On Monday, August 8, 2016 at 9:38:20 PM UTC+2, Daniel Jodlbauer wrote:
>
> - Debug is enabled (at least for dealii, I will have to rebuild trilinos 
> with debug later)
>
i mean deal.II. No need to recompile trilinos.
 

> - I am not sure if I got you correctly, but If I use a regular 
> Triangulation, then every rank owns all dofs and finally the initialization 
> of the distributed vectors fails (as expected)
>
 
you just need to use it carefully. For both parallel::shared::Tria and 
general Tria all mpi cores own the complete triangulation. Same for DoFs, 
each cell on each core 
have info about local DoFs. The difference between the two tria's is that 
internally p::s::Tria calls Metis to partition shared triangulation into 
number of processes and 
then calculates locally owned and relevant dofs. But you can also do it 
manually with normal Triangulation.

see https://www.dealii.org/developer/doxygen/deal.II/step_17.html which 
does exactly this whereas
https://www.dealii.org/developer/doxygen/deal.II/step_18.html does 
p::s::Tria.

 

>
> What I additionally tried is: (with 2 ranks)
>
> 1) assemble the rhs / matrix in serial,
> 2) create a partition by hand: [0, n/2), [n/2, n)
> 3) copy/distribute
> 3) solve in parallel
>
> which works.
>
> However, when I changed the partition into something like 
>  { 0, 2, 4, ... } , { 1, 3, 5, ... } 
> it fails, which makes me believe that non-contiguous partitions are not 
> (completely) supported by dealii or trilinos
>

should be fine for Trilinos and deal.II. 

Regards,
Denis.

-- 
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 [email protected].
For more options, visit https://groups.google.com/d/optout.

Reply via email to