MueLu_RepartitionHeuristicFactory_def.hpp

Go to the documentation of this file.

// @HEADER
// *****************************************************************************
//        MueLu: A package for multigrid based preconditioning
//
// Copyright 2012 NTESS and the MueLu contributors.
// SPDX-License-Identifier: BSD-3-Clause
// *****************************************************************************
// @HEADER

#ifndef PACKAGES_MUELU_SRC_REBALANCING_MUELU_REPARTITIONHEURISTICFACTORY_DEF_HPP_
#define PACKAGES_MUELU_SRC_REBALANCING_MUELU_REPARTITIONHEURISTICFACTORY_DEF_HPP_

#include <algorithm>
#include <iostream>
#include <sstream>

#ifdef HAVE_MPI
#include <Teuchos_DefaultMpiComm.hpp>
#include <Teuchos_CommHelpers.hpp>

//#include <Xpetra_Map.hpp>
#include <Xpetra_Matrix.hpp>

#include "MueLu_RAPFactory.hpp"
#include "MueLu_BlockedRAPFactory.hpp"
#include "MueLu_SubBlockAFactory.hpp"
#include "MueLu_Level.hpp"
#include "MueLu_MasterList.hpp"
#include "MueLu_Monitor.hpp"

#include "MueLu_RepartitionHeuristicFactory_decl.hpp"

namespace MueLu {

template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
RepartitionHeuristicFactory<Scalar, LocalOrdinal, GlobalOrdinal, Node>::RepartitionHeuristicFactory() = default;

template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
RepartitionHeuristicFactory<Scalar, LocalOrdinal, GlobalOrdinal, Node>::~RepartitionHeuristicFactory() = default;

template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
RCP<const ParameterList> RepartitionHeuristicFactory<Scalar, LocalOrdinal, GlobalOrdinal, Node>::GetValidParameterList() const {
  RCP<ParameterList> validParamList = rcp(new ParameterList());

#define SET_VALID_ENTRY(name) validParamList->setEntry(name, MasterList::getEntry(name))
  SET_VALID_ENTRY("repartition: start level");
  SET_VALID_ENTRY("repartition: use map");
  SET_VALID_ENTRY("repartition: node repartition level");
  SET_VALID_ENTRY("repartition: min rows per proc");
  SET_VALID_ENTRY("repartition: target rows per proc");
  SET_VALID_ENTRY("repartition: min rows per thread");
  SET_VALID_ENTRY("repartition: target rows per thread");
  SET_VALID_ENTRY("repartition: max imbalance");
  SET_VALID_ENTRY("repartition: put on single proc");
#undef SET_VALID_ENTRY

  validParamList->set<RCP<const FactoryBase> >("A", Teuchos::null, "Factory of the matrix A");
  validParamList->set<RCP<const FactoryBase> >("Map", Teuchos::null, "Factory of the map Map");
  validParamList->set<RCP<const FactoryBase> >("Node Comm", Teuchos::null, "Generating factory of the node level communicator");

  return validParamList;
}

template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
void RepartitionHeuristicFactory<Scalar, LocalOrdinal, GlobalOrdinal, Node>::DeclareInput(Level& currentLevel) const {
  const Teuchos::ParameterList& pL = GetParameterList();
  if (pL.isParameter("repartition: use map")) {
    const bool useMap = pL.get<bool>("repartition: use map");
    if (useMap)
      Input(currentLevel, "Map");
    else
      Input(currentLevel, "A");
  } else
    Input(currentLevel, "A");
  if (pL.isParameter("repartition: node repartition level")) {
    const int nodeRepartLevel = pL.get<int>("repartition: node repartition level");
    if (currentLevel.GetLevelID() == nodeRepartLevel) {
      Input(currentLevel, "Node Comm");
    }
  }
}

template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
void RepartitionHeuristicFactory<Scalar, LocalOrdinal, GlobalOrdinal, Node>::Build(Level& currentLevel) const {
  FactoryMonitor m(*this, "Build", currentLevel);

