Parameter Tuning of a Hybrid Treecode-FMM on GPUs
Dept. Mechanical Engineering, Boston University, 15 St. Mary’s Street, Boston, Massachusetts
Characterizing Applications for Heterogeneous Exascale Systems (CACHES’11), at the 25th International Conference on Supercomputing (ACM/SIGARCH), 2011
@article{yokota2011parameter,
title={Parameter Tuning of a Hybrid Treecode-FMM on GPUs},
author={Yokota, R. and Barba, L.},
year={2011}
}
Treecodes are O(N log N) hierarchical N-body algorithms, which have traditionally been used for applications in astrophysics, in a low-accuracy regime. Fast multipole methods (FMM) are O(N) hierarchical N-body algorithms that have been used in a variety of applications, often in the high-accuracy regime. Both algorithms are known to perform well on massively parallel heterogeneous systems, with the treecode showing stronger speed-ups due to higher computational intensity. We propose a hybridization of the two algorithms and aim to determine whether such approach can best either algorithm on its own. We study the behavior of the hybrid treecode-FMM for different accuracies, along with its performance on GPUs.
January 8, 2012 by hgpu
