MeltPoolDG::CutUtil::PartitionWeights< dim > Class Template Reference

Cell weights for the repartitioning of a parallel::distributed::Triangulation according to the location of the cells with respect to the level set. More...

#include <repartition.hpp>

Collaboration diagram for MeltPoolDG::CutUtil::PartitionWeights< dim >:

Public Types
using	CellIterator = typename dealii::Triangulation< dim >::cell_iterator

Public Member Functions
	PartitionWeights (const dealii::Triangulation< dim > &tria_in, const CutRepartitionData &repartition_data_in)
	Constructor.
void	update (const dealii::NonMatching::MeshClassifier< dim > &mesh_classifier)
	Compute and cache the weights of the locally owned active cells according to their location with respect to the level set.
unsigned int	get_weight (const CellIterator &cell) const
	Get the cached weight of a cell.
dealii::Utilities::MPI::MinMaxAvg	compute_load_statistics (const MPI_Comm mpi_comm) const
	Compute statistics of the summed cell weights (i.e. the estimated load) of the locally owned cells over all MPI processes.

Private Member Functions
void	connect_to_triangulation ()
	Connect the function returning the cached weights to the weight signal of the triangulation, and the transfer of the cached weights to the signals before and after refinement and repartitioning.
void	prepare_transfer ()
	Attach the cached weights to the cells before the triangulation changes.
void	finish_transfer ()
	Unpack the cached weights after the triangulation has changed.
unsigned int	get_partition_weight (const dealii::NonMatching::LocationToLevelSet location) const
	Get the partitioning weight of a cell according to its location with respect to the level set.

Private Attributes
const dealii::Triangulation< dim > *	tria
	Pointer to the triangulation.
const CutRepartitionData	repartition_data
	Repartitioning parameters including the cell weights.
std::vector< unsigned int >	weights
	Cached weights of the active cells (indexed by the active cell index)
unsigned int	weights_handle = dealii::numbers::invalid_unsigned_int
	Handle of the cell-attached weights during a change of the triangulation.
std::vector< boost::signals2::scoped_connection >	connections
	Connections to the signals of the triangulation.

Detailed Description

template<int dim>
class MeltPoolDG::CutUtil::PartitionWeights< dim >

Cell weights for the repartitioning of a parallel::distributed::Triangulation according to the location of the cells with respect to the level set.

The weights are evaluated with a dealii::NonMatching::MeshClassifier in update() and cached per active cell. Upon the first call of update(), a function returning the cached weights is connected to dealii::Triangulation::Signals::weight. It stays connected for the lifetime of this object.

With the default settings of parallel::distributed::Triangulation, p4est partitions the mesh according to the connected weights not only in an explicit repartitioning (see repartition_triangulation()), but in every call to execute_coarsening_and_refinement(). Such calls also occur if only the active FE indices change, e.g. in the solution transfer for a moved interface (see CutUtil::SolutionTransferOperator), where the level-set vector is not transferred. In these calls, no cells must be moved between the processes. Since the partitioning computed by p4est only depends on the global order of the cells and their weights, this is the case if the connected weights are exactly the ones p4est used for the current partitioning. Therefore, the weights are cached instead of evaluating the mesh classifier on the fly, and the cache is kept consistent with the partitioning as follows:

• update() computes new weights from the current cell classification. They are used by p4est in the next change of the triangulation. Hence, update() must be called directly before an explicit repartitioning (see repartition_triangulation()).
• The cached weights are transferred automatically in every change of the triangulation, i.e. in every call to execute_coarsening_and_refinement() and repartition(), exactly as p4est sees them: The children of a refined cell inherit the weight of their parent, and a coarsened cell gets the maximum weight of its former children (see get_weight()). This is implemented by attaching the weights to the cells (see dealii::Triangulation::register_data_attach()) in the signals pre_distributed_refinement and pre_distributed_repartition, and by unpacking them in the signals post_distributed_refinement and post_distributed_repartition of the triangulation.

Consequently, after adaptive mesh refinement the cached weights do not correspond to a fresh classification of the new cells, but they reproduce the partitioning computed during the refinement. They are updated in the next explicit repartitioning.

Note

update() must be called directly before each explicit repartitioning.

