Performance Evaluation Using LS-DYNA Hybrid Version on the K computer

In order to improve the accuracy of the car crash analysis, the number of elements in the analytical model has been increasing rapidly. A large-scale analysis using model with 10 million elements is slowly becoming popular. Some Companies actually has been adapting large models in their crash analysis nowadays. ́7KH K computer ́> @ D KLJKO\ SDUDOOHO V\VWHP FDQ FDUU\ RXW D FDU FUDVK analysis in several hours which other supercomputers would have taken several days to complete the analysis in the past. However in order to achieve such efficiency, the analytical jobs have to meet following conditions: use LS-DYNA Hybrid version; deploy Groupable contact function of LS-DYNA; and reduce contact definitions as much as possible. In this paper, we investigated the performance and behaviour of LS-DYNA Hybrid version using VHYHUDO WKRXVDQG SURFHVVHV RQ 37KH K computer ́. More specifically, due to the critical role of contact on the performance in a highly parallel system, we mainly focus on the following two aspects of contact calculation part throughout the discussion: the relationship of computational time with the number of contact definitions; and the effectiveness of Groupable contact. This paper gives brief descriptions DERXW 37KH K computer ́ RI 5LNHQ DQG /6-DYNA Hybrid version used in this study in Sections 2 and 3. Several issues that have been encountered when carrying out crash analysis on such highly parallel computing environment are discussed in section 4. The performance bottleneck and factors that hurt scalability are investigated in the next section using simplified model to reveal the effect on the performance with respect to various contact patterns and different number of contact definitions. The effects of Groupable contact function of LS-DYNA versus different contact patterns are studied in this section, too. In the end it is concluded that Groupable contact functions of LS-DYNA help to improve performance of crash analysis on highly parallel environment. A modification of the model to reduce number of contact definitions can boost the performance of LS-DYNA on the K computer.