Fault-tolerant broadcasting

in wormhole-routed torus networks

It is well-known that efficient interprocessor communication is essential to the performance of parallel systems. In a one-to-all broadcast, which is used in almost all applications for parallel systems, a single node sends a message to all other nonfaulty nodes in the system. In this talk I will present a new fault-tolerant broadcast algorithm for wormhole-routed torus networks for which there is only local knowledge of faulty components. This algorithm tolerates up to k-1 faulty nodes in a k-ary n-cube provided the faults do not cause the network to be disconnected. Based on fault-tolerant dimension-order unicast communication, this algorithm is deadlock-free and no additional virtual channels are required.

(This paper will be presented in International Parallel and Distributed Processing Symposium (IPDPS))