Physics aware simulations, often arising in problems of fluid dynamics, aerodynamics and multi-physics areas have demanded the need for computing software that has the capability to resolve the complexities at multiple scales to analyze and visualize the effects of their interactions with the surroundings. Usually the governing dynamics of these phenomenon appear in the form of complex partial differential equations whose numerical solutions impose various constraints on computational complexity, programming time and efficient throughput. In this scenario, the need of computing software that can solve very large problems resolving the physics of these phenomenon at multiple scales is imperative. Despite traditional computing capabilities in today's hardware through massively parallel systems, optimization and tuning of legacy physics code are usually constrained to specific super-computing clusters and often fail to reproduce similar efficiencies on others. With the dawn of heterogeneous computing systems equipped with accelerators, optimized code that is portable on different systems with varying architectures is a necessity. Such code exploits the advantages of specific hardware capabilities and scales sufficiently for very large and highly nonlinear problems. In this context Sandia National Labs' Trilinos and Kokkos libraries with inherently optimized parallelism and performance portability layers provide a suitable abstraction to build APIs (application programming interfaces) that can model complex physics at multiple scales with a very high degree of fidelity while scaling on massively parallel computers and heterogeneous computing architectures requiring little to no modification of source code. This thesis discusses the development of two such APIs that have been built to solve a range of different fluid dynamics problems and demonstrate the physics that they can simulate. At the same time the different performance metrics obtained from testing these APIs on different supercomputing platforms have been discussed.
展开▼
机译：通常在流体动力学，空气动力学和多物理场问题中出现的，具有物理意识的模拟，要求对计算机软件的需求，该软件能够解决多种尺度的复杂性，以分析和可视化其与周围环境相互作用的影响。通常，这些现象的控制动力学以复杂的偏微分方程的形式出现，其数值解对计算复杂度，编程时间和有效吞吐量施加了各种约束。在这种情况下，迫切需要能够解决非常大的问题的计算机软件，以多种尺度解决这些现象的物理问题。尽管通过大规模并行系统在当今硬件中具有传统的计算功能，但对传统物理代码的优化和调整通常仅限于特定的超级计算集群，并且往往无法在其他集群上重现类似的效率。随着配备了加速器的异构计算系统的兴起，有必要在不同体系结构的不同系统上移植优化代码。这样的代码利用了特定硬件功能的优势，并可以针对非常大且高度非线性的问题进行充分扩展。在这种情况下，Sandia National Labs的Trilinos和Kokkos库具有固有优化的并行性和性能可移植性层，提供了合适的抽象来构建API（应用程序编程接口），该API可以以很高的保真度在多个规模上建模复杂的物理模型，同时可以大规模扩展并行计算机和异构计算体系结构，几乎不需要修改源代码。本文讨论了两种这样的API的开发，这些API旨在解决一系列不同的流体动力学问题，并演示了它们可以模拟的物理原理。同时，还讨论了通过在不同的超级计算平台上测试这些API而获得的不同性能指标。
展开▼
