Highly-Scalable Computational-Based Engineering Algorithms for Emerging Parallel Machine Architectures

ABSTRACT: RNET and The Ohio State University propose to use algorithmic modifications and multi-level parallelization techniques and tools to improve the scalability of the CFD/CSD codes relevant to the DoD/AF (e.g., CREATE/Kestrel). The linear system solver is expected to be a primary bottleneck in aero-line/aero-elastic CFD and CSD codes. RNET and its subcontractors will explore effective parallel linear solvers targeted at multi-core nodes and scaling to 10's of thousands of compute cores across a multi-tiered compute architecture. Specifically, in Phase I, the performance and scalability of using the proposed algorithmic improvements and parallelization tools will be demonstrated on existing codes; e.g., NASTRAN and NASA's USM3d. In Phase I, the primary goal will be to evaluate the performance of existing linear solvers (e.g., PETSc) and to explore high performance implementation of the most suitable linear solver(s). Ultimately, in Phase II, a highly-scalable version of the linear solver will be developed and integrated into the tools used in the Kestrel toolkit. BENEFIT: The scalability advancements will help allow the CREATE/Kestrel project to efficiently reach its goals of providing rapid aeronautic simulations to early phase design and acquisition projects. The scalability improvements will enable the effective usage of the DoD HPC modernization systems for the Kestrel users, allowing large scale simulations to achieve the required rapid turnaround times. In addition, the algorithms, tools, and techniques developed in this project will enable similar scalability improvements in other applications, especially other CFD/CSD applications of interest to the DoD and industry.