EVALUATING PERFORMANCE TRADEOFFS BETWEEN FINE-GRAINED AND COARSE-GRAINED ALTERNATIVES

Recent simulation-based studies suggest that while superpipelines and superscalars are equally capable of exploiting fine-grained concurrenc y, multiprocessors are better at exploiting coarse-grained parallelism . An analytical model that is more flexible and less costly in terms o f run time than simulation, is proposed as a tool for analyzing the tr adeoff between superpipelined processors, superscalar processors, and multiprocessors. The duality of superpipelines and superscalars is exa mined in detail, The performance limit for these systems has been deri ved and it supports the fetch bottleneck observation of previous resea rchers. Common characteristics of utilization curves for such systems are examined. Combined systems, such as superpipelined multiprocessors and superscalar multiprocessors, are also analyzed. The model shows t hat the number of pipelines (or processors) at which the maximum throu ghput is obtained is, as memory access time increases, increasingly se nsitive to the ratio of memory access time to network access delay. Fu rther, as a function of interiteration dependence distance, optimum th roughput is shown to vary nonlinearly, whereas the corresponding optim um number of processors varies linearly. The predictions from the anal ytical model agree with similar results published using simulation-bas ed techniques.