7.4M IOPS Achieved Per-Core With Newest Linux Patches
This month Jens Axboe has been making some remarkable improvements to the Linux block code for squeezing out every bit of I/O potential of the system. Yesterday Jens Axboe was hitting 7.2M IOPS with new persistent DMA map patches that also shaved off around 10% of synchronization latency.
As for his approach to squeezing out these gains with the Linux I/O stack, Axboe commented on Monday, "I've just been whack-a-moling the obvious culprits and trying to think about new ways we can do things rather than try and just shave cycles off the existing methods."
When the news can't really keep up:— Jens Axboe (@axboe) October 12, 2021
IOPS=7412800, BW=3619MiB/s, IOS/call=32/31, inflight=(96 103)
IOPS=7446080, BW=3635MiB/s, IOS/call=32/32, inflight=(78 100)
7M? Pffft, easy. https://t.co/7hzSfku2lA
Today he's celebrating 7.4M IOPS per-core using Intel Optane Gen2 storage and an AMD Ryzen 9 5950X processor.
These various improvements are being housed in linux-block's "perf-wip" branch. Hopefully at least most of these changes will be ready in time for the upcoming Linux 5.16 merge window.