Performance evaluation and resource optimization of multiple SIMD computer organizations
Ki Soo Hwang, Lionel Ming-shuan Ni · Rare & Special e-Zone (The Hong Kong University of Science and Technology) · 1979
Multiprocessor computer systems, which can be used to execute multiple number of single-instruction multiple-data (SIMD) vector jobs, are modeled and analyzed. Such a parallel computer organization usually has multiple control units (CU's) sharing a pool of dynamically allocated processing elements (PE's). The multiple SIMD machine is characterized as a finite-storage multi-server M/M/K/l queueing family where the number of active servers, K, is an upper bounded random variable and l is the maximal queue length. Analytic results using conditional probabilities are obtained under equilibrium conditions. These results can be readily applied to evalute the performance of multiple SIMD machines. The system performance is measured by the resource utilization factors CU's and PE's and by the average job response time. Given a fixed number of CU's and a prespecified workload, systematic procedures are given to determine the optimal size of the resource pool of PE's and the sufficient queue capacity of MSIMD operations.