GPU4GEO - Frontier GPU multi-physics solvers in Julia

The accelerating outflow of ice in Antarctica or Greenland due to a warming climate or the geodynamic processes shaping the Earth share common computational challenges: extreme-scale high-performance computing (HPC) which requires the next-generation of numerical models, parallel solvers and supercomputers. We here present a fresh approach to modern HPC and share our experience running Julia on thousands of graphical processing units (GPUs).

Computational Earth sciences leverage numerical modelling to understand and predict the evolution of complex multi-physical systems. Ice sheet dynamics and solid Earth geodynamics are, despite their apparent differences, two domains that build upon analogous physical description and share similar computational challenges. Resolving the interactions among various physical processes in three-dimensions on high spatio-temporal resolution is crucial to capture rapid changes in the system leading to the formation of, e.g., ice streams or mountains ranges.

Within the GPU4GEO project, we propose software tools which provide a way forward in ice dynamics, geodynamics and computational Earth sciences by exploiting two powerful emerging paradigms in HPC: supercomputing with Julia on graphical processing units (GPUs) and massively parallel iterative solvers. We use Julia as the main language because it features high-level and high-performance capabilities and performance portability amongst multiple backends (e.g., multi-core CPUs, and AMD and NVIDIA GPUs).

We will discuss our experience using ParallelStencil.jl and ImplicitGlobalGrid.jl as software building blocks in combination to CUDA.jl, AMDGPU.jl and MPI.jl for designing massively parallel and scalable solvers based on the pseudo-transient relaxation method, namely FastIce.jl and JustRelax.jl. Our work shows great promise for solving a wide range of mechanical multi-physics problems in geoscience, at scale and on GPU-accelerated supercomputers.

Co-authors: Ivan Utkin¹ ², Albert De Montserrat¹, Boris Kaus³, Samuel Omlin⁴

¹ ETH Zurich | ² Swiss Federal Institute for Forest, Snow and Landscape Research (WSL) | ³ Johannes Gutenberg University Mainz | ⁴ Swiss National Supercomputing Centre (CSCS)