A performance model and metrics for fine grain parallel computing systems-finding optimal parallelism

摘要

Parallel computing, and fine grain computing in particular, need criteria to find optimal parallelism. This paper proposes performance models that measure ability to generate and synchronize parallel processes, and to switch control in parallel processing systems. We consider the performance of controlling the number of parallel processes (synchronization capability), the performance of generating parallel processes (generation capability), and the performance of controlling a computing flow of parallel processes (branch capability) in fine grain parallel computing. We also discuss the usefulness of these performance measures, and prove that optimization is possible by measuring the branch capability of instruction level data flow computers. With optimal parallelism, extraction and control of parallel processes is well-balanced, and the balancing point is specified.