Análise de escalabilidade em simulação de reservatórios de petróleo via elementos finitos e programação distribuída (MPI)

Abstract

Computational simulation of petroleum reservoirs is fundamental for strategic decision-making in the oil and gas industry, where numerical accuracy of models and computational efficiency are critical factors. This work proposes a hybrid parallelization implementation for the linear system solver of a numerical code based on the stabilized finite element method, developed to simulate polymer solution injection in reservoirs. The implementation was carried out in Fortran 90, combining OpenMP for shared memory parallelism and MPI for distributed memory. The Hypre library was used for linear system solution and MPI for data distribution among processes. Static load balancing was employed, defined based on the distribution of the number of linear system equations among processes. Experiments were conducted at the High Performance Computing Center (NPAD) of UFRN, using unstructured meshes of different sizes. Results demonstrated significant improvements in the linear system solution stage, with reductions of up to 74% in linear system solution time using 4 computational nodes with 4 MPI processes and 16 threads each. Scalability analysis revealed that the implementation presents both strong and weak scalability, although the speedup was sublinear, indicating opportunities for future optimization.