Calling uring.submitter().submit() on every loop is probably expensive

[JackKelly]

submit calls submit_and_wait.

submit_and_wait always calls atomic::fence(atomic::Ordering::SeqCst) when the kernel is configured to use sqpoll.

So it might be faster to keep an eye on the clock (in our code), and only call submit if, say, it's been at least 800 ms since we last added to the SQ. (And only do this if our code uses sqpoll). But, as always, I should benchmark!

Implementation

I could use crossbeam::channel::after. But supposedly that also requires an atomic load. So it might just be faster to do the simple thing and check the time on every iteration! I probably want to wrap the submission queue in my own type, which will also look after waking the SQ (if enough time has elapsed, and if we're actually using sqpoll!)

Related

• open question about race with kernel that liburing had to fix tokio-rs/io-uring#197