ABSTRACT:
Communication overhead is one of the most important factors affecting
performance of message-passing multicomputers. We present evidence
that there exists communication locality, and that this locality is
"structured". We propose a number of heuristics that can be used to
"predict" the target of subsequent communication requests. Communication
latency is hidden through reconfiguring the network concurrently to the
computation. Quantitative results obtained from standard parallel
benchmarks run on IBM SP systems are also presented.