#include <Cabana_Grid_SparseDimPartitioner.hpp>

Inheritance diagram for Cabana::Grid::SparseDimPartitioner< MemorySpace, CellPerTileDim, NumSpaceDim >:

Collaboration diagram for Cabana::Grid::SparseDimPartitioner< MemorySpace, CellPerTileDim, NumSpaceDim >:

Classes
struct	SubWorkloadFunctor
	functor to compute the sub workload in a given region (from the prefix sum) More...

Public Types
using	memory_space = MemorySpace
	Kokkos memory space.
using	execution_space = typename memory_space::execution_space
	Default execution space.
using	workload_view = Kokkos::View<int***, memory_space>
	Workload device view.
using	partition_view = Kokkos::View<int* [num_space_dim], memory_space>
	Partition device view.
using	workload_view_host
	Workload host view.
using	partition_view_host
	Partition host view.

Public Member Functions
	SparseDimPartitioner (MPI_Comm comm, float max_workload_coeff, int workload_num, int num_step_rebalance, const std::array< int, num_space_dim > &global_cells_per_dim, int max_optimize_iteration=10)
	Constructor - automatically compute ranks_per_dim from MPI communicator.
	SparseDimPartitioner (MPI_Comm comm, float max_workload_coeff, int workload_num, int num_step_rebalance, const std::array< int, num_space_dim > &ranks_per_dim, const std::array< int, num_space_dim > &global_cells_per_dim, int max_optimize_iteration=10)
	Constructor - user-defined ranks_per_dim communicator.
std::array< int, num_space_dim >	ranksPerDimension (MPI_Comm comm)
	Compute the number of MPI ranks in each dimension of the grid from the given MPI communicator.
std::array< int, num_space_dim >	ranksPerDimension (MPI_Comm comm, const std::array< int, num_space_dim > &) const override
	Get the number of MPI ranks in each dimension of the grid from the given MPI communicator.
std::array< int, num_space_dim >	ownedTilesPerDimension (MPI_Comm cart_comm) const
	Get the tile number in each dimension owned by the current MPI rank.
std::array< int, num_space_dim >	ownedCellsPerDimension (MPI_Comm cart_comm) const
	Get the cell number in each dimension owned by the current MPI rank.
void	ownedTileInfo (MPI_Comm cart_comm, std::array< int, num_space_dim > &owned_num_tile, std::array< int, num_space_dim > &global_tile_offset) const
	Get the owned number of tiles and the global tile offset of the current MPI rank.
void	ownedCellInfo (MPI_Comm cart_comm, const std::array< int, num_space_dim > &, std::array< int, num_space_dim > &owned_num_cell, std::array< int, num_space_dim > &global_cell_offset) const override
	Get the owned number of cells and the global cell offset of the current MPI rank.
void	initializeRecPartition (std::vector< int > &rec_partition_i, std::vector< int > &rec_partition_j, std::vector< int > &rec_partition_k)
	Initialize the tile partition; partition in each dimension has the form [0, p_1, ..., p_n, total_tile_num], so the partition would be [0, p_1), [p_1, p_2) ... [p_n, total_tile_num].
std::array< std::vector< int >, num_space_dim >	getCurrentPartition ()
	Get the current partition. Copy partition from the device view to host std::array<vector>
void	resetWorkload ()
	set all elements in _workload_per_tile and _workload_prefix_sum matrix to 0
template<class ParticlePosViewType, typename ArrayType, typename CellUnit>
void	computeLocalWorkLoad (const ParticlePosViewType &view, int particle_num, const ArrayType &global_lower_corner, const CellUnit dx)
	compute the workload in the current MPI rank from particle positions (each particle count for 1 workload value)
template<class SparseMapType>
void	computeLocalWorkLoad (const SparseMapType &sparseMap)
	compute the workload in the current MPI rank from sparseMap (the workload of a tile is 1 if the tile is occupied, 0 otherwise)
void	computeFullPrefixSum (MPI_Comm comm)
	reduce the total workload in all MPI ranks; 2. compute the workload prefix sum matrix for all MPI ranks
template<class ParticlePosViewType, typename ArrayType, typename CellUnit>
int	optimizePartition (const ParticlePosViewType &view, int particle_num, const ArrayType &global_lower_corner, const CellUnit dx, MPI_Comm comm)
	iteratively optimize the partition
template<class SparseMapType>
int	optimizePartition (const SparseMapType &sparseMap, MPI_Comm comm)
	iteratively optimize the partition
void	optimizePartition (bool &is_changed, int iter_seed)
	optimize the partition in three dimensions separately
int	currentRankWorkload (MPI_Comm cart_comm)
	compute the total workload on the current MPI rank
template<typename PartitionViewHost, typename WorkloadViewHost>
int	currentRankWorkload (MPI_Comm cart_comm, PartitionViewHost &rec_view, WorkloadViewHost &prefix_sum_view)
	compute the total workload on the current MPI rank
int	averageRankWorkload ()
	compute the average workload on each MPI rank
template<typename WorkloadViewHost>
int	averageRankWorkload (WorkloadViewHost &prefix_sum_view)
	compute the average workload on each MPI rank
float	computeImbalanceFactor (MPI_Comm cart_comm)
	compute the imbalance factor for the current partition

