ParallelDomainBase.cc

#include <algorithm>
#include <iostream>
#include <fstream>

#include "ParallelDomainBase.h"
#include "ParMetisPartitioner.h"
#include "Mesh.h"
#include "Traverse.h"
#include "ElInfo.h"
#include "Element.h"
#include "MacroElement.h"
#include "PartitionElementData.h"
#include "DOFMatrix.h"
#include "DOFVector.h"
#include "SystemVector.h"
#include "VtkWriter.h"
#include "ElementDofIterator.h"
#include "ProblemStatBase.h"
#include "StandardProblemIteration.h"
#include "ElementFileWriter.h"
#include "VertexVector.h"
#include "StdMpi.h"
#include "MeshStructure.h"
#include "Debug.h"

#include "petscksp.h"

namespace AMDiS {

  PetscErrorCode myKSPMonitor(KSP ksp, PetscInt iter, PetscReal rnorm, void *)
  {    
    if (iter % 100 == 0 && MPI::COMM_WORLD.Get_rank() == 0)
      std::cout << "  Petsc-Iteration " << iter << ": " << rnorm << std::endl;

    return 0;
  }

  inline bool cmpDofsByValue(const DegreeOfFreedom* dof1, const DegreeOfFreedom* dof2)
  {
    return (*dof1 < *dof2);
  }

  ParallelDomainBase::ParallelDomainBase(ProblemIterationInterface *iIF,
					 ProblemTimeInterface *tIF,
					 FiniteElemSpace *fe,
					 RefinementManager *refinementManager)
    : iterationIF(iIF),
      timeIF(tIF),
      name(iIF->getName()),
      feSpace(fe),
      mesh(fe->getMesh()),
      refineManager(refinementManager),
      initialPartitionMesh(true),
      d_nnz(NULL),
      o_nnz(NULL),
      nRankDofs(0),
      rstart(0),
      nComponents(1),
      deserialized(false),
      lastMeshChangeIndex(0)
  {
    FUNCNAME("ParallelDomainBase::ParalleDomainBase()");

    TEST_EXIT(mesh->getNumberOfDOFAdmin() == 1)
      ("Only meshes with one DOFAdmin are supported!\n");
    TEST_EXIT(mesh->getDOFAdmin(0).getNumberOfPreDOFs(0) == 0)
      ("Wrong pre dof number for DOFAdmin!\n");

    mpiRank = MPI::COMM_WORLD.Get_rank();
    mpiSize = MPI::COMM_WORLD.Get_size();
    mpiComm = MPI::COMM_WORLD;
    partitioner = new ParMetisPartitioner(mesh, &mpiComm);
  }

  void ParallelDomainBase::initParallelization(AdaptInfo *adaptInfo)
  {
    FUNCNAME("ParallelDomainBase::initParallelization()");

    TEST_EXIT(mpiSize > 1)
      ("Parallelization does not work with only one process!\n");

    // If the problem has been already read from a file, we do not need to do anything.
    if (deserialized)
      return;

    // Test, if the mesh is the macro mesh only! Paritioning of the mesh is supported
    // only for macro meshes, so it will not work yet if the mesh is already refined
    // in some way.
    testForMacroMesh();

    MSG("START PARTITIONING!\n");

    // create an initial partitioning of the mesh
    partitioner->createPartitionData();
    // set the element weights, which are 1 at the very first begin
    setElemWeights(adaptInfo);
    // and now partition the mesh
    partitionMesh(adaptInfo);   

#if (DEBUG != 0)
    ElementIdxToDofs elMap;
    dbgCreateElementMap(elMap);
    if (mpiRank == 0) {
      int writePartMesh = 1;
      GET_PARAMETER(0, "dbg->write part mesh", "%d", &writePartMesh);

      if (writePartMesh > 0)
	writePartitioningMesh("part.vtu");
      else 
	MSG("Skip write part mesh!\n");
    }
#endif

    // === Create interior boundary information. ===

    createInteriorBoundaryInfo();

    // === Create new global and local DOF numbering. ===

    createLocalGlobalNumbering();

    // === Remove all macro elements that are not part of the rank partition. ===

    removeMacroElements();

    // === Reset all DOFAdmins of the mesh. ===

    updateDofAdmins();

    // === If in debug mode, make some tests. ===

#if (DEBUG != 0)
    MSG("AMDiS runs in debug mode, so make some test ...\n");
    dbgTestElementMap(elMap);
    dbgTestInteriorBoundary();
    dbgTestCommonDofs(true);
    MSG("Debug mode tests finished!\n");

    debug::writeMesh(feSpace, -1, "macromesh");   
#endif

    // === Create periodic dof mapping, if there are periodic boundaries. ===

    createPeriodicMap();