  const Teuchos::ParameterList& pL = GetParameterList();
  // Access parameters here to make sure that we set the parameter entry flag to "used" even in case of short-circuit evaluation.
  // TODO (JG): I don't really know if we want to do this.
  const int startLevel          = pL.get<int>("repartition: start level");
  const int nodeRepartLevel     = pL.get<int>("repartition: node repartition level");
  LO minRowsPerProcess          = pL.get<LO>("repartition: min rows per proc");
  LO targetRowsPerProcess       = pL.get<LO>("repartition: target rows per proc");
  LO minRowsPerThread           = pL.get<LO>("repartition: min rows per thread");
  LO targetRowsPerThread        = pL.get<LO>("repartition: target rows per thread");
  LO putOnSingleProc            = pL.get<LO>("repartition: put on single proc");
  const double nonzeroImbalance = pL.get<double>("repartition: max imbalance");
  const bool useMap             = pL.get<bool>("repartition: use map");

  int thread_per_mpi_rank = 1;
#if defined(KOKKOS_ENABLE_OPENMP)
  using execution_space = typename Node::device_type::execution_space;
  if (std::is_same<execution_space, Kokkos::OpenMP>::value)
    thread_per_mpi_rank = execution_space().concurrency();
#endif

  if (minRowsPerThread > 0)
    // We ignore the value given by minRowsPerProcess and repartition based on threads instead
    minRowsPerProcess = minRowsPerThread * thread_per_mpi_rank;

  if (targetRowsPerThread == 0)
    targetRowsPerThread = minRowsPerThread;

  if (targetRowsPerThread > 0)
    // We ignore the value given by targetRowsPerProcess and repartition based on threads instead
    targetRowsPerProcess = targetRowsPerThread * thread_per_mpi_rank;

  if (targetRowsPerProcess == 0)
    targetRowsPerProcess = minRowsPerProcess;

  // Stick this on the level so Zoltan2Interface can use this later
  Set<LO>(currentLevel, "repartition: heuristic target rows per process", targetRowsPerProcess);

  // Check for validity of the node repartition option
  TEUCHOS_TEST_FOR_EXCEPTION(nodeRepartLevel >= startLevel, Exceptions::RuntimeError, "MueLu::RepartitionHeuristicFactory::Build(): If 'repartition: node repartition level' is set, it must be less than or equal to 'repartition: start level'");

  RCP<Matrix> A;
  RCP<const FactoryBase> Afact;
  RCP<const Map> map;
  if (!useMap) {
    Afact = GetFactory("A");
    if (!Afact.is_null() && Teuchos::rcp_dynamic_cast<const RAPFactory>(Afact) == Teuchos::null &&
        Teuchos::rcp_dynamic_cast<const BlockedRAPFactory>(Afact) == Teuchos::null &&
        Teuchos::rcp_dynamic_cast<const SubBlockAFactory>(Afact) == Teuchos::null) {
      GetOStream(Warnings) << "MueLu::RepartitionHeuristicFactory::Build: The generation factory for A must "
                              "be a RAPFactory or a SubBlockAFactory providing the non-rebalanced matrix information! "
                              "It specifically must not be of type Rebalance(Blocked)AcFactory or similar. "
                              "Please check the input. Make also sure that \"number of partitions\" is provided to "
                              "the Interface class and the RepartitionFactory instance.  Instead, we have a "
                           << Afact->description() << std::endl;
    }
    // TODO: We only need a CrsGraph. This class does not have to be templated on Scalar types.
    A   = Get<RCP<Matrix> >(currentLevel, "A");
    map = A->getRowMap();
  } else
    map = Get<RCP<const Map> >(currentLevel, "Map");

  // ======================================================================================================
  // Determine whether partitioning is needed
  // ======================================================================================================
  // NOTE: most tests include some global communication, which is why we currently only do tests until we make
  // a decision on whether to repartition. However, there is value in knowing how "close" we are to having to
  // rebalance an operator. So, it would probably be beneficial to do and report *all* tests.

