I have not read existing MPI implementations, but there is certainly a
whole slew of MP libraries for the CM5 which rely upon network interrupts,
and a few compilers as well. Unless one is doing a blocking synchronous
receive, it is the (overwhlemingly) preferred approach. Any MPI
implementation for the CM5 which does not talk to the hardware directly is
almost certainly handling progress by interrupting the CPU. I cannot speak
to other platforms.