The effect of application characteristics on performance in a parallel architecture

A shared-memory, hierarchical, clustered architecture is reported. Measurements presented (including speedups) were obtained by running six different symbolic application kernels with varying architectural parameters. The results and graphs shown highlight the manner in which application characteristics place limits on achievable parallel performance, given particular architectural features. The discussion covers processor starvation, fine-grain parallelism, uneven loads, foreign reference, cost scheduling and indeterminate computation with respect to the applications chosen. These measurements provide understanding of some of the interactions that take place between the architectural parameters and the specific applications and suggest ways to improve the architecture. To date, the architecture has shown reasonable speedups on the chosen applications.<>