Previous | Next --- Slide 33 of 49
Back to Lecture Thumbnails
firebb

Why can the invalidation msgs be sent in parallel? Does it means the scalable interconnect is a set of N * (N - 1) one to one connections?

jkorn

@firebb I believe so. There's no reason why the messages need to be serialized (each processor can flush their cache line independently), so it would make sense to just have a separate line of communication to each processor to send these invalidation messages. You probably just need a single bus wire between each processor that's on the interconnect (communication goes both ways), giving you N(N+1)/2 wires for N processors that are connected.

Surely with different vector schemes however, the number of wires in this interconnect can be reduced (i.e. sending the invalidation signal to all processors in a group vs individual processors).

whitelez

@firebb I think it could just implement as a single invalidate specific bus connecting every processor. The other communication could be done one to one. Since invalidation need to inform everyone, a broadcast channel would be more efficient.