Direct Or Iterative ? A Dilemma For The User

The subject of this paper is the comparison of Direct and Iterative solvers in the Solution of large Finite Element Problems with MSC/NASTRAN. The results of such a comparison vary based on problem type and with new developments in both the direct and iterative solution techniques. This paper gives the current state of this comparison with MSC/NASTRAN V70.5. Experiences gained in the past show that analysts are hesitant to use the iterative solver because the direct sparse solver has been in MSC/NASTRAN from the beginning, is very robust and has gone through several optimization phases. MSC's direct sparse solver has very low memory requirements, a very efficient out-of-core logic and very good re-ordering techniques. The iterative solver, on the other hand, was first delivered only early in this decade. It has since been improved significantly by adding more and better preconditioning techniques, the latest one being BIC (Block Incomplete Cholesky) preconditioning which was introduced in V69 . Many clients started to use it for 3-D models and have reported positive results. More recently, in MSC/NASTRAN V70.5 the memory management and spill logic for the BIC preconditioner of the iterative solver have been improved significantly, particularly for jobs where the solution cannot be done completely in-core. We will briefly discuss the methods used for the direct and iterative solvers in MSC/NASTRAN, the effects of new re-ordering schemes in the direct solver and the effects of the BIC preconditioning technique in the iterative solver. Moreover, we will present performance results for large models involving different types of elements to give guidelines on where the advantages of each solver lie and when is it more beneficial to use one solver over the other. For example, for models built from 2D elements it is always recommended to use the direct solver while for models of 3D elements the iterative solver should be preferred. Other issues like memory and disk space requirements and multiple load cases will also be discussed. The focus of the paper is on linear static analysis.