David M. Oliveira
CAUP
Abstract
We explore the new emerging technique of GPGPU (General-Purpose computing on Graphics Processing Units) and its low-cost commodity high computing capacity, in the context of N-Body Cosmological simulations.
We have implemented a standard Particle-Particle (PP), direct summation solver in the style of the well-known Hydra cosmological simulation code in a GPGPU, GLSL over OpenGL 2.1 environment. This results in a GPU hardware vendor independent, platform-agnostic, complete PP direct-summation code with all steps and the full simulation iterative loop running on the graphics hardware in its entirety - fully-leveraging the performance potential of the GPGPU framework.
This endeavour required the reengineering and recasting of the known algorithm in a parallel, vectorial form in order to fully exploit the GPU hardware.
Therefore, a novel implementation of the traditional PP algorithm was developed to address the particularities of this vector processing paradigm, opposed to the standard sequential processing based one.
We perform an analysis of the performance gains of this novel approach, compared to a “classical” implementation of the same algorithm in a traditional x86 architecture CPU.
Finally, we show that by using a GPGPU approach in PP direct-summation simulations, we transform the algorithm runtime scaling from O(N2) to O(~N) for the surveyed problem sizes, enabling us to perform PP simulations of scales not practical before.
2011 July 06, 13:30
IA/U.Porto
Centro de Astrofísica da Universidade do Porto (Classroom)
Rua das Estrelas, 4150-762 Porto
