1.9 Concurrent Maps

[siennathesane]

Was looking at tip for 1.9, there is a new feature called Concurrent Maps, wanted to start a discussion around adding support for concurrent cache ops.

Not sure on the performance of it, but the design seems fairly solid. Looks like it uses a lot of unsafe.Pointer types and atomic ops, so it might be fairly fast.

[janisz]

Looks interesting. We give it a shot and try to replace current maps and locks with concurrent map. Still we need to synchronize writes to array but reads could be faster.

[siennathesane]

Yeah, I think reads could be faster, but I think the writes will be about the same, maybe ±5% since the current implementation is so highly optimized. My thoughts on this would be to try replace the base map[uint64]uint32 implementation and see what happens. I think that and removing the explicit locking would be the only change for testing this.

If it's close in performance (±10%), it might also be useful as a secondary API instead. Something like concurrentCache := bigcache.NewBigCache(bigcache.Config{Concurrent: true}), where concurrentCache utilises the concurrent map feature, but keeps the interface the same. That being said, it would create a dependency on 1.9+, so that use case would have to be addressed.

My time for the next month is fairly limited since I'm moving, and 1.9 doesn't debut until August. I'll try and fiddle with it over the next month and see what I can come up with.

[janisz]

#49 makes possible to do this task pretty easy. Just copy implementation of shard.go to new file (e.g., shard_concurent_map.go) and replace maps and locks with concurrent maps. Finally add build tags on both files: // +build !go1.9 and // +build go1.9. This will make use of concurrent hash map default in go 1.9. If the performance does not increase we can just add optional build tag to enable concurrent hash map instead of regular one.

[siennathesane]

Perfect. I will work on it in the next week or so.

[siennathesane]

Okay, this is actually going to be a big change as concurrent maps do not support indexing. So this will take awhile.