Load balancing requirements in parallel implementations of image feature extraction tasks

Abstract

Load balancing requirements in parallel image analysis are considered and results on the performance of parallel implementations of two image feature extraction tasks on the Connection Machine and the iPSC/2 hypercube are reported and discussed. A load redistribution algorithm, which makes use of parallel prefix operations and one-to-one permutations among the processors, is described and has been used. The expected improvement in performance resulting from load balancing has been determined analytically and is compared to actual performance results obtained from the above implementations. The analytical results demonstrate the specific dependence of the expected improvement in performance on the computational and communication requirements of each task, characteristic machine parameters, a characterization of prior load distribution in terms of parameters which can be computed dynamically at the start of task execution, and the overhead incurred by load redistribution. >