    // === Global refinements. ===

    int globalRefinement = 0;
    GET_PARAMETER(0, mesh->getName() + "->global refinements", "%d", &globalRefinement);

    if (globalRefinement > 0) {
      refineManager->globalRefine(mesh, globalRefinement);

      updateLocalGlobalNumbering();

      // === Update periodic mapping, if there are periodic boundaries. ===

      createPeriodicMap();
    }
  }


  void ParallelDomainBase::exitParallelization(AdaptInfo *adaptInfo)
  {}

  
  void ParallelDomainBase::updateDofAdmins()
  {
    FUNCNAME("ParallelDomainBase::updateDofAdmins()");

    for (int i = 0; i < mesh->getNumberOfDOFAdmin(); i++) {
      DOFAdmin& admin = const_cast<DOFAdmin&>(mesh->getDOFAdmin(i));

      // There must be always more allocated DOFs than used DOFs in DOFAdmin. Otherwise,
      // it is not possible to define a first DOF hole.
      if (static_cast<unsigned int>(admin.getSize()) == mapLocalGlobalDofs.size())
	admin.enlargeDOFLists();
      
      for (int j = 0; j < admin.getSize(); j++)
	admin.setDOFFree(j, true);
      for (unsigned int j = 0; j < mapLocalGlobalDofs.size(); j++)
 	admin.setDOFFree(j, false);

      admin.setUsedSize(mapLocalGlobalDofs.size());
      admin.setUsedCount(mapLocalGlobalDofs.size());
      admin.setFirstHole(mapLocalGlobalDofs.size());
    }
  }


  void ParallelDomainBase::testForMacroMesh()
  {
    FUNCNAME("ParallelDomainBase::testForMacroMesh()");

    int nMacroElements = 0;

    TraverseStack stack;
    ElInfo *elInfo = stack.traverseFirst(mesh, -1, Mesh::CALL_LEAF_EL);
    while (elInfo) {
      TEST_EXIT(elInfo->getLevel() == 0)
	("Mesh is already refined! This does not work with parallelization!\n");
      
      nMacroElements++;

      elInfo = stack.traverseNext(elInfo);
    }

    TEST_EXIT(nMacroElements >= mpiSize)
      ("The mesh has less macro elements than number of mpi processes!\n");
  }


  void ParallelDomainBase::setDofMatrix(DOFMatrix* mat, int dispMult, 
					int dispAddRow, int dispAddCol)
  {
    FUNCNAME("ParallelDomainBase::setDofMatrix()");

    TEST_EXIT(mat)("No DOFMatrix!\n");

    using mtl::tag::row; using mtl::tag::nz; using mtl::begin; using mtl::end;
    namespace traits= mtl::traits;
    typedef DOFMatrix::base_matrix_type Matrix;

    traits::col<Matrix>::type col(mat->getBaseMatrix());
    traits::const_value<Matrix>::type value(mat->getBaseMatrix());

    typedef traits::range_generator<row, Matrix>::type cursor_type;
    typedef traits::range_generator<nz, cursor_type>::type icursor_type;

    std::vector<int> cols;
    std::vector<double> values;
    cols.reserve(300);
    values.reserve(300);

    // === Traverse all rows of the dof matrix and insert row wise the values ===
    // === to the petsc matrix.                                               ===

    for (cursor_type cursor = begin<row>(mat->getBaseMatrix()), 
	   cend = end<row>(mat->getBaseMatrix()); cursor != cend; ++cursor) {

      cols.clear();
      values.clear();

      // Global index of the current row dof.
      int globalRowDof = mapLocalGlobalDofs[*cursor];
      // Test if the current row dof is a periodic dof.
      bool periodicRow = (periodicDof.count(globalRowDof) > 0);


      // === Traverse all non zero entries of the row and produce vector cols ===
      // === with the column indices of all row entries and vector values     ===
      // === with the corresponding values.                                   ===

      for (icursor_type icursor = begin<nz>(cursor), icend = end<nz>(cursor); 
	   icursor != icend; ++icursor) {

	// Set only non null values.
	if (value(*icursor) != 0.0) {
	  // Global index of the current column index.
	  int globalColDof = mapLocalGlobalDofs[col(*icursor)];
	  // Calculate the exact position of the column index in the petsc matrix.
	  int colIndex = globalColDof * dispMult + dispAddCol;

	  // If the current row is not periodic, but the current dof index is periodic,
	  // we have to duplicate the value to the other corresponding periodic columns.
 	  if (periodicRow == false && periodicDof.count(globalColDof) > 0) {
	    // The value is assign to n matrix entries, therefore, every entry 
	    // has only 1/n value of the original entry.
	    double scalFactor = 1.0 / (periodicDof[globalColDof].size() + 1.0);

