Adaptive Routing based on deadlock recovery

Nidhi Agrawal and C.P.Ravikumar

Abstract
This paper presents a deadlock recovery based fully adaptive routing for any interconnection network topology. The routing is simple, adaptive and is based on calculating the probabilities of routing at each node to neighbors, depending upon the static and dynamic conditions of the network. The probability of routing to the i-th neighbor at any node is a function of the traffic and distance from the neighbor to the destination. It is shown that with our routing algorithm deadlocks are rare; thus deadlock recovery is a better solution. We also propose here two deadlock recovery schemes. Since deadlocks occur due to cyclic dependencies, these cycles are broken by allowing one of the mesages involved in deadlock to take an alternate path consisting of buffers reserved for such messages. These buffers can be centralized buffers accessible to all neighboring nodes or can be set of virtual channels. The performance of our algorithm is compared with other recently proposed deadlock r!ecovery schemes. The 2-Phase routing is found to be superior compared to the other schemes in terms of network throughput and mean delay.
Contact
Nidhi Agrawal
Nidhi Agrawal,MobilSat Division, SND,,Hughes Software Systems,Electronic city, Plot 31, Sector 18,,Gurgaon-122015, Haryana,,INDIA
nagrawal@hss.hns.com