  // NOTE: We should probably consider merging all of these sumAll/minAll things at some point.

  // Test0: Should we do node repartitioning?
  if (currentLevel.GetLevelID() == nodeRepartLevel && map->getComm()->getSize() > 1) {
    RCP<const Teuchos::Comm<int> > NodeComm = Get<RCP<const Teuchos::Comm<int> > >(currentLevel, "Node Comm");
    TEUCHOS_TEST_FOR_EXCEPTION(NodeComm.is_null(), Exceptions::RuntimeError, "MueLu::RepartitionHeuristicFactory::Build(): NodeComm is null.");

    // If we only have one node, then we don't want to pop down to one rank
    if (NodeComm()->getSize() != map->getComm()->getSize()) {
      GetOStream(Statistics1) << "Repartitioning?  YES: \n  Within node only" << std::endl;
      int nodeRank = NodeComm->getRank();

      // Do a reduction to get the total number of nodes
      int isZero   = (nodeRank == 0);
      int numNodes = 0;
      Teuchos::reduceAll(*map->getComm(), Teuchos::REDUCE_SUM, isZero, Teuchos::outArg(numNodes));
      Set(currentLevel, "number of partitions", numNodes);
      return;
    }
  }

  // Test1: skip repartitioning if current level is less than the specified minimum level for repartitioning
  if (currentLevel.GetLevelID() < startLevel) {
    GetOStream(Statistics1) << "Repartitioning?  NO:"
                            << "\n  current level = " << Teuchos::toString(currentLevel.GetLevelID()) << ", first level where repartitioning can happen is " + Teuchos::toString(startLevel) << std::endl;

    // a negative number of processors means: no repartitioning
    Set(currentLevel, "number of partitions", -1);

    return;
  }

  RCP<const Teuchos::Comm<int> > origComm = map->getComm();
  RCP<const Teuchos::Comm<int> > comm     = origComm;

  // Test 2: check whether A is actually distributed, i.e. more than one processor owns part of A
  // TODO: this global communication can be avoided if we store the information with the matrix (it is known when matrix is created)
  // TODO: further improvements could be achieved when we use subcommunicator for the active set. Then we only need to check its size

  // TODO: The block transfer factories do not check correctly whether or not repartitioning actually took place.
  {
    if (comm->getSize() == 1 && Teuchos::rcp_dynamic_cast<const RAPFactory>(Afact) != Teuchos::null) {
      GetOStream(Statistics1) << "Repartitioning?  NO:"
                              << "\n  comm size = 1" << std::endl;

      Set(currentLevel, "number of partitions", -1);
      return;
    }

    int numActiveProcesses = 0;
    MueLu_sumAll(comm, Teuchos::as<int>((map->getLocalNumElements() > 0) ? 1 : 0), numActiveProcesses);

    if (numActiveProcesses == 1) {
      GetOStream(Statistics1) << "Repartitioning?  NO:"
                              << "\n  # processes with rows = " << Teuchos::toString(numActiveProcesses) << std::endl;

      Set(currentLevel, "number of partitions", 1);
      return;
    }
  }

  // Test 2.5: Repartition down to one rank if we're below "repatition: put on single proc"
  // This is (mostly) an ML-compatibility thing
  if (putOnSingleProc && map->getGlobalNumElements() < (Xpetra::global_size_t)putOnSingleProc) {
    GetOStream(Statistics1) << "Repartitioning? YES:"
                            << "\n # rows is below the single-proc threshold = " << putOnSingleProc << std::endl;

    Set(currentLevel, "number of partitions", 1);
    return;
  }

  bool test3 = false, test4 = false;
  std::string msg3, msg4;