	    // Insert original entry.
 	    cols.push_back(colIndex);
 	    values.push_back(value(*icursor) * scalFactor);

	    // Insert the periodic entries.
 	    for (std::set<DegreeOfFreedom>::iterator it = 
		   periodicDof[globalColDof].begin();
 		 it != periodicDof[globalColDof].end(); ++it) {
 	      cols.push_back(*it * dispMult + dispAddCol);
 	      values.push_back(value(*icursor) * scalFactor);
	    }
 	  } else {
	    // Neigher row nor column dof index is periodic, simple add entry.
	    cols.push_back(colIndex);
	    values.push_back(value(*icursor));
	  }
	}
      }


      // === Up to now we have assembled on row. Now, the row must be send to the ===
      // === corresponding rows to the petsc matrix.                              ===

      // Calculate petsc row index.
      int rowIndex = globalRowDof * dispMult + dispAddRow;
      
      if (periodicRow) {
	// The row dof is periodic, so send dof to all the corresponding rows.

	double scalFactor = 1.0 / (periodicDof[globalRowDof].size() + 1.0);
	
	int diagIndex = -1;
	for (int i = 0; i < static_cast<int>(values.size()); i++) {
	  // Change only the non diagonal values in the col. For the diagonal test
	  // we compare the global dof indices of the dof matrix (not of the petsc
	  // matrix!).
	  if ((cols[i] - dispAddCol) / dispMult != globalRowDof)
	    values[i] *= scalFactor;
	  else
	    diagIndex = i;
	}
	
	// Send the main row to the petsc matrix.
	MatSetValues(petscMatrix, 1, &rowIndex, cols.size(), 
		     &(cols[0]), &(values[0]), ADD_VALUES);	
 
	// Set diagonal element to zero, i.e., the diagonal element of the current
	// row is not send to the periodic row indices.
	if (diagIndex != -1)
	  values[diagIndex] = 0.0;

	// Send the row to all periodic row indices.
	for (std::set<DegreeOfFreedom>::iterator it = periodicDof[globalRowDof].begin();
	     it != periodicDof[globalRowDof].end(); ++it) {
	  int perRowIndex = *it * dispMult + dispAddRow;
	  MatSetValues(petscMatrix, 1, &perRowIndex, cols.size(), 
		       &(cols[0]), &(values[0]), ADD_VALUES);
	}

      } else {
	// The row dof is not periodic, simply send the row to the petsc matrix.
	MatSetValues(petscMatrix, 1, &rowIndex, cols.size(), 
		     &(cols[0]), &(values[0]), ADD_VALUES);
      }    
    }
  }


  void ParallelDomainBase::setDofVector(Vec& petscVec, DOFVector<double>* vec, 
					int dispMult, int dispAdd)
  {
    // Traverse all used dofs in the dof vector.
    DOFVector<double>::Iterator dofIt(vec, USED_DOFS);
    for (dofIt.reset(); !dofIt.end(); ++dofIt) {
      // Calculate global row index of the dof.
      int globalRow = mapLocalGlobalDofs[dofIt.getDOFIndex()];
      // Calculate petsc index of the row dof.
      int index = globalRow * dispMult + dispAdd;

      if (periodicDof.count(globalRow) > 0) {
	// The dof index is periodic, so devide the value to all dof entries.

	double value = *dofIt / (periodicDof[globalRow].size() + 1.0);
	VecSetValues(petscVec, 1, &index, &value, ADD_VALUES);

	for (std::set<DegreeOfFreedom>::iterator it = periodicDof[globalRow].begin();
	     it != periodicDof[globalRow].end(); ++it) {
	  index = *it * dispMult + dispAdd;
	  VecSetValues(petscVec, 1, &index, &value, ADD_VALUES);
	}

      } else {
	// The dof index is not periodic.
	double value = *dofIt;
	VecSetValues(petscVec, 1, &index, &value, ADD_VALUES);
      }
    }    
  }


  void ParallelDomainBase::createPetscNnzStructure(Matrix<DOFMatrix*> *mat)
  {
    FUNCNAME("ParallelDomainBase::createPetscNnzStructure()");

    TEST_EXIT_DBG(!d_nnz)("There is something wrong!\n");
    TEST_EXIT_DBG(!o_nnz)("There is something wrong!\n");

    d_nnz = new int[nRankRows];
    o_nnz = new int[nRankRows];
    for (int i = 0; i < nRankRows; i++) {
      d_nnz[i] = 0;
      o_nnz[i] = 0;
    }

    using mtl::tag::row; using mtl::tag::nz; using mtl::begin; using mtl::end;
    namespace traits = mtl::traits;
    typedef DOFMatrix::base_matrix_type Matrix;
    typedef std::vector<std::pair<int, int> > MatrixNnzEntry;

