ParallelDomainBase.cc

#include <algorithm>

#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 "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 *refineManager)
    : iterationIF(iIF),
      timeIF(tIF),
      name(iIF->getName()),
      feSpace(fe),
      mesh(fe->getMesh()),
      refinementManager(refineManager),
      initialPartitionMesh(true),
      nRankDofs(0),
      rstart(0),
      nComponents(1),
      deserialized(false)
  {
    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();

    // 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)
      writePartitioningMesh("part.vtu");
#endif

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

    // Set of all DOFs of the rank.
    std::vector<const DegreeOfFreedom*> rankDofs;
    // Number of DOFs in ranks partition that are owned by the rank.
    nRankDofs = 0;
    // Number of all DOFs in the macro mesh.
    int nOverallDOFs = 0;

    createLocalGlobalNumbering(rankDofs, nRankDofs, nOverallDOFs);

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

    createInteriorBoundaryInfo();

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

    removeMacroElements();

#if (DEBUG != 0)
    DbgTestElementMap(elMap);
    DbgTestInteriorBoundary();
#endif

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

    updateDofAdmins();

    // === 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) {
      refinementManager->globalRefine(mesh, globalRefinement);

#if (DEBUG != 0)
      elMap.clear();
      DbgCreateElementMap(elMap);
#endif

      updateLocalGlobalNumbering(nRankDofs, nOverallDOFs);

      updateDofAdmins();

#if (DEBUG != 0)
      DbgTestElementMap(elMap);
#endif

      // === Update periodic mapping, if there are periodic boundaries. ===
      createPeriodicMap();
    }

    //  exit(0);

#if (DEBUG != 0)
    DbgTestCommonDofs(true);
#endif

    nRankRows = nRankDofs * nComponents;
    nOverallRows = nOverallDOFs * nComponents;
  }


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

  
  void ParallelDomainBase::updateDofAdmins()
  {
    int nAdmins = mesh->getNumberOfDOFAdmin();
    for (int i = 0; i < nAdmins; i++) {
      DOFAdmin& admin = const_cast<DOFAdmin&>(mesh->getDOFAdmin(i));
      
      for (int j = 0; j < admin.getSize(); j++)
	admin.setDOFFree(j, true);
      for (int j = 0; j < static_cast<int>(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::fillPetscMatrix(DOFMatrix *mat, DOFVector<double> *vec)
  {
    FUNCNAME("ParallelDomainBase::fillPetscMatrix()");

    ERROR_EXIT("Not yet tested for scalar problem definition!\n");

    MatCreate(PETSC_COMM_WORLD, &petscMatrix);
    MatSetSizes(petscMatrix, nRankRows, nRankRows, nOverallRows, nOverallRows);
    MatSetType(petscMatrix, MATAIJ);

    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);

    setDofMatrix(mat);

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

    setDofVector(petscRhsVec, vec);

    VecAssemblyBegin(petscRhsVec);
    VecAssemblyEnd(petscRhsVec);
  }

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

    clock_t first = clock();

    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);

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

    int d_nnz[nRankRows];
    int o_nnz[nRankRows];

    std::map<int, std::vector< std::pair<int, int> > > sendMatrixEntry;

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

    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) {

	    int r = mapLocalGlobalDOFs[*cursor] * nComponents + i;

	    if (isRankDof[*cursor]) {
	      r -= rstart * nComponents;

#if (DEBUG != 0)    
	      if (r < 0 || r >= nRankRows) {
		std::cout << "ERROR in rank: " << mpiRank << std::endl;
		std::cout << "  Wrong r = " << r << " " << *cursor << " " 
			  << mapLocalGlobalDOFs[*cursor] << " " 
			  << nRankRows << std::endl;
		ERROR_EXIT("Should not happen!\n");
	      }
#endif
	      
	      for (icursor_type icursor = begin<nz>(cursor), 
		     icend = end<nz>(cursor); icursor != icend; ++icursor) {
		if (value(*icursor) != 0.0) {
		  int c = mapLocalGlobalDOFs[col(*icursor)] * nComponents + j;

