Multi-GPU acceleration of a high-order multi-species Vlasov-Poisson solver

Microphysics plays a critical role in collisionless plasma transport and affects macroscopic properties like resistivity. A kinetic description is required to model these anomalous transport phenomena. The challenge of using kinetic simulations is the computational cost associated with tracking the species distribution functions in phase space. By utilizing GPU accelerated supercomputers the scope of kinetic simulations accessible can be increased to handle realistic proton-electron mass ratios. To achieve a meaningful scaling speedup on GPUs requires reevaluating the core design of kinetic codes. Components such as the Poisson solver and inter-node data communication which accounted for an insignificant fraction of execution time on CPU codes require significant changes to avoid being major bottlenecks for scaling performance in GPU accelerated codes. We present theoretical bounds for how multi-GPU parallel algorithms scale for a multi-species fourth-order accurate finite-volume Vlasov-Poisson solver. These bounds are then used to inform the design and implementation of the VCK-GPU code. VCK-GPU is able to achieve up to 54x speedup as well has increasing simulation throughput by 341x over the CPU code. The new capabilities enabled by GPU acceleration are leveraged to characterize collisionless resistivity induced by the lower hybrid drift instability in pulsed power inertial confinement fusion experiments.