- Home
- Documents
*Node and element resequencing using the Laplacian of a ... Heuristic node and element resequencing...*

prev

next

out of 20

View

217Download

1

Embed Size (px)

INTERNATIONAL JOURNAL FOR NUMERICAL METHODS IN ENGINEERING, VOL. 37, 151 1-1530 (1994)

NODE AND ELEMENT RESEQUENCING USING THE

GENERAL CONCEPTS AND ALGORITHM LAPLACIAN OF A FINITE ELEMENT GRAPH: PART I-

GLAUCIO H. PAULINO.

School of Civil and Environmental Engineering, Cornell University, Ithaca, N Y, 14853, U.S.A.

IVAN F. M. MENEZES'

Department of Civil Engineering, PUC-Rio, Rua MarquCs de Sdo Vicente, 225, 22453, Rio de Janeiro, R.J., Brazil

MARCEL0 GATTASS'

Department of Computer Science, PUC-Rio, Rua MarquQs de Sdo Vicente, 225, 22453, Rio de Janeiro, R.J., Brazil

SUBRATA MUKHERJEE'

Department of Theoretical and Applied Mechanics, Kimball Hall, Cornell University, Ithaca, N Y, 14853. U.S.A.

SUMMARY

A Finite Element Graph (FEG) is defined here as a nodal graph (G), a dual graph (G*) , or a communication graph (G') associated with a generic finite element mesh. The Laplacian matrix ((L(G),L(G*) or L(G')), used for the study of spectral properties of an FEG, is constructed from usual vertex and edge connectivities of a graph. An automatic algorithm, based on spectral properties of an FEG (G, G* or G O ) , is proposed to reorder the nodes and/or elements of the associated finite element mesh. The new algorithm is called Spectral FEG Resequencing (SFR). This algorithm uses global information in the graph, it does not depend on a pseudoperipheral vertex in the resequencing process, and it does not use any kind of level structure of the graph. Moreover, the SFR algorithm is of special advantage in computing environments with vector and parallel processing capabilities.

Nodes or elements in the mesh can be reordered depending on the use of an adequate graph representa- tion associated with the mesh. If G is used, then the nodes in the mesh are properly reordered for achieving profile and wavefront reduction of the finite element stiffness matrix. If either G* or G' is used, then the elements in the mesh are suitably reordered for a finite element frontal solver. A unified approach involving FEGs and finite element concepts is presented. Conclusions are inferred and possible extensions of this research are pointed out.

In Part I1 of this work,' the computational implementation of the SFR algorithm is described and several numerical examples are presented. The examples emphasize important theoretical, numerical and practical aspects of the new resequencing method.

1. INTRODUCTION

Heuristic node and element resequencing techniques in the Finite Element Method (FEM) have been a subject of research for a long time. Recognized books on finite elements, such as Desai and Abel,* Irons and Ahmad3, Bathe: Hughes' and Zienkiewicz and Taylor,6 observe the import- ance of this subject. Duff' has surveyed sparse matrix research until 1976. Everstine' has

Ph.D. Student ' Associate Professor 8 Professor

CCC 0029-598 1/94/09 151 1-20$9.00 0 1994 by John Wiley & Sons, Ltd.

Received 30 March 1993 Revised 2 June 1993

1512 G. H. PAULINO ET AL

presented a review of the resequencing algorithms published until 1978. Chinn et aL9 has written a survey paper emphasizing the bandwidth problem for graphs and matrices, and covering the literature until 1981. George has reviewed resequencing of nodes and elements in the context of automatic mesh generation and finite elements.

