go-mutexbench

Personal playground for concurrency benchmarks. This code is purely educational and should not be used in any other project.

Run it

go-mutexbench <goroutines> <loops>

Findings

1. Defer has a higher performance impact than expected
2. Atomic functions are "semantically inlined" by the compiler
3. Writing your own atomic functions does not help much because of missing inlining
4. Mutex with defer is faster than without defer under high congestion

Results

Taken on an Intel Core i7 (Gen4) 2.2 Ghz using Golang 1.8.1.

• Mutex uses sync.Mutex
• Spinlock uses a minimal spinlock implementation via atomic.CompareAndSwap
• Spinlock2 uses a minimal spinlock implementation via the Intel BitTestAndSet instruction

10.000.000 loops single threaded

test	ms	speedup
plain	442.04	1
mutex	543.80	0.81
spinlock	523.28	0.84
spinlock2	627.49	0.7
mutex defer	1053.86	0.41
spinlock defer	1059.52	0.41
spinlock2 defer	1039.80	0.42

1.000.000 loops 8 go routines

test	ms	speedup
plain	1495.79	1
mutex	1789.38	0.83
spinlock	884.75	1.69
spinlock2	926.29	1.61
mutex defer	1818.81	0.82
spinlock defer	1808.71	0.82
spinlock defer2	1852.24	0.80

1.000.000 loops 16 go routines

test	ms	speedup
plain	3210.73	1
mutex	4265.63	0.75
spinlock	1803.39	1.78
spinlock2	1882.31	1.70
mutex defer	3889.62	0.82
spinlock defer	3631.06	0.88
spinlock defer2	3842.22	0.83

1.000.000 loops 32 go routines

test	ms	speedup
plain	6784.43	1
mutex	9348.37	0.72
spinlock	3737.34	1.81
spinlock2	3916.39	1.73
mutex defer	8037.36	0.84
spinlock defer	7976.36	0.85
spinlock defer2	8423.33	0.80