| Author | Message | ||
     aloisdanek Posted From: 216.52.7.226 |
I have an interesting observation regarding experiments on how to iteratively improve numbering for the fill reduction. The idea works as follows: For number_of_trials repeat: call metisnodend. renumber adjacency(either using perm or iperm) End For My intuitive guess was, that such iterative scheme would come up with numbering which has lower fill than the previous one. To my surprise it came up with 10 times more fill (After 5 trials on 3d cube model with 20000 hexes). Similar behavior was repeating on a small model with 4 quads (9 nodes). The fill after 1 trial was about 1/2 bigger than the previous one. The graph adjacency renumbering seems to work fine. (tested both visually on small example, as well as with the GraphCheck). So I wonder, whether you observed something similar in the past (i.e. numbering can not be iteratively improved), or there is something wrong in my reasoning/scheme? |