The particle simulation library AutoPas implements many algorithms with vastly different performance characteristics to solve the pairwise short-range particle interactions in molecular dynamics simulations. During the simulation, it uses black-box optimization techniques to automatically select the fastest algorithm for the current state. While they are able to find good algorithms eventually, they often try highly unsuitable ones at the start due to lack of initial per- formance information. As some algorithms perform orders of magnitude worse than the optimum for a given simulation state, this has a significant negative impact on the time to solution. In this project, we gather knowledge about the performance characteristics of the algorithms through theoretical modeling, profiling, and benchmarking. We make the results available through a new white-box optimization strategy that is able to apply any domain-specific knowledge during optimization. It removes those algorithms from the candidate list that likely perform worst in the current simulation state. In our tests, removing the five percent slowest algorithms reduced the tuning time by up to 80 percent while still finding the best algorithm. Furthermore, we give insights and recommendations what can be done to potentially improve the performance further.      «

The particle simulation library AutoPas implements many algorithms with vastly different performance characteristics to solve the pairwise short-range particle interactions in molecular dynamics simulations. During the simulation, it uses black-box optimization techniques to automatically select the fastest algorithm for the current state. While they are able to find good algorithms eventually, they often try highly unsuitable ones at the start due to lack of initial per- formance information. A...     »