Change cache mutex to sync.RWMutex

[abursavich]

A small tweak to the cache mutex that could help performance. Multiple reads to the cache don't need to exclude each other. Let the lock be held by an arbitrary number of readers or a single writer. Now once the cache is fully primed, there won't be any blocking.

[kisielk]

Have you actually benchmarked this change? I'm not sure it would actually be faster since the actual read operation after a mutex lock is quite quick, and there is a fair bit more overhead to an RWMutex over a standard Mutex.

[abursavich]

Nah. It doesn't make sense for RWMutex to exist if the read lock isn't more performant than a regular lock. I'll throw some benchmarks together.

[kisielk]

It is more performant if the read lock needs to be held for a long time, but that time is much longer than it takes to read an entry from a map.

[abursavich]

https://gist.github.com/abursavich/7729965

$ go test --bench .
PASS
Benchmark_TypeMap_Get-8 50000000            39.0 ns/op
Benchmark_RWMutex_RLock-8   50000000            48.4 ns/op
Benchmark_RWMutex_Lock-8    50000000            66.5 ns/op
Benchmark_Mutex_Lock-8  50000000            49.4 ns/op
Benchmark_Concurrent_RWMutex_RLock-8    50000000            39.5 ns/op
Benchmark_Concurrent_RWMutex_Lock-8 10000000           241 ns/op
Benchmark_Concurrent_Mutex_Lock-8   10000000           228 ns/op
ok      github.com/abursavich/schema    17.593s

RWMutex does seem to perform better. In certain use cases the difference is likely to be negligible, but in others it's more impressive.

[kisielk]

Looks good. I did some testing of my own and it seems RLock is indeed faster, at least if you have GOMAXPROCS of 2 or greater.