sync: improve RWMutex performance
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