Eric Wolter: GPU Simulation of Rigid Fibers

The major objective of this Master's thesis is to accelerate a serial implementation of a numerical algorithm for the simulation of slender fiber dynamics by using Graphical Processing Units (GPU). We focus on rigid fibers sedimenting due to gravity in a free-space Stokes flow. The ability to simulate a large number of fibers in a reasonable computational time on a high-performance parallel platform opens exciting new research opportunities.

The previous serial implementation is rewritten for parallel execution. The algorithm is implemented in single precision using the Compute Unified Device Architecture (CUDA) on NVIDIA GPUs. In addition, we develop an OpenMP version of the parallel implementation to run on multi-core CPUs. Using both implementations, we perform a number of benchmarks to determine the fastest variant of the algorithm. We observe a speedup of 20x to 40x on the NVIDIA GTX 970 compared to an Intel Core i7 4770. The GPU implementation can simulate up to 2000 fibers on a desktop computer and it takes only in the order of 8 seconds to advance one time step.

Furthermore, we have performed a number of simulations of known experiments for sedimenting fibers to validate the algorithm and to explore the numerical precision of the results. The results show an excellent agreement with results from prior experiments in the field.