Abstract

The Generalized Dimension Exchange (GDE) method is a fully distributed load balancing method that is most suitable for multicomputers with a direct communication network. It is extremely easy to implement and can yield optimal performance given a proper tuning. We propose a decentralized remapping method that uses the GDE algorithm periodically to balance (remap) the system's load. We implemented this remapping method in two data parallel applications and found it to be effective in reducing the computation time. The gains in performance (5 \Gamma 15%) due to remapping are reasonably substantial given the fact that the two applications by their very nature do not necessarily favor remapping. 1 Introduction The mapping problem in parallel computations is to distribute the workload or processes of a computation across the available processors so that each processor would end up with more or less the same amount of work to do. In most cases, this is done prior to execution and is done o...

Citations