AutoPas in ls1 mardyn: Massively parallel particle simulations with node-level auto-tuning

摘要

Due to computational cost, simulation software is confronted with the need to always use optimal building blocks — data structures, solver algorithms, parallelization schemes, and so forth — in terms of efficiency, while it typically needs to support a variety of hardware architectures. AutoPas implements the computationally most expensive molecular dynamics (MD) steps (e.g., force calculation) and chooses on-the-fly, i.e., at run time, the optimal combination of the previously mentioned building blocks. We detail decisions made in AutoPas to enable the interplay with MPI-parallel simulations and, to our knowledge, showcase the first MPI-parallel MD simulations that use dynamic tuning. We discuss the benefits of this approach for three simulation scenarios from process engineering, in which we obtain performance improvements of up to 50%, compared to the baseline performance of the highly optimized ls1 mardyn software.