CHEOPS'23, Rome, Italy, May 8, 2023
Linux storage stack offers a variety of storage I/O stacks and APIs such as POSIX I/O, asynchronous I/O (libaio), high-performance asynchronous I/O (emerging io_uring) or SPDK, the last of which completely bypasses the kernel. Despite their availability, there has not been a systematic study of their performance and overheads. In order to aid our understanding, in this work we systematically characterize performance, scalability and microarchitectural properties of popular Linux I/O APIs on high-performance storage hardware (Intel Optane SSDs). Our characterization reveals that: (1) at low I/O loads, all APIs perform competitively with each other, with polling helping the performance by 1.7×, but consuming 2.3× CPU instructions; (2) at high-loads and scale, io_uring is more than an order of magnitude slower than SPDK; (3) at high-loads and scale, the benchmarking tool (fio) itself becomes a bottleneck; (4) state-of-practice Linux block I/O schedulers (BFQ, mq-deadline, and Kyber) introduce significant (up to 50%) overheads, and their use of global locks hinder their scalability. All artifacts from this work are available at https://github.com/atlarge-research/Performance-Characterization-Storage-Stacks.