[Trilinos-Users] [EXTERNAL] slow anasazi in NOX::Epetra

Veltz Romain romain.veltz at inria.fr
Sun Dec 22 08:30:50 MST 2013


Andrew,

Are you saying that shiftedlinsys should take another Epetra_Operator instead of MFJac?

How this new operator would be related to the interface function bool ProblemInterface::applyShiftedMatrix ? 

On Dec 13, 2013, at 12:00 AM, "Salinger, Andrew" <agsalin at sandia.gov> wrote:

> 
> It looks like your shifted linear system is with the Jacobian operator. I think
> it needs to be with the shifted operator. 
> 
> 
> From: Veltz Romain <romain.veltz at inria.fr>
> Date: Wednesday, December 11, 2013 2:38 PM
> To: Andy Salinger <agsalin at sandia.gov>
> Cc: "trilinos-users at software.sandia.gov" <trilinos-users at software.sandia.gov>
> Subject: Re: [Trilinos-Users] [EXTERNAL] slow anasazi in NOX::Epetra
> 
> Andy,
> 
> My mass matrix is the identity matrix. I hope it is not wrong (see below).
> 
> If I use the MatrixFree Jacobian, I get the same results. They are not wrong per se as if the function transformEigenvalue were not working.
> 
> If I use the Shift-Invert settings, I have the following output which insure that the class is called
> 
> WARNING: LOCA::AnasaziOperator::ShiftInvert::apply() - 
> Return Type = NotConverged
> 
> Hence, I have no clue what can be wrong. You suggested that my function ProblemInterface::applyShiftedMatrix is buggy and/or not called but I think it is called because of what is reported above.
> 
> Thank you again for your suggestions,
> 
> Romain
> 
> PS: To help, I copied the part of the code that might relevant. Here are the definitions for the interface/group
> 
> Teuchos::RCP<LOCA::Epetra::Interface::TimeDependentMatrixFree> iReq = interface;
> Teuchos::RCP<FFTJac> MFJac = Teuchos::rcp(new FFTJac(nlPrintParams, interface, *noxSoln, Problem));
> Teuchos::RCP<LOCA::Epetra::Interface::TimeDependentMatrixFree> iReq = interface;
> Teuchos::RCP<NOX::Epetra::Interface::Jacobian> iJac = MFJac;
> // Create the linear systems
> Teuchos::RCP<NOX::Epetra::LinearSystemAztecOO> linsys =
> Teuchos::rcp(new NOX::Epetra::LinearSystemAztecOO(nlPrintParams, lsParams,
> iReq, iJac, MFJac, *soln));
> Teuchos::RCP<NOX::Epetra::LinearSystemAztecOO> shiftedLinsys =
> Teuchos::rcp(new NOX::Epetra::LinearSystemAztecOO(nlPrintParams, lsParams,
> iReq, iJac, MFJac, *soln));
> // access Jacobian from interface
> interface->setJacobianOperator(MFJac);
> 
> // Create the Group
> Teuchos::RCP<LOCA::MultiContinuation::AbstractGroup> grp =
> Teuchos::rcp(newLOCA::Epetra::Group(
> globalData,
> nlPrintParams,
> iReq, // interface
> locaSoln,// initial guess
> linsys,// linear system
> shiftedLinsys,// shifted linear system
> pVector));// parameters vector
> 
> And, in my interface, I have
> bool setJacobianOperator(Teuchos::RCP<Epetra_Operator> Jac)
> {
> Jac_= Jac;
> }
> 
> bool ProblemInterface::applyShiftedMatrix(double alpha, double beta,
> const NOX::Epetra::Vector& input,
> NOX::Epetra::Vector& result) const
> {
> // compute result = alpha * J * input + beta * input
> Epetra_Vector X(input.getEpetraVector());
> Epetra_Vector JX(result.getEpetraVector());
> if (alpha!=0)
> assert(Jac_->Apply(X,JX)==0);
> result.update(alpha,JX,beta,X,0.0);
> returntrue;
> }
> 
> Romain
> 
> On Dec 11, 2013, at 5:40 PM, "Salinger, Andrew" <agsalin at sandia.gov> wrote:
> 
>> 
>> Veltz,
>> 
>> I have two ideas:
>> 
