Newest Linux Optimizations Can Achieve 10M IOPS Per-Core With IO_uring

Just one week ago Linux block subsystem maintainer Jens Axboe was optimizing the kernel to get 8 million IOPS on a single CPU core. He progressed the week hitting around ~8.9M IOPS per-core and began to think he was hitting the hardware limits and running out of possible optimizations. However, this week he is kicking things off by managing to hit 10 million IOPS!

Jens Axboe shared that today he's achieved 10M IOPS off one physical CPU core with his latest Linux kernel patches.

That's it.

10M IOPS, one physical core. #io_uring #linux pic.twitter.com/uzXQekTxMI

— Jens Axboe (@axboe) October 25, 2021

The testing is still done from the same development system of his, an AMD Ryzen 9 5950X with dual Intel P5800X Gen2 Optane solid-state drives.

These optimizations today were primarily within the kernel's block / NVMe / IO_uring code but also some touching the memory management code. All of these optimizations he has been pursuing for achieving the best possible per-core I/O performance can be found via linux-block's perf-wip branch. Many if not all of these performance optimization patches should be landing for the Linux 5.16 cycle.