	match *matcher
   515  
   516  	maxLen int // The largest recorded benchmark name.
   517  	extLen int // Maximum extension length.
   518  }
   519  
   520  // RunBenchmarks is an internal function but exported because it is cross-package;
   521  // it is part of the implementation of the "go test" command.
   522  func RunBenchmarks(matchString func(pat, str string) (bool, error), benchmarks []InternalBenchmark) {
   523  	runBenchmarks("", matchString, benchmarks)
   524  }
   525  
   526  func runBenchmarks(importPath string, matchString func(pat, str string) (bool, error), benchmarks []InternalBenchmark) bool {
   527  	// If no flag was specified, don't run benchmarks.
   528  	if len(*matchBenchmarks) == 0 {
   529  		return true
   530  	}
   531  	// Collect matching benchmarks and determine longest name.
   532  	maxprocs := 1
   533  	for _, procs := range cpuList {
   534  		if procs > maxprocs {
   535  			maxprocs = procs
   536  		}
   537  	}
   538  	ctx := &benchContext{
   539  		match:  newMatcher(matchString, *matchBenchmarks, "-test.bench", *skip),
   540  		extLen: len(benchmarkName("", maxprocs)),
   541  	}
   542  	var bs []InternalBenchmark
   543  	for _, Benchmark := range benchmarks {
   544  		if _, matched, _ := ctx.match.fullName(nil, Benchmark.Name); matched {
   545  			bs = append(bs, Benchmark)
   546  			benchName := benchmarkName(Benchmark.Name, maxprocs)
   547  			if l := len(benchName) + ctx.extLen + 1; l > ctx.maxLen {
   548  				ctx.maxLen = l
   549  			}
   550  		}
   551  	}
   552  	main := &B{
   553  		common: common{
   554  			name:  "Main",
   555  			w:     os.Stdout,
   556  			bench: true,
   557  		},
   558  		importPath: importPath,
   559  		benchFunc: func(b *B) {
   560  			for _, Benchmark := range bs {
   561  				b.Run(Benchmark.Name, Benchmark.F)
   562  			}
   563  		},
   564  		benchTime: benchTime,
   565  		context:   ctx,
   566  	}
   567  	if Verbose() {
   568  		main.chatty = newChattyPrinter(main.w)
   569  	}
   570  	main.runN(1)
   571  	return !main.failed
   572  }
   573  
   574  // processBench runs bench b for the configured CPU counts and prints the results.
   575  func (ctx *benchContext) processBench(b *B) {
   576  	for i, procs := range cpuList {
   577  		for j := uint(0); j < *count; j++ {
   578  			runtime.GOMAXPROCS(procs)
   579  			benchName := benchmarkName(b.name, procs)
   580  
   581  			// If it's chatty, we've already printed this information.
   582  			if b.chatty == nil {
   583  				fmt.Fprintf(b.w, "%-*s\t", ctx.maxLen, benchName)
   584  			}
   585  			// Recompute the running time for all but the first iteration.
   586  			if i > 0 || j > 0 {
   587  				b = &B{
   588  					common: common{
   589  						signal: make(chan bool),
   590  						name:   b.name,
   591  						w:      b.w,
   592  						chatty: b.chatty,
   593  						bench:  true,
   594  					},
   595  					benchFunc: b.benchFunc,
   596  					benchTime: b.benchTime,
   597  				}
   598  				b.run1()
   599  			}
   600  			r := b.doBench()
   601  			if b.failed {
   602  				// The output could be very long here, but probably isn't.
   603  				// We print it all, regardless, because we don't want to trim the reason
   604  				// the benchmark failed.
   605  				fmt.Fprintf(b.w, "%s--- FAIL: %s\n%s", b.chatty.prefix(), benchName, b.output)
   606  				continue
   607  			}
   608  			results := r.String()
   609  			if b.chatty != nil {
   610  				fmt.Fprintf(b.w, "%-*s\t", ctx.maxLen, benchName)
   611  			}
   612  			if *benchmarkMemory || b.showAllocResult {
   613  				results += "\t" + r.MemString()
   614  			}
   615  			fmt.Fprintln(b.w, results)
   616  			// Unlike with tests, we ignore the -chatty flag and always print output for
   617  			// benchmarks since the output generation time will skew the results.
   618  			if len(b.output) > 0 {
   619  				b.trimOutput()
   620  				fmt.Fprintf(b.w, "%s--- BENCH: %s\n%s", b.chatty.prefix(), benchName, b.output)
   621  			}
   622  			if p := runtime.GOMAXPROCS(-1); p != procs {
   623  				fmt.Fprintf(os.Stderr, "testing: %s left GOMAXPROCS set to %d\n", benchName, p)
   624  			}
   625  			if b.chatty != nil && b.chatty.json {
   626  				b.chatty.Updatef("", "=== NAME  %s\n", "")
   627  			}
   628  		}
   629  	}
   630  }
   631  
   632  // If hideStdoutForTesting is true, Run does not print the benchName.
