Wednesday 02 Nov 2022: Driven Turbulence using Discontinuous Galerkin Hydrodynamics on GPUs
Miha Cernetic - MPA Garching
Physics building, 4th floor 14:00-15:00
High-order Discontinuous Galerkin methods are characterised by superior convergence rates and high computational efficiency, making them particularly promising for the high Reynolds numbers and huge dynamic range of astrophysical turbulence as encountered in many different systems, including the gas kinematic and structure of the ISM, the influence turbulence has on star formation and general CGM modelling.
I will present a novel multi-GPU/multi-node realisation of Discontinuous Galerkin hydrodynamics written in CUDA and discuss how we cope with physical discontinuities such as shocks in this method, which has previously proven problematic in DG at high order.
Our new code runs efficiently on different hardware configurations, ranging from a single core on a single node, to all available GPUs and CPUs on large compute clusters using MPI-parallelisation between nodes.
I will describe our development strategy and technical approaches to address drawbacks of GPU computing, including overlapping communication and computation, and minimising non-coalesced memory reads and registry pressure.
I will then discuss the accuracy of the code in test problems, and analyse the weak and strong scaling behaviour of the code, demonstrating that some problems reach a speedup of ~50x on the GPU compared to the CPU version.
Finally, I will present first applications to large, high-order simulations of sub- and supersonic driven turbulence.