...
Run Format

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		"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, 3eX, 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 <= (3 * base):
   169			return 3 * base
   170		case n <= (5 * base):
   171			return 5 * base
   172		default:
   173			return 10 * base
   174		}
   175	}
   176	
   177	// run times the benchmark function in a separate goroutine.
   178	func (b *B) run() BenchmarkResult {
   179		go b.launch()
   180		<-b.signal
   181		return b.result
   182	}
   183	
   184	// launch launches the benchmark function.  It gradually increases the number
   185	// of benchmark iterations until the benchmark runs for the requested benchtime.
   186	// It prints timing information in this form
   187	//		testing.BenchmarkHello	100000		19 ns/op
   188	// launch is run by the run function as a separate goroutine.
   189	func (b *B) launch() {
   190		// Run the benchmark for a single iteration in case it's expensive.
   191		n := 1
   192	
   193		// Signal that we're done whether we return normally
   194		// or by FailNow's runtime.Goexit.
   195		defer func() {
   196			b.signal <- b
   197		}()
   198	
   199		b.runN(n)
   200		// Run the benchmark for at least the specified amount of time.
   201		d := *benchTime
   202		for !b.failed && b.duration < d && n < 1e9 {
   203			last := n
   204			// Predict required iterations.
   205			if b.nsPerOp() == 0 {
   206				n = 1e9
   207			} else {
   208				n = int(d.Nanoseconds() / b.nsPerOp())
   209			}
   210			// Run more iterations than we think we'll need (1.2x).
   211			// Don't grow too fast in case we had timing errors previously.
   212			// Be sure to run at least one more than last time.
   213			n = max(min(n+n/5, 100*last), last+1)
   214			// Round up to something easy to read.
   215			n = roundUp(n)
   216			b.runN(n)
   217		}
   218		b.result = BenchmarkResult{b.N, b.duration, b.bytes, b.netAllocs, b.netBytes}
   219	}
   220	
   221	// The results of a benchmark run.
   222	type BenchmarkResult struct {
   223		N         int           // The number of iterations.
   224		T         time.Duration // The total time taken.
   225		Bytes     int64         // Bytes processed in one iteration.
   226		MemAllocs uint64        // The total number of memory allocations.
   227		MemBytes  uint64        // The total number of bytes allocated.
   228	}
   229	
   230	func (r BenchmarkResult) NsPerOp() int64 {
   231		if r.N <= 0 {
   232			return 0
   233		}
   234		return r.T.Nanoseconds() / int64(r.N)
   235	}
   236	
   237	func (r BenchmarkResult) mbPerSec() float64 {
   238		if r.Bytes <= 0 || r.T <= 0 || r.N <= 0 {
   239			return 0
   240		}
   241		return (float64(r.Bytes) * float64(r.N) / 1e6) / r.T.Seconds()
   242	}
   243	
   244	func (r BenchmarkResult) AllocsPerOp() int64 {
   245		if r.N <= 0 {
   246			return 0
   247		}
   248		return int64(r.MemAllocs) / int64(r.N)
   249	}
   250	
   251	func (r BenchmarkResult) AllocedBytesPerOp() int64 {
   252		if r.N <= 0 {
   253			return 0
   254		}
   255		return int64(r.MemBytes) / int64(r.N)
   256	}
   257	
   258	func (r BenchmarkResult) String() string {
   259		mbs := r.mbPerSec()
   260		mb := ""
   261		if mbs != 0 {
   262			mb = fmt.Sprintf("\t%7.2f MB/s", mbs)
   263		}
   264		nsop := r.NsPerOp()
   265		ns := fmt.Sprintf("%10d ns/op", nsop)
   266		if r.N > 0 && nsop < 100 {
   267			// The format specifiers here make sure that
   268			// the ones digits line up for all three possible formats.
   269			if nsop < 10 {
   270				ns = fmt.Sprintf("%13.2f ns/op", float64(r.T.Nanoseconds())/float64(r.N))
   271			} else {
   272				ns = fmt.Sprintf("%12.1f ns/op", float64(r.T.Nanoseconds())/float64(r.N))
   273			}
   274		}
   275		return fmt.Sprintf("%8d\t%s%s", r.N, ns, mb)
   276	}
   277	
   278	func (r BenchmarkResult) MemString() string {
   279		return fmt.Sprintf("%8d B/op\t%8d allocs/op",
   280			r.AllocedBytesPerOp(), r.AllocsPerOp())
   281	}
   282	
   283	// An internal function but exported because it is cross-package; part of the implementation
   284	// of the "go test" command.
   285	func RunBenchmarks(matchString func(pat, str string) (bool, error), benchmarks []InternalBenchmark) {
   286		// If no flag was specified, don't run benchmarks.
   287		if len(*matchBenchmarks) == 0 {
   288			return
   289		}
   290		for _, Benchmark := range benchmarks {
   291			matched, err := matchString(*matchBenchmarks, Benchmark.Name)
   292			if err != nil {
   293				fmt.Fprintf(os.Stderr, "testing: invalid regexp for -test.bench: %s\n", err)
   294				os.Exit(1)
   295			}
   296			if !matched {
   297				continue
   298			}
   299			for _, procs := range cpuList {
   300				runtime.GOMAXPROCS(procs)
   301				b := &B{
   302					common: common{
   303						signal: make(chan interface{}),
   304					},
   305					benchmark: Benchmark,
   306				}
   307				benchName := Benchmark.Name
   308				if procs != 1 {
   309					benchName = fmt.Sprintf("%s-%d", Benchmark.Name, procs)
   310				}
   311				fmt.Printf("%s\t", benchName)
   312				r := b.run()
   313				if b.failed {
   314					// The output could be very long here, but probably isn't.
