...
Run Format

Source file src/pkg/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		"os"
    11		"runtime"
    12		"sync"
    13		"sync/atomic"
    14		"time"
    15	)
    16	
    17	var matchBenchmarks = flag.String("test.bench", "", "regular expression to select benchmarks to run")
    18	var benchTime = flag.Duration("test.benchtime", 1*time.Second, "approximate run time for each benchmark")
    19	var benchmarkMemory = flag.Bool("test.benchmem", false, "print memory allocations for benchmarks")
    20	
    21	// Global lock to ensure only one benchmark runs at a time.
    22	var benchmarkLock sync.Mutex
    23	
    24	// Used for every benchmark for measuring memory.
    25	var memStats runtime.MemStats
    26	
    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 {
    30		Name string
    31		F    func(b *B)
    32	}
    33	
    34	// B is a type passed to Benchmark functions to manage benchmark
    35	// timing and to specify the number of iterations to run.
    36	type B struct {
    37		common
    38		N                int
    39		previousN        int           // number of iterations in the previous run
    40		previousDuration time.Duration // total duration of the previous run
    41		benchmark        InternalBenchmark
    42		bytes            int64
    43		timerOn          bool
    44		showAllocResult  bool
    45		result           BenchmarkResult
    46		parallelism      int // RunParallel creates parallelism*GOMAXPROCS goroutines
    47		// The initial states of memStats.Mallocs and memStats.TotalAlloc.
    48		startAllocs uint64
    49		startBytes  uint64
    50		// The net total of this test after being run.
    51		netAllocs uint64
    52		netBytes  uint64
    53	}
    54	
    55	// StartTimer starts timing a test.  This function is called automatically
    56	// before a benchmark starts, but it can also used to resume timing after
    57	// a call to StopTimer.
    58	func (b *B) StartTimer() {
    59		if !b.timerOn {
    60			runtime.ReadMemStats(&memStats)
    61			b.startAllocs = memStats.Mallocs
    62			b.startBytes = memStats.TotalAlloc
    63			b.start = time.Now()
    64			b.timerOn = true
    65		}
    66	}
    67	
    68	// StopTimer stops timing a test.  This can be used to pause the timer
    69	// while performing complex initialization that you don't
    70	// want to measure.
    71	func (b *B) StopTimer() {
    72		if b.timerOn {
    73			b.duration += time.Now().Sub(b.start)
    74			runtime.ReadMemStats(&memStats)
    75			b.netAllocs += memStats.Mallocs - b.startAllocs
    76			b.netBytes += memStats.TotalAlloc - b.startBytes
    77			b.timerOn = false
    78		}
    79	}
    80	
    81	// ResetTimer zeros the elapsed benchmark time and memory allocation counters.
    82	// It does not affect whether the timer is running.
    83	func (b *B) ResetTimer() {
    84		if b.timerOn {
    85			runtime.ReadMemStats(&memStats)
    86			b.startAllocs = memStats.Mallocs
    87			b.startBytes = memStats.TotalAlloc
    88			b.start = time.Now()
    89		}
    90		b.duration = 0
    91		b.netAllocs = 0
    92		b.netBytes = 0
    93	}
    94	
    95	// SetBytes records the number of bytes processed in a single operation.
    96	// If this is called, the benchmark will report ns/op and MB/s.
    97	func (b *B) SetBytes(n int64) { b.bytes = n }
    98	
    99	// ReportAllocs enables malloc statistics for this benchmark.
   100	// It is equivalent to setting -test.benchmem, but it only affects the
   101	// benchmark function that calls ReportAllocs.
   102	func (b *B) ReportAllocs() {
   103		b.showAllocResult = true
   104	}
   105	
   106	func (b *B) nsPerOp() int64 {
   107		if b.N <= 0 {
   108			return 0
   109		}
   110		return b.duration.Nanoseconds() / int64(b.N)
   111	}
   112	
   113	// runN runs a single benchmark for the specified number of iterations.
   114	func (b *B) runN(n int) {
   115		benchmarkLock.Lock()
   116		defer benchmarkLock.Unlock()
   117		// Try to get a comparable environment for each run
   118		// by clearing garbage from previous runs.
