当前位置: X-MOL 学术Finite Elem. Anal. Des. › 论文详情
Our official English website, www.x-mol.net, welcomes your feedback! (Note: you will need to create a separate account there.)
A high-performance multiscale space-time approach to high cycle fatigue simulation based on hybrid CPU/GPU computing
Finite Elements in Analysis and Design ( IF 3.5 ) Pub Date : 2019-11-01 , DOI: 10.1016/j.finel.2019.103320
Rui Zhang , Sam Naboulsi , Thomas Eason , Dong Qian

Abstract A multiscale space/time computational framework for high cycle fatigue (HCF) life predictions is established by integrating the extended space-time finite element method (XTFEM) with a multiscale progressive damage model. While the robustness of the multiscale space/time method has been previously demonstrated, the associated high computational cost remains a critical barrier for practical applications. In this work, a novel hybrid iterative/direct linear system solver is first proposed with a unique preconditioner. Computational efficiency is further improved by taking advantage of the high-performance computing platform featuring hierarchy of the distributed- and the shared-memory parallelisms using CPUs and GPUs. Robustness of the accelerated framework is demonstrated through benchmark problems. It is shown that the serial version of the hybrid solver is at least 1–2 orders of magnitude faster in computing time and cheaper in memory consumption than the conventional sparse direct or iterative solver, while the parallel version efficiently handles XTFEM stiffness matrix equations with over 100 million unknowns using 64 CPU cores. Optimal speedups are achieved in the parallel implementations of the multiscale progressive damage model using either CPUs or GPUs. HCF simulations on 3D specimens are performed to quantify key effects due to mean stress and multiaxial load conditions.

中文翻译:

基于混合CPU/GPU计算的高性能多尺度时空高周疲劳仿真方法

摘要 通过将扩展时空有限元法(XTFEM)与多尺度渐进损伤模型相结合,建立了用于高周疲劳(HCF)寿命预测的多尺度时空计算框架。虽然之前已经证明了多尺度空间/时间方法的稳健性,但相关的高计算成本仍然是实际应用的关键障碍。在这项工作中,首先提出了一种具有独特预处理器的新型混合迭代/直接线性系统求解器。通过利用具有使用 CPU 和 GPU 的分布式和共享内存并行性层次结构的高性能计算平台,进一步提高了计算效率。加速框架的稳健性通过基准问题得到了证明。结果表明,与传统的稀疏直接或迭代求解器相比,混合求解器的串行版本在计算时间上至少快 1-2 个数量级,内存消耗更便宜,而并行版本有效地处理 XTFEM 刚度矩阵方程。使用 64 个 CPU 内核的 1 亿个未知数。在使用 CPU 或 GPU 的多尺度渐进式损伤模型的并行实现中实现了最佳加速。对 3D 试样进行 HCF 模拟以量化平均应力和多轴载荷条件造成的关键影响。而并行版本使用 64 个 CPU 内核有效地处理具有超过 1 亿个未知数的 XTFEM 刚度矩阵方程。在使用 CPU 或 GPU 的多尺度渐进式损伤模型的并行实现中实现了最佳加速。对 3D 试样进行 HCF 模拟以量化平均应力和多轴载荷条件造成的关键影响。而并行版本使用 64 个 CPU 内核有效地处理具有超过 1 亿个未知数的 XTFEM 刚度矩阵方程。在使用 CPU 或 GPU 的多尺度渐进式损伤模型的并行实现中实现了最佳加速。对 3D 试样进行 HCF 模拟以量化平均应力和多轴载荷条件造成的关键影响。
更新日期:2019-11-01
down
wechat
bug