Abstract

Spatial databases are used in a wide variety of real-world applications, such as land surveying, urban planning, and environmental assessments, as well as geospatial Web services. As uses of spatial databases become more widespread, there is a growing need for good performance of spatial applications. In spatial workloads, queries tend to be computationally-intensive due to the complex processing of geometric relationships. Furthermore, a significant fraction of spatial query execution time is spent on CPU stalls due to memory accesses, caused by the ever-increasing processor-memory speed gap. With the advent of massively-parallel graphics-processing hardware (GPUs) and frameworks like CUDA, opportunities for speeding up spatial processing have emerged. In addition to massive parallelism, GPUs can also better hide the memory latency. We aim to speed up spatial query execution using CUDA and recent GPU cards. One of the main challenges in using GPUs is the transfer time from main memory to GPU memory. We implement a set of six typical spatial queries and achieve a baseline speedup (without the transfer cost) of 62-318x over the CPU counterparts. We show that the transfer cost can be amortized over the execution of each individual query. For simpler spatial queries, the transfer time is a significant fraction of the query execution time, but we still achieve a 6-10x speedup. For more complex spatial queries, the transfer time becomes negligible compared to the processing time, and we obtain a 62-240x speedup.

 

Manuscript

 

Bibtex