static int run_test(Teuchos::RCP<Epetra_CrsMatrix> matrix,
	  bool verbose,           // display the graph before & after
	  bool contract,          // set global number of partitions to 1/2 num procs
	  int partitioningType,   // hypergraph or graph partitioning, or simple
	  int vertexWeightType,   // use vertex weights?
	  int edgeWeightType,     // use edge/hyperedge weights?
	  int objectType)         // use isorropia's CrsMatrix or CrsGraph
{
  int rc=0, fail = 0;
#ifdef HAVE_EPETRAEXT
  int localProc = 0;
  double balance1, balance2, cutn1, cutn2, cutl1, cutl2;
  double balance3, cutn3, cutl3;
  double cutWgt1, cutWgt2, cutWgt3;
  int numCuts1, numCuts2, numCuts3, valid;
  int numPartitions = 0;
  int keepDenseEdges = 0;
  int numProcs = 1;

#ifdef HAVE_MPI
  const Epetra_MpiComm &Comm = dynamic_cast<const Epetra_MpiComm &>(matrix->Comm());
  localProc = Comm.MyPID();
  numProcs = Comm.NumProc();
#else
  const Epetra_SerialComm &Comm = dynamic_cast<const Epetra_SerialComm &>(matrix->Comm());
#endif

  int numRows = matrix->NumGlobalRows();

  if (numRows < (numProcs * 100)){
    // By default Zoltan throws out dense edges, defined as those
    // whose number of non-zeros exceeds 25% of the number of vertices.
    //
    // If dense edges are thrown out of a small matrix, there may be nothing left.
    keepDenseEdges = 1;
  }

  double myShareBefore = 1.0 / numProcs;
  double myShare = myShareBefore;

  if (contract){
    numPartitions = numProcs / 2;

    if (numPartitions > numRows)
      numPartitions = numRows;

    if (numPartitions > 0){
      if (localProc < numPartitions){
	myShare = 1.0 / numPartitions;
      }
      else{
	myShare = 0.0;
      }
    }
    else{
      contract = 0;
    }
  }

  // If we want Zoltan's or Isorropia's default weights, then we don't
  // need to supply a CostDescriber object to createBalancedCopy,
  // so we get to test the API functions that don't take a CostDescriber.

  bool noCosts = ((vertexWeightType == NO_APPLICATION_SUPPLIED_WEIGHTS) &&
		   (edgeWeightType == NO_APPLICATION_SUPPLIED_WEIGHTS));

  // Test the interface that has no parameters, if possible

  bool noParams =
    ((partitioningType == HYPERGRAPH_PARTITIONING) && // default, so requires no params
     (numPartitions == 0) &&                          // >0 would require a parameter
     (keepDenseEdges == 0));                          // >0 would require a parameter

  // Maps for original object
  const Epetra_Map &sourceRowMap = matrix->RowMap();
  const Epetra_Map &sourceRangeMap = matrix->RangeMap();
//   const Epetra_Map &sourceColMap = matrix->ColMap();
  const Epetra_Map &sourceDomainMap = matrix->DomainMap();

  int numCols = matrix->NumGlobalCols();
  int nMyRows = sourceRowMap.NumMyElements();
  int base = sourceRowMap.IndexBase();

  // Compute vertex and edge weights

  Isorropia::Epetra::CostDescriber costs;

  Teuchos::RCP<Epetra_Vector> vptr;

  Teuchos::RCP<Epetra_CrsMatrix> eptr;

  Teuchos::RCP<Epetra_Vector> hyperEdgeWeights;

  if (edgeWeightType != NO_APPLICATION_SUPPLIED_WEIGHTS){

    if (partitioningType == GRAPH_PARTITIONING){

      // Create graph edge weights.

