high performance computing on graphics processing units: hgpu.org

Fast growing computing power on commodity parallel hardware makes it both an opportunity and a challenge to use modern hardware for large-scale data management. While GPU (Graphics Processing Unit) computing is conceptually an excellent match for spatial data management which is both data and computing intensive, the complexity of multi-dimensional spatial indexing and query processing techniques has made it difficult to port existing serial algorithms to GPUs. In this study, we propose a parallel primitives based strategy for spatial data management. We present data parallel designs for polygon decomposition, quadtree construction and spatial query processing. These designs can be realized on both GPUs and multi-core CPUs as well as future generation hardware when parallel libraries that support the primitives are available. Using a large-scale geo-referenced species distribution dataset as an example, the GPU-based implementations can achieve up to 190X speedups over serial CPU implementations and 14X speedups over 16-core CPU implementations for polygon decomposition, which is the most computing intensive module in the end-to-end spatial data management solution we have provided. For quadtree constructions and spatial range/polygon query modules, which are more data intensive, the speedups over single and multi-core CPUs are up to 27X and 2X, respectively, depending on workloads. Comparing with a similar technique on polygon decomposition that is realized using a native parallel programming language, our parallel primitives based implementation is up to 3X faster on the species distribution dataset. The results may suggest that simplicity and efficiency can be achieved simultaneously using the data parallel design strategy by identifying the inherent data parallelisms in application domains.