Simulating the Ocean on the GPU

Details

Supervisors

Lambrechts, Jonathan ; Legat, Vincent ; Hanert, Emmanuel

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en mathématiques appliquées, à finalité spécialisée

Abstract

Coastal ocean models are of great importance to understand the impact of human activities along the land-sea continuum. These activities often occur at a very small scale and their impact is felt at a much larger scale. Numerical models should therefore be able to bridge the gap between those very different scales. This requires large computational resources and leads to long computing time. Graphics Processing Units (GPUs) offer a way to drastically reduce the computation time by providing a new massively parallel computing paradigm. Here we develop a new version of the multi-scale coastal ocean model SLIM that can run on the latest AMD GPU-compute architectures, such as the one is about to be available in the pre-exascale European supercomputer LUMI. We analyze with great detail the the performance improvements that different changes can have, and how they relate to the low-level architecture of GPUs. We also focus on the performance and scaling of multi-GPU setups. Finally, we give an initial look at some of the he 3D components of SLIM and highlight the specific challenges that they bring to the GPU. Overall our results show very promising speedups of up to 100x that will allow the future applications of SLIM to tackle much more ambitious coastal applications.