Abstract

The efficient execution of numerical simulations with dynamically adaptive mesh refinement (DAMR) belongs to the major challenges in high performance computing (HPC). With simulations demanding for steadily changing grid structures, this imposes efficiency requirements on handling grid structure and managing connectivity data. Large-scale HPC systems furthermore lead to additional requirements such as load balancing and thus data migration on distributed- memory systems, which are non-trivial for simulations running with DAMR.

The first part of this thesis focuses on the optimization and parallelization of simulations with DAMR. Our dynamic grid generation approach is based on the Sierpiński space-filling curve (SFC). We developed a novel and efficient parallel management of the grid structure, simulation data and dynamically changing connectivity information, and introduced the cluster concept for grid partitioning. This cluster-based domain decomposition directly leads to efficient parallelization of DAMR on shared-, distributed- as well as hybrid-memory systems, and further yields optimization methods based on such a clustering.

The second part of this work is on optimization of HPC parallelization models currently assigning compute resources statically during the program start. This yields a perspective for dynamic resource distribution addressing the following issues: First, static resource allocation restricts starting other applications in case of insufficient resources available at program start. Second, changing efficiency of applications with different scalability behaviour is not considered. We solve both issues with a resource manager based on Invasive Computing paradigms, dynamically redistributing resources to applications aiming at higher application throughput and thus efficiency.

For several executions of simulations based on our DAMR, we are now able to redistribute the computation resources dynamically among concurrently running applications on shared-memory systems. With dynamic resource assignment, this results in improved throughput and thus higher efficiency.