A Divide and Conquer Strategy for Scaling Weather Simulations with Multiple Regions of Interest

Abstract

Abstract. Accurate and timely prediction of weather phenomena, such as hurricanes and flash floods, require high-fidelity compute intensive simulations of multiple finer regions of interest within a coarse simulation domain. Current weather applications execute these nested simulations sequentially using all the available processors, which is sub-optimal due to their sub-linear scalability. In this work, we present a strategy for parallel execution of multiple nested domain simulations based on partitioning the 2-D processor grid into disjoint rectangular regions associated with each domain. We propose a novel combination of performance prediction, processor allocation methods and topology-aware mapping of the regions on torus interconnects. Experiments on IBM Blue Gene systems using WRF show that the proposed strategies result in performance improvement of up to 33% with topology-oblivious mapping and up to additional 7% with topology-aware mapping over the default sequential strategy.

A Diffusion-Based Processor Reallocation Strategy for Tracking Multiple Dynamically Varying Weather Phenomena

“…However, inserting a new node near a node with large difference in weights will lead to skewed rectangles. As reported in [1], square-like partitions lead to smaller executions times for the nests, while skewed rectangular partition increases the execution time of a nest. This is illustrated in the example shown in Figure 5.…”

“…In a recent work [1], it was shown that significant performance improvements can be achieved by executing the nests simultaneously on different subsets of the total number of processors, P . We use the performance modeling and Huffman tree based algorithm in [1] to determine the size of the subset of processors for a nest and the position of the subset in the processor grid P x × P y where P x·P y = P .…”

“…We use the performance modeling and Huffman tree based algorithm in [1] to determine the size of the subset of processors for a nest and the position of the subset in the processor grid P x × P y where P x·P y = P . The performance model is used to predict the execution times of nests based on the size and aspect ratio of the nests.…”

“…The actual execution times of these 13 domains are used to interpolate the execution times of the nests formed in our simulation using Delaunay triangulation. The details of these steps are presented in [1]. Additionally, we predict the execution times of the nests for the 10 processor sizes.…”

“…Each nested simulation is executed on disjoint subsets of the total number of processors for high performance [1]. Therefore, with dynamic appearance and disappearance of regions of interest, we need to modify the processor allocation for the nests.…”