| Author | Message | ||
     Karl W. Schulz Posted From: 129.114.2.131 |
Hello, I have a feeling I may be using Metis for an unintended purpose, but I'm having a problem decomposing a small mesh with vertex weights included. Namely, I tend to get empty processors. Is there a way to have Metis guarantee that all partitions are assigned at least one vertex in the graph? As an example, I'm trying to decompose a 40 element mesh onto 16 processors. The 40 elements represent individual blocks from a multi-block finite-volume code, so I'm adding weight to each element based on the number of cells in each block (eg, nx*ny*nz). If I include the weighting, then I see processors with no assigned elements. I've tried both PartGraphRecursive and PartGraphKway with different combinations of options, but so far I haven't been able to have work assigned to each processor. Thanks in advance for any tips/suggestions. Karl | ||
| george Posted From: 160.94.179.138 |
Karl, You are running into limitations with Metis's heuristics. It is designed to operate on graphs in which each final partition will have a non-trivial number of nodes. My suggestion is to paly around with pmetis/kmetis and to also try hmetis to see it will solve your problem. george |