A Deadlock-Free Routing Scheme for Interconnection Networks with Irregular Topologies

Interconnection networks with irregular topologies (or irregular networks) are ideal communication subsystems for workstation clusters owing to their incremental scalability. While many deadlock-free routing schemes have been proposed for regular networks such as mesh, torus, and hypercube, they cannot be applied in irregular networks. This paper presents a cost-effective routing architecture, called TRAIN, to solve the routing problem with irregular networks. We show that TRAIN is a deadlock-free scheme. Furthermore, unlike many other routing schemes proposed previously for irregular networks, TRAIN does not require a routing table in the switch. Due to this feature, a TRAIN switch is small and the routing decision can be made rapidly. In order to evaluate the effectiveness of our routing scheme, analysis and event-driven simulation have been performed for various irregular networks. Our results show that TRAIN outperforms other schemes with a higher maximum throughput and lower average latency consistently. 1. Introduction Multiprocessors and multicomputers achieve high performance using interconnection networks that provide high-bandwidth low-latency interprocessor communication. Point-to-point interconnection networks composed of communication switches have been employed in many experimental and commercial systems. Most of these networks use regular topologies, such as mesh, torus, and hypercube, to connect their switch components. With a regular topology, an interconnection network can be compactly implemented in a circuit board or chassis for massively parallel processors. Many routing algorithms have been developed for regular networks to provide efficient transmission of packets and guarantee deadlock freedom [1, 2, 4, 5, 6, 8]. Lately, workstation clusters have emerged to become a cost-effective approach to build a multiprocessing platform. Traditional local area networks can be used to support inter-node traffic for such parallel systems. However, their performance is not maximized due to the limited bandwidth. Although it is possible to use regular networks originally designed for massively parallel processors, they are not incrementally scalable and cost-effective for workstation clusters. In comparison, an