[toast0]

Is IO in Erlang slow? I never got that impression; iolists are a good fit for scatter/gather I/O, which reduces copying, and Erlang supports kqueue, epoll, and I believe similar on Windows. I don't follow Erlang on Linux closely, so I don't know if there's any movement towards io_uring, which does seem promising to help with throughput.

Either way, I wouldn't expect doing I/O in NIFs or ports to substantially improve throughput, unless you're also moving significant processing into that layer as well, or you're going to end up doing substantially the same level of marshaling work as ERTS does, just in a different language. Setting up a different set of kqueue/epoll descriptors sounds like a lot of work for not much gain too, IMHO; again, maybe io_uring would be useful, but I think you'd be better served to bite the bullet and integrate it with ERTS.