Using Linda to compute spatial autocorrelation in parallel

Abstract

A measure of spatial autocorrelation for polygon data, Moran's I, is decomposed into a set of processes that are executed in a MIMD parallel processing environment. Computational experiments, implemented using C-Linda, were conducted to evaluate the performance of two versions of the parallel algorithm. The first used contiguity-based weights and the second used distance-based weights. C-Linda is a parallel coordination language that enables researchers to process problems in parallel using a wide variety of computer architectures. In our analyses, the number of processors was changed between one and fourteen and the size of the data set was increased from 500 to 3000 polygons. The results indicate that for contiguity-based weighting the use of additional processors failed to improve performance. When distance-based weights were used, however, substantial gains in performance were realized.