Linux 6.0 rocks

[uNetworkingAB]

These numbers are insane,

With epoll (highly optimized usage) I could formerly do TCP localhost ping/pong between two processes, small messages, on one CPU-core per process, with 325k messages/second.

With latest features of Linux 6.0 and liburing, I'm now doing 424k messages per second. This is on a 10 year old mid-range CPU.

I used to have a lot of problems getting io_uring to perform with former kernels but since 6.0 it's just completely obliterating epoll for me.

[da-line]

Great achievement 👍 Would be amazing if you could share us some more details/code, what io_uring features you used SQPOLL, file or buffer registration? Did you compare system cpu/memory usage between epoll/io_uring?

[uNetworkingAB]

I'm still at POC stage but I have something that I will build on in the following weeks. These are the features I've found work best:

• io_uring_register_buf_ring
• io_uring_register_files_sparse
• io_uring_register_ring_fd
• params.flags = IORING_SETUP_COOP_TASKRUN | IORING_SETUP_SINGLE_ISSUER;
• io_uring_submit_and_wait
• io_uring_for_each_cqe
• io_uring_prep_recv_multishot
• io_uring_prep_multishot_accept_direct
• io_uring_prep_send (haven't seen anything better for my case)

I think this is all. But prep_send_zc looks interesting for large sends.