Authors: Huidae Cho – New Mexico State University
Presentation Type: Lightning Talk
Title: Memory Efficiency in Parallel Computation of Continental-Scale Hydrologic Parameters
Abstract: Geospatial data has grown in both its quantity and quality thanks to significant advances in remote sensing technologies. One example is the 3D Elevation Program (3DEP) tiles from USGS. Their spatial resolution varies from 1 arc-second (approximately 30 meters) down to 1 meter depending on the coverage. For the Contiguous United States (CONUS), the 1/3 arc-second (approximately 10 meters) product provides the highest resolution with about 135 billion cells including no data. Just reading in this Digital Elevation Model (DEM) as the 8-byte double type takes up 1TB of memory. My open-source hydrology research focuses on high-resolution macro-scale computations of hydrologic parameters to address these challenges in computational resources including computing power and memory. The very first step in this research agenda is to tackle single-node (one-computer) problems before moving towards multi-node (multiple-computer) ones because I believe that efficient single-node algorithms will pave the solid foundation for multi-node computing. Since none of my lab computers has more than 1TB memory, I used the 1 arc-second CONUS DEM for computation of flow accumulation, delineation of tens of thousands of watersheds, and computation of tens of thousands of longest flow paths. For flow accumulation, I used the entire Texas with about 2 billion cells because the input and output data for the CONUS did not fit in the testing computer’s 64GB memory. This memory restriction remained the same for all benchmark algorithms for watershed delineation and longest flow path. However, I was able to solve this memory issue by developing an input-destructive algorithm. In this talk, I introduce a new OpenMP parallel algorithm for Memory-Efficient Flow Accumulation (MEFA) and briefly present preliminary results from my ongoing research on watershed delineation, longest flow path, and upstream/downstream flow length algorithms.