high performance computing on graphics processing units: hgpu.org

Molecular Dynamics (MD) simulations provide a molecular-resolution picture of the folding and assembly processes of biomolecules, however, the size and timescales of MD simulations are limited by the computational demands of the underlying numerical algorithms. Recently, Graphics Processing Units(GPUs), specialized devices that were originally designed for rendering images, have been repurposed for high performance computing with significant increases in performance for parallel algorithms. In this thesis, we briefly review the history of high performance computing and present the motivation for recasting molecular dynamics algorithms to be optimized for the GPU. We discuss cutoff methods used in MD simulations including the Verlet Neighbor List algorithm, Cell List algorithm, and a recently developed GPU-optimized parallel Verlet Neighbor List algorithm implemented in our simulation code, and we present performance analyses of the algorithm on the GPU. There exists an N-dependent speedup over the CPU-optimized version that is ~30x faster for the full 70s ribosome (N=10,219 beads). We then implement our simulations into HOOMD, a leading general particle dynamics simulation code that is also optimized for GPUs. Our simulation code is slower for systems less than around 400 beads but is faster for systems greater than 400 beads up to about 1,000 beads. After that point, HOOMD is unable to accommodate any more beads, but our simulation code is able to handle systems much larger than 10,000 beads. We then introduce a GPU-optimized parallel Hybrid Neighbor/Cell List algorithm. From our performance benchmark analyses, we observe that it is ~10% faster for the full 70s ribosome than our implementation of the the parallel Verlet Neighbor List algorithm.