		  if (c >= rstart * nComponents && 
		      c < rstart * nComponents + nRankRows)
		    d_nnz[r]++;
		  else
		    o_nnz[r]++;		  
		}    
	      }
	    } else {
	      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) {
		    sendToRank = it->first;
		    break;
		  }
		}

		if (sendToRank != -1)
		  break;
	      }

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

	      for (icursor_type icursor = begin<nz>(cursor), 
		     icend = end<nz>(cursor); icursor != icend; ++icursor) {
		if (value(*icursor) != 0.0) {
		  int c = mapLocalGlobalDOFs[col(*icursor)] * nComponents + j;
		  
		  sendMatrixEntry[sendToRank].push_back(std::make_pair(r, c));
		}
	      }
	    }
	  }
	}
      }
    }

    MPI::Request request[sendDofs.size() + recvDofs.size()];
    int requestCounter = 0;

    std::vector<int*> sendBuffers;
    sendBuffers.reserve(recvDofs.size());

    for (RankToDofContainer::iterator it = recvDofs.begin(); 
	 it != recvDofs.end(); ++it) {
      int nSend = sendMatrixEntry[it->first].size();

      request[requestCounter++] = mpiComm.Isend(&nSend, 1, MPI_INT, it->first, 0);
      
      if (nSend > 0) {
	int *sendbuf = new int[nSend * 2];
	for (int i = 0; i < nSend; i++) {
	  sendbuf[i * 2] = sendMatrixEntry[it->first][i].first;
	  sendbuf[i * 2 + 1] = sendMatrixEntry[it->first][i].second;
	}
	sendBuffers.push_back(sendbuf);
      } else {
	sendBuffers.push_back(NULL);
      }
    }

    std::vector<int> recvSize;
    recvSize.reserve(sendDofs.size());
    
    int i = 0;
    for (RankToDofContainer::iterator it = sendDofs.begin();
	 it != sendDofs.end(); ++it)
      request[requestCounter++] = 
	mpiComm.Irecv(&(recvSize[i++]), 1, MPI_INT, it->first, 0);

    MPI::Request::Waitall(requestCounter, request);

    requestCounter = 0;

    i = 0;
    for (RankToDofContainer::iterator it = recvDofs.begin(); 
	 it != recvDofs.end(); ++it) {
      int nSend = sendMatrixEntry[it->first].size();

      if (nSend > 0)
	request[requestCounter++] = 
	  mpiComm.Isend(sendBuffers[i], nSend * 2, MPI_INT, it->first, 0);

      i++;
    }

    std::vector<int*> recvBuffers;
    recvBuffers.reserve(sendDofs.size());
    
    i = 0;
    for (RankToDofContainer::iterator it = sendDofs.begin();
	 it != sendDofs.end(); ++it) {
      if (recvSize[i] > 0) {
	int *recvbuf = new int[recvSize[i] * 2];
	recvBuffers.push_back(recvbuf);

	request[requestCounter++] =
	  mpiComm.Irecv(recvbuf, recvSize[i] * 2, MPI_INT, it->first, 0);
      } else {
	recvBuffers.push_back(NULL);
      }

      i++;
    }

    MPI::Request::Waitall(requestCounter, request);

    for (int j = 0; j < static_cast<int>(sendBuffers.size()); j++)
      if (sendBuffers[j])
 	delete [] sendBuffers[j];

    i = 0;
    for (RankToDofContainer::iterator it = sendDofs.begin();
	 it != sendDofs.end(); ++it) {
      if (recvSize[i] > 0) {
	for (int j = 0; j < recvSize[i]; j++) {
	  int r = recvBuffers[i][j * 2];
	  int c = recvBuffers[i][j * 2 + 1];

	  r -= rstart * nComponents;

	  TEST_EXIT_DBG(r >= 0 && r < nRankRows)("Should not happen!\n");
	  
	  if (c < rstart * nComponents || c >= rstart * nComponents + nRankRows)
	    o_nnz[r]++;
	  else
	    d_nnz[r]++;
	}

	delete [] recvBuffers[i];
      }

      i++;
    }

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

#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

    using mtl::tag::major; using mtl::tag::nz; using mtl::begin; using mtl::end;
    namespace traits= mtl::traits;
    typedef DOFMatrix::base_matrix_type 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);

