Dynamic Mesh Partitioning for Improving Linear Solver Load Balancing in Compositional Adaptive Implicit Method Simulations

Summary When an adaptive implicit method (AIM) discretization is used in compositional reservoir simulations, the number of equations being solved simultaneously varies significantly depending on whether a cell is treated explicitly or implicitly. This results in implicit cells carrying a much higher computational workload. As the decision of assigning implicit or explicit treatment varies throughout the simulation, static mesh partitioning will result in load imbalance. In this paper, we address this issue by presenting a new dynamic mesh partition algorithm that adapts to the distribution of implicit/explicit cells, aiming to improve load balancing in compositional AIM simulations. The algorithm’s effectiveness is demonstrated in several complex real models, where it obtains a close to even distribution of implicit equations among domains. Performance timings for the main linear solver kernels on multicore hardware reveal an improved distribution of computations among threads, resulting in significant enhancements in linear solver performance.