ABSTRACT:
Communication overhead adversely affects the performance of
multicomputers. In this work, we present evidence (through the
analysis of several parallel benchmarks) that there exists
communications locality, and that it is "structured". We have used
this in a number of heuristics that "predict" the target of
subsequent communications. This technique, can be applied directly
to reconfigurable interconnects (optical or conventional) to hide
the communications latency by reconfiguring the interconnect
concurrently to the computation.