Dynamic load-balancing on a parallel inference system

Reports a dynamic load balancing algorithm on the hierarchical massively parallel inference system (HMPIS). The main concern is to reduce the communication overhead among the processors (inference clusters in the system). The algorithm, called LOOK-LEAP-Balancing, has quadratic polynomial time complexity. It can localize the communication in the system so that the communication overhead can be reduced substantially. The simulation results show that the algorithm can obtain a nearly optimal solution for a communication overhead model of combinatorial complexity.<
>