Abstract
Performance optimization of a shared-resource parallel computer is studied in this correspondence. Such a parallel computer contains multiple control units (CU's) sharing a resource pool of processing elements (PE's) and operating with multiple single-instruction-multiple-data (MSIMD) streams. A formal queueing model is proposed for MSIMD machines used in multiple array processing. Analytic results are obtained to evaluate the performance of MSIMD computers. Systematic procedures are given to optimize the size of PE resource pool and to determine the sufficient job queue size for a given vector workload distribution.

This publication has 6 references indexed in Scilit: