Compute cluster

• Term: Funded in 2022
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 492175459

Funding is requested for an experimental computing cluster. At the Institute of Computational Physics, high performance and supercomputers are the main tool in a large majority of the projects. We work in the areas of soft matter, liquids in restricted geometries, as well as on materials science issues at the quantum level. The central methods are mainly molecular dynamics simulations, as well as hybrid simulations coupling several scales, and quantum mechanical calculations. This combination leads to a high demand for computing time, which, however, can not easily be covered by shared clusters and supercomputers at computing centers. In addition to physics questions, an important focus is on method development. For example, the ESPResSo software package for particle-based simulations is being developed at the institute. Further method development is taking place in connection with the use of machine learning in particle-based simulations. For performance measurement in HPC environments, frequently changing instrumentation has to be used for performance measurement, also frequent software and driver updates for the GPUs are necessary to evaluate the benefit of new functionalities in GPUs. Furthermore, we plan to automatically measure the speed of simulations in HPC environments to detect performance regressions early in the software development process. These simulations and tests need dynamic changes and are relatively small in terms of resource requirements. On larger systems they would cause a disproportionate effort and thus disruption of operations. Having the requested system located directly at the institute will offer us the necessary flexibility. With the processing of more complex and multi-scale questions, the coupling of methods gains importance. Here, for example, grid-based methods such as Lattice-Boltzmann for hydrodynamics as well as machine learning techniques for the approximation of quantum mechanical influences should be mentioned. We use these methods in combination with particle-based simulations. Their numerical solution can be greatly accelerated with graphics cards. Therefore, the computational nodes will be equipped with four GPUs each.

• DFG Programme: Major Research Instrumentation
• Major Instrumentation: Rechencluster
• Instrumentation Group: 7040 Vektorrechner
• Applicant Institution: Universität Stuttgart
• Leader: Professor Dr. Christian Holm