[dune-pdelab] Nonoverlapping solver backend for implicit time-steppers

Jared Okiro jared.okiro at st.ovgu.de
Wed Oct 31 15:03:34 CET 2012


On 10/30/2012 03:05 PM, Jared Okiro wrote:
> On 10/30/2012 11:41 AM, Steffen Müthing wrote:
>> Hi Jared,
>>
>> what do you mean by parallelizing the system? I assume you ran this 
>> in parallel,
>> and that's when the problems cropped up?
>>
>> Having said that, your example uses a nonoverlapping solver, not an 
>> overlapping
>> one as per Christian's suggestion. For DG, that won't work just like 
>> that, as the
>> per-processor domains are not coupled in any way (with nonoverlapping 
>> assembly,
>> the processes only share DOFs with codim>  0, those attached to a 
>> Border entity in
>> Dune-speak, and you don't have any DOFs of that type with a DG 
>> method). I'd suggest
>> you switch to an overlapping method like Christian said.
>>
>> Apart from that, you cannot use the nonoverlapping Jacobi 
>> preconditioner here, as
>> the nonoverlapping Jacobi implementation in PDELab currently is not 
>> block-aware
>> and will thus not invert the dense DG blocks on the diagonal, but 
>> only the diagonal
>> of those blocks.
>>
>> Steffen
>>
>> Am 30.10.2012 um 10:25 schrieb Jared Okiro:
>>
>>> On 10/19/2012 10:15 PM, Christian Engwer wrote:
>>>> UG provides a ghost layer, which should be sufficient for overlamming
>>>> parallelization using DG methods, as they are cell-based.
>>>>
>>>> regarding the edge communication and UG you have to be more specific,
>>>> as "but it doesn't work for me" is not sufficient for us to guess what
>>>> the problem might be.
>>>>
>>>> Christian
>>>>
>>>> On Fri, Oct 19, 2012 at 06:27:56PM +0200, Jared Okiro wrote:
>>>>> Dear PDELab Members,
>>>>>
>>>>> I am solving a reaction-diffusion system with more than two
>>>>> equations in 2D using
>>>>> Interior Penalty DG methods on UG grid. The sequential version works
>>>>> well, but if I try the parallel
>>>>> version, I get the error "missing edge communication". I saw
>>>>> sometime that edge communication was
>>>>> implemented in UG but it doesn't work for me.  The linear solver
>>>>> fails from the first iteration.
>>>>> I am still new to DUNE and in particular parallel implementation. I
>>>>> am inquiring if there is any example
>>>>> available for Nonoverlapping solvers with implicit time integrators
>>>>> which I can refer to. If not, then I will appreciate any advice.
>>>>> Thank you.
>>>>>
>>>>> Kind regards,--
>>>>>
>>>>> Jared Ouma Okiro,
>>>>> Institute for Analysis and Numerics,
>>>>> Otto-von- Guericke Universität,
>>>>> Universitätsplatz 02, Gebaude 02 / 115
>>>>> 39106 Magdeburg, Germany
>>>>> Office Tel: +493916712660
>>>>> Fax:        +493916718073
>>>>> Mobile:     +4917657479787
>>>>> Home Tel:   +4939155917376
>>>>>
>>>>>
>>>>> _______________________________________________
>>>>> dune-pdelab mailing list
>>>>> dune-pdelab at dune-project.org
>>>>> http://lists.dune-project.org/mailman/listinfo/dune-pdelab
>>>>>
>>> Dear Christian,
>>> Thank you for the information. I decided to do a test case for a 
>>> scalar problem before parallelizing the system. Attached is the test 
>>> case. In this test case only one time step is solved and one can see 
>>> that
>>> it produces the wrong solution at the processor boundaries.
>>> I would be grateful for any further assistance.
>>> Thank you.
>>>
>>> -- 
>>> Jared Ouma Okiro,
>>> Institute for Analysis and Numerics,
>>> Otto-von- Guericke Universität,
>>> Universitätsplatz 02, Gebaude 02 / 115
>>> 39106 Magdeburg, Germany
>>> Office Tel: +493916712660
>>> Fax:        +493916718073
>>> Mobile:     +4917657479787
>>> Home Tel:   +4939155917376
>>>
>>> <nonlinearproblem.cc><nonlinearproblem.hh>_______________________________________________ 
>>>
>>> dune-pdelab mailing list
>>> dune-pdelab at dune-project.org
>>> http://lists.dune-project.org/mailman/listinfo/dune-pdelab
>> Steffen Müthing
>> Universität Stuttgart
>> Institut für Parallele und Verteilte Systeme
>> Universitätsstr. 38
>> 70569 Stuttgart
>> Tel: +49 711 685 88429
>> Fax: +49 711 685 88340
>> Email: steffen.muething at ipvs.uni-stuttgart.de
>>
>
> Dear Steffen,
> I have tried the overlapping version also. I followed the steps from 
> example06 code in pdelab-howto and
> did it for DG method, but still did not succeed. I have attached the 
> files.
> The linear solver fails to converge the solution.
> I will appreciate further assistance.
> Thank you.
>
Dear Steffen/PDELab Members,

Sorry, I am writing again because I haven't managed to solve the problem 
I asked.
Steffen, as you and Christian advised, I attached the overlapping 
version of my test parallel case.
I find the same strategy to work with FEM but apparently it doesn't work 
with DG. Does  DG works in this way or do I need to incorporate some 
things into the code?. I actually do not understand why DG fails in this 
similar way. I will appreciate some more tips on this.
Thank you.

-- 
Jared Ouma Okiro,
Institute for Analysis and Numerics,
Otto-von- Guericke Universität,
Universitätsplatz 02, Gebaude 02 / 115
39106 Magdeburg, Germany
Office Tel: +493916712660
Fax:        +493916718073
Mobile:     +4917657479787
Home Tel:   +4939155917376

-------------- next part --------------
A non-text attachment was scrubbed...
Name: nonlinearproblem-1.cc
Type: text/x-c++src
Size: 13157 bytes
Desc: not available
URL: <https://lists.dune-project.org/pipermail/dune-pdelab/attachments/20121031/87f6f12a/attachment-0001.cc>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: nonlinearproblem-1.hh
Type: text/x-c++hdr
Size: 43544 bytes
Desc: not available
URL: <https://lists.dune-project.org/pipermail/dune-pdelab/attachments/20121031/87f6f12a/attachment-0001.hh>


More information about the dune-pdelab mailing list