Highly-Scalable Computational Algorithms for Solving Aerostructural FSI Problems on Emerging Parallel Machine Architectures

ABSTRACT: Traditional multidisciplinary solvers function by solving the individual governing equations on computer clusters. In the case of aero-elastic Fluid-Structures Interaction (FSI) simulations, individual solvers have disparate requirements for optimal performance. CFD codes use iterative schemes and distributed memory, while CSD codes, use direct linear-solvers and utilize shared memory to ensure fast execution, large time steps, and minimal communication. Developers face major challenges in selecting linear algebra tools that can support the above conflicting requirements. The existing libraries such as PETSc are"stretched"to the limits due to (i) adopting the hybrid MPI/OpenMP parallelization approach, (ii) design-failure to take advantage of the multicore processors: since it is designed to mainly operate on parallel clusters. The CFDRC-UoM team proposes to develop a highly scalable solver, partitioning and execution algorithms for large scale aeroelastic applications. The Phase-I will concentrate on assessing the scalability and bottleneck issues in CFD, CSD and FSI calculations, domain decomposition granularity, coupling techniques on emerging multicore HPC platforms, while simultaneously investigating techniques like cache memory management, partitioning for multicore nodes and parallel loop optimization. The CFDRC-UoM team proposes creation of a next generation numerical suite to accomplish the above. The actual testing will be performed using both PETSc and the aforementioned suite. This suite development will be a precursor to the actual application of the ideas developed for a"production-quality"CFD/CSD solver in Phase-II. BENEFIT: AFOSR scientists and engineers are involved in improving the scalability of large scale parallel multidisciplinary computational codes on the next generation multicore platforms. This project will develop and deliver partitioning techniques for multicore nodes, execution algorithms, coupling techniques between the individual solvers, main memory management for the multicore systems, cache memory management for faster convergence and other scalability enhancement methods. The ideas developed during this project will help developers and users of parallel multidisciplinary computational codes (like aeroelastic codes, Fluid-Structural Interaction (FSI) codes) in understanding the coupling between the individual solvers, improve the communications taking into account the multicore technology, while achieving the desired levels of scalability.