Integrated Parallel Simulations and Visualization for Large-scale Weather Applications
Critical applications like cyclone tracking and earthquake modeling require high performance simulations involving large-scale computations. Accurate and timely prediction of such severe weather phenomena requires high fidelity simulations of multiple finer regions of interest within a coarse simulation domain. Such high fidelity simulations are compute intensive and generate huge amounts of data. Hence, it is necessary to perform faster simulations and simultaneous online data analysis and visualization, which can enable scientists to provide real-time guidance to decision makers. In this thesis, we have developed strategies and framework for faster simulations and efficient online data analysis and on-the-fly visualization.
A key challenge in weather simulations is identifying important regions of interest and simulating them in greater detail. We introduce parallel data analysis techniques that can detect important regions of interest in an ongoing simulation and enhance the quality of the simulation. We have developed processor allocation and reallocation strategies for nested simulations that improves performance of dynamic regions of interest. Next, we have built an adaptive framework to enable smooth simulation and continuous visualization in resource-constrained environments. The framework also incorporates the steering inputs of the scientists and reconciles between user-driven steering and automatic tuning. Further, we have proposed heuristics for selection of representative subset of frames for efficient online visualization.
This talk will focus on processor allocation and reallocation strategies, and adaptive framework for online visualization for weather applications. First, we will present a parallel data analysis algorithm for detecting regions of interest in an ongoing simulation. We will introduce a strategy for parallel execution of multiple nested domain simulations based on partitioning the 2D process grid into disjoint rectangular regions associated with each domain. The strategy consists of a novel combination of performance prediction, processor allocation and reallocation methods, and topology-aware mapping of the regions on torus interconnects.
Experiments show 33% performance improvement over the default sequential strategy. Second, we will present an adaptive framework that simultaneously performs smooth simulations and continuous online visualization even in resource-constrained environments. High simulation rates on modern-day processors combined with high I/O bandwidth can lead to rapid accumulation of data at the simulation site and eventual stalling of simulations. We formulate this as an optimization problem to determine optimal execution parameters for enabling smooth simulation and visualization. Our optimization method provides about 30% higher simulation rate and consumes about 25-50% lesser storage space than a naive greedy approach.