VPIC on GPU

Efficient operation of Particle-in-Cell codes on Graphics Processing Units (GPUs) has been a coveted goal since they were adopted by High Performance Computing platforms years ago. While a variety of research exist on this topic, many of the world’s highest performing PIC codes have yet to demonstrate their ability to effectively use large GPU machines at extreme scale. In this work we demonstrate the effort of Los Alamos National Laboratory to port VPIC to run on large-scale DoE GPU super computers. We demonstrate the codes ability to scale to thousands of GPUs, directly compare code performance to previous systems, and present lessons learnt from porting to the performance-portable code framework Kokkos. We document our strategy for data management in the context of the limited memory regime presented by GPUs, and also demonstrate the strategies we employ to minimize data copies between host and device.