A Model of Shared Data References in Parallel Programs
Adrian Holliday · 1989
The reference pattern to shared data has a major effect on the performance of large-scale shared memory machines with distributed memory. We propose a technique for characterizing this pattern which is systematic and application domain-independent. In particular, each process''s trace is viewed as the sample path of a semi-Markov process. We experimentally evaluate the effectiveness of the proposed approach on a range of realistic parallel programs with encouraging results. This technique may be useful for quantitatively comparing ``typical'''' parallel computations and for systematically studying the effect on performance of variations in a particular computation.