Análise de desempenho de aplicações paralelas em arquiteturas multi-core e many-core

Abstract

Numerical simulations assists science advances and require ever more computing resources to their prediction. Higher resolution models will be more accurate and computationally expensive. Current supercomputers are formed by heterogeneous architectures that are divided in multi-core and many-core. Thus, developers can use both processors and accelerators for higher resolution simulations. However, developing applications for these systems is a challenge due hard programming. In this context, this study analyzed the performance of parallel applications on heterogeneous systems, by running them in processors with OpenMP and GPUs with CUDA. Therefore, seven scientific applications are developed in both libraries. Our experiments show that applications with low communication demand has better performance in GPUs, 8 times in comparasion with CPUs running 32 threads.