A Meta-Partitioner for run-time selection and evaluation of multiple partitioning algorithms for SAMR grid hierarchies

Parallel structured adaptive mesh refinement (SAMR) methods increase the efficiency of the numerical solution to partial differential equations. These methods use an adaptive grid hierarchy to dynamically assign computational resources to areas with large solution errors. The grid hierarchy needs to be repeatedly repartitioned and distributed over the processors but no single partitioning algorithm performs well for all hierarchies. This paper presents an extended and improved version of the Meta-Partitioner, a partitioning framework that uses the state of the application to autonomously select, configure, invoke, and evaluate partitioning algorithms during run-time. The performance of the partitioning algorithms are predicted using historical performance data for grid hierarchies similar to the current hierarchy. At each re-partitioning, a user-specified number of partitioning algorithms are selected and invoked. When multiple partitionings are constructed, the performance of each partitioning is evaluated during run-time and the best partitioning is selected. The performance evaluation shows huge improvements for the two most performance-inhibiting factors — the load imbalance and the synchronization delays. On average, the load imbalance is increased by only 11.5% and the synchronization delays by 13.6% compared to the optimal results from 768 different hybrid partitioning algorithms.