      eptr = Teuchos::rcp(new Epetra_CrsMatrix(*matrix));

      if (vertexWeightType == SUPPLY_EQUAL_WEIGHTS){
	eptr->PutScalar(1.0);   // set all nonzeros to 1.0
      }
      else{
	int maxRowSize = eptr->MaxNumEntries();
	double *newVal = NULL;
	if (maxRowSize > 0){
	  newVal = new double [maxRowSize];
	  for (int j=0; j<maxRowSize; j++){
	    newVal[j] = localProc + 1 + j;
	  }
	}
	int numEntries;
	int *idx;
	double *val;
	for (int i=0; i<nMyRows; i++){
	  rc = eptr->ExtractMyRowView(i, numEntries, val, idx);
	  for (int j=0; j<numEntries; j++){
	    val[j] = newVal[j];
	  }
	}
	if (newVal) delete [] newVal;
      }

      eptr->FillComplete(sourceDomainMap, sourceRangeMap);

      costs.setGraphEdgeWeights(eptr);
    }
    else{
      // Create hyperedge weights.  (Note that the list of hyperedges that a
      // process provides weights for has no relation to the columns
      // that it has non-zeroes for, or the rows that is has.  Hypergraphs
      // in general are not square.  Also more than one process can provide
      // a weight for the same edge.  Zoltan combines the weights according
      // to the value of the PHG_EDGE_WEIGHT_OPERATION parameter.  The default
      // for this parameter is to use the maximum edge weight provided by any
      // process for a given hyperedge.)

      Epetra_Map hyperEdgeMap(numCols, base, Comm);

      hyperEdgeWeights = Teuchos::rcp(new Epetra_Vector(hyperEdgeMap));

      int *edgeGIDs = NULL;
      double *weights = NULL;
      int numHEweights = hyperEdgeMap.NumMyElements();

      if (numHEweights){
	edgeGIDs = new int [numHEweights];
	weights = new double [numHEweights];

	if (edgeWeightType == SUPPLY_EQUAL_WEIGHTS){
	  for (int i=0; i<numHEweights; i++){
	    edgeGIDs[i] = hyperEdgeMap.GID(i);
	    weights[i] = 1.0;
	  }
	}
	else{
	  int hiVolumeStart = matrix->NumGlobalCols() / 3;
	  int hiVolumeEnd = hiVolumeStart * 2;
	  for (int i=0; i<numHEweights; i++){
	    edgeGIDs[i] = hyperEdgeMap.GID(i);
	    if ((edgeGIDs[i] < hiVolumeStart) || (edgeGIDs[i] >= hiVolumeEnd)){
	      weights[i] = 1.0;
	    }
	    else{
	      weights[i] = 3.0;
	    }
	  }
	}
	hyperEdgeWeights->ReplaceGlobalValues(numHEweights, weights, edgeGIDs);
      }

      if (weights){
	delete [] weights;
	delete [] edgeGIDs;
      }

      costs.setHypergraphEdgeWeights(hyperEdgeWeights);
    }
  }

  bool need_importer = false;

  if ((vertexWeightType != NO_APPLICATION_SUPPLIED_WEIGHTS)){

    need_importer = true;  // to redistribute row weights

    double *val = NULL;

    if (nMyRows){
      val = new double [nMyRows];

      if (vertexWeightType == SUPPLY_EQUAL_WEIGHTS){
	for (int i=0; i<nMyRows; i++){
	  val[i] = 1.0;
	}
      }
      else if (vertexWeightType == SUPPLY_UNEQUAL_WEIGHTS){
	for (int i=0; i<nMyRows; i++){
	  val[i] = 1.0 + ((localProc+1) / 2);
	}
      }
    }

    vptr = Teuchos::rcp(new Epetra_Vector(Copy, sourceRowMap, val));

    if (val) delete [] val;

    costs.setVertexWeights(vptr);
  }

  // Calculate partition quality metrics before calling Zoltan

  if (partitioningType == GRAPH_PARTITIONING){
    rc = ispatest::compute_graph_metrics(matrix->Graph(), costs,
	     myShare, balance1, numCuts1, cutWgt1, cutn1, cutl1);
    if (contract){
      // balance wrt target of balancing weight over *all* procs
      rc = ispatest::compute_graph_metrics(matrix->Graph(), costs,
	     myShareBefore, balance3, numCuts3, cutWgt3, cutn3, cutl3);
    }
  }
  else{
    rc = ispatest::compute_hypergraph_metrics(matrix->Graph(), costs,
	     myShare, balance1, cutn1, cutl1);
    if (contract){
      // balance wrt target of balancing weight over *all* procs
      rc = ispatest::compute_hypergraph_metrics(matrix->Graph(), costs,
	     myShareBefore, balance3, cutn3, cutl3);
    }
  }

  if (rc){
    ERROREXIT((localProc==0), "Error in computing partitioning metrics")
  }

  Teuchos::ParameterList params;

#ifdef HAVE_ISORROPIA_ZOLTAN

  if (!noParams){

    // We're using Zoltan for partitioning and supplying
    // parameters, overriding defaults.

