Scaling and Performance of Real-Space Electronic Structure Calculations on Exascale Architectures

Electronic structure calculations are hard to scale on massively parallel systems, and these challenges are compounded on CPU-GPU exascale architectures. Real-space formulations enable easy parallelization via domain decomposition, which has been implemented in our open-source real-space multigrid code (RMG). RMG has been designed to perform well on leading-edge supercomputers from inception. However, scaling to extreme sizes and to distributed multi-CPU-GPU architectures requires careful consideration of data distribution and flow, including inter-node transfers as well as between CPUs and GPUs located on the same node. The large mismatches between CPU and GPU clock speeds and FLOP rates provides additional constraints as well as optimization opportunities. We will discuss efficient and scalable implementation of distributed data flow and key electronic structure algorithms on exascale class machines in RMG, as well as methods for addressing MPI scalability constraints and data bottlenecks. 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.