Estudo de técnicas de otimização de desempenho para GPUs utilizando CUDA aplicado a um modelo meteorológico

Abstract

This work studied techniques for optimizing performance for a prototype of the Ocean-Land-Atmosphere Model (OLAM) implemented on CUDA using GPUs. One wide literature review about the CUDA programming model and its architecture of GPUs has beenmade. Performance analyzes were conducted to compare the changes implemented witha version without parallelism in CPU and the initial version of the prototype. The model was subjected to analysis via profiling, where more detailed observations were made through a graphical visualization tool. The changes implemented had little performance gain for the model, but the studies analyzed several factors that can still be exploited to take advantage of a massively parallel execution environment as GPU architectures.