Accelerating atmospheric simulations with mixed-precision high-order methods
- Academic lead
- Massimiliano Fasi, School of Computer Science, m.fasi@leeds.ac.uk
- Industrial lead
- Stéphane Gaudreault, Environment and Climate Change Canada (ECCC), stephane.gaudreault@ec.gc.ca
- Co-supervisor(s)
- Andrew Ross, School of Earth and Environment, A.N.Ross@leeds.ac.uk
- Project themes
- Climate & Weather, Computational & Analytical Tools, Fundamental, Multiphysics & Complex Fluids
The Euler equations describing atmospheric motion can be partitioned into a stiff linear term, which governs fast atmospheric processes, and a nonlinear term, which captures large-scale dynamics. Exponential integrators offer a way to solve the linear part with higher-order accuracy compared to traditional finite-differences approaches.
In this project, we will develop new high-order exponential integrators, and we will study how these can be implemented efficiently using mixed precision. Our algorithms will use different levels of precision to significantly reduce the runtime and energy requirements of a simulation without sacrificing any accuracy, making numerical weather prediction codes faster and lowering their carbon footprint. The idea is to use the faster – but less accurate – low precision for the bulk of the computation, switching to high precision only when necessary to guarantee high accuracy in the final solution. Such algorithms are particularly well suited to modern supercomputers, which heavily rely on mixed-precision hardware accelerators (such as GPUs, TPUs, and DPUs) to achieve their record-breaking performance.
Hardware accelerators rely heavily on parallel execution, but exponential integrators are particularly challenging to parallelize. Therefore, the project will focus on developing scalable strategies that minimise communication between tasks running in parallel.
Top left: Frame from simulation of atmospheric pressure from ECCC’s Global Deterministic Prediction System (GDPS), which uses a 15km grid. Top right: Frame of simulation of atmospheric pressure from ECCC’s Regional Deterministic Prediction System (RDPS), which uses a 10km grid over North America. Bottom left: Frame from simulation of surface temperature from ECCC’s Regional Ice Ocean Prediction System (RIOPS), which uses a 1/12° resolution grid. Bottom right: Frame from simulation of the temperature gradient at 40 m from ECCC’s Coastal Ice Ocean Prediction System (CIOPS), which uses a 1/36° resolution grid.