Application performance on graphical processing units (GPUs), in terms of execution speed and memory usage, depends on the efficient use of hierarchical memory. It is expected that enhancing data locality in molecular dynamic simulations will lower the cost of data movement across the GPU memory hierarchy. The work presented in this article analyses the spatial data locality and data reuse characteristics for row-major, Hilbert and Morton orderings and the impact these have on the performance of molecular dynamics simulations. A simple cache model is presented, and this is found to give results that are consistent with the timing results for the particle force computation obtained on NVidia GeForce GTX960 and Tesla P100 GPUs. Further analysis of the observed memory use, in terms of cache hits and the number of memory transactions, 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.
Final Report on Data Locality Enhancement of Dynamic Simulations for Exascale Computing