Design of a Hierarchical Clos-Benes Optical Network-on-Chip Architecture

摘要

As chip multiprocessors keep growing in capability, on-chip communication efficiency is crucial to the overall performance. However, on-chip networks based on electronic switches suffer from excessive power consumption and limited performance. In order to take advantages of optical interconnect for large-scale on-chip communication in chip multiprocessors, we propose a design of hierarchical Clos-Benes optical network-on-chip (NoC) with an optimized control and routing scheme. The proposed control and routing scheme includes a priority based round-robin virtual output queue selection and a Q-learning based heuristic routing algorithm for the Clos network, and a traffic-aware adaptive routing for the intra-switch Benes network. By taking network load and runtime path allocation into account, the proposed Q-learning based heuristic routing can finally predict the best alternative path among all possible available paths with a much better path allocation success rate. A case study on a 256-core chip multiprocessor under uniform traffic shows that the network throughput is increased by 400%, 60%, and 16% respectively than the mesh, fattree and the baseline Clos-Benes optical NoC. On average of a set of real applications, the application ETE delay is reduced by 48%, 29%, and 20% respectively than the mesh, fattree and the baseline Clos-Benes network.