Optimal multidimensional accuracy-resource trade-off

The goal of this project is to analyze, realize and combine unconventional numerical and high performance computing methods to develop truly adaptive simulation software with an important focus on novel optimality criteria. Together with our partner projects, the entire hardware spectrum from conventional CPU-based clusters, accelerator hardware to mobile resource-poor devices will be targeted. In this first project phase, we will optimize in the underlying multi-dimensional discrete-continuous space. This space ranges from model types and discretization over floating point accuracy and (non-)linear solver types to parallelization schemes. The core result of the project will be (a) highly efficient simulation software modules incorporating a wide range of hardware characteristics in terms of compute and communication, (b) the definition of a high-dimensional parameter space, a selection of objective functions, and initial optimization techniques. This will serve as a basis for advancing the first research question (accuracy-resource trade-off) on PN5 into the fourth one (full accuracy-precision-resource trade-off) in a follow-up project.