Hi

Has anyone noticed that CFX doesn't seem to cope with partitioning across General Grid Interfaces (GGIs)? It appears that if half your model, in terms of cells, reside on one side of the GGI, then this will be held in one partition, while the other half of the model is partitioned between your other nodes. This means that one CPU will be swapping and slowing down all the others. I know Fluent can partition across non-conformals, does anyone know how to do it with CFX?

Thanks

CFX does partition across grid interfaces. Did you check the relative size of the partitions in the OUT file?

-Robin

Yes, and the memory allocation for the partitions varies wildly. This is when my three regions which are separated by the GGIs are combined in the same domain. I've got 'round it by having three separate domains and using the Independent Partitioning method during partition. I'm a Fluent user and am used to the number of elements in a partition being the governing factor, not nodes. But even then, the example I was quoting had large variations in the node counts in the partitions, which looked uncannily like the number of nodes on one side of my GGI.

I've seen exactly the same thing. I have a very large part of a model which is hex meshed and separated from the rest of the domain by a GGI's. The Metis partitioner sticks the hex domain in one partition and subdivides up the tet domain into the remaining partitions. It seems to be a foible of the Metis partitioner. The other methods do not do this.

Jeff

Thanks for the confirmation, I used the Bisection Algorithm and things looked better. I also found that having 3 domains (one for each zone either side of the GGIs) partitioned independently with METIS made for a pretty good distribution. I worry about the efficiency of effectively having 3*n partitions distributed around the place. From my point of view, this is another area in which CFX could be a little more thorough in the manual. Sure they've got parallel, sure they've got GGIs - but let's have some discussion about what happens when the two get put together. Large parallel runs routinely have GGIs because of their complexity.

The default partitioning method uses the MeTiS partitioner.

For some domain topologies the MeTiS partitioner will allocate all the nodes in one 3D region to a given partition, and all the nodes in another 3D region to the next one, etc...

One alternative is coupled partitioning, which accounts for the connections through the GGI. It requires more more memory usage and CPU than independent because the GGIs must be intersected but it will get rid of the original problem.

If you have put all your regions inside 1 domain then the other partitioning methods are also possibilites, and they do not require the memory/cpu hit of GGI intersections, but then scalability may be effected.

Point taken about the documentation as well. More is always better. The fully coupled algorithm is still slightly experimental so it's not well documented yet.

Neale