Porting the Meso-NH atmospheric model to GPU architectures allows simulating extreme weather events across scales

Numerical simulation of the atmosphere plays a crucial role in understanding and anticipating extreme weather events. Steady advances in computing power have made it possible to increase the complexity and range of scales represented by numerical simulation. However, the advent of heterogeneous computing architectures with multi-core central processing units (CPUs) and graphics processing units (GPUs) requires atmospheric codes to be adapted.

Here we describe the adaptation of the Meso-NH non-hydrostatic mesoscale atmospheric model of the French research community. The Fortran code is ported to GPUs by including OpenACC directives to the most computationally expensive parts of the code. This approach allows running the same code on CPUs and on hybrid CPUs/GPUs architectures. To guarantee the accuracy of the port and the absence of bugs, measures have been taken to ensure bit-to-bit reproducibility between executions on these two architectures. A critical point lies in the atmospheric pressure solver, which requires the inversion of an elliptic equation. A geometric multigrid inversion algorithm is integrated, because the fast Fourier transforms approach used in the original version of the code becomes inefficient with a high number of GPUs. Currently, the code runs on different GPU-NVIDIA and GPU-AMD platforms and scales efficiently up to at least 1,024 GPUs, achieving a 3x increase in energy efficiency compared to CPUs only.

First scientific applications focus on the simulation of extreme weather events across scales as part of a Grand Challenge GPU pilot project on the AMD-based Adastra supercomputer of GENCI (same architecture as Frontier, the 1^st exascale supercomputer). Three representative storms are simulated: a North Atlantic windstorm associated with a mid-latitude cyclone, a Mediterranean convective storm characterized as a derecho, and a mesoscale convective system over the Amazon rainforest. Representation of the North Atlantic storm requires downscaling from the synoptic cyclone scale (>100 km) down to local wind gust formation (<1 km). Inversely, the representation of the Amazon storm requires upscaling from the local triggering of convective cells (<1 km), which organize and maintain the system at the mesoscale (>100 km). Finally, the Mediterranean storm involves both up- and downscaling. We show that Meso-NH successfully represents the cascade of scales for the three representative storms for horizontal grid spacing down to 100 m and grid size up to 4096x4096x128 points.

Porting Meso-NH to GPUs opens up new opportunities to simulate extreme weather events across scales and paves the way for future European exascale supercomputers.