Locality data properties of 3D data orderings with application to parallel molecular dynamics simulations

General-purpose computing on GPUs is widely adopted for scientific applications, providing inexpensive platforms for massively parallel computation. This has motivated us to investigate GPU performance in terms of speed and memory usage, specifically in relation to data locality in molecular dynamics simulations. The assumption is that enhancing data locality of these applications will lower the cost of data movement across the GPU memory hierarchy. In this research, we analyse spatial data locality and data reuse (temporal data locality) characteristics for row-major, Hilbert, and Morton data orderings, and hybrid variants of these, and assess their impact on the performance of molecular dynamics simulations (MDS). Data locality in MDS applications, based on the relationship between a bin and its neighbouring bins, that are generated using an approximately spherical stencil, previously has not been widely studied. In this research, a simple cache model is presented, and this is found to yield results that are consistent with timing results for the particle force computation obtained on NVIDIA Geforce GTX960 and Tesla P100 graphical processing units (GPUs). The NVIDIA profiling tool is used to investigate the execution time results and to observe the memory usage in terms of cache hits and the number of memory transactions. The analysis also provides a more detailed explanation of execution behaviour for the different orderings. To the best of our knowledge, this is the first study to investigate memory analysis and data locality issues for molecular dynamics simulations of Lennard-Jones fluids on NVIDIA’s Maxwell and Tesla architectures