High-performance zonal histogramming on large-scale geospatial rasters using GPUs and GPU-accelerated clusters

Hardware Accelerators are playing increasingly important roles in achieving desired performance from desktop to cluster computing. While General Purpose computing on Graphics Processing Units (GPGPU) technologies have been widely applied to computing intensive applications, there is relatively little work on using GPUs and GPU-accelerated clusters for data intensive computing that typically involves significant irregular data accesses. In this study, we report our designs and implementations of a popular geospatial operation called Zonal Histogramming on Nvidia GPUs. Given a zonal dataset in the form of a collection of polygons and a geospatial raster that can be considered as a 2D grid, for each polygon, Zonal Histogramming computes a histogram of the values of raster cells that fall within the polygon. Our experiments on 3000+ US counties (polygons) over 20+ billion NASA Shuttle Radar Topography Mission (SRTM) 30 meter resolution Digital Elevation Model (DEM) raster cells have shown that, an impressive 46 seconds end-to-end runtime can be achieved using a single Nvidia GTX Titan GPU device. The runtime is further reduced to 10 seconds using 8 nodes on ORNL's Titan GPU-accelerated cluster. The desired high performance opens many possibilities for large-scale geospatial computing that is important for environmental and climate research.