    Teuchos::ParameterList &sublist = params.sublist("Zoltan");

    if (partitioningType == GRAPH_PARTITIONING){
      params.set("PARTITIONING METHOD", "GRAPH");
      sublist.set("GRAPH_PACKAGE", "PHG");
    }
    else{
      params.set("PARTITIONING METHOD", "HYPERGRAPH");
      sublist.set("LB_APPROACH", "PARTITION");
      sublist.set("PHG_CUT_OBJECTIVE", "CONNECTIVITY");  // "cutl"
    }

    if (keepDenseEdges){
      // only throw out rows that have no zeroes, default is to
      // throw out if .25 or more of the columns are non-zero
      sublist.set("PHG_EDGE_SIZE_THRESHOLD", "1.0");
    }
     if (numPartitions > 0){
	// test #Partitions < #Processes
	std::ostringstream os;
	os << numPartitions;
	std::string s = os.str();
	//	sublist.set("NUM_GLOBAL_PARTS", s);
	params.set("NUM PARTS", s);
      }

      //sublist.set("DEBUG_LEVEL", "1"); // Zoltan will print out parameters
      //sublist.set("DEBUG_LEVEL", "5");   // proc 0 will trace Zoltan calls
      //sublist.set("DEBUG_MEMORY", "2");  // Zoltan will trace alloc & free
  }

#else
    ERROREXIT((localProc==0),
      "Zoltan partitioning required but Zoltan not available.")
#endif

  // Function scope values

  Teuchos::RCP<Epetra_Vector> newvwgts;
  Teuchos::RCP<Epetra_CrsMatrix> newewgts;

  // Function scope values required for LinearProblem

  Epetra_LinearProblem *problem = NULL;
  Epetra_Map *LHSmap = NULL;
  Epetra_MultiVector *RHS = NULL;
  Epetra_MultiVector *LHS = NULL;

  // Reference counted pointer to balanced object

  Epetra_CrsMatrix *matrixPtr=NULL;
  Epetra_CrsGraph *graphPtr=NULL;
  Epetra_RowMatrix *rowMatrixPtr=NULL;
  Epetra_LinearProblem *problemPtr=NULL;

  // Row map for balanced object
  const Epetra_BlockMap *targetBlockRowMap=NULL;  // for input CrsGraph
  const Epetra_Map *targetRowMap=NULL;            // for all other inputs

  // Column map for balanced object
  const Epetra_BlockMap *targetBlockColMap=NULL;  // for input CrsGraph
  const Epetra_Map *targetColMap=NULL;            // for all other inputs

  if (objectType == EPETRA_CRSMATRIX){
    if (noParams && noCosts){
      matrixPtr = Isorropia::Epetra::createBalancedCopy(*matrix);
    }
    else if (noCosts){
      matrixPtr = Isorropia::Epetra::createBalancedCopy(*matrix, params);
    }
    targetRowMap = &(matrixPtr->RowMap());
    targetColMap = &(matrixPtr->ColMap());
  }
  else if (objectType == EPETRA_CRSGRAPH){
    const Epetra_CrsGraph graph = matrix->Graph();
    if (noParams && noCosts){
      graphPtr = Isorropia::Epetra::createBalancedCopy(graph);
    }
    else if (noCosts){
      graphPtr = Isorropia::Epetra::createBalancedCopy(graph, params);
    }
    targetBlockRowMap = &(graphPtr->RowMap());
    targetBlockColMap = &(graphPtr->ColMap());
  }
  else if (objectType == EPETRA_ROWMATRIX){
    if (noParams && noCosts){
      rowMatrixPtr = Isorropia::Epetra::createBalancedCopy(*matrix);
    }
    else if (noCosts){
      rowMatrixPtr = Isorropia::Epetra::createBalancedCopy(*matrix, params);
    }
    targetRowMap = &(rowMatrixPtr->RowMatrixRowMap());
    targetColMap = &(rowMatrixPtr->RowMatrixColMap());
  }
  else if (objectType == EPETRA_LINEARPROBLEM){

    // Create a linear problem with this matrix.