   119		runtime.GC()
   120		b.N = n
   121		b.parallelism = 1
   122		b.ResetTimer()
   123		b.StartTimer()
   124		b.benchmark.F(b)
   125		b.StopTimer()
   126		b.previousN = n
   127		b.previousDuration = b.duration
   128	}
   129	
   130	func min(x, y int) int {
   131		if x > y {
   132			return y
   133		}
   134		return x
   135	}
   136	
   137	func max(x, y int) int {
   138		if x < y {
   139			return y
   140		}
   141		return x
   142	}
   143	
   144	// roundDown10 rounds a number down to the nearest power of 10.
   145	func roundDown10(n int) int {
   146		var tens = 0
   147		// tens = floor(log_10(n))
   148		for n >= 10 {
   149			n = n / 10
   150			tens++
   151		}
   152		// result = 10^tens
   153		result := 1
   154		for i := 0; i < tens; i++ {
   155			result *= 10
   156		}
   157		return result
   158	}
   159	
   160	// roundUp rounds x up to a number of the form [1eX, 2eX, 5eX].
   161	func roundUp(n int) int {
   162		base := roundDown10(n)
   163		switch {
   164		case n <= base:
   165			return base
   166		case n <= (2 * base):
   167			return 2 * base
   168		case n <= (5 * base):
   169			return 5 * base
   170		default:
   171			return 10 * base
   172		}
   173	}
   174	
   175	// run times the benchmark function in a separate goroutine.
   176	func (b *B) run() BenchmarkResult {
   177		go b.launch()
   178		<-b.signal
   179		return b.result
   180	}
   181	
   182	// launch launches the benchmark function.  It gradually increases the number
   183	// of benchmark iterations until the benchmark runs for a second in order
   184	// to get a reasonable measurement.  It prints timing information in this form
   185	//		testing.BenchmarkHello	100000		19 ns/op
   186	// launch is run by the fun function as a separate goroutine.
   187	func (b *B) launch() {
   188		// Run the benchmark for a single iteration in case it's expensive.
   189		n := 1
   190	
   191		// Signal that we're done whether we return normally
   192		// or by FailNow's runtime.Goexit.
   193		defer func() {
   194			b.signal <- b
   195		}()
   196	
   197		b.runN(n)
   198		// Run the benchmark for at least the specified amount of time.
   199		d := *benchTime
   200		for !b.failed && b.duration < d && n < 1e9 {
   201			last := n
   202			// Predict iterations/sec.
   203			if b.nsPerOp() == 0 {
   204				n = 1e9
   205			} else {
   206				n = int(d.Nanoseconds() / b.nsPerOp())
   207			}
   208			// Run more iterations than we think we'll need for a second (1.5x).
   209			// Don't grow too fast in case we had timing errors previously.
   210			// Be sure to run at least one more than last time.
   211			n = max(min(n+n/2, 100*last), last+1)
   212			// Round up to something easy to read.
   213			n = roundUp(n)
   214			b.runN(n)
   215		}
   216		b.result = BenchmarkResult{b.N, b.duration, b.bytes, b.netAllocs, b.netBytes}
   217	}
   218	
   219	// The results of a benchmark run.
