Tiling Strategies for Distributed Point Cloud Databases

Abstract

Many large point clouds -- such as cosmological N-body simulations, intersections of road networks etc. -- are strongly clustered on a hierarchy of scales. In shared nothing distributed environments, optimized tiling of data is crucial to minimize cross-server communication and balance IO and processing load. We propose histogram-based tiling algorithms, a hierarchical tiling and a spectral clustering algorithm, that can be incorporated into the data extraction or transformation phase of a typical Extraction--Transformation--Loading (ETL) procedure. We define measures to characterize the performance of these tiling techniques with respect to typical spatial search operations, and evaluate the algorithms based on these measures using hierarchically clustered data sets.