Avaliação de desempenho de algoritmos paralelos de busca de vizinhos em cenários com distribuições espaciais distintas

Abstract

Contact detection algorithms are needed in different areas of science and technology. From digital games and computer graphics to high-performance simulations and robotics. These algorithms require great computational effort and are prone to become the bottlenecks of its applications, even more when this computation must be done in real-time or large-scale systems. With the popularization of GPU cards use for both science and business, it is only natural that parallel implementations for this problem arise in the scientific community. In this work the main contact detection algorithms are analyzed and a numerical experiment is performed, with the goal of finding out which algorithm has better computational performance and memory use, or if they efficiency depends on different scenario features. For performing the experiment, a parallel Discrete ElementMethod application was developed using CUDA/C++ with the main algorithms presented in literature, besides these, the author proposes and implements the Sorting Contact Detection algorithm parallelization, that hadn’t been parallelized until now. The tests have found that the parallel Sorting Contact Detection algorithm is the most efficient in all studied scenarios, achieving a good performance and a superiormemory usage than its peers.