Source file src/testing/benchmark.go

Documentation: testing

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

View as plain text