Integração de GPGPU ao modelo ETA sem common blocks

Abstract

Eta is a weather prediction model used in several countries, including Brazil, by CPTEC/INPE. In some cases, the execution of weather prediction models like Eta can take hours, or even days. This dissertation’s main objective was the implementation of GPGPU technologies to the Eta model, in order to reduce its execution time. Analyzing the model code, a subroutine was chosen to implement two new versions of it, using CUDA and OpenACC. Analyzing the results, it can be observed that the new implementations show better performance in comparison to the original version when the model is executed with a low number of MPI tasks, however, their performance worsens as the number of threads increases. The CUDA version has better performance compared to the OpenACC version, however, the development time in OpenACC is considerably less.