A Multi-Layered Finite Element Application and Runtime System for Scalable High-End Computer Architectures

本研究的目标是研究能够充分利用当前Teraflop的高端并行有限元网格生成软件 和下一代千万亿次浮点运算高端架构。 除了对超级计算的直接影响外，这项研究还将对许多（生物）工程应用产生更广泛的影响，这些应用可以从可扩展的现场求解器和Teraflop/ Petaflop超级计算机中受益。 这种软件还可以提供具有挑战性和现实性的基准，这可以帮助理解多层体系结构上的端到端动态应用程序的行为，因此对于性能评估社区和新超级计算机的设计将是非常宝贵的。 和有效率地执行 自适应有限元 高端并行架构上的网格生成和细化。 然而，并行网格生成属于更广泛的一类自适应和不规则的应用程序，这是具有挑战性的和劳动密集型的开发和维护。因此，并行网格生成是安装在并行架构上的最后应用之一。这项研究将通过以下方式解决高端架构的这一问题： 着重于以下目标。为多层并行自适应和不规则应用程序开发运行时系统，可以有效地编排层内和层间数据移动和负载平衡。使用运行时系统来设计一种进化方法，用于开发多层并行网格生成软件，该软件能够从执行的一开始就暴露并发性，并维持多达一个四分之一的并发工作单元。

The goal of this research involves the study of high-end parallel finite element mesh generation software capable to deliver and sustain concurrency which will fully utilize current Teraflop and next generation Petaflop high-end architectures. Other than the direct impact to supercomputing this research will have a broader impact to many (bio-) engineering applications that can benefit from scalable field solvers and Teraflop/ Petaflop supercomputers. Such software can also provide challenging and realistic benchmarks which can help to understand the behavior of end-to-end dynamic applications on multi-layered architectures, and thus will be invaluable for the performance evaluation community and design of new supercomputers.The expected outcome of this research is an effective and efficient implementation of adaptive finite element mesh generation and refinement on high-end parallel architectures. However, parallel mesh generation belongs to the broader class of adaptive and irregular applications which are challenging and labor intensive to develop and maintain. As a result, parallel mesh generation is one of the last applications to be installed on parallel architectures. This research will remedy this problem for high-end architectures by focusing on the following objectives. Develop a runtime system for multi-layered parallel adaptive and irregular applications that can efficiently orchestrate intra- and inter-layer data movement and load balancing. Use the runtime system to design an evolutionary approach for developing multi-layered parallel mesh generation software capable of exposing concurrency from the very beginning of the execution and sustaining up to one quadrillion concurrent work units.