Projecting Performance of LS-DYNA Implicit for Large Multiprocessor Systems

We use a benchmark dataset of a jet-engine impeller to study the performance characteristics of the LS-DYNA Implicit solver, as a function model size and the number of processors available in the system. We analyze a range of models with an increasing number of fans in the jet-engine impeller, from the smallest model of three fans composed of 300,000 nodes to the largest model of ten fans with 1 Million nodes. The resulting stiffness matrices are of sizes .9 and 3 Million Degrees of Freedom (DOF) respectively. These sizes are typical for today's LS-DYNA runs, but in the coming five years, model sizes will to grow upwards of 50 Million DOF. In addition, the computational system available at that time will offer much higher degree of parallelism. In this paper we estimate the performance and scalability of large LS-DYNA runs on these future machines.