Presentation
A High-Order GPU-Accelerated Dynamical Core Based on Discontinous Galerkin Methods
Presenters
DescriptionHigh-order discontinuous Galerkin (DG) methods offer the same level of accuracy as traditional methods while using fewer degrees of freedom. At the same time, their increase in arithmetic complexity occurs locally, leading to efficient algorithms with good scalability and throughput, even on modern GPUs. In this talk we present our current efforts in developing a novel, exascale-ready dynamical core for weather and climate simulations, in which a DG scheme provides increased accuracy and performance and supports efficient utilization of next generation supercomputing hardware. Our approach is based on the flow solver Trixi.jl, which is built around a state-of-the-art spectral element method with entropy conserving flux differencing for improved robustness and with adaptive mesh refinement for reduced time-to-solution. Written in the modern high level language Julia, Trixi.jl provides rapid prototyping capabilities, HPC-grade scalability, and vendor-agnostic GPU support. Legacy applications can utilize our methods via our interface library libtrixi, which provides APIs in C and Fortran. We will present results of simulations steered by the Fortran code MESSy and show performance analyses of our CPU and GPU codes.
TimeMonday, June 1615:00 - 15:30 CEST
LocationRoom 5.0B15 & 16
Session Chair
Event Type
Minisymposium
Climate, Weather, and Earth Sciences
Computational Methods and Applied Mathematics
