Studying the Effects of Asynchronous I/O on HPC I/O Patterns
Arnav A. Gupta, Druva Dhakshinamoorthy, Arnab Kumar Paul · 2024
High-performance computing (HPC) applications have long faced the issue of I/O as a bottleneck. This work analyzes how read/write ratios, I/O block sizes, and sequential versus random access patterns affect a workload's I/O characteristics when using POSIX synchronous I/O (psync), and the relatively new io _uring asynchronous interface for single-node performance. Based on our findings, we introduce a preliminary interceptor that intercepts POSIX synchronous I/O calls and replaces them with io_uring operations. Initial results demonstrate performance gains with the IOR benchmark, indicating the potential of our approach.