[Gmsh] Mesh partition - point and line elements

sguzik at arrow.utias.utoronto.ca sguzik at arrow.utias.utoronto.ca
Tue Feb 10 21:47:27 CET 2009


Quoting "B. Lazarov" <bl at p3mc.net>:

> On Friday 06 February 2009 20:34:29 sguzik at arrow.utias.utoronto.ca wrote:
...
>>
>> Currently, Gmsh can only partition the highest-dimension elements in a
>> mesh.  The partition information is then propagated to lower-dimension
>> elements but only those on the extents of the mesh.  For
>> lower-dimension elements in the interior, it is (possibly) ambiguous
>> to which partition they belong so they are not labeled.
>>
>> It is entirely possible to add the lower-dimension elements to the
>> partitioning graph -- however, I imagine this would also require
>> weighting the elements, depending on what is needed.  This
>> functionality is not currently implemented through the Gmsh interface
>> but can be achieved by using Metis or Chaco directly.
>>
>> Boyan, you aren't the first to ask about partitioning lines and
>> points.  If there is some consensus on the partitioning requirements,
>> it probably can be implemented in Gmsh.  How would you need the lines
>> and points partitioned relative to the surface elements?
>>
>> Stephen
>
> I have separate implementation based on metis to partition everything. The
> partitioning is done on the nodal graph. One can mix all type of elements
> solid, surface, line and point elements for creating the nodal connectivity
> graph. Once the partitioning of the nodes is done , partitioning is performed
> over the elements based on the nodes. If an element has nodes which belong to
> domain 1 and some nodes which belong to domain 2 and so on, a decision  based
> on some rule can be made. The simplest rule would be that the element should
> belong to the domain with the smaller number of elements.
>
> Something similar can be implemented in GMSH. I really like the   
> possibility to
> mesh and partion at the same time in the program. Doing it separately opens
> possibilities for more mistakes.
>
> Boyan
>

Right now an element-based graph is used with the intent of minimizing  
the communication between elements (and supporting all elements of the  
same dimension).  But there are separate functions for creating the  
graph and partitioning in Gmsh so it wouldn't be too hard to add an  
option for a vertex-based graph (if anyone feels like getting their  
hands dirty).

Stephen


----------------------------------------------------------------
This message was sent using IMP, the Internet Messaging Program.