Parallel performance of large scale impact simulations on Linux cluster super computer

摘要

Three-dimensional analyses with fine solid meshes have become quite common in large deformation impact problems. As the complexity of the problem increases in size and resolution, it is necessary to utilize a parallel dynamic code on parallel systems. The purpose of this paper is to present the parallel performance of a new FEM impact code tested on self-made Linux cluster supercomputer for high speed impact problems. Almost ideal speed-up is achieved for the fixed and scaled Taylor bar model even up to 256 CPUs, although some inefficiency is still identified in parallel FE calculation resulting from the increased communication overhead. In the case of metal sphere impacting to oblique plate, the overall speed-up increases continuously even up to 128 CPUs. Investigation of elapsed time of each part reveals that most of the inefficiency comes from the load imbalance of contact. The two benchmark problems demonstrate the possibility of large scale three-dimensional dynamic simulations within a reasonable computational time.