Popis: |
A wide variety of algorithms and parallelism strategies have been developed to solve short-range particle simulations efficiently. However, there is no ``silver bullet" strategy that is the most efficient in all scenarios and for all hardware. As such, selecting an optimal algorithm for a given problem can present a significant task that may be outside the domain of knowledge of many domain scientists who wish to use particle simulations in their work. Furthermore, even if the optimal algorithm is chosen for the start of the simulation, this may become sub-optimal throughout the simulation. AutoPas is a node-level particle simulation library designed to solve these problems. By dynamically tuning the algorithm that evaluates the pairwise forces throughout the simulation, AutoPas takes the task of selecting the fastest algorithm out of the hands of the user. Included within the library are a wide variety of optimized particle containers and cell traversals; as well as a variety of tuning strategies. In this talk, we present AutoPas with new results on how AutoPas treats multi-site molecules in molecular dynamics, as an example of tuning on a complex particle type. In addition, we discuss the latest developments for the internal aspects of the library, including the changes in implementation required to provide GPU support for AutoPas using the Kokkos programming model and the first results from Kokkos implementations of containers. Finally, we consider the future directions of the library. |