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

    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(DOFVector<double> *vec)
  {
    FUNCNAME("ParallelDomainBase::solvePetscMatrix()");

    ERROR_EXIT("Not yet tested for scalar problem definition!\n");

    KSP ksp;
    KSPCreate(PETSC_COMM_WORLD, &ksp);
    KSPSetOperators(ksp, petscMatrix, petscMatrix, SAME_NONZERO_PATTERN);
    KSPSetTolerances(ksp, 1.e-7, PETSC_DEFAULT, PETSC_DEFAULT, PETSC_DEFAULT);
    KSPSetType(ksp, KSPBCGS);
    KSPMonitorSet(ksp, myKSPMonitor, PETSC_NULL, 0);
    KSPSolve(ksp, petscRhsVec, petscSolVec);

    PetscScalar *vecPointer;
    VecGetArray(petscSolVec, &vecPointer);

    for (int i = 0; i < nRankDofs; i++)
      (*vec)[mapLocalToDofIndex[i]] = vecPointer[i];

    VecRestoreArray(petscSolVec, &vecPointer);

    std::vector<double*> sendBuffers(sendDofs.size());
    std::vector<double*> recvBuffers(recvDofs.size());

    MPI::Request request[sendDofs.size() + recvDofs.size()];
    int requestCounter = 0;
    
    int i = 0;
    for (RankToDofContainer::iterator sendIt = sendDofs.begin();
	 sendIt != sendDofs.end(); ++sendIt, i++) {
      int nSendDOFs = sendIt->second.size();
      sendBuffers[i] = new double[nSendDOFs];

      for (int j = 0; j < nSendDOFs; j++)
	sendBuffers[i][j] = (*vec)[*((sendIt->second)[j])];

      request[requestCounter++] =
	mpiComm.Isend(sendBuffers[i], nSendDOFs, MPI_DOUBLE, sendIt->first, 0);
    }

    i = 0;
    for (RankToDofContainer::iterator recvIt = recvDofs.begin();
	 recvIt != recvDofs.end(); ++recvIt, i++) {
      int nRecvDOFs = recvIt->second.size();
      recvBuffers[i] = new double[nRecvDOFs];

      request[requestCounter++] =
	mpiComm.Irecv(recvBuffers[i], nRecvDOFs, MPI_DOUBLE, recvIt->first, 0);
    }


    MPI::Request::Waitall(requestCounter, request);

    i = 0;
    for (RankToDofContainer::iterator recvIt = recvDofs.begin();
	 recvIt != recvDofs.end(); ++recvIt, i++) {
      for (int j = 0; j < static_cast<int>(recvIt->second.size()); j++)
	(*vec)[*(recvIt->second)[j]] = recvBuffers[i][j];

      delete [] recvBuffers[i];
    }
    
    for (int i = 0; i < static_cast<int>(sendBuffers.size()); i++)
      delete [] sendBuffers[i];

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


  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.                                                 ===
    synchVectors(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::synchVectors(SystemVector &vec)
  {
#if 0
    StdMpi<std::vector<double>, std::vector<double> > stdMpi(mpiComm);
    stdMpi.prepareCommunication(false);

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

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

    stdMpi.startCommunication();
#endif

#if 1

    std::vector<double*> sendBuffers(sendDofs.size());
    std::vector<double*> recvBuffers(recvDofs.size());

