GPU Accelerated Particle-in-cell Simulations with Charge-Conserving Current Deposition

Particle-in-Cell (PIC) methods are a well-established first-principle model that can provide a kinetic description of a plasma by following trajectories of an ensemble of charged particles in self-consistent electromagnetic fields. To the extent that quantum mechanical effects can be neglected, the PIC model makes no physics approximations and is a key tool in the study of plasma physics. The first-principle nature of the PIC model determines that PIC simulations require intense computation. Modern graphic processing units (GPU's) provide a significant amount of raw compute power and bandwidth, both about an order of magnitude more than a conventional CPU. In this thesis, we have developed an implementation of an electromagnetic PIC code, with charge-conserving current deposition, on a GPU cluster with CUDA. We have developed a new charge-conserving current deposition scheme with little thread divergence and a new particle sorting algorithm that is especially efficient for explicit PIC codes. The implementation takes advantage of the fast on-chip shared memory and coalesced data access. The thread racing technique used also can provide a general method of resolving write conflict among computation threads on GPU. Particle sorting and boundary update methods are carefully designed to minimize data movement. The code has good scalability where the latency of MPI communication between nodes is the main reason for the performance decrease in weak scaling. Depending on plasma temperatures, the GPU implementation has achieved a processing speed of 2.2-4.5 ns per particle-step in two-dimensional (2D) simulations using 1-225 GPUs, and 4.3-15.8 ns