Abstract

Contributed Talk - Splinter Computation

Friday, 25 September 2020, 14:02   (virtual room B)

OpenGadget3 on GPUS

Klaus Dolag and Antonio Ragagnin
USM/LMU

I will present our work on porting OpenGadget3 onto PGUs with the goal, to perform large, multi physics cosmological simulations. The key aspect for such complex simulation code to perform well on GPUs is an approach, which overlaps computations on the CPUs and GPUs in a clever way. Thereby we are able to keep scalability to very large HPC systems, while keeping re-factoring the code base to a minimum, without having to re-write the various physical modules which are needed to represent the complex physical processes describing galaxy formation in cosmological context. Here the key element is that while GPUs asynchronously compute local interactions at the node level, CPUs perform tree-walks and MPI communications globally which involves other nodes. Thereby significant speed-up can be obtained and the code can make effectively use of accelerator units on HPC systems.