repartitioning_policy_tools.cc

Go to the documentation of this file.

// ------------------------------------------------------------------------
//
// SPDX-License-Identifier: LGPL-2.1-or-later
// Copyright (C) 2021 - 2023 by the deal.II authors
//
// This file is part of the deal.II library.
//
// Part of the source code is dual licensed under Apache-2.0 WITH
// LLVM-exception OR LGPL-2.1-or-later. Detailed license information
// governing the source code and code contributions can be found in
// LICENSE.md and CONTRIBUTING.md at the top level directory of deal.II.
//
// ------------------------------------------------------------------------
 
 
#include <deal.II/base/mpi_compute_index_owner_internal.h>
#include <deal.II/base/mpi_consensus_algorithms.h>
 
#include <deal.II/distributed/repartitioning_policy_tools.h>
#include <deal.II/distributed/tria_base.h>
 
#include <deal.II/grid/cell_id_translator.h>
#include <deal.II/grid/filtered_iterator.h>
 
DEAL_II_NAMESPACE_OPEN
 
 
namespace RepartitioningPolicyTools
{
  namespace
  {
    template <int dim, int spacedim>
    void
    add_indices_recursively_for_first_child_policy(
      const TriaIterator<CellAccessor<dim, spacedim>> &cell,
      const internal::CellIDTranslator<dim>           &cell_id_translator,
      IndexSet                                        &is_fine)
    {
      is_fine.add_index(cell_id_translator.translate(cell));
 
      if (cell->level() > 0 && cell->parent()->child(0) == cell)
        add_indices_recursively_for_first_child_policy(cell->parent(),
                                                       cell_id_translator,
                                                       is_fine);
    }
  } // namespace
 
 
  template <int dim, int spacedim>
  DefaultPolicy<dim, spacedim>::DefaultPolicy(const bool tighten)
    : tighten(tighten)
  {}
 
  template <int dim, int spacedim>
  LinearAlgebra::distributed::Vector<double>
  DefaultPolicy<dim, spacedim>::partition(
    const Triangulation<dim, spacedim> &tria_in) const
  {
    if (tighten == false)
      return {}; // nothing to do
 
    const auto tria =
      dynamic_cast<const parallel::TriangulationBase<dim, spacedim> *>(
        &tria_in);
 
    if (tria == nullptr)
      return {}; // nothing to do, since serial triangulation
 
#ifndef DEAL_II_WITH_MPI
    Assert(false, ExcNeedsMPI());
    return {};
#else
 
    const auto comm = tria->get_communicator();
 
    const unsigned int process_has_active_locally_owned_cells =
      tria->n_locally_owned_active_cells() > 0;
    const unsigned int n_processes_with_active_locally_owned_cells =
      Utilities::MPI::sum(process_has_active_locally_owned_cells, comm);
 
    if (n_processes_with_active_locally_owned_cells ==
        Utilities::MPI::n_mpi_processes(comm))
      return {}; // nothing to do, since all processes have cells
 
    unsigned int offset = 0;
 
    const int ierr =
      MPI_Exscan(&process_has_active_locally_owned_cells,
                 &offset,
                 1,
                 Utilities::MPI::mpi_type_id_for_type<
                   decltype(process_has_active_locally_owned_cells)>,
                 MPI_SUM,
                 comm);
    AssertThrowMPI(ierr);
 
    LinearAlgebra::distributed::Vector<double> partition(
      tria->global_active_cell_index_partitioner().lock());
 
    partition = offset;
 
    return partition;
#endif
  }
 
 
 
  template <int dim, int spacedim>
  FirstChildPolicy<dim, spacedim>::FirstChildPolicy(
    const Triangulation<dim, spacedim> &tria_fine)
    : n_coarse_cells(tria_fine.n_global_coarse_cells())
    , n_global_levels(tria_fine.n_global_levels())
  {
    Assert(
      tria_fine.all_reference_cells_are_hyper_cube(),
      ExcMessage(
        "FirstChildPolicy is only working for pure hex meshes at the moment."));
 
    const internal::CellIDTranslator<dim> cell_id_translator(n_coarse_cells,
                                                             n_global_levels);
    is_level_partitions.set_size(cell_id_translator.size());
 
    for (const auto &cell : tria_fine.active_cell_iterators())
      if (cell->is_locally_owned())
        add_indices_recursively_for_first_child_policy(cell,
                                                       cell_id_translator,
                                                       is_level_partitions);
  }
 
 
 
  template <int dim, int spacedim>
  LinearAlgebra::distributed::Vector<double>
  FirstChildPolicy<dim, spacedim>::partition(
    const Triangulation<dim, spacedim> &tria_coarse_in) const
  {
    const auto communicator = tria_coarse_in.get_communicator();
 
    const internal::CellIDTranslator<dim> cell_id_translator(n_coarse_cells,
                                                             n_global_levels);
 
    IndexSet is_coarse(cell_id_translator.size());
 
    for (const auto &cell : tria_coarse_in.active_cell_iterators())
      if (cell->is_locally_owned())
        is_coarse.add_index(cell_id_translator.translate(cell));
 
    std::vector<unsigned int> owning_ranks_of_coarse_cells(
      is_coarse.n_elements());
    {
      Utilities::MPI::internal::ComputeIndexOwner::ConsensusAlgorithmsPayload
        process(is_level_partitions,
                is_coarse,
                communicator,
                owning_ranks_of_coarse_cells,
                false);
 
      Utilities::MPI::ConsensusAlgorithms::Selector<
        std::vector<
          std::pair<types::global_cell_index, types::global_cell_index>>,
        std::vector<unsigned int>>
        consensus_algorithm;
      consensus_algorithm.run(process, communicator);
    }
 
    const auto tria =
      dynamic_cast<const parallel::TriangulationBase<dim, spacedim> *>(
        &tria_coarse_in);
 
