Separated domains in load balanding

[shahab.g...@gmail.com]

Dear all,

I am using load balancing and I noticed after load balancing, that the cells owned by each processor are sometimes separated from each other. In other words, some processors may own cell domains that are not connected to each other.

As this increases the computational cost in my case, I was wondering whether it would be possible to limit the load balancing to define only adjacent cells?

Thank you for your helps in advance.

Best regards,

Shahab

[Wolfgang Bangerth]

On 9/14/20 7:29 PM, shahab.g...@gmail.com wrote:
> I am using load balancing and I noticed after load balancing, that the cells
> owned by each processor are sometimes separated from each other. In other
> words, some processors may own cell domains that are not connected to each other.
> As this increases the computational cost in my case, I was wondering whether
> it would be possible to limit the load balancing to define only adjacent cells?

Not with parallel::distributed::Triangulation. That class uses a partitioning
algorithm that optimizes for the data structures used in storing
triangulations, sometimes at the expense of creating these kinds of
disconnected sub-domains. In practice, however, this has relatively little
effect on the performance of programs to the best of our knowledge: Yes, it is
not *optimal*, but it is good enough to not be a major problem in most cases.
You state that it increases the computational cost -- that's true, but do you
have evidence that that creates a bottleneck?

If you do need a different partitioning algorithm, you can use
parallel::shared::Triangulation or, since deal.II 9.2, the
parallel::fullydistributed::Triangulation class.

Best
W.

--
------------------------------------------------------------------------
Wolfgang Bangerth email: bang...@colostate.edu
www: http://www.math.colostate.edu/~bangerth/

[Jean-Paul Pelteret]

Ah, I was just busy writing what Wolfgang said. You can find more info on custom partitions for parallel::shared::Triangulation here:

https://dealii.org/current/doxygen/deal.II/classparallel_1_1shared_1_1Triangulation.html#a686a3453dfec098eb64d1510aa1716e1

You might also be able to “encourage” another partitioning using cell weights, but this was more aimed at load balancing in an hp-FEM context:

https://dealii.org/current/doxygen/deal.II/classparallel_1_1CellWeights.html

https://dealii.org/current/doxygen/deal.II/structTriangulation_1_1Signals.html#af58294e40c64257c9de55e78b6443e36

I hope that this helps,

Jean-Paul

--
The deal.II project is located at http://www.dealii.org/
For mailing list/forum options, see https://groups.google.com/d/forum/dealii?hl=en
--- You received this message because you are subscribed to the Google Groups "deal.II User Group" group.
To unsubscribe from this group and stop receiving emails from it, send an email to dealii+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/dealii/e8d17b06-a23c-c00a-1d8c-c3cc11464bb4%40colostate.edu.