Author |
Message |
Anonymous Posted From: 12.163.90.10
| Posted on Tuesday, May 24, 2005 - 08:53 am: | |
I am puzzled by the way a finite element mesh is converted to nodes (i.e., xadj, adjncy). For 2D (triangle or quad) or 3D (tet or hex), all nodes of the element are mathematically connected (by the stiffness matrix) whether there is a "physical element edge" between the nodes or not. It therefore seems to me that a "mathematical edge" should exist between all nodes of an element and not just at physical element edges. Randy Stonesifer rbstonesifer@pennswoods.net |
george Posted From: 160.94.179.138
| Posted on Tuesday, May 24, 2005 - 12:11 pm: | |
This is true. However, for partitioning purposes, the graph obtained from the physical edges provide sufficient information to ensure that nodes connected via these "mathematical edges" end up in the same partition (at least most of the time). |
Anonymous Posted From: 12.163.93.223
| Posted on Tuesday, May 24, 2005 - 03:14 pm: | |
So are you saying that it would be more consistent to include "edges" for all node connections due to the element? If so, I will write a routine to create the adjncy info. My computer has only one processor, so I am currently less interested in partitioning than in reducing fill. |
george Posted From: 128.101.189.210
| Posted on Tuesday, May 24, 2005 - 04:09 pm: | |
yes, you may get a benefit by ordering the graph that corresponds to the final matrix that you need to solve. |
alois danek Posted From: 64.174.150.210
| Posted on Wednesday, May 25, 2005 - 09:24 pm: | |
Is there a routine in Metis, which would permute a graph when given vector perm and xadj,adjncy and number vertices? Alois Danek aloisdanek@yahoo.com
|