>> The identity matrix, while seemingly simple, has all eigenvalues repeated. This can cause problems for the iterative eigensolver. Can you try again with diagonal entries 1, 1.1, 1.2, 1.3,..
>> 
>> There might be a problem with your shifted/mass matrix calculation, since that is not used in the Jacobian Inverse method, but is used in the ones with a wrong results. You can
>> try some tests of applying it agains simple vectors to make sure it is correct.
>> 
>> Andy
>> 
>> From: Veltz Romain <romain.veltz at inria.fr>
>> Date: Wednesday, December 11, 2013 5:59 AM
>> To: Eric Phipps <etphipp at sandia.gov>
>> Cc: "trilinos-users at software.sandia.gov" <trilinos-users at software.sandia.gov>
>> Subject: Re: [Trilinos-Users] [EXTERNAL] slow anasazi in NOX::Epetra
>> 
>> I thought I would be done with issues but I found the following output from Anasazi rather puzzling.
>> Note that, for some reasons, I am using trillions-10.8.4 (updating to the current version give me some trouble under macosx).
>> 
>> I set my continuation code in a configuration where the Jacobian is the operator: -Id
>> 
>> Using the LOCA with Epetra - Anasazi with:
>> aList.set("Sorting Order", "LM");
>> aList.set("Operator","Jacobian Inverse");
>> 
>> the following EV are returned
>> 
>> Untransformed eigenvalues (since the operator was Jacobian Inverse)
>> Eigenvalue 0 : -1.000e+00  -0.000e+00 i    :  RQresid 3.331e-16  0.000e+00 i
>> Eigenvalue 1 : -1.000e+00  -0.000e+00 i    :  RQresid 6.439e-15  0.000e+00 i
>> Eigenvalue 2 : -1.000e+00  -0.000e+00 i    :  RQresid 7.327e-15  0.000e+00 i
>> Eigenvalue 3 : -1.000e+00  -0.000e+00 i    :  RQresid 5.995e-15  0.000e+00 i
>> 
>> 
>> Using the LOCA with Epetra - Anasazi with:
>> aList.set("Sorting Order", "LM");
>> aList.set("Operator","Shift-Invert");
>> aList.set("Shift",0.001);
>> 
>> the following ev are returned (The RQresid is large too, equal to the shift…):
>> 
>> Untransformed eigenvalues (since the operator was Shift-Invert)
>> Eigenvalue 0 : -9.990e-01  -0.000e+00 i    :  RQresid 1.000e-03  0.000e+00 i
>> Eigenvalue 1 : -9.990e-01  -0.000e+00 i    :  RQresid 1.000e-03  0.000e+00 i
>> 
>> Note that I have checked that the solver effectively enter the following function
>> ->LOCA::AnasaziOperator::ShiftInvert::apply()
>> 
>> I don't understand these results… I thought the entry  in
>> anasaziOp->transformEigenvalue((*evals_r)[i], (*evals_i)[i]);
>> from LOCA::Eigensolver::AnasaziStrategy::computeEigenvalues would do the job…
>> 
>> Please note that I have the same issues with the Cayley transform.
>> 
>> Thank you for your help and advices,
>> 
>> Romain
>> 
>> 
>> On Dec 10, 2013, at 5:33 PM, "Phipps, Eric T" <etphipp at sandia.gov> wrote:
>> 
>>> I don't understand what you mean be "I don't have access to my Jacobian operator".  You need to be able to implement y = (a*J + b*M)*x for a give (multi-) vector x and scalars a and b, where J is the Jacobian operator and M the mass matrix.  If you have a matrix-free operator to implement J*x, you can use the same approach to implement this, e.g., via y = a*J*x + b*M*x.
>>> 
>>> -Eric
>>> 
>>> On Dec 10, 2013, at 5:03 AM, "Veltz Romain" <romain.veltz at inria.fr> wrote:
>> 
> 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: https://software.sandia.gov/pipermail/trilinos-users/attachments/20131222/4aa37d49/attachment.html 


More information about the Trilinos-Users mailing list