Nonlocal and Fractional Order Methods for Near-wall Turbulence, Large-eddy Simulation, and Fluid-structure Interaction
Abstract
Major Goals: Task 1: We propose a nonlocal modeling technique that expresses the spatial derivatives of the solution as a convolution with a suitably chosen kernel that may be defined independently of the underlying discretization. Because this convolution is essentially a filtering operation, we see a natural connection with large- eddy simulation (LES) and a possibility to define a new class of LES models where one has control over both the discretization and filter size. Nonlocal modeling may be a pathway to achieving an optimal combination of explicit and implicit LES with further improvements in per-degree-of-freedom accuracy, which will bring LES one step closer to being an industrially relevant computational tool. Task 2: Our efforts will build upon the experience of the PIÕs in both developing and using massively parallel computational codes as well as other research codes. Computational libraries built to efficiently exploit distributing memory parallel computing will be utilized and modified to take full advantage of the newer shared memory accelerators, these include graphics processing units (GPUs) capable of general purpose numerics as well as the Many Integrated Core (MIC) coprocessor of Intel. These devices offer tremendous potential for performance and efficiency in the large-scale simulations that will be undertaken in the proposed research. How-ever, exploiting this potential can be challenging, as one must adapt to the specialized and rapidly evolving application programming interfaces (APIs) of these new architectures while maintaining compatibility with more traditional and wide spread forms of parallelism (e.g. distributed memory message passing). These accelerators offer a wide range of APIs to exploit their wide vector width and single-instruction, multiple data architecture, which include OpenMP, OpenCL, OpenACC, and CUDA (for Nvidia GPUs). These APIs will be used in conjunction with advanced compilers and traditional message passing (for distributed memory computing) to generate the fastest algorithms that will be applicable for computations that range to several million degrees of freedom. It will also be important to have a long-term vision for the code/algorithm development as well, looking towards exascale computing in the next decade. One approach that will likely be a general way to keep-pace with the complexity and every-changing APIs that exploit hybrid architecture is though utilization of the Trilinos Kokkos package. Kokkos is a uniform API for both linear and nonlinear solvers that utilizes specific kernels for many arbitrary accelerators (e.g. MIC, threads, CUDA cores) that are identified at compile time. Codes that are written with Trilinos/Kokkos are highly portable to different hybrid architectures. The co-PIÕs have experience with Trilinos/Kokkos in developing application codes and this knowledge will be utilized in the code development efforts undertaken in the projects.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Oct 30, 2018
- Source ID
- W911NF1510552
Entities
People
- John E. Foster
Organizations
- Army Contracting Command
- United States Army
- University of Texas at Austin