MLB: Multilevel Load Balancing for Structured Grid Applications

The Multilevel Load Balancing algorithm (MLB) is a parallel algorithm that determines the communication schedule that is necessary to balance a distributed discrete load function. The MLB algorithm focuses on structured grid computations and their load balancing requirements, which we feel are largely unsupported within the load balancing community. The interface to MLB is inherently simple; a distributed discrete load function is provided by the user and a communication schedule is returned. The load function can, for example, map to one or more distributed arrays. So far the implementation includes a parallel version of only the one dimensional MLB algorithm and produces a communication schedule that requires at most log(p) communication steps, where p is the number of processors (log() stands for the logarithm of base two). MLB was first described by Dan Quinlan and Steve McCormick [l, 2, 71. This work forms just one of the object-oriented class libraries within the OVERTURE Framework [5], an object-oriented environment for the numerical solution of partial differential equations in serial and parallel environments.