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 "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(std::string name,
					 ProblemIterationInterface *iIF,
					 ProblemTimeInterface *tIF,
					 FiniteElemSpace *fe,
					 RefinementManager *refineManager)
    : iterationIF(iIF),
      timeIF(tIF),
      feSpace(fe),
      mesh(fe->getMesh()),
      refinementManager(refineManager),
      initialPartitionMesh(true),
      nRankDOFs(0),
      rstart(0),
      nComponents(1)
  {
    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)
  {
    if (mpiSize <= 1)
      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(rankDOFs);

    // === 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();


    // === Global refinements. ===

    int globalRefinement = 0;
    GET_PARAMETER(0, "testgr", "%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
    }


#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 nor 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);

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

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

      int r = mapLocalGlobalDOFs[*cursor] * dispMult + dispAddRow;

      for (icursor_type icursor = begin<nz>(cursor), 
	     icend = end<nz>(cursor); icursor != icend; ++icursor)
	if (value(*icursor) != 0.0) {
 	  cols.push_back(mapLocalGlobalDOFs[col(*icursor)] * dispMult + dispAddCol);
 	  values.push_back(value(*icursor));
	}

      MatSetValues(petscMatrix, 1, &r, cols.size(), &(cols[0]), &(values[0]), ADD_VALUES);
    }
  }


  void ParallelDomainBase::setDofVector(DOFVector<double>* vec, 
					int dispMult, int dispAdd)
  {
    DOFVector<double>::Iterator dofIt(vec, USED_DOFS);
    for (dofIt.reset(); !dofIt.end(); ++dofIt) {
      int index = mapLocalGlobalDOFs[dofIt.getDOFIndex()] * dispMult + dispAdd;
      double value = *dofIt;

      VecSetValues(petscRhsVec, 1, &index, &value, ADD_VALUES);
    }    
  }


  void ParallelDomainBase::fillPetscMatrix(DOFMatrix *mat, DOFVector<double> *vec)
  {
    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(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(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()");

    KSP ksp;
    PC pc;

    KSPCreate(PETSC_COMM_WORLD, &ksp);
    KSPSetOperators(ksp, petscMatrix, petscMatrix, SAME_NONZERO_PATTERN);
    KSPGetPC(ksp, &pc);
    //PCSetType(pc, PCJACOBI);
    PCSetType(pc, PCILU);
    KSPSetTolerances(ksp, 1.e-7, PETSC_DEFAULT, PETSC_DEFAULT, PETSC_DEFAULT);
    KSPSetType(ksp, KSPBCGS);
    //KSPSetType(ksp, KSPCG);
    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()");

    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);

    KSPSolve(solver, petscRhsVec, petscSolVec);

    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);

    synchVectors(vec);

    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);

    MatDestroy(petscMatrix);
    VecDestroy(petscRhsVec);
    VecDestroy(petscSolVec);
    VecDestroy(petscTmpVec);
    KSPDestroy(solver);
  }
  
  void ParallelDomainBase::synchVectors(SystemVector &vec)
  {
    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];
  }


  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(DofContainer& rankDOFs)
  {
    FUNCNAME("ParallelDomainBase::createInteriorBoundaryInfo()");

    // === First, create all the information about the interior boundaries. ===

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

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

      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 at an interior boundary. 

	    int otherElementRank = partitionVec[elInfo->getNeighbour(i)->getIndex()];
	    bool ranksBoundary = (mpiRank > otherElementRank);

	    AtomicBoundary& bound = 
	      (ranksBoundary ?
	       myIntBoundary.getNewAtomicBoundary(otherElementRank) :
	       otherIntBoundary.getNewAtomicBoundary(otherElementRank));

	    bound.rankObject.el = element;
	    bound.rankObject.elIndex = element->getIndex();
	    bound.rankObject.subObjAtBoundary = EDGE;
	    bound.rankObject.ithObjAtBoundary = 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.neighbourObject.el = NULL;
	    bound.neighbourObject.elIndex = elInfo->getNeighbour(i)->getIndex();
	    bound.neighbourObject.subObjAtBoundary = EDGE;
	    bound.neighbourObject.ithObjAtBoundary = elInfo->getSideOfNeighbour(i);

	    // i == 2  =>  getSideOfNeighbour(i) == 2
	    TEST_EXIT_DBG(i != 2 || elInfo->getSideOfNeighbour(i) == 2)
	      ("Should not happen!\n");
 	  }
	}
      }

      elInfo = stack.traverseNext(elInfo);
    }


    // === Once we have this information, we must care about their order. Eventually ===
    // === all the boundaries have to be in the same order on both ranks that share  ===
    // === the bounday.                                                              ===

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

    MPI::Request request[myIntBoundary.boundary.size() + 
			 otherIntBoundary.boundary.size()];
    int requestCounter = 0;

