Distributed Computation and Stressed Interconnection Networks

The effects of stress on large, distributed networks of processors performing collective computations are investigated. Networks considered in this paper are defined to be systolic arrays arranged in such a fashion so as to be robust against component failure. Variations in stress level (such as damage, congestion and intermittent outages), topologies, algorithms, and bandwidth suggest certain interconnection network architectures which are more resistant to stress and which are efficient in implementing the systolic algorithms studied. Algorithm-dependent control of the networks with regard to synchronization, data flow regulation and adaptivity is discussed.