  // Test3: check whether number of rows on any processor satisfies the minimum number of rows requirement
  // NOTE: Test2 ensures that repartitionning is not done when there is only one processor (it may or may not satisfy Test3)
  if (minRowsPerProcess > 0) {
    LO numMyRows = Teuchos::as<LO>(map->getLocalNumElements()), minNumRows, LOMAX = Teuchos::OrdinalTraits<LO>::max();
    LO haveFewRows = (numMyRows < minRowsPerProcess ? 1 : 0), numWithFewRows = 0;
    MueLu_sumAll(comm, haveFewRows, numWithFewRows);
    MueLu_minAll(comm, (numMyRows > 0 ? numMyRows : LOMAX), minNumRows);

    // TODO: we could change it to repartition only if the number of processors with numRows < minNumRows is larger than some
    // percentage of the total number. This way, we won't repartition if 2 out of 1000 processors don't have enough elements.
    // I'm thinking maybe 20% threshold. To implement, simply add " && numWithFewRows < .2*numProcs" to the if statement.
    if (numWithFewRows > 0)
      test3 = true;

    msg3 = "\n  min # rows per proc = " + Teuchos::toString(minNumRows) + ", min allowable = " + Teuchos::toString(minRowsPerProcess);
  }

  // Test4: check whether the balance in the number of nonzeros per processor is greater than threshold
  if (!test3) {
    if (useMap)
      msg4 = "";
    else {
      GO minNnz, maxNnz, numMyNnz = Teuchos::as<GO>(A->getLocalNumEntries());
      MueLu_maxAll(comm, numMyNnz, maxNnz);
      MueLu_minAll(comm, (numMyNnz > 0 ? numMyNnz : maxNnz), minNnz);  // min nnz over all active processors
      double imbalance = Teuchos::as<double>(maxNnz) / minNnz;

      if (imbalance > nonzeroImbalance)
        test4 = true;

      msg4 = "\n  nonzero imbalance = " + Teuchos::toString(imbalance) + ", max allowable = " + Teuchos::toString(nonzeroImbalance);
    }
  }

  if (!test3 && !test4) {
    GetOStream(Statistics1) << "Repartitioning?  NO:" << msg3 + msg4 << std::endl;

    // A negative number of partitions means: no repartitioning
    Set(currentLevel, "number of partitions", -1);
    return;
  }

  GetOStream(Statistics1) << "Repartitioning? YES:" << msg3 + msg4 << std::endl;

  // ======================================================================================================
  // Calculate number of partitions
  // ======================================================================================================
  // FIXME Quick way to figure out how many partitions there should be (same algorithm as ML)
  // FIXME Should take into account nnz? Perhaps only when user is using min #nnz per row threshold.

  // The number of partitions is calculated by the RepartitionFactory and stored in "number of partitions" variable on
  // the current level. If this variable is already set (e.g., by another instance of RepartitionFactory) then this number
  // is used. The "number of partitions" variable serves as basic communication between the RepartitionFactory (which
  // requests a certain number of partitions) and the *Interface classes which call the underlying partitioning algorithms
  // and produce the "Partition" array with the requested number of partitions.
  const auto globalNumRows = Teuchos::as<GO>(map->getGlobalNumElements());
  int numPartitions        = 1;
  if (globalNumRows >= targetRowsPerProcess) {
    // Make sure that each CPU thread has approximately targetRowsPerProcess
    numPartitions = std::max(Teuchos::as<int>(globalNumRows / targetRowsPerProcess), 1);
  }
  numPartitions = std::min(numPartitions, comm->getSize());

  Set(currentLevel, "number of partitions", numPartitions);

  GetOStream(Statistics1) << "Number of partitions to use = " << numPartitions << std::endl;
}  // Build
}  // namespace MueLu

#endif  // ifdef HAVE_MPI
#endif  /* PACKAGES_MUELU_SRC_REBALANCING_MUELU_REPARTITIONHEURISTICFACTORY_DEF_HPP_ */