    // Stores to each rank a list of nnz entries (i.e. pairs of row and column index)
    // that this rank will send to. This nnz entries will be assembled on this rank,
    // but because the row DOFs are not DOFs of this rank they will be send to the
    // owner of the row DOFs.
    std::map<int, MatrixNnzEntry> sendMatrixEntry;

    for (int i = 0; i < nComponents; i++) {
      for (int j = 0; j < nComponents; j++) {
 	if ((*mat)[i][j]) {
	  Matrix bmat = (*mat)[i][j]->getBaseMatrix();

	  traits::col<Matrix>::type col(bmat);
	  traits::const_value<Matrix>::type value(bmat);
	  
	  typedef traits::range_generator<row, Matrix>::type cursor_type;
	  typedef traits::range_generator<nz, cursor_type>::type icursor_type;
	  
	  for (cursor_type cursor = begin<row>(bmat), 
		 cend = end<row>(bmat); cursor != cend; ++cursor) {

	    // Map the local row number to the global DOF index and create from it
	    // the global PETSc row index of this DOF.
	    int petscRowIdx = mapLocalGlobalDofs[*cursor] * nComponents + i;

	    if (isRankDof[*cursor]) {

	      // === The current row DOF is a rank dof, so create the corresponding ===
	      // === nnz values directly on rank's nnz data.                        ===

	      // This is the local row index of the local PETSc matrix.
	      int localPetscRowIdx = petscRowIdx - rstart * nComponents;

#if (DEBUG != 0)    
	      if (localPetscRowIdx < 0 || localPetscRowIdx >= nRankRows) {
		std::cout << "ERROR in rank: " << mpiRank << std::endl;
		std::cout << "  Wrong r = " << localPetscRowIdx << " " << *cursor 
			  << " " << mapLocalGlobalDofs[*cursor] << " " 
			  << nRankRows << std::endl;
		ERROR_EXIT("Should not happen!\n");
	      }
#endif
	      
	      // Traverse all non zero entries in this row.
	      for (icursor_type icursor = begin<nz>(cursor), 
		     icend = end<nz>(cursor); icursor != icend; ++icursor) {
		if (value(*icursor) != 0.0) {
		  int petscColIdx = mapLocalGlobalDofs[col(*icursor)] * nComponents + j;

		  // The row DOF is a rank DOF, if also the column is a rank DOF, 
		  // increment the d_nnz values for this row, otherwise the o_nnz value.
		  if (petscColIdx >= rstart * nComponents && 
		      petscColIdx < rstart * nComponents + nRankRows)
		    d_nnz[localPetscRowIdx]++;
		  else
		    o_nnz[localPetscRowIdx]++;
		}    
	      }
	    } else {
	      
	      // === The current row DOF is not a rank dof, i.e., it will be created ===
	      // === on this rank, but after this it will be send to another rank    ===
	      // === matrix. So we need to send also the corresponding nnz structure ===
	      // === of this row to the corresponding rank.                          ===

	      // Find out who is the member of this DOF.
	      int sendToRank = -1;
	      for (RankToDofContainer::iterator it = recvDofs.begin();
		   it != recvDofs.end(); ++it) {
		for (DofContainer::iterator dofIt = it->second.begin();
		     dofIt != it->second.end(); ++dofIt) {
		  if (**dofIt == *cursor) {

		    if (petscRowIdx == 6717) {
		      debug::writeDofMesh(mpiRank, *cursor, feSpace);
		      MSG("SEND DOF TO: %d/%d\n", it->first, *cursor);
		    }

		    sendToRank = it->first;
		    break;
		  }
		}

		if (sendToRank != -1)
		  break;
	      }

	      TEST_EXIT_DBG(sendToRank != -1)("Should not happen!\n");

	      // Send all non zero entries to the member of the row DOF.
	      for (icursor_type icursor = begin<nz>(cursor), 
		     icend = end<nz>(cursor); icursor != icend; ++icursor) {
		if (value(*icursor) != 0.0) {
		  int petscColIdx = mapLocalGlobalDofs[col(*icursor)] * nComponents + j;
		  
		  sendMatrixEntry[sendToRank].
		    push_back(std::make_pair(petscRowIdx, petscColIdx));
		}
	      }

	    } // if (isRankDof[*cursor]) ... else ...
	  } // for each row in mat[i][j]
	} // if mat[i][j]
      } 
    }

    // === Send and recv the nnz row structure to/from other ranks. ===

    StdMpi<MatrixNnzEntry> stdMpi(mpiComm, true);
    stdMpi.send(sendMatrixEntry);
    stdMpi.recv(sendDofs);
    stdMpi.startCommunication<int>(MPI_INT);

