How good single uring scales across multiple cpu cores

How good single uring scales across multiple cpu cores

[Dmitry Sychov]

Hey,

With the intro of uring there is finally a sane aio interface which
allows straight porting of Win IOCP-based apps to Linux - hooray.

The question though which I've raised in the prev. thread but was
kinda unanswered is how good
single uring submits+retrieves are at scaling across multiple active
user threads for socket io, still bound(to prevent excessive ctx
switches) by the overall number of system cores.

For IOCP I'am observing raw socket TPS increase with every added core
thread up to like 32..64 of them using single submit+completion queue
pair.

All uring benchmarks I've found lack specific context like the number
of running threads and there is no data on the gains from multicore
submits/retrieves against a single uring that make them kinda not to
the point...

Dmitry