This work revolves around dynamic workloads on distributed systems. Simulations that require finer resolutions in subdomains of interest induce load imbalances that cannot be rectified with static load-balancing. This work aims to analyze available dynamic loadbalancing strategies of ExaHyPE 2. ExaHyPE 2 is an HPC simulation software that depends on Peano 4, a dynamically adaptive grid traversal framework. We deduced and justified custom load-balancing metrics for ExaHyPE 2. These metrics can be easily tailored for analyzing other applications experiencing MPI+X scaling problems. The metrics are then used for evaluating previously analyzed strategies. During testing, problematic loadbalancing behavior is highlighted. Intra- and inter-rank metrics give insights into load distribution and density. Conclusions about the quality of the load-balancing strategies were drawn with multiple load-balancing algorithms yielding deficient results. Finally, a fine-granular post-processing tool for plotting workload over time is provided that will aid future work with analyzing dynamic load-balancing statistics of ExaHyPE 2 and Peano 4.     «

This work revolves around dynamic workloads on distributed systems. Simulations that require finer resolutions in subdomains of interest induce load imbalances that cannot be rectified with static load-balancing. This work aims to analyze available dynamic loadbalancing strategies of ExaHyPE 2. ExaHyPE 2 is an HPC simulation software that depends on Peano 4, a dynamically adaptive grid traversal framework. We deduced and justified custom load-balancing metrics for ExaHyPE 2. These metrics can be...     »