Persson, Per. Parallel Numerical Procedures for the Solution of Contact-Impact Problems. Linköping Studies in Sci. Tech. Dissertation No. 620, 2000
Finite element (FE) simulations of complex problems tend to yield very large models if detailed knowledge of the problems is required. As the model becomes larger, the time needed for solving the problem increases. The solution time is even longer if contact is present in the model. In order to achieve results within a reasonable time frame, High Performance Computers (HPC) are used. Current HPC are often of a parallel type, and to be utilised efficiently they need parallel algorithms.
In this thesis an efficient parallel contact algorithm for explicit FE simulations is presented. The contact searching is based on a hierarchical concept where also velocity information about the objects is considered. An augmented Lagrange-multiplier method is used to enforce the impenetrability constraints. The parallel algorithm is based on the so-called HITA-DENA algorithm and it is implemented for use in the general FE program mpp/LS-DYNA.
The parallelisation of an FE method using explicit time integration comes mainly from a domain decomposition of the problem. As the solution is carried out, each processor has to communicate certain information to other processors. The objective of most domain decomposition methods is to reduce the fictitious interior border length, i.e. the cut size, We have shown that this criterion is not necessarily the best for a general problem when contact is present in the calculations. Contacts may easily destroy the communication pattern, leading to an increased cost for communication, and decreased performance of the algorithm. In order to increase the efficiency, two separate domain decompositions are proposed: A static decomposition is constructed to give a good load balance for most element-related calculations, while a dynamic decomposition includes only contact information. The redecompsition of the domain for contact calculations is performed regularly with a newly developed parallel Recursive Coordinate Bisection (pRCB) method.
The presented algorithms are general, they can handle complex contact situations, and they have shown good efficiency on several problems.
Page responsible: Bo Torstenfelt
Last updated: 2008-02-07