Evaluating the Cost-effectiveness of Dynamic-balanced Adaptive Wormhole Routers

The cost-effectiveness of wormhole torus-networks is systematically evaluated with emphasis on new dynamically-balanced wormhole routing algorithms. These algorithms use balanced assignments of virtual channels to distribute the workload evenly throughout the network, resulting in significant performance improvement. In particular we enhance previous algorithms by removing the non-packet-co-residence restriction on the escape channel. We demonstrate that the algorithms are deadlock-free. A two part evaluation environment is developed consisting of a cycle-driven simulator for high-level measurements such as network capacity and an ASIC design package for low-level measurements such as operating frequency and chip area. This environment is used to compare our routing algorithms with their counterparts under various network constructions and workload parameters. Routers are modeled, verified, synthesized and optimized targeting a state-of-the-art ASIC technology. It is demonstrated that the complicated control of adaptive routing is not necessary on the critical path of the router if sophisticated techniques, such as parallel decoding with optimal path generation and a hierarchical skipping design of the round-robin arbiter, are applied.