Flow simulation and visualization on GPU clusters

In recent years, the performance of graphics processing units (GPUs) has been increasing dramatically. Modern GPUs have surpassed CPUs in raw computational power by an order of magnitude. Because of the explicit data parallelism in the graphics pipeline, the GPU can efficiently use hundreds of thread processors to process data in parallel. Moreover, the GPU is becoming more and more flexible and programmable. As a result, accelerating general-purpose computation on the GPU (GPGPU) has become an active area of research.;This dissertation presents efficient ways to use GPUs and GPU clusters for GPGPU in general and flow simulation and visualization in visual applications in particular. We focus on a GPU-friendly method, called the Lattice Boltzmann Model (LBM), a mesoscopic method that applies linear and local operations at discrete lattice sites. We describe an optimized LBM implementation on a single GPU and its applications in real-time modeling of natural phenomena, such as fire, smoke, wind, and heat shimmering. We also present a novel GPU-based adapted unstructured LBM algorithm for simulating flow on arbitrary 3D triangular surfaces. We further extend the LBM implementation from a single GPU to a GPU cluster and describe how to efficiently manage the communication among the multiple GPUs. We also present an application of the GPU cluster simulation in urban dispersion modeling. We further present an LBM implementation of irregular-shaped simulation domain on a GPU cluster and its application for thermal fluid dynamics in a pressurized water reactor of a nuclear power plant. Finally, Zippy, a general framework for GPU cluster programming, is presented. Zippy abstracts the GPU cluster architecture with two characteristics important to high performance---two-level parallelism hierarchy and non-uniform memory access (NUMA)---and hides other architecture details. It simplifies the programming of a GPU cluster while maintaining high performance. We also present three example applications developed using Zippy and show how simulation and visualization modules can be seamlessly integrated on a GPU cluster.