Dynamic Load Balancing for Adaptive Mesh Refinement Applications: Improvements and Sensitivity Analysis

Adaptive Mesh Refinement (AMR) is a type of multiscale algorithm that dynamically achieves high resolution in localized regions of multidimensional numerical simulations. A dynamic load balancing(DLB) scheme for structured AMR applications was proposed in [19]. Unfortunately, the overhead introduced by this DLB scheme is significant. Further, a parameter called threshold is used in this scheme which determines whether a load balancing process should be invoked; its value directly influence the efficiency of the overall DLB scheme. In this paper, we first present two improvements on this DLB scheme to reduce its overhead. Then a detailed sensitivity analysis is provided to identify an optimal value for the parameter threshold. Experiments show that by interleaving grid splitting with direct grid movement and by employing nonblocking communication, the execution time can be improved by up to 32% and the overhead can be reduced as much as 18%; and setting threshold to be around 1:25 is optimal.