Publications Details
Reducing data migration in the context of adaptive partitioning for AMR
Parallel adaptive mesh refinement methods potentially lead to realistic modeling of complex three-dimensional physical phenomena. However, they also present significant challenges in data partitioning and load balancing. As the mesh adapts to the solution, the partitioning requirements change. By explicitly considering these dynamic conditions, the scalability for large, realistic simulations could possibly be significantly improved. Our hypothesis is that adaptive partitioning, meaning dynamic and automatic switching of partitioning techniques, based on the current run-time state, can be beneficial for these simulations. However, switching partitioners can be expensive due to differences in the algorithms' native mapping of data onto processors. We suggest forcing a uniform starting point for all included partitioners. We present a penalty-based method for determining whether switching is beneficial. We study the effects on data migration, as well as on overall cost, of using the uniform starting point and the switching-penalties to select the best partitioning algorithm, among a set of graph-based and geometric partitioning algorithms, for each adaptive time-step for four different adaptive scientific applications. The results show that data migration can be significantly reduced and that adaptive partitioning indeed can be effective for unstructured adaptive applications.