    // === Evaluate the nnz structure this rank got from other ranks and add it to ===
    // === the PETSc nnz data structure.                                           ===

    for (std::map<int, MatrixNnzEntry>::iterator it = stdMpi.getRecvData().begin();
	 it != stdMpi.getRecvData().end(); ++it) {
      if (it->second.size() > 0) {
	for (unsigned int i = 0; i < it->second.size(); i++) {
	  int r = it->second[i].first;
	  int c = it->second[i].second;

	  int localRowIdx = r - rstart * nComponents;

	  TEST_EXIT_DBG(localRowIdx >= 0 && localRowIdx < nRankRows)
	    ("Got row index %d/%d (nRankRows = %d) from rank %d. Should not happen!\n",
	     r, localRowIdx, nRankRows, it->first);
	  
	  if (c < rstart * nComponents || c >= rstart * nComponents + nRankRows)
	    o_nnz[localRowIdx]++;
	  else
	    d_nnz[localRowIdx]++;
	}
      }
    }  
  }


  void ParallelDomainBase::fillPetscMatrix(Matrix<DOFMatrix*> *mat, SystemVector *vec)
  {
    FUNCNAME("ParallelDomainBase::fillPetscMatrix()");

    clock_t first = clock();

    // === Create PETSc vector (rhs, solution and a temporary vector). ===

    VecCreate(PETSC_COMM_WORLD, &petscRhsVec);
    VecSetSizes(petscRhsVec, nRankRows, nOverallRows);
    VecSetType(petscRhsVec, VECMPI);

    VecCreate(PETSC_COMM_WORLD, &petscSolVec);
    VecSetSizes(petscSolVec, nRankRows, nOverallRows);
    VecSetType(petscSolVec, VECMPI);

    VecCreate(PETSC_COMM_WORLD, &petscTmpVec);
    VecSetSizes(petscTmpVec, nRankRows, nOverallRows);
    VecSetType(petscTmpVec, VECMPI);

    if (!d_nnz)
      createPetscNnzStructure(mat);

    // === Create PETSc matrix with the computed nnz data structure. ===

    MatCreateMPIAIJ(PETSC_COMM_WORLD, nRankRows, nRankRows, nOverallRows, nOverallRows,
		    0, d_nnz, 0, o_nnz, &petscMatrix);

    INFO(info, 8)("Fill petsc matrix 1 needed %.5f seconds\n", TIME_USED(first, clock()));

#if (DEBUG != 0)
    int a, b;
    MatGetOwnershipRange(petscMatrix, &a, &b);
    TEST_EXIT(a == rstart * nComponents)("Wrong matrix ownership range!\n");
    TEST_EXIT(b == rstart * nComponents + nRankRows)("Wrong matrix ownership range!\n");
#endif

    // === Transfer values from DOF matrices to the PETSc matrix. === 

    for (int i = 0; i < nComponents; i++)
      for (int j = 0; j < nComponents; j++)
	if ((*mat)[i][j])
	  setDofMatrix((*mat)[i][j], nComponents, i, j);

    INFO(info, 8)("Fill petsc matrix 2 needed %.5f seconds\n", TIME_USED(first, clock()));

    MatAssemblyBegin(petscMatrix, MAT_FINAL_ASSEMBLY);
    MatAssemblyEnd(petscMatrix, MAT_FINAL_ASSEMBLY);

    // === Transfer values from DOF vector to the PETSc vector. === 

    for (int i = 0; i < nComponents; i++)
      setDofVector(petscRhsVec, vec->getDOFVector(i), nComponents, i);

    VecAssemblyBegin(petscRhsVec);
    VecAssemblyEnd(petscRhsVec);

    INFO(info, 8)("Fill petsc matrix needed %.5f seconds\n", TIME_USED(first, clock()));
  }


  void ParallelDomainBase::solvePetscMatrix(SystemVector &vec)
  {
    FUNCNAME("ParallelDomainBase::solvePetscMatrix()");

#if 0
    // Set old solution to be initiual guess for petsc solver.
    for (int i = 0; i < nComponents; i++)
      setDofVector(petscSolVec, vec->getDOFVector(i), nComponents, i);

    VecAssemblyBegin(petscSolVec);
    VecAssemblyEnd(petscSolVec);
#endif

    // === Init Petsc solver. ===

    KSP solver;
    KSPCreate(PETSC_COMM_WORLD, &solver);
    KSPSetOperators(solver, petscMatrix, petscMatrix, SAME_NONZERO_PATTERN); 
    KSPSetTolerances(solver, 0.0, 1e-8, PETSC_DEFAULT, PETSC_DEFAULT);
    KSPSetType(solver, KSPBCGS);
    KSPMonitorSet(solver, myKSPMonitor, PETSC_NULL, 0);
    KSPSetFromOptions(solver);
    // Do not delete the solution vector, use it for the initial guess.
    //    KSPSetInitialGuessNonzero(solver, PETSC_TRUE);


