Assessment of the Performance of Cluster Computing for the simulation of High-Speed Dynamics Problems with LS-Dyna

Thanks to the increase in computational and storage power it is possible to develop extremely refined and complex FE models. With the development of cluster computing and parallel solver it is also possible to cut computational time without losing accuracy (instead of for instance adopting a coarser yet computational more efficient mesh) . This is particularly important when an explicit solver is adopted for the resolution of dynamic problems - typically related to high strain rate events, such as impacts and detonations – since the Courant condition might led to extremely reduced timesteps. Many workarounds and tricks can be adopted (mass scaling, erosion criteria, cutoffs) at the cost of introducing nonphysical phenomena. However, it is also well known that the solvers and the algorithms might introduce errors that are purely numeric. A study was carried out to assess the effect of several computational variables on high dynamic problems. In the paper two cases are analysed: a simpler penetration test of a metal plate and a more complex simulation of a shockwave tube test involving a concrete plate. The simulations are implemented into the commercial software LS-Dyna and a pipeline is created for the systematic analysis of several scenarios. The considered variables are the software version (13 vs. 16), the number of cores running (1, 2, 4, 8, and 16 cores), the solver (MPP, SMP, and hybrid), and the consistency parameter. Results are then compared in terms of accuracy in comparison with experimental results, CPU running time, and effect on the simulation energy content. The study might be useful for leading users through the vast environment of cluster computing and for understanding the effects of FE analysis parallelization.