Date of Award


Degree Type


Degree Name

Doctor of Philosophy (PhD)


Computer Science

First Advisor

Sushil K. Prasad

Second Advisor

Shamkant B. Navathe

Third Advisor

Rajshekhar Sunderraman

Fourth Advisor

Rafal Angryk


The frequently employed spatial join processing over two large layers of polygonal datasets to detect cross-layer polygon pairs (CPP) satisfying a join-predicate faces challenges common to ill-structured sparse problems, namely, that of identifying the few intersecting cross-layer edges out of the quadratic universe. The algorithmic engineering challenge is compounded by GPGPU SIMT architecture. Spatial join involves lightweight filter phase typically using overlap test over minimum bounding rectangles (MBRs) to discard majority of CPPs, followed by refinement phase to rigorously test the join predicate over the edges of the surviving CPPs. In this dissertation, we develop new techniques - algorithms, data structure, i/o, load balancing and system implementation - to accelerate the two-phase spatial-join processing. We present a new filtering technique, called Common MBR Filter (CMF), which changes the overall characteristic of the spatial join algorithms wherein the refinement phase is no longer the computational bottleneck. CMF is designed based on the insight that intersecting cross-layer edges must lie within the rectangular intersection of the MBRs of CPPs, their common MBRs (CMBR). We also address a key limitation of CMF for class of spatial datasets with either large or dense active CMBRs by extended CMF, called CMF-grid, that effectively employs both CMBR and grid techniques by embedding a uniform grid over CMBR of each CPP, but of suitably engineered sizes for different CPPs. To show efficiency of CMF-based filters, extensive mathematical and experimental analysis is provided. Then, two GPU-based spatial join systems are proposed based on two CMF versions including four components: 1) sort-based MBR filter, 2) CMF/CMF-grid, 3) point-in-polygon test, and, 4) edge-intersection test. The systems show two orders of magnitude speedup over the optimized sequential GEOS C++ library. Furthermore, we present a distributed system of heterogeneous compute nodes to exploit GPU-CPU computing in order to scale up the computation. A load balancing model based on Integer Linear Programming (ILP) is formulated for this system. We also provide three heuristic algorithms to approximate the ILP. Finally, we develop MPI-cuda-GIS system based on this heterogeneous computing model by integrating our CUDA-based GPU system into a newly designed distributed framework designed based on Message Passing Interface (MPI). Experimental results show good scalability and performance of MPI-cuda-GIS system.