    MPI::Request request[sendDofs.size() + recvDofs.size()];
    int requestCounter = 0;
    
    int i = 0;
    for (RankToDofContainer::iterator sendIt = sendDofs.begin();
	 sendIt != sendDofs.end(); ++sendIt, i++) {
      int nSendDOFs = sendIt->second.size();
      sendBuffers[i] = new double[nSendDOFs * nComponents];

      int counter = 0;
      for (int j = 0; j < nComponents; j++) {
	DOFVector<double> *dofvec = vec.getDOFVector(j);
	for (int k = 0; k < nSendDOFs; k++)
	  sendBuffers[i][counter++] = (*dofvec)[*((sendIt->second)[k])];
      }

      request[requestCounter++] = mpiComm.Isend(sendBuffers[i], nSendDOFs * nComponents,
						MPI_DOUBLE, sendIt->first, 0);
    }

    i = 0;
    for (RankToDofContainer::iterator recvIt = recvDofs.begin();
	 recvIt != recvDofs.end(); ++recvIt, i++) {
      int nRecvDOFs = recvIt->second.size() * nComponents;
      recvBuffers[i] = new double[nRecvDOFs];

      request[requestCounter++] =
	mpiComm.Irecv(recvBuffers[i], nRecvDOFs, MPI_DOUBLE, recvIt->first, 0);
    }


    MPI::Request::Waitall(requestCounter, request);

    i = 0;
    for (RankToDofContainer::iterator recvIt = recvDofs.begin();
	 recvIt != recvDofs.end(); ++recvIt, i++) {
      int nRecvDOFs = recvIt->second.size();

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

      delete [] recvBuffers[i];
    }
    
    for (int i = 0; i < static_cast<int>(sendBuffers.size()); i++)
      delete [] sendBuffers[i];
#endif
  }

  
  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) 
  {
    double localWeightSum = 0.0;
    int elNum = -1;

    elemWeights.clear();

    TraverseStack stack;
    ElInfo *elInfo = stack.traverseFirst(mesh, -1,
					 Mesh::CALL_EVERY_EL_PREORDER);
    while (elInfo) {
      Element *element = elInfo->getElement();

      // get partition data
      PartitionElementData *partitionData = dynamic_cast<PartitionElementData*>
	(element->getElementData(PARTITION_ED));

      if (partitionData && partitionData->getPartitionStatus() == IN) {
	if (partitionData->getLevel() == 0)
	  elNum = element->getIndex();
	
	TEST_EXIT_DBG(elNum != -1)("invalid element number\n");
	if (element->isLeaf()) {
	  elemWeights[elNum] += 1.0;
	  localWeightSum += 1.0;
	}
      }

      elInfo = stack.traverseNext(elInfo);
    }

    return localWeightSum;
  }


  void ParallelDomainBase::partitionMesh(AdaptInfo *adaptInfo)
  {
    if (initialPartitionMesh) {
      initialPartitionMesh = false;
      partitioner->fillCoarsePartitionVec(&oldPartitionVec);
      partitioner->partition(&elemWeights, INITIAL);
    } else {
      oldPartitionVec = partitionVec;
      partitioner->partition(&elemWeights, ADAPTIVE_REPART, 100.0 /*0.000001*/);
    }    

    partitioner->fillCoarsePartitionVec(&partitionVec);
  }


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

    // To be general, we have to use elInfo->getBoundary(EDGE/FACE, i) instead of
    // elInfo->getBoundar(i).
    TEST_EXIT(mesh->getDim() == 2)
      ("getBoundary(i) returns always i-th edge, generalize for 3d!\n");


    // === First, traverse the mesh and search for all elements that have an  ===
    // === boundary with an element within another partition.                 ===

    TraverseStack stack;
    ElInfo *elInfo = 
      stack.traverseFirst(mesh, -1, 
			  Mesh::CALL_LEAF_EL | Mesh::FILL_NEIGH | Mesh::FILL_BOUND);
    while (elInfo) {
      Element *element = elInfo->getElement();

      PartitionElementData *partitionData = 
	dynamic_cast<PartitionElementData*>(element->getElementData(PARTITION_ED));   

      // Check, if the element is within rank's partition.
      if (partitionData->getPartitionStatus() == IN) {
	for (int i = 0; i < 3; i++) {
	  if (!elInfo->getNeighbour(i))
	    continue;

	  PartitionElementData *neighbourPartitionData =
	    dynamic_cast<PartitionElementData*>(elInfo->getNeighbour(i)->getElementData(PARTITION_ED));

 	  if (neighbourPartitionData->getPartitionStatus() == OUT) {

	    // We have found an element that is in rank's partition, but has a 
	    // neighbour outside of the rank's partition.

