Predictive Switching in 2-D Torus Routers

This paper proposes predictive switching in 2D torus routers to reduce the number of pipeline stages for low-latency communication. By utilizing the communication regularity in parallel applications, a dynamic predicting mechanism presets packet traversal paths inside the router before packet arrivals. Hence, we can bypass the pipeline stages of routing computation, virtual channel allocation and switch allocation when the prediction hits. We considered the predictor architecture and accuracy for several traffic patterns in NAS parallel benchmarks. Our experiments show that a sampled pattern matching (SPM) predictor achieves 77% to 96% of the prediction hit rates when we use the dimension-order routing algorithm. We also discuss a method to improve the prediction accuracy of SPM by examining the frequency of occurrence for the prediction values in the communication history