Scalability Tuning of the Load Balancing and Coupling Framework FD4

In this paper, we discuss the scalability tuning of the HPC software framework FD4. The framework provides dynamic load balancing and model coupling for multiphase and multiphysics simulations. We first investigate scalability bottlenecks using the Vampir performance analysis tool-set on the BlueGene/Q system at JSC. Then, we describe and evaluate our optimized algorithms: A new hierarchical 1D partitioning algorithm for SFC-based dynamic load balancing and a new method to organize the coupling metadata. The final scalability benchmark shows that the overhead of FD4 has been reduced by a factor of 3.3 at 262 144 ranks, which leads to a considerably increased scalability of the whole application.