Data-parallel implementations of map analysis and animal movement for landscape ecology models

In this thesis, improved sequential and data-parallel implementations of landscape ecology model components are presented. Parallelization efforts on the (SIMD) MasPar MP-2 focus on three model components: cluster identification, mean squared radius computation (cluster geometry), and animal movement. The NOrthern YELlowstone Park ungulate model (NOYELP), developed by Drs. Monica Turner and Yegang Wu of the Environmental Sciences Division, Oak Ridge National Laboratory, serves as the example landscape ecology model for the model components studied. Modifications made to the original Fortran-77 NOYELP program as part of this thesis project resulted in a revised serial version which executes 11 times faster than the original (CPU time on a Sun SPARCstation 2). Parallel implementations were tested and compared to functionally comparable serial algorithms using both random maps and maps extracted from runs of the NOYELP model. Speed improvements of MasPax MP-2 parallel kernels over serial implementations on Sun SPARCstations on the order of 9 and 150 for cluster identification and mean squared radius computation, respectively, were measured on 512 x 512 random maps with a resource probability of 0.85. Speed improvements generally increased with map size and density. For landscape maps tested, speed improvements were somewhat lower, due largely to the inclusion of map pixels outside the study area (54% of total map pixels) in the data maps analyzed. Results of this study indicate that parallel adaptation of kernels for cluster identification and geometry is straightforward, but that effective parallelization of animal movements in the NOYELP model and similar individual-based models will involve reconceptualizing the movement rule. Issues involved in the parallelization of landscape ecology models are discussed and suggestions are made for future work in this area.