Autonomic Management of Partitioners for SAMR Grid Hierarchies

Parallel structured adaptive mesh refinement methods decrease the execution time and memory usage of partial differential equation solvers by adaptively assigning computational resources to regions with large solution errors. These methods result in a dynamic grid hierarchy. To get good parallel performance, the grid hierarchy is frequently re-partitioned and distributed over the processors. Optimally, the partitioner should minimize all performance-inhibiting factors like load imbalance, communication volumes, synchronization delays, and data migration. No single partitioner performs well for all hierarchies and parallel computers. Because the partitioning conditions change during run-time, dynamically selecting a partitioner is non-trivial. In this thesis, we present the Meta-Partitioner: a partitioning framework that autonomously selects, configures, invokes, and evaluates partitioning algorithms during run-time. For the implementation, we use component-based software-engineering. We predict the performance of the candidate partitioning algorithms with historical performance data for grid hierarchies similar to the current hierarchy. We focus the partitioning effort on the most performance-inhibiting factors — the load imbalance and the synchronization delays. At re-partitioning, a user-specified number of partitioning algorithms is selected and invoked. The performance of each partitioning is evaluated during run-time and the best one is selected. The performance of the selected partitioning algorithms was compared both to the average performance of 768 algorithms and the global minimum at each re-partitioning. The results showed huge improvements both for the load imbalance and the synchronization delays. Compared to the average partitioning, the load imbalance was decreased by 28.2%. The synchronization delays were decreased by 21.5%. Compared to the global optimum, the load imbalance was increased by only 11.5%. For the synchronization delays, the increase was 13.6%. Often, the Meta-Partitioner selected the best algorithm among all candidate algorithms