A hybrid parallel Barnes-Hut algorithm for GPU and multicore architectures

Abstract

Using the Barnes-Hut algorithm as an example we deal with the design of parallel algorithms that are able to exploit multicore CPUs and GPUs conjointly. Specifically, we demonstrate how to modularize a parallel application according to specific aspects of parallel execution. This allows for a flexible assignment of individual modules to the two parallel architectures based on their actual performance characteristics. Furthermore, we discuss a hybrid module for the most time consuming part of the algorithm that utilizes CPU and GPU simultaneously employing a novel load balancing heuristic. Our experimental evaluation shows that our method greatly increases overall efficiency by allowing to deploy the optimal configuration of modules for each individual computer system. © 2013 Springer-Verlag.