[Trilinos-Users] Komplex vs. Epetra_LocalMap

[Heroux, Michael A]

Przemek,

Although what you have done is possible, it is not a valid usage.  Epetra
assumes (it would be possible to check this, but very expensive) that an
object based on a LocalMap will have identical values on each processor.  If
such an object is the output of an operation, Epetra will also perform the
global operations to make sure that values are the same across all
processors.

Mike


On 10/23/10 10:01 AM, "Przemysław Kłosiewicz"
<przemyslaw.klosiewicz at ua.ac.be> wrote:

> Oh, I see. I didn't realize mixing different types of communicators could be a
> valid solution; I'll try that.
> One thing though:
> Suppose I create a vector v with a LocalMap, fill it with
> v.PutScalar(comm.MyPID()) and run it on two mpi nodes.
> If I now do a cout << v the output suggests that I end up having two different
> vectors v: on node 0 it's a v full of zeros and on node 1 v is full of ones;
> thus not identical.
> Am I actually creating two different vectors or is it one and the same vector
> which just happens to be out of sync somehow?
> 
> Thanks,
> Przemek.
> 
> On 23 Oct 2010, at 16:26, Heroux, Michael A wrote:
> 
>> Przemek,
>> 
>> I think you want to create an Epetra_SerialComm object on each processor and
>> then construct regular Epetra_Map objects using that comm object.  This will
>> give you the independence you want.
>> 
>> LocalComm is intended to be used with data that is identical and replicated
>> on each processor.
>> 
>> I hope this helps.
>> 
>> Mike
>> 
>> 
>> On 10/23/10 7:49 AM, "Przemysław Kłosiewicz"
>> <przemyslaw.klosiewicz at ua.ac.be> wrote:
>> 
>>> Dear all
>>> 
>>> I am solving a moderately sized complex linear system Ax = b using Amesos.
>>> In
>>> order to obtain an equivalent real-valued formulation I use Komplex which
>>> needs re(A), im(A), re(b), etc... separately.
>>> As long as I use a regular Epetra_Map for these matrices & vectors all goes
>>> well. However, for some well-defined reason I don't want these matrices to
>>> be
>>> distributed so I tried using an Epetra_LocalMap (or an Epetra_Map with
>>> NumGlobalElements = NumMyElements).
>>> This is where Komplex fails: During the creation of a Komplex_LinearProblem
>>> from matrices created with such maps I get the following exception:
>>> 
>>> /home/pklosiew/trilinos-10.6.0-Source/packages/komplex/src/Komplex_LinearPro
>>> bl
>>> em.cpp:291:
>>> Throw number = 1
>>> Throw test that evaluated to true: KomplexMatrix_->InsertGlobalValues(Row,
>>> NumIndices, Values, Indices)<0
>>> 
>>> This only happens when running with mpi and where the number of processors >
>>> 1.
>>> 
>>> I have the impression Komplex tries to create double-sized distributed maps
>>> from the original map, even though they should stay local.
>>> Is this a bug, a feature or am I misusing the Epetra_LocalMap?
>>> 
>>> Thanks,
>>> Przemek.
>>> 
>>> 
>>> (ps: the reason why I don't want Ax=b to be distributed is that I need to
>>> solve these systems for a range of parameters p, as in: A(p)x = b(p) which
>>> can
>>> all be done in parallel with minimal communication.)
>>> 
>>> 
>>> --
>>> Przemyslaw Klosiewicz
>>> PhD Student @ CoMP Research Group
>>> Department of Mathematics and Computer Science, University of Antwerp
>>> Middelheimlaan 1, 2020 Antwerp, Belgium
>>> 
>>> Office (G2.07):  T (+32) 3 265 32 80, F (+32) 3 265 37 77, E
>>> Przemyslaw.Klosiewicz at ua.ac.be
>>> Home: T (+32) 497 24 18 82, E przemek.klosiewicz at gmail.com, Skype
>>> przemek_klosiewicz
>>> 
>>> 
>>> 
>>> 
>>> 
>>> 
>>> _______________________________________________
>>> Trilinos-Users mailing list
>>> Trilinos-Users at software.sandia.gov
>>> http://software.sandia.gov/mailman/listinfo/trilinos-users
>>> 
>> 
>> 
> 
> --
> Przemyslaw Klosiewicz
> PhD Student @ CoMP Research Group
> Department of Mathematics and Computer Science, University of Antwerp
> Middelheimlaan 1, 2020 Antwerp, Belgium
> 
> Office (G2.07):  T (+32) 3 265 32 80, F (+32) 3 265 37 77, E
> Przemyslaw.Klosiewicz at ua.ac.be
> Home: T (+32) 497 24 18 82, E przemek.klosiewicz at gmail.com, Skype
> przemek_klosiewicz
> 
> 
> 
> 
> 
> 
> _______________________________________________
> Trilinos-Users mailing list
> Trilinos-Users at software.sandia.gov
> http://software.sandia.gov/mailman/listinfo/trilinos-users