Computing pairwise short-range particle interactions is a time-consuming and unavoidable part of Molecular Dynamics simulations. For the parallelization of the Linked Cell Algorithm there exist different traversal patterns having varying performance depending on the parameters and state of the simulation. We use Extra-P to create performance models from benchmarks and use them to apply auto-tuning in ls1-mardyn. Thus we predict the optimal traversal given the current state of the simulation and use it until reevaluation. For single-parameter models the prediction is quite good but the prediction quality declines fast for multi-parameter models. Additionally, the time required for benchmarks also gets infeasible quickly when adding more and more dimensions to the parameter space because the number of samples required grows exponentially. We thus model the runtime depending only on the density. Auto-tuning with this simple model on each MPI rank separately already delivers close to optimal performance.     «

Computing pairwise short-range particle interactions is a time-consuming and unavoidable part of Molecular Dynamics simulations. For the parallelization of the Linked Cell Algorithm there exist different traversal patterns having varying performance depending on the parameters and state of the simulation. We use Extra-P to create performance models from benchmarks and use them to apply auto-tuning in ls1-mardyn. Thus we predict the optimal traversal given the current state of the simulation and...     »