    LHSmap = new Epetra_Map(numCols, base, Comm);

    int myRHSsize = sourceRowMap.NumMyElements();
    int myLHSsize = LHSmap->NumMyElements();

    int valSize = ((myRHSsize > myLHSsize) ? myRHSsize : myLHSsize);

    double *vals = NULL;

    if (valSize){
      vals = new double [valSize];
    }

    if (valSize){
      for (int i=0; i < valSize; i++){
	// put my rank in my portion of LHS and my portion of RHS
	vals[i] = localProc;
      }
    }

    RHS = new Epetra_MultiVector(Copy, sourceRowMap, vals, 1, 1);

    LHS = new Epetra_MultiVector(Copy, *LHSmap, vals, 1, 1);

    if (valSize){
      delete [] vals;
    }

    problem = new Epetra_LinearProblem(matrix.get(), LHS, RHS);

    Epetra_LinearProblem lp = *problem;

    if (lp.CheckInput()){
      ERROREXIT((localProc==0), "Error creating a LinearProblem");
    }
    if (noParams && noCosts){
      problemPtr = Isorropia::Epetra::createBalancedCopy(lp);
    }
    else if (noCosts){
      problemPtr = Isorropia::Epetra::createBalancedCopy(lp, params);
    }

    targetRowMap = &(problemPtr->GetMatrix()->RowMatrixRowMap());
    targetColMap = &(problemPtr->GetMatrix()->RowMatrixColMap());
  }

  // Redistribute the edge weights
  // Comment this out since we don't redistribute columns

  if (edgeWeightType != NO_APPLICATION_SUPPLIED_WEIGHTS){

    if (partitioningType == GRAPH_PARTITIONING){

      Epetra_Import *importer = NULL;

      if (objectType == EPETRA_CRSGRAPH){
	newewgts = Teuchos::rcp(new Epetra_CrsMatrix(Copy, *graphPtr));
	targetRowMap = &(newewgts->RowMap());
	targetColMap = &(newewgts->ColMap());
      }
      else{
	newewgts = Teuchos::rcp(new Epetra_CrsMatrix(Copy, *targetRowMap, *targetColMap, 0));
      }

      importer = new Epetra_Import(*targetRowMap, sourceRowMap);
      newewgts->Import(*eptr, *importer, Insert);
      newewgts->FillComplete(*targetColMap, *targetRowMap);

      costs.setGraphEdgeWeights(newewgts);
    }
  }

  // Redistribute the vertex weights

  if ((vertexWeightType != NO_APPLICATION_SUPPLIED_WEIGHTS)){

    Epetra_Import *importer = NULL;

    if (objectType == EPETRA_CRSGRAPH){
      newvwgts = Teuchos::rcp(new Epetra_Vector(*targetBlockRowMap));
      importer = new Epetra_Import(*targetBlockRowMap, sourceRowMap);
    }
    else{
      newvwgts = Teuchos::rcp(new Epetra_Vector(*targetRowMap));
      importer = new Epetra_Import(*targetRowMap, sourceRowMap);
    }

    newvwgts->Import(*vptr, *importer, Insert);
    costs.setVertexWeights(newvwgts);
  }

  if (localProc == 0){
    test_type(numPartitions, partitioningType, vertexWeightType, edgeWeightType, objectType);
  }

  if (verbose){

    // Picture of problem before balancing

    if (objectType == EPETRA_LINEARPROBLEM){

      ispatest::show_matrix("Before load balancing", *problem, Comm);
    }
    else{
      ispatest::show_matrix("Before load balancing", matrix->Graph(), Comm);
    }