   315					// We print it all, regardless, because we don't want to trim the reason
   316					// the benchmark failed.
   317					fmt.Printf("--- FAIL: %s\n%s", benchName, b.output)
   318					continue
   319				}
   320				results := r.String()
   321				if *benchmarkMemory || b.showAllocResult {
   322					results += "\t" + r.MemString()
   323				}
   324				fmt.Println(results)
   325				// Unlike with tests, we ignore the -chatty flag and always print output for
   326				// benchmarks since the output generation time will skew the results.
   327				if len(b.output) > 0 {
   328					b.trimOutput()
   329					fmt.Printf("--- BENCH: %s\n%s", benchName, b.output)
   330				}
   331				if p := runtime.GOMAXPROCS(-1); p != procs {
   332					fmt.Fprintf(os.Stderr, "testing: %s left GOMAXPROCS set to %d\n", benchName, p)
   333				}
   334			}
   335		}
   336	}
   337	
   338	// trimOutput shortens the output from a benchmark, which can be very long.
   339	func (b *B) trimOutput() {
   340		// The output is likely to appear multiple times because the benchmark
   341		// is run multiple times, but at least it will be seen. This is not a big deal
   342		// because benchmarks rarely print, but just in case, we trim it if it's too long.
   343		const maxNewlines = 10
   344		for nlCount, j := 0, 0; j < len(b.output); j++ {
   345			if b.output[j] == '\n' {
   346				nlCount++
   347				if nlCount >= maxNewlines {
   348					b.output = append(b.output[:j], "\n\t... [output truncated]\n"...)
   349					break
   350				}
   351			}
   352		}
   353	}
   354	
   355	// A PB is used by RunParallel for running parallel benchmarks.
   356	type PB struct {
   357		globalN *uint64 // shared between all worker goroutines iteration counter
   358		grain   uint64  // acquire that many iterations from globalN at once
   359		cache   uint64  // local cache of acquired iterations
   360		bN      uint64  // total number of iterations to execute (b.N)
   361	}
   362	
   363	// Next reports whether there are more iterations to execute.
   364	func (pb *PB) Next() bool {
   365		if pb.cache == 0 {
   366			n := atomic.AddUint64(pb.globalN, pb.grain)
   367			if n <= pb.bN {
   368				pb.cache = pb.grain
   369			} else if n < pb.bN+pb.grain {
   370				pb.cache = pb.bN + pb.grain - n
   371			} else {
   372				return false
   373			}
   374		}
   375		pb.cache--
   376		return true
   377	}
   378	
   379	// RunParallel runs a benchmark in parallel.
   380	// It creates multiple goroutines and distributes b.N iterations among them.
   381	// The number of goroutines defaults to GOMAXPROCS. To increase parallelism for
   382	// non-CPU-bound benchmarks, call SetParallelism before RunParallel.
   383	// RunParallel is usually used with the go test -cpu flag.
   384	//
   385	// The body function will be run in each goroutine. It should set up any
   386	// goroutine-local state and then iterate until pb.Next returns false.
   387	// It should not use the StartTimer, StopTimer, or ResetTimer functions,
   388	// because they have global effect.
   389	func (b *B) RunParallel(body func(*PB)) {
   390		// Calculate grain size as number of iterations that take ~100µs.
   391		// 100µs is enough to amortize the overhead and provide sufficient
   392		// dynamic load balancing.
   393		grain := uint64(0)
   394		if b.previousN > 0 && b.previousDuration > 0 {
   395			grain = 1e5 * uint64(b.previousN) / uint64(b.previousDuration)
   396		}
   397		if grain < 1 {
   398			grain = 1
   399		}
   400		// We expect the inner loop and function call to take at least 10ns,
   401		// so do not do more than 100µs/10ns=1e4 iterations.
   402		if grain > 1e4 {
   403			grain = 1e4
   404		}
   405	
   406		n := uint64(0)
   407		numProcs := b.parallelism * runtime.GOMAXPROCS(0)
   408		var wg sync.WaitGroup
   409		wg.Add(numProcs)
   410		for p := 0; p < numProcs; p++ {
   411			go func() {
   412				defer wg.Done()
   413				pb := &PB{
   414					globalN: &n,
   415					grain:   grain,
   416					bN:      uint64(b.N),
   417				}
   418				body(pb)
   419			}()
   420		}
   421		wg.Wait()
   422		if n <= uint64(b.N) && !b.Failed() {
   423			b.Fatal("RunParallel: body exited without pb.Next() == false")
   424		}
   425	}
   426	
   427	// SetParallelism sets the number of goroutines used by RunParallel to p*GOMAXPROCS.
   428	// There is usually no need to call SetParallelism for CPU-bound benchmarks.
   429	// If p is less than 1, this call will have no effect.
   430	func (b *B) SetParallelism(p int) {
   431		if p >= 1 {
   432			b.parallelism = p
   433		}
   434	}
   435	
   436	// Benchmark benchmarks a single function. Useful for creating
   437	// custom benchmarks that do not use the "go test" command.
   438	func Benchmark(f func(b *B)) BenchmarkResult {
   439		b := &B{
   440			common: common{
   441				signal: make(chan interface{}),
   442			},
   443			benchmark: InternalBenchmark{"", f},
   444		}
   445		return b.run()
   446	}
   447	

View as plain text