The Performance of Car Crash Simulation by LS-DYNA Hybrid Parallel Version on Fujitsu FX1

The number of elements of car simulation for crash analysis has increased rapidly over recent years in order to achieve better accuracy. Several MPP versions of LS-DYNA using variations of MPI have been widely applied to car crash simulation for better job turnaround. The modern computing hardware has been exploring multi-core CPU technology since the performance of single-core can not meet the current and future demands. The Hybrid Parallelization version of LS-DYNA, taking the advantage of multi-core computing platforms, provides an efficient tool for large- model car crash simulation. This hybrid version combines threads parallelization using OpenMP across multiple cores within a node as well as process parallelization using MPI between nodes to achieve maximum parallelization of the analysis. In this paper, we revealed performance bottle neck of MPP version for highly parallelization first; then we showed performance and examined accuracy of solution of LS-DYNA Hybrid version. We tested both 2-million and 10-million elements Caravan crash models on both LS-DYNA MPP and Hybrid versions. The superior performance of Hybrid version demonstrated the feasibility of car crash simulation using massively parallel processors.