hi,

how to control 'Overlap' in partitioning?

I want that all partitions(partition1, partition2, partition3....) have same nodes.

I have 3 computers, and it has 8GB ram, all same.

and my mesh has about 7000000 nodes.

I want that one computer get 2600000 nodes, how to devide partitions? how to control weight factor?

thanks all.

Is your problem that you have a multi-domain simulation? In that case, in the solver manager go to tab "partitioner" and switch on "coupled partitioning" (as opposed to "independent partitioning").

thanks, Timon

But I try to that method (coupled partioning and independent partitionion also)

my simulation has two domains, and they have long 'interface(boundary condition)'. so the simulation's overlap is 20%(two partitions) to 47%(six partitions). because, If I try to remove 'interface boudary condition', they have few(0.7~3%) overlap.

I give up reducing overlap. but I want devide the partitions same size(same node, so it spands same memory)

thanks

Hi,

Unless the overlap area being reported is degrading performance I would not bother about fiddling with it. Most of the time the parallel performance is still pretty good.

Glenn Horrocks

thanks Glenn

I need parallel.

the mesh has about 7000000 nodes

I have 3 computers of 8GB ram, same cpu

if I will parallel run normally, the mesh didn't devide same. so part 1 use 5GB, part 2 use 10GB, part 3 use 10GB example.

I don't know why same weight factor has defferent partition size, how control partition size,

thanks

Hi,

Read the documentation regarding partitioning. When the solver partitions the grid it gives an increased weight to nodes on interfaces. This can lead to partitions have different sizes. However if the partitions are as different as you suggest either something is wrong or you have lots of interfaces.

Are you using metis partitioning? If you have sufficient memory to run it this is always the best choice. If it runs out of memory you have to look at the other partitioning options but which one is best becomes problem dependant. A poor choice can lead to dissimilar partitions. What partitioning method are you using?

You should also consider the coupled partitioning suggested previously.

Glenn Horrocks

thanks

I use the Metis method. other methods make overlap, more large. like documentation, Metis was the best partitioning method. I find this by trial and error

yes I have a lot of interface, maybe it make the overlap.

I can't find the control partitioning of Metis in documentation. anyone know this?

Hi

metis has very few controls. Really the only option is coupled or uncoupled.

If you have lots of interfaces then poor parallel performance is hard to avoid. If it is possible to reduce the number or size of the interfaces (or make some interfaces 1 to 1 rather than GGIs) then your performance will increase significantly.

But my original point is this affecting the parallel speed of the simulation? You may well be getting a good speedup despite the poor partitioning.

Glenn Horrocks