Performance analysis & optimization of DLR High-Performance Computing codes

Abstract

The importance of High Performance Computing (HPC) in the field of aircraft design is significantly growing. However, capabilities of current HPC systems are inadequate for scale-resolving simulations of in-flight aircraft. Consequently, coordinated advances in algorithms, hardware, and software are needed. For instance, with the end of Moore’s law, processors evolve towards an ever-increasing core count which raises requirements on the scalability of scientific software. Therefore, this submission provides an in-depth scalability analysis of the DLR flow solvers CODA and Musubi. Performance profiles guide the identification of critical aspects of both codes. The presented analysis reveals potentially sub-optimal communication patterns, proposes possible solutions, and shows first improvements of the codes.