A high-performance computing applied to composition reservoir simulation using distributed memory and 3D hybrid unstructured grids

Reservoir models have become more complex over time, and accurate models also require refined grids for obtaining accurate results. The memory required for the simulation of reservoir models with refined grids containing dozens of millions of grid blocks is a hard task to be accomplished in small workstations. Parallel processing in reservoir simulation has gained repute over the past decades as a solution to tackle this challenging problem. This work shows how the distributed memory parallelization is applied to an in-house compositional simulator called UTCOMP in conjunction with an implicit pressure explicit composition approach using unstructured grids and the element-based finite volume method. To achieve high-performance computing, it was developed an in-house library named automatic distributed mesh database that employs open-source libraries, like Zoltan and ParMETIS, to manage the distributed grid information and Petsc as the parallel solver. The simulator can handle four different 3D element types: hexahedrons, tetrahedrons, prisms, and pyramids. Several case scenarios with grids ranging from 200 thousand to 26 million nodes using up to 512 processes were successfully simulated. Results showed excellent speedup, with values very close to the ideal speedup for most simulated cases.