Implementation and optimisation of GPU-based oceanic particle transport processes using a Lagrangian Particle Tracker approach

Details

Supervisors

Hanert, Emmanuel ; Lambrechts, Jonathan

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil électromécanicien, à finalité spécialisée: énergie

Abstract

enOceanic particle transport is indispensable for understanding various underlying advection and diffusion phenomena taking place in the oceans. The particles transported can vary in size and properties, but need often to be studied on the individual scale. This need for an individual study is not well served by an Eulerian approach, relying on density gradients. The particulate study allowed by a Lagrangian approach enables taking advantage off the high parallelisability of Graphics Processing Units - GPUs -. In this thesis I implemented and developed this approach on an NVIDIA AD104 GN21-X9 (GeForce RTX 4080 Mobile) from an existing marching model developed for Central Processing Units - CPUs - for the Second-generation Louvain-la-Neuve Ice-ocean Model - SLIM -. I developed, using Python, C++, and NVIDIA’s proprietary Compute Unified Device Architecture - CUDA -, the GPU code that will enable the parallelisation of the computation. The use of the in-built SLIM functions is heavily implied, as well as the high tunability of SLIM in terms of computation needs and format to have a seamless integration. The different modifications done for the adaptation from CPU to GPU of the marching model, their results, and subsequent speedup and power efficiency compared to the CPU implementation are discussed with great details. The utilisation of the GPU’s memory and computing capability are discussed throughout the versions. The final results show a more than twenty-fold speedup with 15% reductionin electrical consumption for a single GPU.