Dan Luu

This paper shows that in single-threaded writes to Redis, the kernel IO mediation takes 84% of processing time. https://www.usenix.org/system/files/conference/osdi14/osdi14-paper-peter_simon.pdf …

For single-threaded reads, it's ~62% of the time. In both cases, ~70% of time in the kernel is spent in just the networking part.

They argue hardware provides most of things you want the OS to do anyway, so they yank the IO stack and configure hardware themselves.

e.g., they set up the virtual network cards to do multiplexing, which makes the multiplexing 0-copy, since the kernel normally copies it.

I wonder why they didn't just go with RDMA. It seems like a natural fit and requires less hardware-specific config. Is it security issues?