Improving Vectorization of a Crash worthiness Code

Crashworthiness programs to simulate frontal impact of a full vehicle structure into a rigid barrier can require 20 to 30 CPU hours on a single processor of a CRAY X-MP. Any techniques which significantly diminish that time enable engineers to produce economical crash simulations which can, in turn, reduce the necessary physical crashes which often can cost between $50,000 and $750,000 per test on a prototype vehicle. The focus of this paper is on a technique which can improve the vectorization of crashworthiness codes and thereby cut the CPU time for full car crash simulations as well as for structural subassembly problems. The technique involves sorting shell and solid elements by rearranging them by material properties first into separate sets and then by rearranging each material set into subsets of disconnected elements. By so doing, certain Fortran DO loops which dominate CPU activity can be vectorized and thus speed up code execution. Examples are given with an explicit nonlinear finite element code, DYNA3D, on a CRAY X-MP.