Independent performance modeling of parallel architectures andalgorithms

摘要

A key requirement for the effective use of multiprocessor systemsnin real-world applications is an ability to accurately predict thenperformance of a specific algorithm on a specific architecture. Suchnperformance prediction tools assist the system designer in initiallynselecting, and then modifying, both the algorithm and the architecturento obtain acceptable performance. In this paper, we present a modelingnapproach that permits separate evaluation of algorithm and architecturenperformance with only a small number of “cross” parametersnrequired to link the two models. An example application of thisntechnique to a Gaussian elimination algorithm on two dissimilarnmultiprocessor architectures shows good agreement with actualnperformance figures obtained from measurement and simulation