Member Typedef Documentation

CellIterator

template<int dim>

using MeltPoolDG::CutUtil::PartitionWeights< dim >::CellIterator = typename dealii::Triangulation<dim>::cell_iterator

Constructor & Destructor Documentation

PartitionWeights()

template<int dim>

MeltPoolDG::CutUtil::PartitionWeights< dim >::PartitionWeights ( const dealii::Triangulation< dim > &  tria_in, const CutRepartitionData &  repartition_data_in  )

inlineexplicit

Constructor.

Parameters

tria_in	Triangulation to be repartitioned.
repartition_data_in	Repartitioning parameters including the cell weights.

Member Function Documentation

compute_load_statistics()

template<int dim>

dealii::Utilities::MPI::MinMaxAvg MeltPoolDG::CutUtil::PartitionWeights< dim >::compute_load_statistics ( const MPI_Comm  mpi_comm ) const

inline

Compute statistics of the summed cell weights (i.e. the estimated load) of the locally owned cells over all MPI processes.

Parameters

mpi_comm

MPI communicator.

Returns

Minimum, maximum and average load per MPI process.

connect_to_triangulation()

template<int dim>

void MeltPoolDG::CutUtil::PartitionWeights< dim >::connect_to_triangulation ( )

inlineprivate

Connect the function returning the cached weights to the weight signal of the triangulation, and the transfer of the cached weights to the signals before and after refinement and repartitioning.

finish_transfer()

template<int dim>

void MeltPoolDG::CutUtil::PartitionWeights< dim >::finish_transfer ( )

inlineprivate

Unpack the cached weights after the triangulation has changed.

The children of a refined cell inherit the weight of their parent.

get_partition_weight()

template<int dim>

unsigned int MeltPoolDG::CutUtil::PartitionWeights< dim >::get_partition_weight ( const dealii::NonMatching::LocationToLevelSet  location ) const

inlineprivate

Get the partitioning weight of a cell according to its location with respect to the level set.

Parameters

location

Location of the cell with respect to the level set.

Returns

Partitioning weight of the cell.

get_weight()

template<int dim>

unsigned int MeltPoolDG::CutUtil::PartitionWeights< dim >::get_weight ( const CellIterator &  cell ) const

inline

Get the cached weight of a cell.

Note

For cells whose children will be coarsened, the maximum weight of the children is returned.

Parameters

cell	Cell iterator.

Returns

Cached weight of the current cell.

prepare_transfer()

template<int dim>

void MeltPoolDG::CutUtil::PartitionWeights< dim >::prepare_transfer ( )

inlineprivate

Attach the cached weights to the cells before the triangulation changes.

The packed weight of a cell is exactly the weight returned to p4est by get_weight(), i.e. the cached weight for persisting cells and cells to be refined, and the maximum weight of the children for cells whose children will be coarsened.

update()

template<int dim>

void MeltPoolDG::CutUtil::PartitionWeights< dim >::update ( const dealii::NonMatching::MeshClassifier< dim > &  mesh_classifier )

inline

Compute and cache the weights of the locally owned active cells according to their location with respect to the level set.

The new weights are used by p4est in the next change of the triangulation. Therefore, this function must be called directly before an explicit repartitioning (see repartition_triangulation()).

Upon the first call, the function returning the cached weights is connected to the weight signal of the triangulation, and the transfer of the cached weights is connected to the refinement and repartitioning signals.

Parameters

mesh_classifier

Mesh classifier, which must be up-to-date with the current triangulation, i.e. reclassify() must have been called after the last change of the triangulation.

Member Data Documentation

connections

template<int dim>

std::vector<boost::signals2::scoped_connection> MeltPoolDG::CutUtil::PartitionWeights< dim >::connections

private

Connections to the signals of the triangulation.

repartition_data

template<int dim>

const CutRepartitionData MeltPoolDG::CutUtil::PartitionWeights< dim >::repartition_data

private

Repartitioning parameters including the cell weights.

tria

template<int dim>

const dealii::Triangulation<dim>* MeltPoolDG::CutUtil::PartitionWeights< dim >::tria

private

Pointer to the triangulation.

weights

template<int dim>

std::vector<unsigned int> MeltPoolDG::CutUtil::PartitionWeights< dim >::weights

private

Cached weights of the active cells (indexed by the active cell index)

weights_handle

template<int dim>

unsigned int MeltPoolDG::CutUtil::PartitionWeights< dim >::weights_handle = dealii::numbers::invalid_unsigned_int

private

Handle of the cell-attached weights during a change of the triangulation.

---

The documentation for this class was generated from the following file:

• include/meltpooldg/cut/repartition.hpp