    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].rankObject.elIndex;
	buffer[i * 2 + 1] = (rankIt->second)[i].rankObject.ithObjAtBoundary;
      }
      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 *buffer = new int[nRecvInt * 2];
      recvBuffers.push_back(buffer);

      request[requestCounter++] = 
	mpiComm.Irecv(buffer, nRecvInt * 2, MPI_INT, rankIt->first, 0);
    }


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


    int i = 0;
    for (RankToBoundMap::iterator rankIt = otherIntBoundary.boundary.begin();
	 rankIt != otherIntBoundary.boundary.end(); ++rankIt) {

      // === We have received from rank "rankIt->first" the ordered list of element ===
      // === indices. We now have to sort the corresponding list in this rank to    ===
      // === get the same order.                                                    ===
      
      for (int j = 0; j < static_cast<int>(rankIt->second.size()); j++) {
	// If the expected object is not at place, search for it.
	if ((rankIt->second)[j].neighbourObject.elIndex != recvBuffers[i][j * 2] ||
	    (rankIt->second)[j].neighbourObject.ithObjAtBoundary != recvBuffers[i][j * 2 + 1]) {
	  int k = j + 1;
	  for (; k < static_cast<int>(rankIt->second.size()); k++)
	    if ((rankIt->second)[k].neighbourObject.elIndex == recvBuffers[i][j * 2] && 
		(rankIt->second)[k].neighbourObject.ithObjAtBoundary == recvBuffers[i][j * 2 + 1])
	      break;

	  // The element must always be found, because the list is just in another order.
	  TEST_EXIT_DBG(k < static_cast<int>(rankIt->second.size()))
	    ("Should never happen!\n");

	  // Swap the current with the found element.
	  AtomicBoundary tmpBound = (rankIt->second)[k];
	  (rankIt->second)[k] = (rankIt->second)[j];
	  (rankIt->second)[j] = tmpBound;	
	}
      }

      delete [] recvBuffers[i++];
    }

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


  void ParallelDomainBase::removeMacroElements()
  {
    std::vector<MacroElement*> macrosToRemove;
    for (std::deque<MacroElement*>::iterator it = mesh->firstMacroElement();
	 it != mesh->endOfMacroElements();
	 ++it) {
      PartitionElementData *partitionData = 
	dynamic_cast<PartitionElementData*>
	((*it)->getElement()->getElementData(PARTITION_ED));
      if (partitionData->getPartitionStatus() != IN)
	macrosToRemove.push_back(*it);
    }

    mesh->removeMacroElements(macrosToRemove, feSpace);
  }


  void ParallelDomainBase::createLocalGlobalNumbering(DofContainer& rankDOFs,
						      int& nRankDOFs, 
						      int& nOverallDOFs)
  {
    FUNCNAME("ParallelDomainBase::createLocalGlobalNumbering()");

    // === Get rank information about DOFs. ===

    // Stores to each DOF pointer the set of ranks the DOF is part of.
    std::map<const DegreeOfFreedom*, std::set<int> > partitionDOFs;
    DofContainer rankAllDofs;
    DofToRank boundaryDofs;

    createDOFMemberInfo(partitionDOFs, rankDOFs, rankAllDofs, boundaryDofs, vertexDof);

    nRankDOFs = rankDOFs.size();
    nOverallDOFs = partitionDOFs.size();


    // === Get starting position for global rank dof ordering. ====

    rstart = 0;
    mpiComm.Scan(&nRankDOFs, &rstart, 1, MPI_INT, MPI_SUM);
    rstart -= nRankDOFs;

    
    // === Create for all dofs in rank new indices. The new index must start at ===
    // === index 0, must be continues and have the same order as the indices    ===
    // === had before.                                                          ===


    // Do not change the indices now, but create a new indexing a store it here.
    DofIndexMap rankDofsNewLocalIndex;
    isRankDof.clear();
    int i = 0;
    for (DofContainer::iterator dofIt = rankAllDofs.begin();
	 dofIt != rankAllDofs.end(); ++dofIt) {
      rankDofsNewLocalIndex[*dofIt] = i;
      // First, we set all dofs in ranks partition to be owend by the rank. Later,
      // the dofs in ranks partition that are owned by other rank are set to false.
      isRankDof[i] = true;
      i++;
    }


    // === Create for all rank owned dofs a new global indexing.                ===

    // Stores for dofs in rank a new global index.
    DofIndexMap rankDofsNewGlobalIndex;
    // Stores for all rank owned dofs a continues local index.
    DofIndexMap rankOwnedDofsNewLocalIndex;

    i = 0;
    for (DofContainer::iterator dofIt = rankDOFs.begin();
	 dofIt != rankDOFs.end(); ++dofIt) {
      rankDofsNewGlobalIndex[*dofIt] = i + rstart;
      rankOwnedDofsNewLocalIndex[*dofIt] = i;
      i++;
    }

 
    // === Create information which dof indices must be send and which must  ===
    // === be received.                                                      ===

    // Stores to each rank a map from DOF pointers (which are all owned by the rank
    // and lie on an interior boundary) to new global DOF indices.
    std::map<int, DofIndexMap > sendNewDofs;

    // Maps to each rank the number of new DOF indices this rank will receive from.
    // All these DOFs are at an interior boundary on this rank, but are owned by
    // another rank.
    std::map<int, int> recvNewDofs;

    for (DofToRank::iterator it = boundaryDofs.begin(); it != boundaryDofs.end(); ++it) {

      if (it->second == mpiRank) {
	// If the boundary dof is a rank dof, it must be send to other ranks.

