]> Cypherpunks repositories - gostls13.git/commitdiff
sync: improve RWMutex performance
authorDmitriy Vyukov <dvyukov@google.com>
Tue, 12 Jul 2011 16:24:21 +0000 (09:24 -0700)
committerRuss Cox <rsc@golang.org>
Tue, 12 Jul 2011 16:24:21 +0000 (09:24 -0700)
The new implementation features wait-free
fast path for readers which significantly improves
performance/scalability on read-mostly workloads.
Benchmark results on HP Z600 (2 x Xeon E5620, 8 HT cores, 2.40GHz)
are as follows:
benchmark                              old ns/op    new ns/op    delta
BenchmarkRWMutexUncontended               179.00        96.60  -46.03%
BenchmarkRWMutexUncontended-2              89.10        49.10  -44.89%
BenchmarkRWMutexUncontended-4              44.70        24.70  -44.74%
BenchmarkRWMutexUncontended-8              23.30        12.90  -44.64%
BenchmarkRWMutexUncontended-16             16.80         8.75  -47.92%
BenchmarkRWMutexWrite100                   79.60        26.80  -66.33%
BenchmarkRWMutexWrite100-2                305.00        33.00  -89.18%
BenchmarkRWMutexWrite100-4                245.00       113.00  -53.88%
BenchmarkRWMutexWrite100-8                330.00       147.00  -55.45%
BenchmarkRWMutexWrite100-16               371.00       152.00  -59.03%
BenchmarkRWMutexWrite10                    78.30        29.80  -61.94%
BenchmarkRWMutexWrite10-2                 348.00       165.00  -52.59%
BenchmarkRWMutexWrite10-4                 447.00       199.00  -55.48%
BenchmarkRWMutexWrite10-8                 564.00       180.00  -68.09%
BenchmarkRWMutexWrite10-16                492.00       192.00  -60.98%
BenchmarkRWMutexWorkWrite100             1077.00      1037.00   -3.71%
BenchmarkRWMutexWorkWrite100-2            659.00       596.00   -9.56%
BenchmarkRWMutexWorkWrite100-4            509.00       361.00  -29.08%
BenchmarkRWMutexWorkWrite100-8            603.00       351.00  -41.79%
BenchmarkRWMutexWorkWrite100-16           750.00       607.00  -19.07%
BenchmarkRWMutexWorkWrite10               990.00       951.00   -3.94%
BenchmarkRWMutexWorkWrite10-2            1119.00      1070.00   -4.38%
BenchmarkRWMutexWorkWrite10-4            1300.00      1199.00   -7.77%
BenchmarkRWMutexWorkWrite10-8            1424.00      1291.00   -9.34%
BenchmarkRWMutexWorkWrite10-16           1981.00      1786.00   -9.84%

R=rsc
CC=golang-dev
https://golang.org/cl/4671051

src/pkg/sync/rwmutex.go
src/pkg/sync/rwmutex_test.go

index 9248b4b03700207758f0214c450e658164353a40..cb1a47720b21e9fa6af0c002208b77753c41b81b 100644 (file)
@@ -4,7 +4,10 @@
 
 package sync
 
-import "sync/atomic"
+import (
+       "runtime"
+       "sync/atomic"
+)
 
 // An RWMutex is a reader/writer mutual exclusion lock.
 // The lock can be held by an arbitrary number of readers
@@ -12,35 +15,22 @@ import "sync/atomic"
 // RWMutexes can be created as part of other
 // structures; the zero value for a RWMutex is
 // an unlocked mutex.
-//
-// Writers take priority over Readers: no new RLocks
-// are granted while a blocked Lock call is waiting.
 type RWMutex struct {
-       w           Mutex // held if there are pending readers or writers
-       r           Mutex // held if the w is being rd
-       readerCount int32 // number of pending readers
+       w           Mutex  // held if there are pending writers
+       writerSem   uint32 // semaphore for writers to wait for completing readers
+       readerSem   uint32 // semaphore for readers to wait for completing writers
+       readerCount int32  // number of pending readers
+       readerWait  int32  // number of departing readers
 }
 
+const rwmutexMaxReaders = 1 << 30
+
 // RLock locks rw for reading.
-// If the lock is already locked for writing or there is a writer already waiting
-// to release the lock, RLock blocks until the writer has released the lock.
 func (rw *RWMutex) RLock() {
-       // Use rw.r.Lock() to block granting the RLock if a goroutine
-       // is waiting for its Lock. This is the prevent starvation of W in
-       // this situation:
-       //   A: rw.RLock()  // granted
-       //   W: rw.Lock()   // waiting for rw.w().Lock()
-       //   B: rw.RLock()  // granted
-       //   C: rw.RLock()  // granted
-       //   B: rw.RUnlock()
-       //   ... (new readers come and go indefinitely, W is starving)
-       rw.r.Lock()
-       if atomic.AddInt32(&rw.readerCount, 1) == 1 {
-               // The first reader locks rw.w, so writers will be blocked
-               // while the readers have the RLock.
-               rw.w.Lock()
+       if atomic.AddInt32(&rw.readerCount, 1) < 0 {
+               // A writer is pending, wait for it.
+               runtime.Semacquire(&rw.readerSem)
        }
-       rw.r.Unlock()
 }
 
