Part 6 TLDR; Non scientific ballpark benchmarks on isolated backends comparing io_uring, epoll and hybrid models. Load generators used; wrk (epoll) gcannon (io_uring) Didn't find any other viable option - oha, bombardier, hey, k6 and friends are painfully heavy and slow and end up being the benchmark bottleneck by a long mile. If you know any other options I'd appreciate feedback on it, gcannon is still a research tool. The io_uring bechmarks are based on the Minima project which is the base of this series. The epoll benchmarks are based on the Unhinged project which was an overall top 3 on the latter techempower benchmarks. The Socket benchmarks are just a plain System.Net.Socket You can find results on the post link, but resuming; In raw numbers io_uring wins but this advantage disappears when threadpool enters the equation. Either a mix of io_uring with libc send or just plain old epoll take the cake when connection handler has async workloads that are handed to the threadpool.…