    // Picture of problem after balancing

    if (objectType == EPETRA_LINEARPROBLEM){
      ispatest::show_matrix("After load balancing (x in Ax=b is not redistributed)", *problemPtr, Comm);
    }
    else if (objectType == EPETRA_ROWMATRIX){
      ispatest::show_matrix("After load balancing", *rowMatrixPtr, Comm);
    }
    else if (objectType == EPETRA_CRSMATRIX){
      ispatest::show_matrix("After load balancing", matrixPtr->Graph(), Comm);
    }
    else if (objectType == EPETRA_CRSGRAPH){
      ispatest::show_matrix("After load balancing", *graphPtr, Comm);
    }
  }

  // After partitioning, recompute the metrics

  if (partitioningType == GRAPH_PARTITIONING){
    if (objectType == EPETRA_LINEARPROBLEM){
      rc = ispatest::compute_graph_metrics(*(problemPtr->GetMatrix()), costs,
	     myShare, balance2, numCuts2, cutWgt2, cutn2, cutl2);
    }
    else if (objectType == EPETRA_ROWMATRIX){
      rc = ispatest::compute_graph_metrics(*rowMatrixPtr, costs,
	     myShare, balance2, numCuts2, cutWgt2, cutn2, cutl2);
    }
    else if (objectType == EPETRA_CRSMATRIX){
      rc = ispatest::compute_graph_metrics(matrixPtr->Graph(), costs,
	     myShare, balance2, numCuts2, cutWgt2, cutn2, cutl2);
    }
    else {
      rc = ispatest::compute_graph_metrics(*graphPtr, costs,
	     myShare, balance2, numCuts2, cutWgt2, cutn2, cutl2);
    }
  }
  else{
    if (objectType == EPETRA_LINEARPROBLEM){
      rc = ispatest::compute_hypergraph_metrics(*(problemPtr->GetMatrix()), costs,
	     myShare, balance2, cutn2, cutl2);
    }
    else if (objectType == EPETRA_ROWMATRIX){
      rc = ispatest::compute_hypergraph_metrics(*rowMatrixPtr, costs,
	     myShare, balance2, cutn2, cutl2);
    }
    else if (objectType == EPETRA_CRSMATRIX){
      rc = ispatest::compute_hypergraph_metrics(matrixPtr->Graph(), costs,
	     myShare, balance2, cutn2, cutl2);
    }
    else{
      rc = ispatest::compute_hypergraph_metrics(*graphPtr, costs,
	     myShare, balance2, cutn2, cutl2);
    }
  }

  if (rc){
    ERROREXIT((localProc==0), "Error in computing partitioning metrics")
  }

  std::string why;

  if (partitioningType == GRAPH_PARTITIONING){
    fail = (cutWgt2 > cutWgt1);
    why = "New weighted edge cuts are worse";

    if (localProc == 0){
      std::cout << "Before partitioning: Balance " << balance1 ;
      std::cout << " cutn " << cutn1 ;
      std::cout << " cutl " << cutl1 ;

      if (contract){
	std::cout << "  (wrt balancing over " << numPartitions << " partitions)" << std::endl;
	std::cout << "Before partitioning: Balance " << balance3 ;
	std::cout << " cutn " << cutn3 ;
	std::cout << " cutl " << cutl3 ;
	std::cout << "  (wrt balancing over " << numProcs << " partitions)" ;
      }
      std::cout << std::endl;

      std::cout << " Total edge cuts: " << numCuts1;
      std::cout << " Total weighted edge cuts: " << cutWgt1 << std::endl;
      std::cout << "After partitioning: Balance " << balance2 ;
      std::cout << " cutn " << cutn2 ;
      std::cout << " cutl " << cutl2 << std::endl;
      std::cout << " Total edge cuts: " << numCuts2;
      std::cout << " Total weighted edge cuts: " << cutWgt2 << std::endl;
    }
  }
  else{
      fail = (cutl2 > cutl1);
      why = "New cutl is worse";

