Date of Award
Doctor of Philosophy (PhD)
Geospatial datasets are growing in volume, complexity, and heterogeneity. For efficient execution of geospatial computations and analytics on large scale datasets, parallel processing is necessary. To exploit fine-grained parallel processing on large scale compute clusters, partitioning of skewed datasets in a load-balanced way is challenging. The workload in spatial join is data dependent and highly irregular. Moreover, wide variation in the size and density of geometries from one region of the map to another, further exacerbates the load imbalance. This dissertation focuses on spatial join operation used in Geographic Information Systems (GIS) and spatial databases, where the inputs are two layers of geospatial data, and the output is a combination of the two layers according to join predicate.This dissertation introduces a novel spatial data partitioning algorithm geared towards load balancing the parallel spatial join processing. Unlike existing partitioning techniques, the proposed partitioning algorithm divides the spatial join workload instead of partitioning the individual datasets separately to provide better load-balancing. This workload partitioning algorithm has been evaluated on a high-performance computing system using real-world datasets. An intermediate output-sensitive duplication avoidance technique is proposed that decreases the external memory space requirement for storing spatial join candidates across the partitions. GPU acceleration is used to further reduce the spatial partitioning runtime. For dynamic load balancing in spatial join, a novel framework for fine-grained work stealing is presented. This framework is efficient and NUMA-aware. Performance improvements are demonstrated on shared and distributed memory architectures using threads and message passing. Experimental results show effective mitigation of data skew. The framework supports a variety of spatial join predicates and spatial overlay using partitioned and un-partitioned datasets.