Recently Added Features, Scaling and Performance of the Real-Space MultiGrid (RMG) Code on Exascale Architectures

lectronic structure calculations are hard to scale on massively parallel systems, and these challenges are compounded on CPU-GPU exascale architectures. The open-source RMG code exploits its real-space formulation to enable easy parallelization via domain decomposition but scalability to large node numbers of CPUs and GPUs requires careful attention to data structures and flow between nodes and between CPUs and GPUs located on the same node. We describe the solutions to these problems implemented in RMG as well as the results for large scale ab-initio calculations that use the recently implemented features of RMG: hybrid functionals, semi-local pseudopotentials and spin orbit coupling. For example, hybrid functional (HSE) calculations for up to 512-atom antifferromagnetic NiO supercells with 3584 semi-core and valence electrons exhibit excellent scaling up to 192 Summit CPU-GPU nodes, utilizing all 44 CPU cores and 6 GPUs per node. The RMG source code and build scripts for pre-exascale Summit, Cray XE-XK, clusters, Linux, Windows, and MacOS workstations are available at www.rmgdft.org together with help files and examples.