perf: Set async_io default value to yes

[xiaobiaozhao]

async_io had ready, I think kvrocks should use async_io as default value.

[git-hulk]

@xiaobiaozhao You need to turn the default value of rocksdb.async_io to true as well. Another quick question: is the rocksdb community marking this feature as stable?

[xiaobiaozhao]

The feature has been unlabeled for experimentation, but rocksdb has not enabled it by default.

[xiaobiaozhao]

This is LGTM. Seems for Scan it depends on advanced fs api and could decrease latency on some disk. I don't know on some fast NVMe disk, would this helps a lot? Also maybe it doesn't works well on some Linux kernel without io_uring?

Anyway this is ok

Do you mean on older kernels? I just test on ubuntu 20/22/24 , it works well.

[mapleFU]

Do you mean on older kernels? I just test on ubuntu 20/22/24 , it works well.

On legacy os it uses "Read" rather than ReadAsync api

[mapleFU]

I've check the code of async_io:

document:

  // If async_io is enabled, RocksDB will prefetch some of data asynchronously.
  // RocksDB apply it if reads are sequential and its internal automatic
  // prefetching.
  bool async_io = false;

For rocksdb, related macro is

• WITH_LIBURING ROCKSDB_IOURING_PRESENT: io_uring related
• USE_COROUTINES USE_FOLLY: MultiGet related, it will using coroutine to schedule issued task

I didn't check do we enabled this.

For fs io, a code below would be triggered:

inline bool CheckFSFeatureSupport(FileSystem* fs, FSSupportedOps feat) {
  int64_t supported_ops = 0;
  fs->SupportedOps(supported_ops);
  if (supported_ops & (1ULL << feat)) {
    return true;
  }
  return false;
}

If fs doesn't supports async (.i.e no io_uring ) I guess this might making sense in MultiGet.

The ForwardIterator will also check the async feature is support.

I also checked the code for iterator, BlockPrefetcher might be issued io for this, so, without io_uring, scan might be optimized(with larger io prefetch-depth)? I didn't run the test, maybe it work ( or maybe not )

[PragmaTwice]

Yeah we need to check the implementation for its behavior.

In docker image, seems we don't install liburing at all. Not sure if the option has effect.

[xiaobiaozhao]

Yeah we need to check the implementation for its behavior.

In docker image, seems we don't install liburing at all. Not sure if the option has effect.

here is a test report

https://github.com/apache/kvrocks/pull/1215

[sonarqubecloud[bot]]

Quality Gate failed

Failed conditions
11.9% Coverage on New Code (required ≥ 50%)

See analysis details on SonarCloud

[PragmaTwice]

## 8.1.0 (03/18/2023)
### Behavior changes
* Compaction output file cutting logic now considers range tombstone start keys. For example, SST partitioner now may receive ParitionRequest for range tombstone start keys.
* If the async_io ReadOption is specified for MultiGet or NewIterator on a platform that doesn't support IO uring, the option is ignored and synchronous IO is used.