Scaling staggered grid code on pre-exascale machines

The transition to exascale (>1000 Petaflops) computing necessitates the adaptation of numerical modeling tools to efficiently utilize emerging high-performance computing architectures. Within the ChEESE-2P project, the further development of LaMEM (Lithosphere and Mantle Evolution Model) focuses on achieving scalable performance on advanced systems, including the EuroHPC supercomputer LUMI (currently #3 in Europe). Taking advantage of using the PETSc library, LaMEM demonstrates strong and weak scalability, achieving linear performance up to 512 compute nodes and supporting high-resolution simulations with grids up to 1024³

In response to the increasing emphasis on GPU-based computing, ongoing efforts are directed towards optimizing LaMEM for GPU architectures, including both NVIDIA and AMD systems. Preliminary results highlight significant progress in enabling GPU-accelerated runs and improving resource utilization. This work highlights LaMEM's ability to perform large-scale geodynamic simulations, contributing to the broader goal of integrating physics-based models with available data.