Improved algorithms for load balancing in circuit-switched hypercubes

The authors present efficient parallel algorithms for balancing load in circuit-switched hypercubes, which use the e-cube algorithm to route messages between the processing elements. Performance of the hypercube system can be significantly improved by transferring load, at regular intervals of time, from the overloaded processors to the underloaded processors. A centralized algorithm, based on a minimum-cost network flow model, for balancing load in the hypercube systems was developed by Bokhari. This algorithm requires O(N/sup 2/n/sup 2/) time to balance load in an n-cube with N=2/sup n/ processing elements. The authors develop efficient network models-network models in which every arc has unit capacity-to solve the load balancing problem. They also present distributed algorithms for balancing load in the hypercube systems. They show that the distributed algorithms perform better than the centralized algorithms in most practical situations.<<ETX>>