sync: use RunParallel in benchmarks

LGTM=bradfitz
R=golang-codereviews, bradfitz
CC=golang-codereviews
https://golang.org/cl/68050043

src/pkg/sync/mutex_test.go

@@ -9,7 +9,6 @@ package sync_test
 import (
 	"runtime"
 	. "sync"
-	"sync/atomic"
 	"testing"
 )
 
@@ -90,63 +89,34 @@ func BenchmarkMutexUncontended(b *testing.B) {
 		Mutex
 		pad [128]uint8
 	}
-	const CallsPerSched = 1000
+	b.RunParallel(func(pb *testing.PB) {
-	procs := runtime.GOMAXPROCS(-1)
+		var mu PaddedMutex
-	N := int32(b.N / CallsPerSched)
+		for pb.Next() {
-	c := make(chan bool, procs)
+			mu.Lock()
-	for p := 0; p < procs; p++ {
+			mu.Unlock()
-		go func() {
+		}
-			var mu PaddedMutex
+	})
-			for atomic.AddInt32(&N, -1) >= 0 {
-				runtime.Gosched()
-				for g := 0; g < CallsPerSched; g++ {
-					mu.Lock()
-					mu.Unlock()
-				}
-			}
-			c <- true
-		}()
-	}
-	for p := 0; p < procs; p++ {
-		<-c
-	}
 }
 
 func benchmarkMutex(b *testing.B, slack, work bool) {
-	const (
-		CallsPerSched  = 1000
-		LocalWork      = 100
-		GoroutineSlack = 10
-	)
-	procs := runtime.GOMAXPROCS(-1)
-	if slack {
-		procs *= GoroutineSlack
-	}
-	N := int32(b.N / CallsPerSched)
-	c := make(chan bool, procs)
 	var mu Mutex
-	for p := 0; p < procs; p++ {
+	if slack {
-		go func() {
+		b.SetParallelism(10)
-			foo := 0
+	}
-			for atomic.AddInt32(&N, -1) >= 0 {
+	b.RunParallel(func(pb *testing.PB) {
-				runtime.Gosched()
+		foo := 0
-				for g := 0; g < CallsPerSched; g++ {
+		for pb.Next() {
-					mu.Lock()
+			mu.Lock()
-					mu.Unlock()
+			mu.Unlock()
-					if work {
+			if work {
-						for i := 0; i < LocalWork; i++ {
+				for i := 0; i < 100; i++ {
-							foo *= 2
+					foo *= 2
-							foo /= 2
+					foo /= 2
-						}
-					}
 				}
 			}
-			c <- foo == 42
+		}
-		}()
+		_ = foo
-	}
+	})
-	for p := 0; p < procs; p++ {
-		<-c
-	}
 }
 
 func BenchmarkMutex(b *testing.B) {

src/pkg/sync/once_test.go

@@ -5,9 +5,7 @@
 package sync_test
 
 import (
-	"runtime"
 	. "sync"
-	"sync/atomic"
 	"testing"
 )
 
@@ -62,24 +60,11 @@ func TestOncePanic(t *testing.T) {
 }
 
 func BenchmarkOnce(b *testing.B) {
-	const CallsPerSched = 1000
-	procs := runtime.GOMAXPROCS(-1)
-	N := int32(b.N / CallsPerSched)
 	var once Once
 	f := func() {}
-	c := make(chan bool, procs)
+	b.RunParallel(func(pb *testing.PB) {
-	for p := 0; p < procs; p++ {
+		for pb.Next() {
-		go func() {
+			once.Do(f)
-			for atomic.AddInt32(&N, -1) >= 0 {
+		}
-				runtime.Gosched()
+	})
-				for g := 0; g < CallsPerSched; g++ {
-					once.Do(f)
-				}
-			}
-			c <- true
-		}()
-	}
-	for p := 0; p < procs; p++ {
-		<-c
-	}
 }