 // RUnlock undoes a single RLock call;
@@ -48,9 +38,12 @@ func (rw *RWMutex) RLock() {
 // It is a run-time error if rw is not locked for reading
 // on entry to RUnlock.
 func (rw *RWMutex) RUnlock() {
-       if atomic.AddInt32(&rw.readerCount, -1) == 0 {
-               // last reader finished, enable writers
-               rw.w.Unlock()
+       if atomic.AddInt32(&rw.readerCount, -1) < 0 {
+               // A writer is pending.
+               if atomic.AddInt32(&rw.readerWait, -1) == 0 {
+                       // The last reader unblocks the writer.
+                       runtime.Semrelease(&rw.writerSem)
+               }
        }
 }
 
@@ -61,9 +54,14 @@ func (rw *RWMutex) RUnlock() {
 // a blocked Lock call excludes new readers from acquiring
 // the lock.
 func (rw *RWMutex) Lock() {
-       rw.r.Lock()
+       // First, resolve competition with other writers.
        rw.w.Lock()
-       rw.r.Unlock()
+       // Announce to readers there is a pending writer.
+       r := atomic.AddInt32(&rw.readerCount, -rwmutexMaxReaders) + rwmutexMaxReaders
+       // Wait for active readers.
+       if r != 0 && atomic.AddInt32(&rw.readerWait, r) != 0 {
+               runtime.Semacquire(&rw.writerSem)
+       }
 }
 
 // Unlock unlocks rw for writing.  It is a run-time error if rw is
@@ -72,7 +70,16 @@ func (rw *RWMutex) Lock() {
 // As with Mutexes, a locked RWMutex is not associated with a particular
 // goroutine.  One goroutine may RLock (Lock) an RWMutex and then
 // arrange for another goroutine to RUnlock (Unlock) it.
-func (rw *RWMutex) Unlock() { rw.w.Unlock() }
+func (rw *RWMutex) Unlock() {
+       // Announce to readers there is no active writer.
+       r := atomic.AddInt32(&rw.readerCount, rwmutexMaxReaders)
+       // Unblock blocked readers, if any.
+       for i := 0; i < int(r); i++ {
+               runtime.Semrelease(&rw.readerSem)
+       }
+       // Allow other writers to proceed.
+       rw.w.Unlock()
+}
 
 // RLocker returns a Locker interface that implements
 // the Lock and Unlock methods by calling rw.RLock and rw.RUnlock.
index 0480a6601838b27608b575890238197ddcf99bcf..dc8ce9653ce1dcec5a86e049609efc8a0053fcd8 100644 (file)
@@ -154,3 +154,84 @@ func TestRLocker(t *testing.T) {
                wl.Unlock()
        }
 }
+
+func BenchmarkRWMutexUncontended(b *testing.B) {
+       type PaddedRWMutex struct {
+               RWMutex
+               pad [32]uint32
+       }
+       const CallsPerSched = 1000
+       procs := runtime.GOMAXPROCS(-1)
+       N := int32(b.N / CallsPerSched)
+       c := make(chan bool, procs)
+       for p := 0; p < procs; p++ {
+               go func() {
+                       var rwm PaddedRWMutex
+                       for atomic.AddInt32(&N, -1) >= 0 {
+                               runtime.Gosched()
+                               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++ {
+               go func() {
+                       foo := 0
+                       for atomic.AddInt32(&N, -1) >= 0 {
+                               runtime.Gosched()
+                               for g := 0; g < CallsPerSched; g++ {
+                                       foo++
+                                       if foo%writeRatio == 0 {
+                                               rwm.Lock()
+                                               rwm.Unlock()
+                                       } else {
+                                               rwm.RLock()
+                                               for i := 0; i != localWork; i += 1 {
+                                                       foo *= 2
+                                                       foo /= 2
+                                               }
+                                               rwm.RUnlock()
+                                       }
+                               }
+                       }
+                       c <- foo == 42
+               }()
+       }
+       for p := 0; p < procs; p++ {
+               <-c
+       }
+}
+
+func BenchmarkRWMutexWrite100(b *testing.B) {
+       benchmarkRWMutex(b, 0, 100)
+}
+
+func BenchmarkRWMutexWrite10(b *testing.B) {
+       benchmarkRWMutex(b, 0, 10)
+}
+
+func BenchmarkRWMutexWorkWrite100(b *testing.B) {
+       benchmarkRWMutex(b, 100, 100)
+}
+
+func BenchmarkRWMutexWorkWrite10(b *testing.B) {
+       benchmarkRWMutex(b, 100, 10)
+}