   220	type BenchmarkResult struct {
   221		N         int           // The number of iterations.
   222		T         time.Duration // The total time taken.
   223		Bytes     int64         // Bytes processed in one iteration.
   224		MemAllocs uint64        // The total number of memory allocations.
   225		MemBytes  uint64        // The total number of bytes allocated.
   226	}
   227	
   228	func (r BenchmarkResult) NsPerOp() int64 {
   229		if r.N <= 0 {
   230			return 0
   231		}
   232		return r.T.Nanoseconds() / int64(r.N)
   233	}
   234	
   235	func (r BenchmarkResult) mbPerSec() float64 {
   236		if r.Bytes <= 0 || r.T <= 0 || r.N <= 0 {
   237			return 0
   238		}
   239		return (float64(r.Bytes) * float64(r.N) / 1e6) / r.T.Seconds()
   240	}
   241	
   242	func (r BenchmarkResult) AllocsPerOp() int64 {
   243		if r.N <= 0 {
   244			return 0
   245		}
   246		return int64(r.MemAllocs) / int64(r.N)
   247	}
   248	
   249	func (r BenchmarkResult) AllocedBytesPerOp() int64 {
   250		if r.N <= 0 {
   251			return 0
   252		}
   253		return int64(r.MemBytes) / int64(r.N)
   254	}
   255	
   256	func (r BenchmarkResult) String() string {
   257		mbs := r.mbPerSec()
   258		mb := ""
   259		if mbs != 0 {
   260			mb = fmt.Sprintf("\t%7.2f MB/s", mbs)
   261		}
   262		nsop := r.NsPerOp()
   263		ns := fmt.Sprintf("%10d ns/op", nsop)
   264		if r.N > 0 && nsop < 100 {
   265			// The format specifiers here make sure that
   266			// the ones digits line up for all three possible formats.
   267			if nsop < 10 {
   268				ns = fmt.Sprintf("%13.2f ns/op", float64(r.T.Nanoseconds())/float64(r.N))
   269			} else {
   270				ns = fmt.Sprintf("%12.1f ns/op", float64(r.T.Nanoseconds())/float64(r.N))
   271			}
   272		}
   273		return fmt.Sprintf("%8d\t%s%s", r.N, ns, mb)
   274	}
   275	
   276	func (r BenchmarkResult) MemString() string {
   277		return fmt.Sprintf("%8d B/op\t%8d allocs/op",
   278			r.AllocedBytesPerOp(), r.AllocsPerOp())
   279	}
   280	
   281	// An internal function but exported because it is cross-package; part of the implementation
   282	// of the "go test" command.
   283	func RunBenchmarks(matchString func(pat, str string) (bool, error), benchmarks []InternalBenchmark) {
   284		// If no flag was specified, don't run benchmarks.
   285		if len(*matchBenchmarks) == 0 {
   286			return
   287		}
   288		for _, Benchmark := range benchmarks {
   289			matched, err := matchString(*matchBenchmarks, Benchmark.Name)
   290			if err != nil {
   291				fmt.Fprintf(os.Stderr, "testing: invalid regexp for -test.bench: %s\n", err)
   292				os.Exit(1)
   293			}
   294			if !matched {
   295				continue
   296			}
   297			for _, procs := range cpuList {
   298				runtime.GOMAXPROCS(procs)
   299				b := &B{
   300					common: common{
   301						signal: make(chan interface{}),
   302					},
   303					benchmark: Benchmark,
   304				}
   305				benchName := Benchmark.Name
   306				if procs != 1 {
   307					benchName = fmt.Sprintf("%s-%d", Benchmark.Name, procs)
   308				}
   309				fmt.Printf("%s\t", benchName)
   310				r := b.run()
   311				if b.failed {
   312					// The output could be very long here, but probably isn't.
   313					// We print it all, regardless, because we don't want to trim the reason
   314					// the benchmark failed.
   315					fmt.Printf("--- FAIL: %s\n%s", benchName, b.output)
   316					continue
   317				}
   318				results := r.String()
   319				if *benchmarkMemory || b.showAllocResult {
   320					results += "\t" + r.MemString()
   321				}
   322				fmt.Println(results)
   323				// Unlike with tests, we ignore the -chatty flag and always print output for
   324				// benchmarks since the output generation time will skew the results.
   325				if len(b.output) > 0 {
   326					b.trimOutput()
   327					fmt.Printf("--- BENCH: %s\n%s", benchName, b.output)
   328				}
   329				if p := runtime.GOMAXPROCS(-1); p != procs {
   330					fmt.Fprintf(os.Stderr, "testing: %s left GOMAXPROCS set to %d\n", benchName, p)
   331				}
   332			}
   333		}
   334	}
   335	
   336	// trimOutput shortens the output from a benchmark, which can be very long.
   337	func (b *B) trimOutput() {
   338		// The output is likely to appear multiple times because the benchmark
   339		// is run multiple times, but at least it will be seen. This is not a big deal
   340		// because benchmarks rarely print, but just in case, we trim it if it's too long.
