Particle simulations are widely used in the natural sciences for a broad range of applications—from replicating experiments to investigate physical properties, to observing the evolution of system states. These simulations can yield insights that are otherwise difficult or impossible to obtain using conventional experimental methods. As a result, various simulation techniques have been developed, each tailored to specific scenarios. As the number of particles increases, the pairwise computation of forces becomes computationally intensive, necessitating the development of optimized algorithms. Most existing particle simulation engines rely on static optimization, selecting a single highly optimized algorithm
at the beginning of the simulation.
AutoPas, an open-source C++ performance library for node-level particle simulations, addresses this challenge by dynamically selecting suitable settings and algorithms based on the current simulation state. The selection of the optimal configuration begins with a sampling phase, during which the runtime of each iteration is measured. A weighted average of these samples is then used to balance the influence of iterations involving a container rebuild versus regular iterations, depending on the rebuild frequency.
Unlike static containers, the rebuild frequency in dynamic containers is neither user-defined nor constant. This thesis investigates a method for accurately estimating the rebuild frequency in dynamic containers within AutoPas. This estimation is integral to runtime prediction and plays a key role in the selection process for determining the optimal simulation configuration.