SCCS Colloquium - Oct 28, 2020
|Date:||Oct 28, 2020|
|Time:||15:00 - 16:00|
Julian Pelloth: Implementing a predictive tuning strategy in AutoPas using extrapolation
The AutoPas library is capable of simulating computationally intensive molecular dynamic simulations. To reduce the run time, it provides many configurations to select the fastest for the specific scenario. AutoPas tries to achieve this task by testing the whole search space frequently. In this thesis, a new approach is implemented that uses extrapolation to predict the efficiency and only tests the best predictions to avoid inefficient configurations. The predicting strategy can reduce the run time by up to 74%.
Keywords: Molecular Dynamics, AutoPas, Auto-Tuning, Extrapolation
Marco Papula: Implementing the Linked Cell Algorithm inAutoPas using References
AutoPas is an open-source library implemented in C++ that provides a variety of md algorithms optimized for most modern hardware. The library supports auto-tuning, meaning the library will periodically evaluate the whole domain and if another algorithm is expected to increase performance, switch containers accordingly. This process currently involves copying every particle in the domain into a new container, cancelling out some if not all of the performance gained from switching containers. This thesis describes a proof of concept implementation of the Linked Cells container, whose cells operate on references with the actual particles being stored in another structure. The thesis demonstrates the correctness of the implementation and gives a rough estimate of its performance.
Keywords: HPC, MD, Linked Cells