Optimizing Molecular Dynamics Simulations with Dynamic Auto-Tuning

Molecular Dynamics simulations are highly diverse by nature and require different optimization strategies depending on the scenario. For example, the simulation of a homogeneous gas poses different challenges than droplets. To make matters worse this optimum can change over the course of a simulation if said gas begins to form droplets. To tackle this challenge we published the C++ library project AutoPas, which acts as a node-level performance library for arbitrary N-Body simulations. It implements multiple types of particle containers, parallelization strategies, data layouts, and further optimization techniques. During runtime AutoPas then chooses the fastest combination of aforementioned aspects for an optimal time to solution of the pairwise force calculation in the system. Throughout the simulation, the library periodically reevaluates the current state of the simulation and is capable to adapt all of its internal algorithm configurations to sustain optimal performance for the whole duration of the simulation. In this talk, we present recent optimization options and parallelization strategies implemented in the library and their impact while integrated into the simulation software ls1-mardyn. The introduction of even more parameters and choices inevitably increases the search space for the auto-tuning process. Therefore techniques are discussed how to efficiently find a near-optimal configuration and first results are presented.     «

Molecular Dynamics simulations are highly diverse by nature and require different optimization strategies depending on the scenario. For example, the simulation of a homogeneous gas poses different challenges than droplets. To make matters worse this optimum can change over the course of a simulation if said gas begins to form droplets. To tackle this challenge we published the C++ library project AutoPas, which acts as a node-level performance library for arbitrary N-Body simulations. It imp...     »