High-Performance Spatial Simulations and Optimisations on 64-Bit Architectures.
Ken A. Hawick, Heath A. James, Chris J. Scogings · 2006
Many numerical simulations applications continue to require large computing budgets to allow their use in stateof-the-art parameter regimes. We report on a number of optimisation techniques that are especially applicable to spatial simulations. We employ trade-o ↵ techniques based on the use of high-memory and especially 64-bit addressable memory to boost the performance of simulations that need extensive geometric maths function evaluations. We also describe data structures and indirect addressing techniques to improve spatially-oriented simulation algorithms. We present performance data and a discussion and analysis on these techniques which we believe generalise to many spatially-based scientific simulations codes. We illustrate these methods in the context of simulations for: di↵usion; cluster-cluster aggregation; Monte-Carlo annealing; and artificial-life models.