GPU-based parallel indexing for concurrent spatial query processing

In most spatial database applications, the input data is very large. Previous work has shown the importance of using spatial indexing and parallel computing to speed up such tasks. In recent years, GPUs have become a mainstream platform for massively parallel data processing. On the other hand, due to the complex hardware architecture and programming model, developing programs optimized towards high performance on GPUs is non-trivial, and traditional wisdom geared towards CPU implementations is often found to be ineffective. Recent work on GPU-based spatial indexing focused on parallelizing one individual query at a time. In this paper, we argue that current one-query-at-a-time approach has low work efficiency and cannot make good use of GPU resources. To address such challenges, we present a framework named G-PICS for parallel processing of large number of concurrent spatial queries over big datasets on GPUs. G-PICS is motivated by the fact that many spatial query processing applications are busy systems in which a large number of queries arrive per unit of time. G-PICS encapsulates an efficient parallel algorithm for constructing spatial trees on GPUs and supports major spatial query types such as spatial point search, range search, within-distance search, k-nearest neighbors, and spatial joins. While support for dynamic data inputs missing from existing work, G-PICS provides an efficient parallel update procedure on GPUs. With the query processing, tree construction, and update procedure introduced, G-PICS shows great performance boosts over best-known parallel GPU and parallel CPU-based spatial processing systems.