	// Search for all ranks that have this dof too.
	for (std::set<int>::iterator itRanks = partitionDOFs[it->first].begin();
	     itRanks != partitionDOFs[it->first].end();
	     ++itRanks) {
	  if (*itRanks != mpiRank) {
	    TEST_EXIT_DBG(rankDofsNewGlobalIndex.count(it->first) == 1)
	      ("DOF Key not found!\n");

	    sendNewDofs[*itRanks][it->first] = rankDofsNewGlobalIndex[it->first];
	  }
	}
      } else {
	// If the boundary dof is not a rank dof, its new dof index (and later
	// also the dof values) must be received from another rank.
	if (recvNewDofs.find(it->second) == recvNewDofs.end()) 
	  recvNewDofs[it->second] = 1;
	else
	  recvNewDofs[it->second]++;
      }
    }


    // === Send and receive the dof indices at boundary. ===

    std::vector<int*> sendBuffers(sendNewDofs.size()), recvBuffers(recvNewDofs.size());

    sendDofs.clear();
    recvDofs.clear();

#if 0
    StlMpi stlMpi(mpiComm);
    stlMpi.prepareCommunication(false);
    std::map<

    for (std::map<int, DofIndexMap >::iterator sendIt = sendNewDofs.begin();
	 sendIt != sendNewDofs.end(); ++sendIt, i++)
      stlMpi.send(it->second, it->first);

    for (std::map<int, int>::iterator recvIt = recvNewDofs.begin();
	 recvIt != recvNewDofs.end(); ++recvIt, i++)
      stlMpi.recv(it->first, it->second * 2); 

    stlMpi.startCommunication();
#endif
    
    MPI::Request request[sendNewDofs.size() + recvNewDofs.size()];
    int requestCounter = 0;

    i = 0;
    for (std::map<int, DofIndexMap >::iterator sendIt = sendNewDofs.begin();
	 sendIt != sendNewDofs.end(); ++sendIt, i++) {
      int nSendDofs = sendIt->second.size() * 2;
      sendBuffers[i] = new int[nSendDofs];

      int c = 0;
      for (DofIndexMap::iterator dofIt = sendIt->second.begin();
	   dofIt != sendIt->second.end(); ++dofIt) {
	sendBuffers[i][c++] = *(dofIt->first);
	sendBuffers[i][c++] = dofIt->second;

	sendDofs[sendIt->first].push_back(dofIt->first);
      }

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

    i = 0;
    for (std::map<int, int>::iterator recvIt = recvNewDofs.begin();
	 recvIt != recvNewDofs.end(); ++recvIt, i++) {
      int nRecvDOFs = recvIt->second * 2;
      recvBuffers[i] = new int[nRecvDOFs];

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


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

    
    // === Delete send buffers. ===

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


    // === Change dof indices at boundary from other ranks. ===

    // Within this small data structure we track which dof index was already changed.
    // This is used to avoid the following situation: Assume, there are two dof indices
    // a and b in boundaryDofs. Then we have to change index a to b and b to c. When
    // the second rule applies, we have to avoid that not the first b, resulted from
    // changing a to b, is set to c, but the second one. Therefore, after the first
    // rule was applied, the dof pointer is set to false in this data structure and 
    // is not allowed to be changed anymore.
    std::map<const DegreeOfFreedom*, bool> dofChanged;
    for (DofToRank::iterator dofIt = boundaryDofs.begin(); dofIt != boundaryDofs.end();
	 ++dofIt)
      dofChanged[dofIt->first] = false;

    i = 0;
    for (std::map<int, int>::iterator recvIt = recvNewDofs.begin();
	 recvIt != recvNewDofs.end();
	 ++recvIt, i++) {

      for (int j = 0; j < recvIt->second; j++) {

	DegreeOfFreedom oldDof = recvBuffers[i][j * 2];
	DegreeOfFreedom newGlobalDof = recvBuffers[i][j * 2 + 1];

	bool found = false;

	// Iterate over all boundary dofs to find the dof, which index we have to change.

	for (DofToRank::iterator dofIt = boundaryDofs.begin(); 
	     dofIt != boundaryDofs.end(); ++dofIt) {

	  if (*(dofIt->first) == oldDof && !dofChanged[dofIt->first]) {
	    dofChanged[dofIt->first] = true;

	    recvDofs[recvIt->first].push_back(dofIt->first);