Opendal support sync read/write API

[sundy-li]

For local fs, we can switch into sync read/write API to improve the performance

[Xuanwo]

tokio is much slower

• fs is based on tokio::fs
• bfs is a simple wrapper by blocking

On NVMe SSD

On tmpfs (in memory)

[Xuanwo]

So the conclusion is:

• tokio::fs is so slow that maybe we should not use it for our fs backend.
• It's worth to provide a sync version API
• We need to develop a real async io fs backend based on aio or io_uring: see Async fs support #19

[sundy-li]

I think we can try the second one to provide a sync version API.

[PsiACE]

Hi @Xuanwo , I noticed we have a lot of benchmarks at the moment, which ones are more reflective of Databend's workload?

Another interesting fact is that if read_full / read_part is improved, then read_parallel may have a different degree of regression.

[Xuanwo]

Yep, the performance of read_parallel is highly related to runtime.

I'm working on providing a blocking API of OpenDAL, and let's evaluate the benchmarks until then.

[sundy-li]

When blocking API is introduced, we can advance a few places on the benchmark list