September 24, 2010, 12:00 pm - 1:00 pm

September 24, 2010, 12:00 pm - 1:00 pm

GPU Accelerated Simulations of Magnetic Reconnection

John Dorelli (GSFC, Geospace Physics Lab)

Graphics Processing Units (GPUs) are fast becoming a credible alternative to multicore CPU programming for a variety of high performance computing applications. We report on our recent experience using GPUs to accelerate fluid simulations of magnetic reconnection. We describe our approach to porting a non-trivial sequential Hall magnetohydrodynamics (Hall MHD) solver to an NVIDIA GTX 480 (Fermi architecture) using the CUDA (Compute Unified Device Architecture) C programming language. The sequential algorithm is an explicit second order TVD MUSCL-Hancock scheme which makes use of an HLL approximate Riemann solver to compute numerical fluxes. Strang splitting is used to maintain second order accuracy for multidimensional problems, and Dedner's hyperbolic divergence cleaning method is used to control errors associated with the divergence of the magnetic field. On a 1024x1024 computational mesh, we achieve a speedup (relative to the sequential algorithm on a single Nehalem core) of over a factor of 58 (double precision) and 87 (single precision). This gives us the power of a small CPU cluster on a single desktop for under $3K. We present initial applications of our new code to some two-dimensional magnetic reconnection problems. We discuss hybrid CPU-GPU approaches for solving larger (e.g., three-dimensional) problems.