Allow to enable the async_io option to improve the performance

[xiaobiaozhao]

Here are some comparisons between async_io and normal

big value 500B

small value 50B

monitor diff

The green line is kvrocks compiled by async io，this pr.
ts* command using seek function to scan data from db.

summarize

	cpu	mem	latency	qps
async	110%	100%	100%	120%
normal	100%	100%	130-400%	100%

async_io is better in seek read, higher qps, lower latency, and slightly higher cpu usage. Perfect for introducing projects

[git-hulk]

Cool, I think it'd be better if we have the option to let users determine whether to enable it or not.

[torwig]

Since this option is marked as Experimental in the rocksdb codebase, I agree to make this option configurable.
@xiaobiaozhao Thank you for the benchmarks. Nice job!

[git-hulk]

@xiaobiaozhao It's not right to involve the server in the storage layer. You should add the async_io flag in storage or implement like the write option https://github.com/apache/incubator-kvrocks/blob/unstable/src/storage/storage.cc#L76

[aleksraiden]

Good feature, any updates? In latest rocksdb update fixed some bugs in async_io support, so this PR is very good point to improve our performance.

[git-hulk]

Good feature, any updates? In latest rocksdb update fixed some bugs in async_io support, so this PR is very good point to improve our performance.

@xiaobiaozhao has updated the PR, to see if @torwig and @PragmaTwice have comments.

[xiaobiaozhao]

I've tested mget and hgetall and async_io obviously gives hgetall a big boost.
If rocksdb has better updates in the future, we can continue to follow up.

[torwig]

@xiaobiaozhao Good research.
If you update your branch, and make a final change as mentioned @git-hulk, I think we can merge this PR.

[torwig]

LGTM

[git-hulk]

Thanks all, merging...