Efficient dynamic load balancing strategies for parallel active set optimization methods

Abstract

In this paper three strategies are described to restore dynamically the load balancing in parallel active set optimization algorithms. The efficiency of our proposals is shown by comparison with other heuristics described in related works, such as the classical Bestfit and Worstfit methods. The computational cost due to the load unbalancing in the parallel code and the communication overheads associated with the most efficient load balancing strategy are analyzed and compared in order to establish whether the distribution is convenient or not. Experimental results on a distributed memory system, the Fujitsu AP3000, highlight the accuracy of our estimations. © Springer-Verlag 2003.