    if (localProc == 0){
      std::cout << "Before partitioning: Balance " << balance1 ;
      std::cout << " cutn " << cutn1 ;
      std::cout << " cutl " << cutl1 ;
      if (contract){
	std::cout << "  (wrt balancing over " << numPartitions << " partitions)" << std::endl;
	std::cout << "Before partitioning: Balance " << balance3 ;
	std::cout << " cutn " << cutn3 ;
	std::cout << " cutl " << cutl3 ;
	std::cout << "  (wrt balancing over " << numProcs << " partitions)" ;
      }
      std::cout << std::endl;
      std::cout << "After partitioning: Balance " << balance2 ;
      std::cout << " cutn " << cutn2 ;
      std::cout << " cutl " << cutl2 << std::endl;
    }
  }

  if (fail){
    if (localProc == 0) std::cout << "ERROR: "+why << std::endl;
  }

  // Check that input matrix is valid.  This test constructs an "x"
  // with the matrix->DomainMap() and a "y" with matrix->RangeMap()
  // and then calculates y = Ax.

  if (objectType == EPETRA_LINEARPROBLEM){
    valid = ispatest::test_matrix_vector_multiply(*problemPtr);
  }
  else if (objectType == EPETRA_ROWMATRIX){
    valid = ispatest::test_row_matrix_vector_multiply(*rowMatrixPtr);
  }
  else if (objectType == EPETRA_CRSMATRIX){
    valid = ispatest::test_matrix_vector_multiply(*matrixPtr);
  }
  else{
    valid = ispatest::test_matrix_vector_multiply(*graphPtr);
  }

  if (!valid){
    if (localProc == 0) std::cout << "Rebalanced matrix is not a valid Epetra matrix" << std::endl;
    fail = 1;
  }
  else{
    if (localProc == 0) std::cout << "Rebalanced matrix is a valid Epetra matrix" << std::endl;
  }

  if (localProc == 0)
    std::cout << std::endl;



#else
  std::cout << "test_simple main: currently can only test "
	 << "with Epetra and EpetraExt enabled." << std::endl;
  rc = -1;
#endif

  return fail;
}
Example #2
0
int main(int argc, char *argv[]) 
{
  // Initialize MPI
#ifdef HAVE_MPI
  MPI_Init(&argc,&argv);
#endif

  // Create a communicator for Epetra objects
#ifdef HAVE_MPI
  Epetra_MpiComm Comm( MPI_COMM_WORLD );
#else
  Epetra_SerialComm Comm;
#endif
 
  bool verbose = false;

  if (argc > 1)
    if (argv[1][0]=='-' && argv[1][1]=='v')
      verbose = true;

  // Get the process ID and the total number of processors
  int MyPID = Comm.MyPID();
#ifdef HAVE_MPI
  int NumProc = Comm.NumProc();
#endif

  // define the parameters of the nonlinear PDE problem
  int nx = 5;
  int ny = 6;  
  double lambda = 1.0;

  PDEProblem Problem(nx,ny,lambda,&Comm);
 
  // starting solution, here a zero vector
  Epetra_Vector InitialGuess(Problem.GetMatrix()->Map());
  InitialGuess.PutScalar(0.0);

  // random vector upon which to apply each operator being tested
  Epetra_Vector directionVec(Problem.GetMatrix()->Map());
  directionVec.Random();

  // Set up the problem interface
  Teuchos::RCP<SimpleProblemInterface> interface = 
    Teuchos::rcp(new SimpleProblemInterface(&Problem) );
  
  // Set up theolver options parameter list
  Teuchos::RCP<Teuchos::ParameterList> noxParamsPtr = Teuchos::rcp(new Teuchos::ParameterList);
  Teuchos::ParameterList & noxParams = *(noxParamsPtr.get());

  // Set the nonlinear solver method
  noxParams.set("Nonlinear Solver", "Line Search Based");

