Speaker
Description
Simulating the formation of large-scale structure in the Universe - from cosmic filaments to galaxy clusters - requires solving gravity for hundreds of millions of particles across a vast range of spatial scales. As HPC systems increasingly rely on GPU accelerators, ensuring that cosmological codes can exploit this hardware efficiently has become a critical challenge.
Within the SPACE-CoE project, we at CINECA are working to actively collaborate and contribute to modernise and GPU-accelerate OpenGadget, one of the reference codes for cosmological N-body and hydrodynamical simulations. OpenGadget handles gravity through a combination of two complementary methods: a Barnes-Hut oct-tree for nearby interactions, and a Particle-Mesh solver based on FFTs for large-scale forces.
We present the progresses achieved by the teams involved - namely CINECA, INAF-Trieste and ExactLab - on porting both components to GPUs on the pre-exascale Leonardo system. For the Particle-Mesh solver, the HeFFTe library has been leveraged to offload distributed FFTs to GPUs, achieving better scalability compared to traditional CPU implementations. Concerning the tree solver, the Barnes-Hut algorithm has been redesigned from the ground up with GPU execution in mind, exploiting Peano-Hilbert space-filling curves to improve memory locality and reduce thread divergence, two of the main bottlenecks when running tree algorithms on massively parallel hardware. We will show the first preliminary scaling tests that confirm the viability of these approaches.
These efforts represent SPACE-CoE commitment to providing the astrophysics and cosmology community with simulation tools ready for the next generation of European supercomputers.