Accelerating the FlowSimulator: Performance Analysis of Finite Element Methods on High-Performance Computers

Abstract

Finite element methods have a crucial role in many engineering applications. With the utilization of high-performance computing, the scope of realizable models has expanded exponentially, while maintaining reasonable time-to-solution expectations. In the context of this study, we dive into the performance analysis of two distinct applications the (i) structural mechanics with the b2000++pro software, and (ii) mesh deformation with the FSMeshDeformation plugin from the FlowSimulator universe. Our investigation begins with the extraction of traces from code runs using Score-P. Traces are subsequently subjected to comprehensive scrutiny using visualization tools such as Vampir and CUBE. To further refine our analysis, we strategically employ LIKWID to examine the most time-intensive segments of the code. Obtained results are then translated into the Roofline model. The investigation reveals that the performance of the analyzed hot-loops is noticeably lower than the theoretical Roofline performance. This finding signifies that a noteworthy portion of the total runtime is neither limited by floating-point performance nor available memory bandwidth. Options to close this performance gap are discussed in this contribution.