   633  // This avoids a spurious print during 'go test' on package testing itself,
   634  // which invokes b.Run in its own tests (see sub_test.go).
   635  var hideStdoutForTesting = false
   636  
   637  // Run benchmarks f as a subbenchmark with the given name. It reports
   638  // whether there were any failures.
   639  //
   640  // A subbenchmark is like any other benchmark. A benchmark that calls Run at
   641  // least once will not be measured itself and will be called once with N=1.
   642  func (b *B) Run(name string, f func(b *B)) bool {
   643  	// Since b has subbenchmarks, we will no longer run it as a benchmark itself.
   644  	// Release the lock and acquire it on exit to ensure locks stay paired.
   645  	b.hasSub.Store(true)
   646  	benchmarkLock.Unlock()
   647  	defer benchmarkLock.Lock()
   648  
   649  	benchName, ok, partial := b.name, true, false
   650  	if b.context != nil {
   651  		benchName, ok, partial = b.context.match.fullName(&b.common, name)
   652  	}
   653  	if !ok {
   654  		return true
   655  	}
   656  	var pc [maxStackLen]uintptr
   657  	n := runtime.Callers(2, pc[:])
   658  	sub := &B{
   659  		common: common{
   660  			signal:  make(chan bool),
   661  			name:    benchName,
   662  			parent:  &b.common,
   663  			level:   b.level + 1,
   664  			creator: pc[:n],
   665  			w:       b.w,
   666  			chatty:  b.chatty,
   667  			bench:   true,
   668  		},
   669  		importPath: b.importPath,
   670  		benchFunc:  f,
   671  		benchTime:  b.benchTime,
   672  		context:    b.context,
   673  	}
   674  	if partial {
   675  		// Partial name match, like -bench=X/Y matching BenchmarkX.
   676  		// Only process sub-benchmarks, if any.
   677  		sub.hasSub.Store(true)
   678  	}
   679  
   680  	if b.chatty != nil {
   681  		labelsOnce.Do(func() {
   682  			fmt.Printf("goos: %s\n", runtime.GOOS)
   683  			fmt.Printf("goarch: %s\n", runtime.GOARCH)
   684  			if b.importPath != "" {
   685  				fmt.Printf("pkg: %s\n", b.importPath)
   686  			}
   687  			if cpu := sysinfo.CPU.Name(); cpu != "" {
   688  				fmt.Printf("cpu: %s\n", cpu)
   689  			}
   690  		})
   691  
   692  		if !hideStdoutForTesting {
   693  			if b.chatty.json {
   694  				b.chatty.Updatef(benchName, "=== RUN   %s\n", benchName)
   695  			}
   696  			fmt.Println(benchName)
   697  		}
   698  	}
   699  
   700  	if sub.run1() {
   701  		sub.run()
   702  	}
   703  	b.add(sub.result)
   704  	return !sub.failed
   705  }
   706  
   707  // add simulates running benchmarks in sequence in a single iteration. It is
   708  // used to give some meaningful results in case func Benchmark is used in
   709  // combination with Run.
   710  func (b *B) add(other BenchmarkResult) {
   711  	r := &b.result
   712  	// The aggregated BenchmarkResults resemble running all subbenchmarks as
   713  	// in sequence in a single benchmark.
   714  	r.N = 1
   715  	r.T += time.Duration(other.NsPerOp())
   716  	if other.Bytes == 0 {
   717  		// Summing Bytes is meaningless in aggregate if not all subbenchmarks
   718  		// set it.
   719  		b.missingBytes = true
   720  		r.Bytes = 0
   721  	}
   722  	if !b.missingBytes {
   723  		r.Bytes += other.Bytes
   724  	}
   725  	r.MemAllocs += uint64(other.AllocsPerOp())
   726  	r.MemBytes += uint64(other.AllocedBytesPerOp())
   727  }
   728  
   729  // trimOutput shortens the output from a benchmark, which can be very long.
   730  func (b *B) trimOutput() {
   731  	// The output is likely to appear multiple times because the benchmark
   732  	// is run multiple times, but at least it will be seen. This is not a big deal
   733  	// because benchmarks rarely print, but just in case, we trim it if it's too long.
   734  	const maxNewlines = 10
   735  	for nlCount, j := 0, 0; j < len(b.output); j++ {
   736  		if b.output[j] == '\n' {
   737  			nlCount++
   738  			if nlCount >= maxNewlines {
   739  				b.output = append(b.output[:j], "\n\t... [output truncated]\n"...)
