To link your code to NOX using the epetra interface, you must write a concrete class derived from the NOX::Epetra::Interface::Required object. This object defines the method for evaluating the set of nonlinear equations to be solved:

computeF(const Epetra_Vector& x, Epetra_Vector& RHS, const FillType flag) - Computes the set of nonlinear equations, , to be solved by NOX. This method must be supplied by the user.

- x - solution vector specified from NOX.
- RHS - the Epetra_Vector to be filled with the values that correspond to the input solution vector x.
- flag - enumerated type (see NOX::Epetra::FillType) that tells a users interface why computeF() was called. NOX has the ability to generate Jacobians based on numerical differencing using calls to computeF(). In this case, the user may want to compute an inexact (and hopefully cheaper) since it is only used in the Jacobian (or preconditioner).

Additonally if the user wishes to provide a Jacobian operator, they will need to implement a concrete class derived from the NOX::Epetra::Interface::Jacobian object. This object defined the method for evaluating the Jacobian:

computeJacobian(const Epetra_Vector& x, Epetra_Operator& Jac) - this is an optional method that the user can implement if they wish to supply their own evaluation of the Jacobian. If the user does not wish to supply their own Jacobian, they should implement this method so that it throws an error if it is called. This method should update the Jac operator so that subsequent Epetra_Operator::Apply() calls on that operator correspond to the Jacobian at the current solution vector x.

- x - solution vector specified from NOX.
- Jac - a reference to the Jacobian operator that the user supplied in the NOX::Epetra::Group constructor.

Additonally if the user wishes to provide a preconditioner operator, they will need to implement a concrete class derived from the NOX::Epetra::Interface::Preconditioner object. This object defined the method for evaluating the Preconditioner or a Matrix to be used with an interal preconditioner:

computePreconditioner(const Epetra_Vector& x, Epetra_Operator& M, Teuchos::ParameterList* precParams) - This method allows a user to supply their own preconditioner. The method should compute a preconditioner based upon the solution vector x and store it in the Epetra_Operator M. Subsequent calls to the Epetra_Operator::Apply method will apply this user supplied preconditioner to epetra vectors. The Epetra_Operator M can also be an Epetra_RowMatrix that can be used by internally constructed preconditioner objects such as AztecOO, Ipfack, and ML.

- x - solution vector specified from NOX.
- M - a reference to the operator that is to be filled. This operator should be updated to correspond to the current solution vector x.
- precParams - a pointer to a parameter list that can be used by the preconditioner.

- Jacobian-Free Newton-Krylov Operator (see NOX::Epetra::MatrixFree)
- Finite Difference Operator (see NOX::Epetra::FiniteDifference)
- Colored Finite Difference Operator (see NOX::Epetra::FiniteDifferenceColoring)
- Broyden Operator (see NOX::Epetra::BroydenOperator)

These operators are built automatically by the linear system object depending on the constuctor used.

- The concrete Epetra implementations for the NOX Group and Vector are in a separate library from the nox solver algorithms. To build the library for nox epetra support use the flag --enable-nox-epetra in NOX's configure script. This will generate a separate library called libnoxepetra.a

- In addition to the NOX headers, be sure to include the NOX Epetra support specific headers in your interface:
// NOX headers for an epetra interface #include "NOX.H" #include "NOX_Epetra.H"

- Multivector Support. NOX has an implementation of multivector support that is typically used for solving multiple RHS vectors in an efficeint manner. See NOX::Abstract::Multivector for more information.
- Scaling. NOX supports a scaling object that allows users to specify diagonal scaling matrices as Epetra_Vectors that will scale the linear system using either right or left scaling. See NOX::Epetra::Scaling for more information.
- Arbitrary vector space. The NOX algorithms were written to support arbitrary vectors spaces. This allows users to define scaled norms and dot products and have them used consistently throughout the code. See NOX::Epetra::VectorSpace for more details.
- Model evaluator interface. NOX supports a concrete implementation of the NOX::Epetra::Interface objects that wraps an EpetraExt::ModelEvaluator. If a code implements their residual evaluation, Jacobian, or preconditioner via the EpetraExt::ModelEvaluator class, then it can be automatically used by the NOX's Epetra Group. See NOX::Epetra::ModelEvaluatorInterface for more details.
- The NOX::Epetra::Group implementation allows users to define their own linear solver via the NOX::Epetra::LinearSystem pure virtual class. NOX comes with the concrete NOX::Epetra::LinearSystemAztecOO class the uses iterative Krylov subspace methods. See NOX::Epetra::LinearSystem for more details.

The directory Trilinos/packages/nox/test/epetra/1Dfem contains many examples and is the best resource on how to use NOX with Epetra.

Generated on Wed May 12 21:41:53 2010 for NOX by 1.4.7