Abstract
Optimal partitioning of structured adaptive mesh applications necessitates dynamically determining and optimizing for the most time-inhibiting factor, such as load imbalance and communication volume. However, any trivial monitoring of an application evaluates the current partitioning rather than the inherent properties of the grid hierarchy. We present an analytical model that given a structured adaptive grid determines, ab initio, to what extent the partitioner should focus on optimizing load imbalance or communication volume to reduce execution time. This model contributes to the meta-partitioner, able to select and configure the optimal partitioner based on the mesh configuration, the simulation and computer characteristics. We validate the predictions of this model by comparing them with actual measurements (via traces) from four different adaptive simulations. The results show that the proposed model generally captures the inherent optimization-need in structured adaptive mesh refinement applications. We conclude that our model is a useful contribution, since tracking and adapting to the dynamic behavior of such applications potentially lead to a large decrease in execution times.
Get full access to this article
View all access options for this article.