DESIGN AND IMPLEMENTATION OF AN INTERCONNECTION NETWORK FOR THE AP1000

This paper discusses the design and implementation of an interconnection network for the distributed-memory highly parallel computer AP1000. The interconnection network involves routing, topology, and communications. Wormhole routing is the most promising switching technique used in multicomputers. Its disadvantage is that blocked messages remain in the network causing deadlocks and deteriorated throughput. We present a high-throughput and deadlock-free wormhole routing scheme that incorporates a structured buffer pool algorithm. We chose a two-dimensional torus topology for cost effectiveness and high performance. The interconnection network provides both point-to-point and broadcast communications. We describe the hardware implementation of the routing chip that composes the network. The results of experiments on the AP1000 show that this network provides high-throughput communication.