| Hi-index | 0.00 |
When geographers and other geo-spatial practitioners are benefiting from the advancements in GIScience and GeoComputation, the gap between the ever-increasing demands and the limited resources of computing power is emerging quickly. High performance computing (HPC), i.e., parallel computing, is needed to handle the massive volume of high-resolution geo-spatial data, and perform increasingly sophisticated and complex Geoprocessing algorithms and models. In this study, a parallel statistical areal interpolation algorithm, which includes two parallel processes, was developed to improve the performance of the computationally intensive FFT-based areal interpolation. A generic parallel Raster Processing programming Library (pRPL) was developed for general geographers to parallelize almost any raster processing algorithms and models with any arbitrary neighborhood configurations. pRPL effectively reduces the development complexity, and efficiently reduces the computing time of computationally intensive algorithms. Finally, the well-known urban land-use simulation Cellular Automata (CA) model, SLEUTH, was parallelized using pRPL. By fully utilizing the advanced features of pRPL, pSLEUTH is able to significantly shorten the computing time for the calibration process.Experiments with real-world datasets suggest: (1) Vector processing can be parallelized using spatial object de-clustering, and cyclic-based mapping schemes deliver better performances than block-based schemes; (2) Regular decomposition methods, i.e., row/column/block-wise decomposition, yield good performance when the workload is rather homogeneous than heterogeneous over the space, and the spatial-adaptive decomposition method, i.e., quad-tree decomposition, deliver the best performance only when the workload is highly clustered over the space; (3) Reducing the decomposition granularity helps improve the performance; (4) data-task-hybrid parallelization delivers better performance than data parallelization when an iterative algorithm is to be parallelized; (5) dynamic tasking yields better performance than static tasking when dealing with a large number of tasks, especially on a parallel computing system where the computing and transfer rates vary among the processors.Furthermore, with the largely improved computing power, the simplifying assumptions that had to be used in sequential computing to make the algorithm computationally feasible on a single-processor computer can be removed, and the findings obtained based on the simplifying assumptions may be altered. Parallelizing Geoprocessing provides geographers a computational test-bed to validate and discover new theories.