Performance of Open-Source Real-Space Multigrid DFT Code on Pre-Exascale Architecture

RMG (www.rmgdft.org) is an open-source suite of codes for performing large-scale, high-throughput electronic structure calculations. Designed for scalability, it discretizes the DFT equations on real-space grids that are distributed over the nodes of a massively parallel system via domain decomposition. The Kohn-Sham and Poisson equations are solved using multigrid techniques that dramatically accelerate convergence while only requiring nearest neighbor communications. In addition to the multigrid algorithms, the main parts of the calculations consist of dense matrix multiplications and iterative solutions of a partitioned eigenvalue problem that are particularly well suited for GPU accelerators. RMG makes very efficient use of GPUs, including multiple GPUs per node, if they are available. On the IBM/NVIDIA Summit supercomputer at ORNL, RMG utilizes all 6 GPUs per node and Cuda-managed memory to reach 83x performance improvement over the previous generation Cray XK7 Titan supercomputer at ORNL, which contains 1 prior-generation GPU per node. We will also discuss algorithmic improvements enabled by large-memory, high memory bandwidth nodes in view of future exascale architectures.