Parallel Computer Architecture

11.5 Proceeding Past Long-Latency Events 863

■ the overheads, delays, and bandwidths in the system

■ the spatial locality in the program and how it interacts with the granularity of

data transfer

If we can really make all messages large enough, then the delay components of

communication may not be the problem; bandwidth becomes a more significant

constraint. But if not, we still need to hide the latency of data accesses by overlap-

ping them with computation or with other accesses. The other three latency toler-

ance approaches do this; however, to be successful at cache block granularity, they

require support in the microprocessor as well. Instea ...