    // === Run Petsc. ===

    KSPSolve(solver, petscRhsVec, petscSolVec);

    // === Transfere values from Petsc's solution vectors to the dof vectors.
    PetscScalar *vecPointer;
    VecGetArray(petscSolVec, &vecPointer);

    for (int i = 0; i < nComponents; i++) {
      DOFVector<double> *dofvec = vec.getDOFVector(i);
      for (int j = 0; j < nRankDofs; j++)
	(*dofvec)[mapLocalToDofIndex[j]] = vecPointer[j * nComponents + i];      
    }

    VecRestoreArray(petscSolVec, &vecPointer);


    // === Synchronize dofs at common dofs, i.e., dofs that correspond to more ===
    // === than one partition.                                                 ===
    synchVector(vec);


    // === Print information about solution process. ===

    int iterations = 0;
    KSPGetIterationNumber(solver, &iterations);
    MSG("  Number of iterations: %d\n", iterations);
    
    double norm = 0.0;
    MatMult(petscMatrix, petscSolVec, petscTmpVec);
    VecAXPY(petscTmpVec, -1.0, petscRhsVec);
    VecNorm(petscTmpVec, NORM_2, &norm);
    MSG("  Residual norm: %e\n", norm);


    // === Destroy Petsc's variables. ===

    MatDestroy(petscMatrix);
    VecDestroy(petscRhsVec);
    VecDestroy(petscSolVec);
    VecDestroy(petscTmpVec);
    KSPDestroy(solver);
  }
  

  void ParallelDomainBase::synchVector(DOFVector<double> &vec)
  {
    StdMpi<std::vector<double> > stdMpi(mpiComm);

    for (RankToDofContainer::iterator sendIt = sendDofs.begin();
	 sendIt != sendDofs.end(); ++sendIt) {
      std::vector<double> dofs;
      int nSendDofs = sendIt->second.size();
      dofs.reserve(nSendDofs);
      
      for (int i = 0; i < nSendDofs; i++)
	dofs.push_back(vec[*((sendIt->second)[i])]);

      stdMpi.send(sendIt->first, dofs);
    }

    for (RankToDofContainer::iterator recvIt = recvDofs.begin();
	 recvIt != recvDofs.end(); ++recvIt)
      stdMpi.recv(recvIt->first, recvIt->second.size());

    stdMpi.startCommunication<double>(MPI_DOUBLE);

    for (RankToDofContainer::iterator recvIt = recvDofs.begin();
	 recvIt != recvDofs.end(); ++recvIt)
      for (unsigned int i = 0; i < recvIt->second.size(); i++)
	vec[*(recvIt->second)[i]] = stdMpi.getRecvData(recvIt->first)[i];
  }


  void ParallelDomainBase::synchVector(SystemVector &vec)
  {
    StdMpi<std::vector<double> > stdMpi(mpiComm);

    for (RankToDofContainer::iterator sendIt = sendDofs.begin();
	 sendIt != sendDofs.end(); ++sendIt) {
      std::vector<double> dofs;
      int nSendDofs = sendIt->second.size();
      dofs.reserve(nComponents * nSendDofs);
      
      for (int i = 0; i < nComponents; i++) {
	DOFVector<double> *dofvec = vec.getDOFVector(i);
	for (int j = 0; j < nSendDofs; j++)
	  dofs.push_back((*dofvec)[*((sendIt->second)[j])]);
      }

      stdMpi.send(sendIt->first, dofs);
    }

    for (RankToDofContainer::iterator recvIt = recvDofs.begin();
	 recvIt != recvDofs.end(); ++recvIt)
      stdMpi.recv(recvIt->first, recvIt->second.size() * nComponents);

    stdMpi.startCommunication<double>(MPI_DOUBLE);

    for (RankToDofContainer::iterator recvIt = recvDofs.begin();
	 recvIt != recvDofs.end(); ++recvIt) {
      int nRecvDofs = recvIt->second.size();

      int counter = 0;
      for (int i = 0; i < nComponents; i++) {
	DOFVector<double> *dofvec = vec.getDOFVector(i);
 	for (int j = 0; j < nRecvDofs; j++)
	  (*dofvec)[*(recvIt->second)[j]] = 
	    stdMpi.getRecvData(recvIt->first)[counter++];
      }
    }
  }


  void ParallelDomainBase::checkMeshChange()
  {
    FUNCNAME("ParallelDomainBase::checkMeshChange()");

