Accelerating Linear Solvers on NVIDIA Grace

Reservoir simulation is a key tool in the energy industry, providing insights into fluid flow and stress behavior within oil and gas reservoirs. Linear system solving is the most critical and time-consuming aspect of these simulations, demanding massive computing power. The Brazilian energy company Petrobras, in collaboration with academic institutions, has developed SolverBR — a specialized linear solver designed for integration into reservoir simulators. SolverBR aims to achieve high efficiency and scalability in HPC environments.

This presentation showcases the collaborative efforts of Petrobras, CESAR, and NVIDIA porting and benchmarking SolverBR in ARM. The goal is to evaluate the performance impact of ARM in the linear system-solving of matrices derived from real and complex reservoir simulations. We conducted tests in both cloud and on-premises environments, comparing the performance of Intel Xeon Platinum processors versus AWS Graviton3 and NVIDIA Grace. Initial experiments showed a cloud price-performance advantage of ARM processors in comparison with x86, while NVIDIA Grace presented a speedup of up to 60% against the same processor for a particular set of matrices. In this presentation, we also discuss the challenges encountered in compiling the x86 code to ARM and present the corresponding solutions.