Quantum Monte Carlo, Data Oriented Design, and GPUs

Software engineering is a core tenet of theoretical physics; that is, when we want to explore new physics, this usually involves extending and testing our code to study the new dynamics. Although nuclear physics has long been at the forefront of computational excellence, over recent decades, the software engineering aspect of theoretical nuclear physics has lagged behind standard practices in computing. Until recently, this was overwhelmingly true for conventional coordinate space Quantum Monte Carlo (QMC). To explore new physics with QMC, such as reactions and larger systems, we needed to restructure the entire framework. From the beginning, we adopted a Data-Oriented Design that has enabled us to map our specific MPI topology to any given hardware, allowing us to transition seamlessly from your laptop to Aurora. As a result of this work, we have successfully advanced the evaluation of the trial wavefunction to A=13 systems using Argonne's Supercomputer Aurora by computing Carbon-13. During this calculation, we obtained a speedup of 1.52 when offloading critical kernels to the GPUs. In this talk, we outline the development process that is allowing QMC to explore new science.