src/pkg/sync/pool_test.go

@@ -128,42 +128,24 @@ func TestPoolStress(t *testing.T) {
 
 func BenchmarkPool(b *testing.B) {
 	var p Pool
-	var wg WaitGroup
+	b.RunParallel(func(pb *testing.PB) {
-	n0 := uintptr(b.N)
+		for pb.Next() {
-	n := n0
+			p.Put(1)
-	for i := 0; i < runtime.GOMAXPROCS(0); i++ {
+			p.Get()
-		wg.Add(1)
+		}
-		go func() {
+	})
-			defer wg.Done()
-			for atomic.AddUintptr(&n, ^uintptr(0)) < n0 {
-				for b := 0; b < 100; b++ {
-					p.Put(1)
-					p.Get()
-				}
-			}
-		}()
-	}
-	wg.Wait()
 }
 
 func BenchmarkPoolOverlflow(b *testing.B) {
 	var p Pool
-	var wg WaitGroup
+	b.RunParallel(func(pb *testing.PB) {
-	n0 := uintptr(b.N)
+		for pb.Next() {
-	n := n0
+			for b := 0; b < 100; b++ {
-	for i := 0; i < runtime.GOMAXPROCS(0); i++ {
+				p.Put(1)
-		wg.Add(1)
-		go func() {
-			defer wg.Done()
-			for atomic.AddUintptr(&n, ^uintptr(0)) < n0 {
-				for b := 0; b < 100; b++ {
-					p.Put(1)
-				}
-				for b := 0; b < 100; b++ {
-					p.Get()
-				}
 			}
-		}()
+			for b := 0; b < 100; b++ {
-	}
+				p.Get()
-	wg.Wait()
+			}
+		}
+	})
 }

src/pkg/sync/runtime_sema_test.go

@@ -7,7 +7,6 @@ package sync_test
 import (
 	"runtime"
 	. "sync"
-	"sync/atomic"
 	"testing"
 )
 
@@ -16,72 +15,44 @@ func BenchmarkSemaUncontended(b *testing.B) {
 		sem uint32
 		pad [32]uint32
 	}
-	const CallsPerSched = 1000
+	b.RunParallel(func(pb *testing.PB) {
-	procs := runtime.GOMAXPROCS(-1)
+		sem := new(PaddedSem)
-	N := int32(b.N / CallsPerSched)
+		for pb.Next() {
-	c := make(chan bool, procs)
+			Runtime_Semrelease(&sem.sem)
-	for p := 0; p < procs; p++ {
+			Runtime_Semacquire(&sem.sem)
-		go func() {
+		}
-			sem := new(PaddedSem)
+	})
-			for atomic.AddInt32(&N, -1) >= 0 {
-				runtime.Gosched()
-				for g := 0; g < CallsPerSched; g++ {
-					Runtime_Semrelease(&sem.sem)
-					Runtime_Semacquire(&sem.sem)
-				}
-			}
-			c <- true
-		}()
-	}
-	for p := 0; p < procs; p++ {
-		<-c
-	}
 }
 