	    // === Create information about the boundary between the two elements. ===

	    AtomicBoundary bound;	    	    
	    bound.rankObj.el = element;
	    bound.rankObj.elIndex = element->getIndex();
	    bound.rankObj.subObj = EDGE;
	    bound.rankObj.ithObj = i;
	    // Do not set a pointer to the element, because if will be deleted from
	    // mesh after partitioning and the pointer would become invalid.
	    bound.neighObj.el = NULL;
	    bound.neighObj.elIndex = elInfo->getNeighbour(i)->getIndex();
	    bound.neighObj.subObj = EDGE;
	    bound.neighObj.ithObj = elInfo->getSideOfNeighbour(i);
	    
	    // i == 2  =>  getSideOfNeighbour(i) == 2
	    TEST_EXIT_DBG(i != 2 || elInfo->getSideOfNeighbour(i) == 2)
	      ("Should not happen!\n");

	    // Get rank number of the neighbouring element.
	    int otherElementRank = partitionVec[elInfo->getNeighbour(i)->getIndex()];


	    // === Add the boundary information object to the corresponding overall ===
	    // === boundary. There are three possibilities: if the boundary is a    ===
	    // === periodic boundary, just add it to \ref periodicBounadry. Here it ===
	    // === does not matter which rank is responsible for this boundary.     ===
	    // === Otherwise, the boundary is added either to \ref myIntBoundary or ===
	    // === to \ref otherIntBoundary. It dependes on which rank is respon-   ===
	    // === sible for this boundary.                                         ===

	    if (BoundaryManager::isBoundaryPeriodic(elInfo->getBoundary(i))) {	      
	      // We have found an element that is at an interior, periodic boundary.
	      AtomicBoundary& b = periodicBoundary.getNewAtomic(otherElementRank);
	      b = bound;
	    } else {
	      // We have found an element that is at an interior, non-periodic boundary.
	      AtomicBoundary& b = ((mpiRank > otherElementRank) ?
				   myIntBoundary.getNewAtomic(otherElementRank) :
				   otherIntBoundary.getNewAtomic(otherElementRank));
	      b = bound;	      
	    }
 	  }
	}
      }

      elInfo = stack.traverseNext(elInfo);
    }


    // === Once we have this information, we must care about the order of the atomic ===
    // === bounds in the three boundary handling object. Eventually all the bound-   ===
    // === aries have to be in the same order on both ranks that share the bounday.  ===

    std::vector<int*> sendBuffers, recvBuffers;

    // First we communicate myInt/otherIntBoundary, and than the periodic boundaries.
    MPI::Request request[max(myIntBoundary.boundary.size() + 
			     otherIntBoundary.boundary.size(),
			     periodicBoundary.boundary.size())];
    int requestCounter = 0;


    // === The owner of the boundaries send their atomic boundary order to the ===
    // === neighbouring ranks.                                                 ===

    for (RankToBoundMap::iterator rankIt = myIntBoundary.boundary.begin();
	 rankIt != myIntBoundary.boundary.end(); ++rankIt) {    
      int nSendInt = rankIt->second.size();
      int* buffer = new int[nSendInt * 2];
      for (int i = 0; i < nSendInt; i++) {
	buffer[i * 2] = (rankIt->second)[i].rankObj.elIndex;
	buffer[i * 2 + 1] = (rankIt->second)[i].rankObj.ithObj;
      }
      sendBuffers.push_back(buffer);
      
      request[requestCounter++] =
	mpiComm.Isend(buffer, nSendInt * 2, MPI_INT, rankIt->first, 0);
    }

    for (RankToBoundMap::iterator rankIt = otherIntBoundary.boundary.begin();
	 rankIt != otherIntBoundary.boundary.end(); ++rankIt) {
      int nRecvInt = rankIt->second.size();
      int *