   341		const maxNewlines = 10
   342		for nlCount, j := 0, 0; j < len(b.output); j++ {
   343			if b.output[j] == '\n' {
   344				nlCount++
   345				if nlCount >= maxNewlines {
   346					b.output = append(b.output[:j], "\n\t... [output truncated]\n"...)
   347					break
   348				}
   349			}
   350		}
   351	}
   352	
   353	// A PB is used by RunParallel for running parallel benchmarks.
   354	type PB struct {
   355		globalN *uint64 // shared between all worker goroutines iteration counter
   356		grain   uint64  // acquire that many iterations from globalN at once
   357		cache   uint64  // local cache of acquired iterations
   358		bN      uint64  // total number of iterations to execute (b.N)
   359	}
   360	
   361	// Next reports whether there are more iterations to execute.
   362	func (pb *PB) Next() bool {
   363		if pb.cache == 0 {
   364			n := atomic.AddUint64(pb.globalN, pb.grain)
   365			if n <= pb.bN {
   366				pb.cache = pb.grain
   367			} else if n < pb.bN+pb.grain {
   368				pb.cache = pb.bN + pb.grain - n
   369			} else {
   370				return false
   371			}
   372		}
   373		pb.cache--
   374		return true
   375	}
   376	
   377	// RunParallel runs a benchmark in parallel.
   378	// It creates multiple goroutines and distributes b.N iterations among them.
   379	// The number of goroutines defaults to GOMAXPROCS. To increase parallelism for
   380	// non-CPU-bound benchmarks, call SetParallelism before RunParallel.
   381	// RunParallel is usually used with the go test -cpu flag.
   382	//
   383	// The body function will be run in each goroutine. It should set up any
   384	// goroutine-local state and then iterate until pb.Next returns false.
   385	// It should not use the StartTimer, StopTimer, or ResetTimer functions,
   386	// because they have global effect.
   387	func (b *B) RunParallel(body func(*PB)) {
   388		// Calculate grain size as number of iterations that take ~100µs.
   389		// 100µs is enough to amortize the overhead and provide sufficient
   390		// dynamic load balancing.
   391		grain := uint64(0)
   392		if b.previousN > 0 && b.previousDuration > 0 {
   393			grain = 1e5 * uint64(b.previousN) / uint64(b.previousDuration)
   394		}
   395		if grain < 1 {
   396			grain = 1
   397		}
   398		// We expect the inner loop and function call to take at least 10ns,
   399		// so do not do more than 100µs/10ns=1e4 iterations.
   400		if grain > 1e4 {
   401			grain = 1e4
   402		}
   403	
   404		n := uint64(0)
   405		numProcs := b.parallelism * runtime.GOMAXPROCS(0)
   406		var wg sync.WaitGroup
   407		wg.Add(numProcs)
   408		for p := 0; p < numProcs; p++ {
   409			go func() {
   410				defer wg.Done()
   411				pb := &PB{
   412					globalN: &n,
   413					grain:   grain,
   414					bN:      uint64(b.N),
   415				}
   416				body(pb)
   417			}()
   418		}
   419		wg.Wait()
   420		if n <= uint64(b.N) && !b.Failed() {
   421			b.Fatal("RunParallel: body exited without pb.Next() == false")
   422		}
   423	}
   424	
   425	// SetParallelism sets the number of goroutines used by RunParallel to p*GOMAXPROCS.
   426	// There is usually no need to call SetParallelism for CPU-bound benchmarks.
   427	// If p is less than 1, this call will have no effect.
   428	func (b *B) SetParallelism(p int) {
   429		if p >= 1 {
   430			b.parallelism = p
   431		}
   432	}
   433	
   434	// Benchmark benchmarks a single function. Useful for creating
   435	// custom benchmarks that do not use the "go test" command.
   436	func Benchmark(f func(b *B)) BenchmarkResult {
   437		b := &B{
   438			common: common{
   439				signal: make(chan interface{}),
   440			},
   441			benchmark: InternalBenchmark{"", f},
   442		}
   443		return b.run()
   444	}
   445	

View as plain text