Agent-Based versus Data-Streamed Big-Data Parallelization in Computational Geometry

Abstract

While some spatial analytics in GIS are well-supported by combinations of data-streaming tools and machine-learning algorithms, (e.g., Spark and KNN on a geographic map), the others are tightly coupled with computational geometry. They handle multi-dimensional or graph structures that should be better maintained in distributed memory for the purpose of dynamic spatial analysis and repetitive geometric queries. We expect that dispatching agents as active data analyzers into datasets would work smoother than disassembling and conveying datasets into passive data-streaming tools. This research compared agent-based and data-streaming big-data parallelization in computational geometry. We parallelized the closest pair of points, the convex hull, the Euclidean shortest path, the point location, and the range search problems, using our MASS (multi-agent spatial simulation) library, MapReduce, and Spark. MASS performed best in three of the programs, which demonstrated agents’ strength in spatial analysis over many data points and repetitive geometric queries with deep tree traverses.