Enhancing load balancing in a data-parallel GSM network simulation through application-specific information

Abstract

Load balancing plays an important role in achieving good performance in any parallel application. In order to produce a proper load distribution and to minimize overhead, a load balancing algorithm needs to take into account characteristics of the application. Three load balancing algorithms that have been enhanced to optimize the performance of a data-parallel simulator are presented. The enhanced load balancing algorithms are provided with additional information about the workunits, such as an estimate of computational requirements. Results show that the additional information has a significant impact on the load balancing algorithms. The algorithms are capable of producing a more optimal load balance, which leads to the improved performance of an application. A closer study of the overhead created by the processing of the information by the load balancing algorithms indicates that the improvement in the performance significantly outweighs the negative impact of the overhead in most cases.