RFC: Introduce io-uring for localfile storage

[zuston]

Motivation

According to the paper io_uring: Rethinking Asynchronous I/O for Storage Systems, io_uring can significantly improve CPU utilization and I/O latency for storage systems, especially those with a buffer manager.

Based on this observation, Riffle introduces an io_uring-based I/O handler (UringIO) to optimize local shuffle storage performance.

Prerequisites

1. Linux kernel version >= 5.10
2. Currently verified on Anolis OS 8

Conclusions

1. With io_uring enabled and 16 threads per disk (still under tuning), CPU load is reduced by approximately 3× compared to the non-io_uring implementation.
   Update: Further tuning shows that 2 threads per disk are sufficient, achieving comparable overall throughput.
2. After enabling io_uring:

1. Write throughput reaches 5 GB/s, compared to 3.75 GB/s without io_uring (~25% improvement).
2. Read throughput shows no significant improvement, which is expected because:
3. Read requests are not intensive
4. Each read operation transfers relatively large data blocks
5. io_uring provides limited benefits for this access pattern

Benchmark

The benchmark runs a 2.2 TB TeraSort application, with a single Riffle server handling all shuffle data.

Performance Results

Type	Write Time	Read Time
With io_uring	3.4 min	6.3 min
Without io_uring	4.1 min	8.3 min

cpu load comparison

Subtasks

• feat(localfile): Add support of io_uring engine (experimental) #549
• feat(io-uring): Respect io-uring config options #551
• fix(io-uring): Data inconsistency due to hanging pointer of written bytes #553
• feat(io-uring): Refactor and introduce the urpcReadIoMode for client #560