Molecular dynamics simulations are very compute-intensive. Together with the fact, that MD-based insight typically arises from ensemble calculations, this underpins the importance of highly efficient, parallelizable MD software. From a software perspective, this is particularly challenging against the background of rapidly changing high performance computing (HPC) hardware, e.g. in terms of the upcoming exascale platforms. Besides, considering highly heterogeneous MD systems such as interfacial flows or phase decompositions, it is far from trivial which parallelized MD algorithm is the best choice -- if there is a "single best choice" for a long-term simulation at all! In our workshop contribution, I will revisit the well-known algorithms for short-range MD calculations, discuss their features and shortcomings in terms of their algorithmics, implementations and parallelizability. This analysis has fed into our software project AutoPas, a node-level particle library that provides many realizations of particle data structures, particle pair traversals and parallelizations thereof. To always run at optimal performance, AutoPas is capable of detecting and choosing "on-the-fly" the fastest short-range MD algorithm (auto-tuning). AutoPas will also be the focus of our support session in which we will demonstrate how AutoPas works, which steps are to be taken to include your own particle description or to use AutoPas as kernel (e.g. for force calculations) in an existing MD software package.     «

Molecular dynamics simulations are very compute-intensive. Together with the fact, that MD-based insight typically arises from ensemble calculations, this underpins the importance of highly efficient, parallelizable MD software. From a software perspective, this is particularly challenging against the background of rapidly changing high performance computing (HPC) hardware, e.g. in terms of the upcoming exascale platforms. Besides, considering highly heterogeneous MD systems such as interfacial...     »