  // Set up the printing utilities
  // Only print output if the "-v" flag is set on the command line
  Teuchos::ParameterList& printParams = noxParams.sublist("Printing");
  printParams.set("MyPID", MyPID); 
  printParams.set("Output Precision", 5);
  printParams.set("Output Processor", 0);
  if( verbose )
    printParams.set("Output Information", 
		NOX::Utils::OuterIteration + 
		NOX::Utils::OuterIterationStatusTest + 
		NOX::Utils::InnerIteration +
		NOX::Utils::Parameters + 
		NOX::Utils::Details + 
		NOX::Utils::Warning +
		NOX::Utils::TestDetails);
  else
    printParams.set("Output Information", NOX::Utils::Error +
		NOX::Utils::TestDetails);

  NOX::Utils printing(printParams);


  // Identify the test problem
  if (printing.isPrintType(NOX::Utils::TestDetails))
    printing.out() << "Starting epetra/NOX_Operators/NOX_Operators.exe" << std::endl;

  // Identify processor information
#ifdef HAVE_MPI
  if (printing.isPrintType(NOX::Utils::TestDetails)) {
    printing.out() << "Parallel Run" << std::endl;
    printing.out() << "Number of processors = " << NumProc << std::endl;
    printing.out() << "Print Process = " << MyPID << std::endl;
  }
  Comm.Barrier();
  if (printing.isPrintType(NOX::Utils::TestDetails))
    printing.out() << "Process " << MyPID << " is alive!" << std::endl;
  Comm.Barrier();
#else
  if (printing.isPrintType(NOX::Utils::TestDetails))
    printing.out() << "Serial Run" << std::endl;
#endif

  int status = 0;

  Teuchos::RCP<NOX::Epetra::Interface::Required> iReq = interface;

  // Need a NOX::Epetra::Vector for constructor
  NOX::Epetra::Vector noxInitGuess(InitialGuess, NOX::DeepCopy);

  // Analytic matrix
  Teuchos::RCP<Epetra_CrsMatrix> A = Teuchos::rcp( Problem.GetMatrix(), false );

  Epetra_Vector A_resultVec(Problem.GetMatrix()->Map());
  interface->computeJacobian( InitialGuess, *A );
  A->Apply( directionVec, A_resultVec );

  // FD operator
  Teuchos::RCP<Epetra_CrsGraph> graph = Teuchos::rcp( const_cast<Epetra_CrsGraph*>(&A->Graph()), false );
  Teuchos::RCP<NOX::Epetra::FiniteDifference> FD = Teuchos::rcp(
    new NOX::Epetra::FiniteDifference(printParams, iReq, noxInitGuess, graph) );
  
  Epetra_Vector FD_resultVec(Problem.GetMatrix()->Map());
  FD->computeJacobian(InitialGuess, *FD);
  FD->Apply( directionVec, FD_resultVec );

  // Matrix-Free operator
  Teuchos::RCP<NOX::Epetra::MatrixFree> MF = Teuchos::rcp(
    new NOX::Epetra::MatrixFree(printParams, iReq, noxInitGuess) );

  Epetra_Vector MF_resultVec(Problem.GetMatrix()->Map());
  MF->computeJacobian(InitialGuess, *MF);
  MF->Apply( directionVec, MF_resultVec );

  // Need NOX::Epetra::Vectors for tests
  NOX::Epetra::Vector noxAvec ( A_resultVec , NOX::DeepCopy );
  NOX::Epetra::Vector noxFDvec( FD_resultVec, NOX::DeepCopy );
  NOX::Epetra::Vector noxMFvec( MF_resultVec, NOX::DeepCopy );

  // Create a TestCompare class
  NOX::Epetra::TestCompare tester( printing.out(), printing);
  double abstol = 1.e-4;
  double reltol = 1.e-4 ;
  //NOX::TestCompare::CompareType aComp = NOX::TestCompare::Absolute;

  status += tester.testVector( noxFDvec, noxAvec, reltol, abstol,
                              "Finite-Difference Operator Apply Test" );
  status += tester.testVector( noxMFvec, noxAvec, reltol, abstol,
                              "Matrix-Free Operator Apply Test" );

  // Summarize test results  
  if( status == 0 )
    printing.out() << "Test passed!" << std::endl;
  else 
    printing.out() << "Test failed!" << std::endl;

#ifdef HAVE_MPI
  MPI_Finalize();
#endif

  // Final return value (0 = successfull, non-zero = failure)
  return status;

}