Cost - E ective Data - Parallel Load Balancing

Load balancing algorithms improve a program's performance on unbalanced datasets, but can degrade performance on balanced datasets, because unnecessary load redistributions occur. This paper presents a cost-e ective data-parallel load balancing algorithm which performs load redistributions only when the possible savings outweigh the redistribution costs. Experiments with a data-parallel polygon renderer show a performance improvement of up to a factor of 33 on unbalanced datasets and a maximum performance loss of only 27 percent on balanced datasets when using this algorithm.