Performance Comparison of Four State-of-the-Art Computer Architectures

摘要

This study determined the relative performance trends of sequential, vector, and multiprocessor architectures solving computationally intense problems. The problem time complexity of sequential matrix multiplication and inversion algorithms on a Cyber 845, a Vax-11/785, and a Cray X-MP/12 were compared with parallelizations of the algorithms on an Intel Personal Supercomputer (iPSC). In addition, sequential and parallel sort algorithms were run on the Vax and iPSC, respectively. The implementations were run with a range of data sizes. The relative time complexities were determined by computing the factor of execution time increase in relation to the factor of data size increase. By analyzing the problem time complexity, the machine-specific influences on execution speed were partially accounted for, thus enabling generalizations to the corresponding architectural classes to be drawn. It was shown that the multiprocessor was the only architecture to allow a reduction in the problem time complexity, and speedups from vector operations were constant. Keywords: theses; Fortran; parallel architectures, parallel processing.