Some recent papers have been published on resequencing techniques in the FEM. Shephard et af. have presented a node queue algorithm which directly uses the mesh data structure (finite quadtree and finite octree mesh generators) to reorder nodes or elements in the mesh. The advantage of using the mesh data structure instead of the standard connectivity tables to obtain adjacency information is a substantial reduction of storage requirements. Sloan has presented an algorithm and a FORTRAN program for profile and wavefront reduction of sparse symmetric matrices. Livesley and Sabin have studied numbering procedures to reduce the bandwidth of grids derived from 2-D and 3-D finite element meshes of rectangular form. They have reported that the algorithm by Gibbs et ~ 1 . ~ performs badly on some of these grids, particularly the ones that are approximately cubic in form. They have also presented a new algorithm which generates near-optimal numberings for this class of grids. Kaveh has presented a connectivity co-ordinate system for node and element ordering. In Kavehs paper, four algorithms are presented for selecting a good starting node for nodal numbering of a structure. J.-C. L u o ~ has presented an algorithm for reducing bandwidth and profile of a symmetric and sparse matrix. The idea of Luos algorithm is to decompose a graph associated with the matrix into a group of isolated sets by general level structures. These are used to construct a maximal-depth partitioned structure, each level of which has as equal width as possible. Compared to Gibbs et ~ 1 . ~ algorithm, Luos algorithm is more complicated, and, in general, it produces smaller bandwidths but bigger profiles than the former algorithm. Koo and Lee have developed a profile reduction algorithm based on the frontal ordering scheme and graph theory. They have developed a two-step approach where the finite elements are ordered first by the Cuthill and McKee algorithm and then the nodes are reordered based on the concept of frontal ordering and the adjacency measure of graph theory. An earlier version of the two-step approach for finite element ordering has been presented by Fenves and Law. Medeiros ef have just published an algorithm for profile and wavefront reduction which is based on some ideas of the Gibbs, King and Sloan methods. They have shown that, in the numerical tests, their algorithm has a better performance than the reverse C~thill-Mckee~ Gibb~-King~ and Sloan algorithms.*

In general, the FEM leads to a linear system of coupled algebraic equations of the form

AX = b (1) where A = [ a i j l N x N is a known sparse symmetric positive-definite matrix of order N , b is a known N-vector and the unknown N-vector x is sought. The system of equations (1) can be solved by Gauss elimination (see, for example, References 3,4). To take advantage of the sparsity of the coefficient matrix A requires that the equations be organized in a special order, which depends on the solution scheme (bandwidth, profile or frontal) being utilized. The bandwidth and the profile methods construct the coefficient matrix explicitly, while the frontal methods alter- nates the assembly and elimination processes-the coefficient matrix is never assembled expli- citly. In terms of the FEM, efficiency of a bandwidth or profile solver depends upon the ordering of the nodes, while efficiency of a frontal solver depends upon the ordering of the elements (equations are processed by element numbering). Therefore, in this paper, bandwidth and profile methods are classified as node-based ones, while a frontal method is classified as an elemenr-based one.

*The last three references relate to the computational implementation of the respective algorithms

LAPLACIAN OF FINITE E L E M E N T GRAPHS, PART I 1513

Nodal resequencing algorithms are designed to produce a permutation matrix P = [ p i j ] N N such that each row or column has only one non-null component which is equal to unity. The identity matrix (I) is the simplest permutation matrix. Moreover PT = P-. The problem of reordering the system of linear equations (1) can be viewed as determining a permutation matrix that produces a convenient ordering of the system matrix A by the following transformation:

(PAPT) (Px) = Pb (2)

where the orthogonality of the matrix P has been utilized. Since it is impractical to verify each of the N ! possible sequences associated with a given system matrix A (equation (2)), fast heuristic nodal resequencing algorithms for reducing bandwidth, profile or wavefront have been de- veloped. Moreover, it has been shown that the general nature of optimal resequencing algorithms is NP c ~ m p l e t e . ~ ~ - ~ ~

There is an intrinsic correspondence between a finite element stiffness matrix and the topology of the mesh. These topological properties can be explored by associating graphs with meshes and matrices.23* 2 8 * 29 P ermutation of rows and columns in a matrix corresponds to relabelling the vertices of the associated graph or to resequencing the nodes of the mesh. This approach is suitable for studying nodal resequencing techniques.

Association of graphs with finite elements is also useful to provide an adequate element ordering for a frontal solution A convenient approach consists of associating the elements in the original mesh to the vertices of the graph. Also, the connectivity among adjacent elements of the original mesh (common nodes, sides or faces among elements) is used to determine the edges of the graph. Relabelling the vertices of this graph is equivalent to resequencing the elements of the associated mesh. This approach is suitable for studying element resequencing techniques. In this case, the ordering of the nodal variables may be irrelevant.

In this paper, the terms nodes and elements relate to finite element meshes, and the terms vertices and edges relate to graphs. This terminology has been motivated in the previous paragraphs and it will be used from now on.

Consider a Finite Element Graph (FEG), i.e. a graph associated with a finite element mesh. A resequencing method, based on spectral properties of an FEG, is proposed here to renumber nodes and/or elements of a mesh. Moreover, numerical techniques for treating non-connected FEGs are also presented. The proposed method can handle large and generic (arbitrary) finite element meshes, i.e. meshes without any geometrical or top