    // === If mesh has not been changed on all ranks, return. ===

    int recvAllValues = 0;
    int sendValue = static_cast<int>(mesh->getChangeIndex() != lastMeshChangeIndex);
    mpiComm.Allreduce(&sendValue, &recvAllValues, 1, MPI_INT, MPI_SUM);

    if (recvAllValues == 0)
      return;

    // === At least one rank mesh has been changed, so the boundaries must be ===
    // === adapted to the new mesh structure.                                 ===

    clock_t first = clock();
    
    do {
      // To check the interior boundaries, the ownership of the boundaries is not 
      // important. Therefore, we add all boundaries to one boundary container.
      RankToBoundMap allBound;
      for (RankToBoundMap::iterator it = myIntBoundary.boundary.begin(); 
	   it != myIntBoundary.boundary.end(); ++it)
	for (std::vector<AtomicBoundary>::iterator bit = it->second.begin();
	     bit != it->second.end(); ++bit)
	  if (bit->rankObj.subObj == FACE)
	    allBound[it->first].push_back(*bit);

      for (RankToBoundMap::iterator it = otherIntBoundary.boundary.begin(); 
	   it != otherIntBoundary.boundary.end(); ++it)
	for (std::vector<AtomicBoundary>::iterator bit = it->second.begin();
	     bit != it->second.end(); ++bit)
	  if (bit->rankObj.subObj == FACE)
	    allBound[it->first].push_back(*bit);

      // === Check the boundaries and adapt mesh if necessary. ===

      bool meshChanged = checkAndAdaptBoundary(allBound);

      // === Check on all ranks if at least one rank's mesh has changed. ===

      int sendValue = static_cast<int>(!meshChanged);
      recvAllValues = 0;
      mpiComm.Allreduce(&sendValue, &recvAllValues, 1, MPI_INT, MPI_SUM);
      MSG("LOOP\n");
    } while (recvAllValues != 0);


#if (DEBUG != 0)
    debug::writeMesh(feSpace, -1, "mesh");
#endif

    INFO(info, 8)("Parallel mesh adaption needed %.5f seconds\n", 
		  TIME_USED(first, clock()));

    // === Because the mesh has been changed, update the DOF numbering and mappings. ===
   
    updateLocalGlobalNumbering();

    // === If there is a non zero pattern computed for Petsc matrix, delete it. So ===
    // === it will be recomputed after next assembling.                            ===

    if (d_nnz) {
      delete [] d_nnz;
      d_nnz = NULL;
    }

    if (o_nnz) {
      delete [] o_nnz;
      o_nnz = NULL;
    }
  }

  
  bool ParallelDomainBase::checkAndAdaptBoundary(RankToBoundMap &allBound)
  {
    FUNCNAME("ParallelDomainBase::checkAndAdaptBoundary()");

    MSG("CHECK_AND_ADAPT 1!\n");

    // === Create mesh structure codes for all ranks boundary elements. ===
       
    std::map<int, MeshCodeVec> sendCodes;
   
    for (RankToBoundMap::iterator it = allBound.begin(); it != allBound.end(); ++it) {
      for (std::vector<AtomicBoundary>::iterator boundIt = it->second.begin();
	   boundIt != it->second.end(); ++boundIt) {
	MeshStructure elCode;
	elCode.init(boundIt->rankObj);
	sendCodes[it->first].push_back(elCode);
      }
    }

    MSG("CHECK_AND_ADAPT 2!\n");

    StdMpi<MeshCodeVec> stdMpi(mpiComm, true);
    stdMpi.send(sendCodes);
    stdMpi.recv(allBound);
    stdMpi.startCommunication<unsigned long int>(MPI_UNSIGNED_LONG);
 
    MSG("CHECK_AND_ADAPT 3!\n");

    clock_t first = clock();

    // === Compare received mesh structure codes. ===
    
    bool meshFitTogether = true;
    int superCounter = 0;

    for (RankToBoundMap::iterator it = allBound.begin(); it != allBound.end(); ++it) {
      
      MeshCodeVec &recvCodes = stdMpi.getRecvData()[it->first];
      int i = 0;
      
      for (std::vector<AtomicBoundary>::iterator boundIt = it->second.begin();
	   boundIt != it->second.end(); ++boundIt) {

	MeshStructure elCode;	
	elCode.init(boundIt->rankObj);
	
	if (elCode.getCode() != recvCodes[i].getCode()) {
	  TEST_EXIT_DBG(refineManager)("Refinement manager is not set correctly!\n");
	  superCounter++;
	  clock_t fff = clock();

	  int refCounter = 0;
	  bool b = fitElementToMeshCode(recvCodes[i], 
					boundIt->rankObj.el,
					boundIt->rankObj.ithObj, 
					boundIt->rankObj.elType, refCounter);  

