/proc/sys/fs/aio-nr is never higher than 1024 (AIO on linux)

Question

I'm trying to use async io on linux. As far as i know there're 3 options:

• kernel calls (io_submit and friends)
• libRT - uses threads in user space
• libRTKAIO - wrapper of kernel calls which does not use threads

I'm using the last option, and i see, that in my unit test that runs a lot of async io requests in multiple threads, /proc/sys/fs/aio-nr is never higher than 1024. I wonder where lays such limitation.

I've set /proc/sys/fs/aio-max-nr to 16M, so it's not an issue.

A related question (also mine) http://stackoverflow.com/questions/1763712/aiowrite-on-linux-with-rtkaio-is-sometimes-long

Answer 1

There's a system call named "io_setup". In librtkaio this call is receiving 1024 as an argument. This, leads to the constant 1024 as i described and hurts performance if there're more than 1024 ios in parallel. Note that 1024 is hardcoded.

Only calling lio_listio in blocking mode will allow you to raise this 1024 bar.