Volumes of urban sensing data captured by consumer electronic devices are increasing exponentially and current disk-resident database systems are becoming increasingly incapable of handling such large-scale data efficiently. In this study, we report our design and implementation of U2SOD-DB, a column-oriented, GPU-accelerated, in-memory data management system targeted at large-scale ubiquitous urban sensing origin-destination data. Experiment results show that U2SOD-DB is capable of handling hundreds of millions of taxi-trip records with GPS recorded pickup and drop-off locations and times efficiently. Spatial and temporal aggregations on 150 million pickup locations and times in middle-town and downtown Manhattan area in the New York City (NYC) can be completed in a fraction of a second which is 10-30X faster than a serial CPU implementation due to Graphics Processing Unit (GPU) accelerations. Spatially joining the 150 million taxi pickup locations with 43 thousands tax lot polygons in identifying trip purposes has reduced the runtime from 30.5 hours to around 1000 seconds and achieved a two orders (100X) speedup using a hybrid CPU-GPU approach.


Related Publications:

1) Jianting Zhang, Hongmian Gong, Camille Kamga, Le Gruenwald.  U2SOD-DB: Design of an Efficient Database System to Manage Large-Scale Ubiquitous Urban Sensing Origin-Destination Data.  To appear in ACM SIGKDD workshop on Urban Computing [Link]