[PeCaN]

I thought I'd seen a rather dramatic benchmark of this but I can't find it now. (I think it was in the context of io_uring, whose documentation mentions that polled IO is much lower latency and I'd trust Jens Axeboe on that one.)

This intuitively makes sense especially with SSDs and especially with NVMe SSDs: you're not going to have to wait that long, especially compared to the overhead of an interrupt. Also, if you're reading sequentially, once you get one I/O completion you're going to get a lot more in a short timeframe—it makes sense to just poll for those instead of handling every one as an interrupt. Interrupts are pretty expensive (I honestly don't have a good grasp of how Linux interrupt handling works but it can easily be on the order of microseconds.)

It also has some side benefits like preventing the CPU from going into a lower power state for only a short period of time.