Task Failed Successfully: Saturating NIC and Disk Bandwidth

Author here. This post is a write-up of a performance-debugging rabbit hole I hit while trying to saturate NICs with NVMe reads using io_uring and RDMA.

The short version: READ_FIXED fixed the obvious per-I/O GUP overhead in a small demo, but the larger deployment still got stuck at roughly half of line rate. After ruling out io-wq backlog, request splitting, fd lookup, and CRC arithmetic, the actual wall turned out to be dTLB misses from scanning 1,028 KiB buffers backed by 4 KiB pages. Moving the read arena to hugepages brought the system close to NIC saturation.

The funny part is that an AI agent suggested hugepages early and got the optimization right, but its explanation was wrong. This post is mostly about reconstructing the evidence for why it worked.

I’d be very interested in feedback from people who have used AI to debug performance issues in a complex system.

It seems that you could have reached this conclusion faster by elaborating on your use of the profiler. Don't assume that cycles are spent on instructions. Look at your IPC and drill down into what CPU-bound means for your workload. In your case I think a standard top down analysis would have made the virtual memory management cost jump right out.

This sounds like a strong statement with little backing. The author does infra at DeepSeek if his LinkedIn is to be trusted, and is the author of Foyer.

It is obvious that the blog is good quality if you have moderate knowledge on the subject and read the blog post.

[flagged]