   740  				break
   741  			}
   742  		}
   743  	}
   744  }
   745  
   746  // A PB is used by RunParallel for running parallel benchmarks.
   747  type PB struct {
   748  	globalN *uint64 // shared between all worker goroutines iteration counter
   749  	grain   uint64  // acquire that many iterations from globalN at once
   750  	cache   uint64  // local cache of acquired iterations
   751  	bN      uint64  // total number of iterations to execute (b.N)
   752  }
   753  
   754  // Next reports whether there are more iterations to execute.
   755  func (pb *PB) Next() bool {
   756  	if pb.cache == 0 {
   757  		n := atomic.AddUint64(pb.globalN, pb.grain)
   758  		if n <= pb.bN {
   759  			pb.cache = pb.grain
   760  		} else if n < pb.bN+pb.grain {
   761  			pb.cache = pb.bN + pb.grain - n
   762  		} else {
   763  			return false
   764  		}
   765  	}
   766  	pb.cache--
   767  	return true
   768  }
   769  
   770  // RunParallel runs a benchmark in parallel.
   771  // It creates multiple goroutines and distributes b.N iterations among them.
   772  // The number of goroutines defaults to GOMAXPROCS. To increase parallelism for
   773  // non-CPU-bound benchmarks, call SetParallelism before RunParallel.
   774  // RunParallel is usually used with the go test -cpu flag.
   775  //
   776  // The body function will be run in each goroutine. It should set up any
   777  // goroutine-local state and then iterate until pb.Next returns false.
   778  // It should not use the StartTimer, StopTimer, or ResetTimer functions,
   779  // because they have global effect. It should also not call Run.
   780  //
   781  // RunParallel reports ns/op values as wall time for the benchmark as a whole,
   782  // not the sum of wall time or CPU time over each parallel goroutine.
   783  func (b *B) RunParallel(body func(*PB)) {
   784  	if b.N == 0 {
   785  		return // Nothing to do when probing.
   786  	}
   787  	// Calculate grain size as number of iterations that take ~100µs.
   788  	// 100µs is enough to amortize the overhead and provide sufficient
   789  	// dynamic load balancing.
   790  	grain := uint64(0)
   791  	if b.previousN > 0 && b.previousDuration > 0 {
   792  		grain = 1e5 * uint64(b.previousN) / uint64(b.previousDuration)
   793  	}
   794  	if grain < 1 {
   795  		grain = 1
   796  	}
   797  	// We expect the inner loop and function call to take at least 10ns,
   798  	// so do not do more than 100µs/10ns=1e4 iterations.
   799  	if grain > 1e4 {
   800  		grain = 1e4
   801  	}
   802  
   803  	n := uint64(0)
   804  	numProcs := b.parallelism * runtime.GOMAXPROCS(0)
   805  	var wg sync.WaitGroup
   806  	wg.Add(numProcs)
   807  	for p := 0; p < numProcs; p++ {
   808  		go func() {
   809  			defer wg.Done()
   810  			pb := &PB{
   811  				globalN: &n,
   812  				grain:   grain,
   813  				bN:      uint64(b.N),
   814  			}
   815  			body(pb)
   816  		}()
   817  	}
   818  	wg.Wait()
   819  	if n <= uint64(b.N) && !b.Failed() {
   820  		b.Fatal("RunParallel: body exited without pb.Next() == false")
   821  	}
   822  }
   823  
   824  // SetParallelism sets the number of goroutines used by RunParallel to p*GOMAXPROCS.
   825  // There is usually no need to call SetParallelism for CPU-bound benchmarks.
   826  // If p is less than 1, this call will have no effect.
   827  func (b *B) SetParallelism(p int) {
   828  	if p >= 1 {
   829  		b.parallelism = p
   830  	}
   831  }
   832  
   833  // Benchmark benchmarks a single function. It is useful for creating
   834  // custom benchmarks that do not use the "go test" command.
   835  //
   836  // If f depends on testing flags, then Init must be used to register
   837  // those flags before calling Benchmark and before calling flag.Parse.
   838  //
   839  // If f calls Run, the result will be an estimate of running all its
   840  // subbenchmarks that don't call Run in sequence in a single benchmark.
   841  func Benchmark(f func(b *B)) BenchmarkResult {
   842  	b := &B{
   843  		common: common{
   844  			signal: make(chan bool),
   845  			w:      discard{},
   846  		},
   847  		benchFunc: f,
   848  		benchTime: benchTime,
   849  	}
   850  	if b.run1() {
   851  		b.run()
   852  	}
   853  	return b.result
   854  }
   855  
   856  type discard struct{}
   857  
   858  func (discard) Write(b []byte) (n int, err error) { return len(b), nil }
   859  

View as plain text