 func benchmarkSema(b *testing.B, block, work bool) {
-	const CallsPerSched = 1000
-	const LocalWork = 100
-	procs := runtime.GOMAXPROCS(-1)
-	N := int32(b.N / CallsPerSched)
-	c := make(chan bool, procs)
-	c2 := make(chan bool, procs/2)
 	sem := uint32(0)
 	if block {
-		for p := 0; p < procs/2; p++ {
+		done := make(chan bool)
-			go func() {
-				Runtime_Semacquire(&sem)
-				c2 <- true
-			}()
-		}
-	}
-	for p := 0; p < procs; p++ {
 		go func() {
-			foo := 0
+			for p := 0; p < runtime.GOMAXPROCS(0)/2; p++ {
-			for atomic.AddInt32(&N, -1) >= 0 {
+				Runtime_Semacquire(&sem)
-				runtime.Gosched()
-				for g := 0; g < CallsPerSched; g++ {
-					Runtime_Semrelease(&sem)
-					if work {
-						for i := 0; i < LocalWork; i++ {
-							foo *= 2
-							foo /= 2
-						}
-					}
-					Runtime_Semacquire(&sem)
-				}
 			}
-			c <- foo == 42
+			done <- true
-			Runtime_Semrelease(&sem)
+		}()
+		defer func() {
+			<-done
 		}()
 	}
-	if block {
+	b.RunParallel(func(pb *testing.PB) {
-		for p := 0; p < procs/2; p++ {
+		foo := 0
-			<-c2
+		for pb.Next() {
+			Runtime_Semrelease(&sem)
+			if work {
+				for i := 0; i < 100; i++ {
+					foo *= 2
+					foo /= 2
+				}
+			}
+			Runtime_Semacquire(&sem)
 		}
-	}
+		_ = foo
-	for p := 0; p < procs; p++ {
+		Runtime_Semrelease(&sem)
-		<-c
+	})
-	}
 }
 
 func BenchmarkSemaSyntNonblock(b *testing.B) {

src/pkg/sync/rwmutex_test.go

@@ -160,64 +160,39 @@ func BenchmarkRWMutexUncontended(b *testing.B) {
 		RWMutex
 		pad [32]uint32
 	}
-	const CallsPerSched = 1000
+	b.RunParallel(func(pb *testing.PB) {
-	procs := runtime.GOMAXPROCS(-1)
+		var rwm PaddedRWMutex
-	N := int32(b.N / CallsPerSched)
+		for pb.Next() {
-	c := make(chan bool, procs)
+			rwm.RLock()
-	for p := 0; p < procs; p++ {
+			rwm.RLock()
-		go func() {
+			rwm.RUnlock()
-			var rwm PaddedRWMutex
+			rwm.RUnlock()
-			for atomic.AddInt32(&N, -1) >= 0 {
+			rwm.Lock()
-				runtime.Gosched()
+			rwm.Unlock()
-				for g := 0; g < CallsPerSched; g++ {
+		}
-					rwm.RLock()
+	})
-					rwm.RLock()
-					rwm.RUnlock()
-					rwm.RUnlock()
-					rwm.Lock()
-					rwm.Unlock()
-				}
-			}
-			c <- true
-		}()
-	}
-	for p := 0; p < procs; p++ {
-		<-c
-	}
 }
 
 func benchmarkRWMutex(b *testing.B, localWork, writeRatio int) {
-	const CallsPerSched = 1000
-	procs := runtime.GOMAXPROCS(-1)
-	N := int32(b.N / CallsPerSched)
-	c := make(chan bool, procs)
 	var rwm RWMutex
-	for p := 0; p < procs; p++ {
+	b.RunParallel(func(pb *testing.PB) {
-		go func() {
+		foo := 0
-			foo := 0
+		for pb.Next() {
-			for atomic.AddInt32(&N, -1) >= 0 {
+			foo++
-				runtime.Gosched()
+			if foo%writeRatio == 0 {
-				for g := 0; g < CallsPerSched; g++ {
+				rwm.Lock()
-					foo++
+				rwm.Unlock()
-					if foo%writeRatio == 0 {
+			} else {
-						rwm.Lock()
+				rwm.RLock()
-						rwm.Unlock()
+				for i := 0; i != localWork; i += 1 {
-					} else {
+					foo *= 2
-						rwm.RLock()
+					foo /= 2
-						for i := 0; i != localWork; i += 1 {
-							foo *= 2
-							foo /= 2
-						}
-						rwm.RUnlock()
-					}
 				}
+				rwm.RUnlock()
 			}
-			c <- foo == 42
+		}
-		}()
+		_ = foo
-	}
+	})
-	for p := 0; p < procs; p++ {
-		<-c
-	}
 }
 
 func BenchmarkRWMutexWrite100(b *testing.B) {

src/pkg/sync/waitgroup_test.go

@@ -5,9 +5,7 @@
 package sync_test
 
 import (
-	"runtime"
 	. "sync"
-	"sync/atomic"
 	"testing"
 )
 