	  MSG("SIZE OF ELINFO3d = %d\n", sizeof(ElInfo3d));
	  MSG("Code length %d with %d refs needed %.5f seconds\n", 
	      recvCodes[i].getNumElements(), 
	      refCounter,
	      TIME_USED(fff, clock()));

	  if (b)
	    meshFitTogether = false;
	}
	
	i++;
      }
    }

    MSG("time for %d needed %.5f seconds\n", superCounter, TIME_USED(first, clock()));
    MSG("CHECK_AND_ADAPT 4!\n");

    return meshFitTogether;
  }


  bool ParallelDomainBase::fitElementToMeshCode(MeshStructure &code, 
						Element *el, 
						int ithSide, 
						int elType,
						int &refCounter)
  {
    FUNCNAME("fitElementToMeshCode()");

    if (code.empty())
      return false;
    
    int s1 = el->getSideOfChild(0, ithSide, elType);
    int s2 = el->getSideOfChild(1, ithSide, elType);

    TEST_EXIT_DBG(s1 != -1 || s2 != -1)("This should not happen!\n");

    if (s1 != -1 && s2 != -1)
      return fitElementToMeshCode2(code, el, ithSide, elType, refCounter);

    if (el->isLeaf()) {
      if (code.getNumElements() == 1 && code.isLeafElement())
	return false;
      
      el->setMark(1);
      refineManager->refineElement(el->getMesh(), el);
      refCounter++;
    }
   
    if (s1 != -1)
      return fitElementToMeshCode2(code, 
				   el->getFirstChild(), s1, el->getChildType(elType), refCounter);
    else
      return fitElementToMeshCode2(code, 
				   el->getSecondChild(), s2, el->getChildType(elType), refCounter);
  }

  bool ParallelDomainBase::fitElementToMeshCode2(MeshStructure &code, 
						 Element *el, 
						 int ithSide, 
						 int elType, int &refCounter)
  {
    if (code.isLeafElement())
      return false;

    bool value = false;

    if (el->isLeaf()) {
      el->setMark(1);
      refineManager->refineElement(el->getMesh(), el);     
      value = true;
      refCounter++;
    }

    int s1 = el->getSideOfChild(0, ithSide, elType);
    int s2 = el->getSideOfChild(1, ithSide, elType);
    
    if (s1 != -1) {
      code.nextElement();
      value |= fitElementToMeshCode2(code, el->getFirstChild(), 
				     s1, el->getChildType(elType), refCounter);
    }  

    if (s2 != -1) {
      code.nextElement();	
      value |= fitElementToMeshCode2(code, el->getSecondChild(), 
				     s2, el->getChildType(elType), refCounter);
    }

    return value;
  }

  
  void ParallelDomainBase::serialize(std::ostream &out, DofContainer &data)
  {
    int vecSize = data.size();
    SerUtil::serialize(out, vecSize);
    for (int i = 0; i < vecSize; i++) {
      int dofIndex = (*data[i]);
      SerUtil::serialize(out, dofIndex);
    }
  }


  void ParallelDomainBase::deserialize(std::istream &in, DofContainer &data,
				       std::map<int, const DegreeOfFreedom*> &dofMap)
  {
    FUNCNAME("ParallelDomainBase::deserialize()");

    int vecSize = 0;
    SerUtil::deserialize(in, vecSize);
    data.resize(vecSize);
    for (int i = 0; i < vecSize; i++) {
      int dofIndex = 0;
      SerUtil::deserialize(in, dofIndex);

      TEST_EXIT_DBG(dofMap.count(dofIndex) != 0)
	("Dof index could not be deserialized correctly!\n");

      data[i] = dofMap[dofIndex];
    }
  }


  void ParallelDomainBase::serialize(std::ostream &out, RankToDofContainer &data)
  {
    int mapSize = data.size();
    SerUtil::serialize(out, mapSize);
    for (RankToDofContainer::iterator it = data.begin(); it != data.end(); ++it) {
      int rank = it->first;
      SerUtil::serialize(out, rank);
      serialize(out, it->second);
    }
  }

  
  void ParallelDomainBase::deserialize(std::istream &in, RankToDofContainer &data,
				       std::map<int, const DegreeOfFreedom*> &dofMap)
  {
    int mapSize = 0;
    SerUtil::deserialize(in, mapSize);
    for (int i = 0; i < mapSize; i++) {
      int rank = 0;
      SerUtil::deserialize(in, rank);
      deserialize(in, data[rank], dofMap);      
    }
  }


  double ParallelDomainBase::setElemWeights(AdaptInfo *adaptInfo) 
  {
    FUNCNAME("ParallelDomainBase::setElemWeights()");