Optimising Cosmological N-body Simulations in GPU Clusters
Super Computer Education and Research Centre, Indian Institute of Science, Bangalore – 560 012, India
Indian Institute of Science, 2012
@phdthesis{santanu2012optimising,
title={Optimising Cosmological N-body Simulations in GPU Clusters},
author={Santanu, T.},
year={2012},
school={Indian Institute of Science}
}
Cosmological simulations play an important role in understanding the evolution of our universe. Since the experiments on the formation of galaxies cannot be performed in laboratory, simulation is the only way to understand this phenomenon. The cosmological simulations are usually modelled as N-body problems. The Barnes-Hut (BH) tree code algorithm is one of the popular N-body algorithms used in cosmological simulations. BH tree code reduces the force computation complexity from O(n^2) to O(n log n). Though the parallel version of BH is in existence for a long time, with the advent of the GPU clusters, there is a renewed interest to port this algorithm to GPU. The previous works in this field did the tree traversal in CPU and the force computation in GPU, but this method involves the overheads of multiple memory copy operations between CPU and GPU, which slows down the computation. In this work, we have developed an algorithm that shifts the local tree traversal along with the force computation to GPU and asynchronously performs the remote work in parallel on CPU. We have ported our algorithm to GADGET [1] simulator. By means of experiments on Tesla cluster with real cosmology data of 16, 772, 216 particles, we show 30% improvement in performance over existing algorithm.
July 25, 2012 by hgpu
