AbstractsMathematics

Scientific simulations offer the potential for accurate solutions of realistic models of complex physical phenomena. These simulations are based on systems of partial differential equations and are playing an increasingly important role in science and engineering. However, the phenomena underlying scientific simulations are inherently multi-phased, have heterogeneous state, and span multiple time and space scales. The resulting dynamism coupled with the spatiotemporal and computational heterogeneity make parallel implementations of these scientific simulations extremely challenging. Key issues that need to be addressed include algorithmic efficiency, load balancing, coordination and performance management, which lead to conflicting objectives and trade-offs at runtime. In cases when analytical approaches are not feasible, this requires an understanding of application and system characteristics, and the impact of dynamism and heterogeneity on simulation performance. The overarching goal of this research is to enable large-scale investigative and exploratory science using high-performance "smart" simulations. The key innovation in this research is a multiobjective approach that provides several distribution, coordination, and adaptation strategies to address dynamism and heterogeneity. The partitioning schemes perform dynamic domain decomposition based on either the application geometry or load characteristics to address spatiotemporal and computational heterogeneity. The synchronization algorithm improves communication overheads by reducing messaging frequency in favor of additional computation, when the application is communication-dominated. A runtime infrastructure integrates these strategies and supports the efficient and scalable execution of parallel scientific simulations. Experimental evaluation of the presented strategies using simulations from several application domains demonstrate improvement in overall performance on large systems. Includes abstract