A Performance Model for Parallel Programs
Appelbe, William F.
Doddapaneni, Srinivas P.
MetadataShow full item record
In this paper, we describe a model for determining the optimal data and computation decomposition for a parallel program by predicting its execution time. The model takes into account various types of data and computation decompositions for each loop nest and combines these to determine a global optimum. The unique features of the model are its accuracy, platform independence, and ability to take potential dynamic decompositions and interleaving of computation and communication into account. We give performance results for the application of the model to standard benchmarks on the IBM SP/2 and a Network of Workstations.