A model for performance prediction of message passing multiprocessors achieving concurrency by domain decomposition

Abstract

This paper discusses an analytical model for evaluating the performance of any message passing multiprocessor achieving concurrency by domain decomposition i.e., by partitioning the data domain into sub-domains, computation on each of which is carried out on a different processor.

Studies on different numerical algorithms reveal that execution of an algorithm can be modelled as repetitive cycles where a cycle consists of computation followed by communication or it may consist of computation with overlapped communication. The expression for parallel execution times and speed-ups are given for different models of algorithm execution when the data domain is partitioned into sub-domains of equal size and all the processors execute the same program.

The model has been validated by measurements on a parallel processing system based on four Transputers and the results reveal that the model proposed here can predict the performance with high accuracy. The model is being used in C-DAC to predict the performance of matrix algorithms for a variable number of Transputers and for different sizes of the data sub-domain assigned to each processor.