Leveraging Node-level Performance for Molecular Dynamics Through Auto-tuning

Total time to solution in molecular dynamics (MD) simulations is highly sensitive to simulation parameters, including material as well as algorithmic settings. For example, there exist multiple data storage concepts and traversal algorithms, which show advantages in different scenarios. Besides, MD settings can change over the course of the simulation, for example, the homogeneity of the particle distribution. The library AutoPas is an easy to use foundation for building arbitrary N-body simulations aiming for optimal node-level performance. Therefore, AutoPas employs auto-tuning during runtime to find the optimal combination of algorithms for the current state of the simulation and adapts when the situation changes. As a proof of concept, we integrated AutoPas into the established highly scalable MD simulation code ls1 mardyn. We showcase the library's ease of use, scalability, and to what extent ls1 mardyn can benefit from the auto-tuning process. This talk gives an overview of the AutoPas library interface, implemented algorithms, and the auto-tuning process. We are further highlighting early results of the coupling with ls1 mardyn.      «

Total time to solution in molecular dynamics (MD) simulations is highly sensitive to simulation parameters, including material as well as algorithmic settings. For example, there exist multiple data storage concepts and traversal algorithms, which show advantages in different scenarios. Besides, MD settings can change over the course of the simulation, for example, the homogeneity of the particle distribution. The library AutoPas is an easy to use foundation for building arbitrary N-body simu...     »