high performance computing on graphics processing units: hgpu.org

Modern GPU architectures closely resemble supercomputers. Commodity GPUs that have already been equipped with personal and cluster computers can be used to boost the performance of spatial databases and GIS. In this study, we report our preliminary work on designing and implementing a spatial join algorithm on GPUs by using generic parallel primitives that have been well understood and efficiently implemented in many parallel libraries. In addition to help understand the inherent data parallelisms in spatial join operations, our experiments have shown that the reference implementation, which represents a tradeoff between code efficiency and code complexity, is able to achieve a 6.7X speedup when compared to an optimized CPU serial implementation. The result is encouraging in the sense that native implementation of spatial joins directly on top of GPU accelerators can potentially achieve much higher speedups for spatial joins which are fundamental to spatial databases and vector GIS. We also believe that the implementations of parallel spatial algorithms on top of generic parallel primitives can be an important first step towards designing and developing high-performance spatial-specific parallel primitives to make it easier to build parallel spatial databases and GIS.