Fine-grained parallelization of automated multilevel substructuring method for modal frequency response analysis

In this work, a fine-grained parallel automated multilevel substructuring (AMLS) method is proposed to improve the efficiency of modal frequency response analysis for large-scale structural models. AMLS is a multilevel extension of component mode synthesis (CMS), offering a hierarchical structure that enables extensive opportunities for parallel computation. To leverage this structure, a task-based parallel model is first introduced for the substructure transformation phase. This model breaks the synchronization barrier, allowing parent nodes to execute without waiting for descendant nodes. A directed acyclic graph (DAG) model is then used to enable parallel execution of tasks, further enhancing computational efficiency. Moreover, a priority-based task pool scheduling strategy is employed to optimize task management and execution. In addition, damping matrix and residual vector computations are integrated into the parallelization framework to further enhance overall efficiency. The accuracy and computational efficiency of the proposed parallel AMLS method are verified through a series of numerical experiments. Specifically, the method demonstrates significant performance improvements in both weak and strong scalability tests. Furthermore, the proposed parallel AMLS method is validated through engineering case studies. These results confirm that it is well-suited for large-scale structural analyses with practical engineering applications.