Advances in Performance-Portable Solver Strategies for Large-Scale Structural Dynamics on Emerging Heterogeneous Supercomputers
Please login to view abstract download link
Sierra Structural Dynamics (SD) is a mature, massively-parallel finite element application developed for a wide range of structural dynamics analyses, including modal, random vibration, shock response, and coupled multi-physics simulations. Over the past several years, the emergence of heterogeneous exascale platforms has motivated a systematic rearchitecting of the code and reinvention of solver strategies to achieve performance portability across diverse hardware from many vendors. New challenges presented by engineers utilizing SD have also highlighted performance bottlenecks and thus new opportunities to improve core code methods and algorithms. This talk describes recent efforts to improve SD's capabilities in modal (eigenvalue) analysis as well as linear solver technology for structural analysis in a performance-portable manner, leveraging packages and abstractions based on Kokkos and Trilinos [1, 2], allowing a single code base to run efficiently on a variety of GPU architectures while employing optimized solver algorithms and settings for each. The presentation will focus on the performance and scalability of solver architecture improvements, demonstrating the advantages of implementing portable algorithms and methods as a practical path for legacy applications to remain performant on current and future heterogeneous architectures. SNL is managed and operated by NTESS under DOE NNSA contract DE-NA0003525. REFERENCES [1] Trilinos Project Team. The Trilinos Project Website, 2026. https://trilinos.github.io. [2] C. Trott et al. The Kokkos ecosystem: Comprehensive performance portability for high performance computing. Computing in Science and Engineering, 23(5):10–18, 2021. doi:10.1109/MCSE.2021.3098509.
