Scholarlink

<p>In this work, a hybrid OpenMP/MPI parallel automated multilevel substructuring (AMLS) method is proposed to efficiently perform modal analysis of large-scale finite element models. The method begins with a static mapping strategy that assigns substructures in the substructure tree to individual MPI processes, enabling scalable distributed computation. Within each process, the transformation of substructures is decomposed into subtasks of varying granularity. Tasks are immediately executed once data dependencies are resolved, maximizing core utilization and minimizing idle time. To further improve scalability, non-blocking MPI communication is introduced to overlap data transfers with computations. The proposed method is evaluated on a four-node HPC cluster with 96 CPU cores. Strong scalability tests show that it achieves up to 23 × speedup when scaling from 1 to 96 cores. Compared with a reference parallel AMLS solver, the total runtime is further reduced by 25% in distributed configurations. Additional comparisons with the SLEPc solver confirm the accuracy and effectiveness of the method in modal analysis.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A hybrid OpenMP/MPI parallel automated multilevel substructuring method for finite element modal analysis

  • 期刊论文

  • 发表日期：

    2026年4月11日

查看全文
  • Guidong Wang,
  • Yujie Wang,
  • Xin Hu,
  • Xiangyang Cui,
  • Xianzhong Yu,
  • Senhai Liu

摘要

In this work, a hybrid OpenMP/MPI parallel automated multilevel substructuring (AMLS) method is proposed to efficiently perform modal analysis of large-scale finite element models. The method begins with a static mapping strategy that assigns substructures in the substructure tree to individual MPI processes, enabling scalable distributed computation. Within each process, the transformation of substructures is decomposed into subtasks of varying granularity. Tasks are immediately executed once data dependencies are resolved, maximizing core utilization and minimizing idle time. To further improve scalability, non-blocking MPI communication is introduced to overlap data transfers with computations. The proposed method is evaluated on a four-node HPC cluster with 96 CPU cores. Strong scalability tests show that it achieves up to 23 × speedup when scaling from 1 to 96 cores. Compared with a reference parallel AMLS solver, the total runtime is further reduced by 25% in distributed configurations. Additional comparisons with the SLEPc solver confirm the accuracy and effectiveness of the method in modal analysis.