High-Performance Spatial Query Processing on Big Taxi Trip Data using GPGPUs

City-wide GPS recorded taxi trip data contains rich information for traffic and travel analysis to facilitate transportation planning and urban studies. However, traditional data management techniques are largely incapable of processing big taxi trip data at the scale of hundreds of millions. In this study, we aim at utilizing the General Purpose computing on Graphics Processing Units (GPGPUs) technologies to speed up processing complex spatial queries on big taxi data on inexpensive commodity GPUs. By using the land use types of tax lot polygons as a proxy for trip purposes at the pickup and drop-off locations, we formulate a taxi trip data analysis problem as a large-scale nearest neighbor spatial query problem based on point-to-polygon distance. Experiments on nearly 170 million taxi trips in the New York City (NYC) in 2009 and 735,488 tax lot polygons with 4,698,986 vertices have demonstrated the efficiency of the proposed techniques: the GPU implementations is about 10-20X faster than the host system and completes the spatial query in about a minute by using a low-end workstation equipped with an Nvidia GTX Titan GPU device with a total equipment cost of below $2,000. We further discuss several interesting patterns discovered from the query results which warrant further study. The proposed approach can be an interesting alternative to traditional MapReduce/Hadoop based approaches to processing big data with respect to performance and cost.