    Assert(tria, ExcNotImplemented());
 
    LinearAlgebra::distributed::Vector<double> partition(
      tria->global_active_cell_index_partitioner().lock());
 
    for (const auto &cell : tria_coarse_in.active_cell_iterators() |
                              IteratorFilters::LocallyOwnedCell())
      partition[cell->global_active_cell_index()] =
        owning_ranks_of_coarse_cells[is_coarse.index_within_set(
          cell_id_translator.translate(cell))];
 
    return partition;
  }
 
 
  template <int dim, int spacedim>
  MinimalGranularityPolicy<dim, spacedim>::MinimalGranularityPolicy(
    const unsigned int n_min_cells)
    : n_min_cells(n_min_cells)
  {}
 
 
 
  template <int dim, int spacedim>
  LinearAlgebra::distributed::Vector<double>
  MinimalGranularityPolicy<dim, spacedim>::partition(
    const Triangulation<dim, spacedim> &tria_in) const
  {
    const auto tria =
      dynamic_cast<const parallel::TriangulationBase<dim, spacedim> *>(
        &tria_in);
 
    Assert(tria, ExcNotImplemented());
 
    LinearAlgebra::distributed::Vector<double> partition(
      tria->global_active_cell_index_partitioner().lock());
 
    // step 1) check if all processes have enough cells
 
    const unsigned int n_locally_owned_active_cells =
      std::count_if(tria_in.begin_active(),
                    typename Triangulation<dim, spacedim>::active_cell_iterator(
                      tria_in.end()),
                    [](const auto &cell) { return cell.is_locally_owned(); });
 
    const auto comm = tria_in.get_communicator();
 
    if (Utilities::MPI::min(n_locally_owned_active_cells, comm) >= n_min_cells)
      return {}; // all processes have enough cells
 
    // step 2) there are processes which do not have enough cells so that
    // a repartitioning kicks in with the aim that all processes that own
    // cells have at least the specified number of cells
 
    const types::global_cell_index n_global_active_cells =
      tria_in.n_global_active_cells();
 
    const unsigned int n_partitions =
      std::max<unsigned int>(1,
                             std::min<types::global_cell_index>(
                               n_global_active_cells / n_min_cells,
                               Utilities::MPI::n_mpi_processes(comm)));
 
    const unsigned int min_cells = n_global_active_cells / n_partitions;
 
    const auto convert = [&](const unsigned int i) {
      // determine the owner of a given index: we try to assign each process
      // the same number of cells; if there is a remainder, we assign the
      // first processes an additional cell (so that the difference of number of
      // locally owned cells is never larger than one between processes).
      const unsigned int n_partitions_with_additional_cell =
        n_global_active_cells - min_cells * n_partitions;
 
      const unsigned int rank =
        (i < (min_cells + 1) * n_partitions_with_additional_cell) ?
          (i / (min_cells + 1)) :
          ((i - n_partitions_with_additional_cell) / min_cells);
 
      AssertIndexRange(rank, n_partitions);
 
      return rank;
    };
 
    for (const auto i : partition.locally_owned_elements())
      partition[i] = convert(i);
 
    return partition;
  }
 
 
 
  template <int dim, int spacedim>
  CellWeightPolicy<dim, spacedim>::CellWeightPolicy(
    const std::function<
      unsigned int(const typename Triangulation<dim, spacedim>::cell_iterator &,
                   const CellStatus)> &weighting_function)
    : weighting_function(weighting_function)
  {}
 
 
 
  template <int dim, int spacedim>
  LinearAlgebra::distributed::Vector<double>
  CellWeightPolicy<dim, spacedim>::partition(
    const Triangulation<dim, spacedim> &tria_in) const
  {
#ifndef DEAL_II_WITH_MPI
    (void)tria_in;
    return {};
#else
 
    const auto tria =
      dynamic_cast<const parallel::TriangulationBase<dim, spacedim> *>(
        &tria_in);
 
    Assert(tria, ExcNotImplemented());
 
    const auto partitioner =
      tria->global_active_cell_index_partitioner().lock();
 
    std::vector<unsigned int> weights(partitioner->locally_owned_size());
 
    const auto mpi_communicator = tria_in.get_communicator();
    const auto n_subdomains = Utilities::MPI::n_mpi_processes(mpi_communicator);
 
    // determine weight of each cell
    for (const auto &cell :
         tria->active_cell_iterators() | IteratorFilters::LocallyOwnedCell())
      weights[partitioner->global_to_local(cell->global_active_cell_index())] =
        weighting_function(cell, CellStatus::cell_will_persist);
 
    // determine weight of all the cells locally owned by this process
    std::uint64_t process_local_weight = 0;
    for (const auto &weight : weights)
      process_local_weight += weight;
 
    // determine partial sum of weights of this process, as well as the total
    // weight
    const auto [process_local_weight_offset, total_weight] =
      Utilities::MPI::partial_and_total_sum(process_local_weight,
                                            tria->get_communicator());
 
    // set up partition
    LinearAlgebra::distributed::Vector<double> partition(partitioner);
 
    for (std::uint64_t i = 0, weight = process_local_weight_offset;
         i < partition.locally_owned_size();
         weight += weights[i], ++i)
      partition.local_element(i) =
        static_cast<double>(weight * n_subdomains / total_weight);
 
    return partition;
#endif
  }
 
 
} // namespace RepartitioningPolicyTools
 
 
 
/*-------------- Explicit Instantiations -------------------------------*/
#include "repartitioning_policy_tools.inst"
 
DEAL_II_NAMESPACE_CLOSE