@@ -66,55 +64,30 @@ func BenchmarkWaitGroupUncontended(b *testing.B) {
 		WaitGroup
 		pad [128]uint8
 	}
-	const CallsPerSched = 1000
+	b.RunParallel(func(pb *testing.PB) {
-	procs := runtime.GOMAXPROCS(-1)
+		var wg PaddedWaitGroup
-	N := int32(b.N / CallsPerSched)
+		for pb.Next() {
-	c := make(chan bool, procs)
+			wg.Add(1)
-	for p := 0; p < procs; p++ {
+			wg.Done()
-		go func() {
+			wg.Wait()
-			var wg PaddedWaitGroup
+		}
-			for atomic.AddInt32(&N, -1) >= 0 {
+	})
-				runtime.Gosched()
-				for g := 0; g < CallsPerSched; g++ {
-					wg.Add(1)
-					wg.Done()
-					wg.Wait()
-				}
-			}
-			c <- true
-		}()
-	}
-	for p := 0; p < procs; p++ {
-		<-c
-	}
 }
 
 func benchmarkWaitGroupAddDone(b *testing.B, localWork int) {
-	const CallsPerSched = 1000
-	procs := runtime.GOMAXPROCS(-1)
-	N := int32(b.N / CallsPerSched)
-	c := make(chan bool, procs)
 	var wg WaitGroup
-	for p := 0; p < procs; p++ {
+	b.RunParallel(func(pb *testing.PB) {
-		go func() {
+		foo := 0
-			foo := 0
+		for pb.Next() {
-			for atomic.AddInt32(&N, -1) >= 0 {
+			wg.Add(1)
-				runtime.Gosched()
+			for i := 0; i < localWork; i++ {
-				for g := 0; g < CallsPerSched; g++ {
+				foo *= 2
-					wg.Add(1)
+				foo /= 2
-					for i := 0; i < localWork; i++ {
-						foo *= 2
-						foo /= 2
-					}
-					wg.Done()
-				}
 			}
-			c <- foo == 42
+			wg.Done()
-		}()
+		}
-	}
+		_ = foo
-	for p := 0; p < procs; p++ {
+	})
-		<-c
-	}
 }
 
 func BenchmarkWaitGroupAddDone(b *testing.B) {
@@ -126,34 +99,18 @@ func BenchmarkWaitGroupAddDoneWork(b *testing.B) {
 }
 
 func benchmarkWaitGroupWait(b *testing.B, localWork int) {
-	const CallsPerSched = 1000
-	procs := runtime.GOMAXPROCS(-1)
-	N := int32(b.N / CallsPerSched)
-	c := make(chan bool, procs)
 	var wg WaitGroup
-	wg.Add(procs)
+	b.RunParallel(func(pb *testing.PB) {
-	for p := 0; p < procs; p++ {
+		foo := 0
-		go wg.Done()
+		for pb.Next() {
-	}
+			wg.Wait()
-	for p := 0; p < procs; p++ {
+			for i := 0; i < localWork; i++ {
-		go func() {
+				foo *= 2
-			foo := 0
+				foo /= 2
-			for atomic.AddInt32(&N, -1) >= 0 {
-				runtime.Gosched()
-				for g := 0; g < CallsPerSched; g++ {
-					wg.Wait()
-					for i := 0; i < localWork; i++ {
-						foo *= 2
-						foo /= 2
-					}
-				}
 			}
-			c <- foo == 42
+		}
-		}()
+		_ = foo
-	}
+	})
-	for p := 0; p < procs; p++ {
-		<-c
-	}
 }
 
 func BenchmarkWaitGroupWait(b *testing.B) {