Discussion: performance about reading/writing a socket

[CarterLi]

When benchmarking an echo server written with io-uring, I found adding a poll_add sqe before readv/recvmsg could result in about 30% performance boost:

https://github.com/CarterLi/io_uring-echo-server/blob/switch/io_uring_echo_server.c#L14

131729 request/sec VS 98694 request/sec using rust_echo_bench.

That was unexpected. AFAIK readv/recvmsg is async operation itself, adding a poll_add sqe won't help but result in extra context switch ( because it will awake io_uring_enter ).

After some investigation, I found program without poll_add will create lots of kernel processes called io_wqe_worker. But program with poll_add won't.

Don't know how poll_add works, but it seems that poll_add has much lower cost then async read. Is it expected? And maybe a silly question, could we implement async read as poll-add and nonblocking read?