ls1-mardyn is a molecular dynamics code specialized in simulations of very large scale of up to 4.125 × 10 12 particles [EHB + 13]. Although the established approach for its force calculation, which is using coloring and OpenMP loop parallelization, proves to be very efficient, it still leaves potential for optimization because of its use of barriers. A promising method to replace these barriers are task-based scheduling concepts due to their higher flexibility and finer granularity. As such a library, the highly dynamic Quicksched code [GCS16] has here been used for implementing task-based shared-memory parallelization approaches for the force calculations in ls1-mardyn. Their performance has been tested on an Intel Xeon Phi Knights Corner accelerator and is compared to established approaches that use domain coloring and OpenMP loop parallelization. When directly comparing the results it is hard to beat the efficiency of the highly optimized established approaches. Yet, through an in-depth analysis of the scheduling, possible performance potentials can be identified and shown to be exploited by the task-based approaches. However, task-based approaches introduce new challenges of their own. These are analyzed and discussed and possible solutions are proposed.     «

ls1-mardyn is a molecular dynamics code specialized in simulations of very large scale of up to 4.125 × 10 12 particles [EHB + 13]. Although the established approach for its force calculation, which is using coloring and OpenMP loop parallelization, proves to be very efficient, it still leaves potential for optimization because of its use of barriers. A promising method to replace these barriers are task-based scheduling concepts due to their higher flexibility and finer granularity. As su...     »