Static Public Attributes
static constexpr std::size_t	num_space_dim = NumSpaceDim
	dimension
static constexpr unsigned long long	cell_bits_per_tile_dim
	Number of bits (per dimension) needed to index the cells inside a tile.
static constexpr unsigned long long	cell_num_per_tile_dim
Static Public Attributes inherited from Cabana::Grid::BlockPartitioner< 3 >
static constexpr std::size_t	num_space_dim
	Spatial dimension.

Detailed Description

template<typename MemorySpace, unsigned long long CellPerTileDim = 4, std::size_t NumSpaceDim = 3>
class Cabana::Grid::SparseDimPartitioner< MemorySpace, CellPerTileDim, NumSpaceDim >

Sparse mesh block partitioner. (Current Version: Support 3D only)

Template Parameters

MemorySpace	Kokkos memory space.
CellPerTileDim	Cells per tile per dimension.
NumSpaceDim	Dimemsion (The current version support 3D only)

Member Typedef Documentation

◆ partition_view_host

template<typename MemorySpace, unsigned long long CellPerTileDim = 4, std::size_t NumSpaceDim = 3>

using Cabana::Grid::SparseDimPartitioner< MemorySpace, CellPerTileDim, NumSpaceDim >::partition_view_host

Initial value:

 
        Kokkos::View<int* [num_space_dim],
                     typename execution_space::array_layout, Kokkos::HostSpace>

Partition host view.

◆ workload_view_host

template<typename MemorySpace, unsigned long long CellPerTileDim = 4, std::size_t NumSpaceDim = 3>

using Cabana::Grid::SparseDimPartitioner< MemorySpace, CellPerTileDim, NumSpaceDim >::workload_view_host

Initial value:

 
        Kokkos::View<int***, typename execution_space::array_layout,
                     Kokkos::HostSpace>

Workload host view.

Constructor & Destructor Documentation

◆ SparseDimPartitioner() [1/2]

template<typename MemorySpace, unsigned long long CellPerTileDim = 4, std::size_t NumSpaceDim = 3>

Cabana::Grid::SparseDimPartitioner< MemorySpace, CellPerTileDim, NumSpaceDim >::SparseDimPartitioner	(	MPI_Comm	comm,
		float	max_workload_coeff,
		int	workload_num,
		int	num_step_rebalance,
		const std::array< int, num_space_dim > &	global_cells_per_dim,
		int	max_optimize_iteration = 10 )

inline

Constructor - automatically compute ranks_per_dim from MPI communicator.

Parameters

comm	MPI communicator to decide the rank nums in each dimension
max_workload_coeff	threshold factor for re-partition
workload_num	total workload(particle/tile) number, used to compute workload_threshold
num_step_rebalance	the simulation step number after which one should check if repartition is needed
global_cells_per_dim	3D array, global cells in each dimension
max_optimize_iteration	max iteration number to run the optimization

◆ SparseDimPartitioner() [2/2]

template<typename MemorySpace, unsigned long long CellPerTileDim = 4, std::size_t NumSpaceDim = 3>

Cabana::Grid::SparseDimPartitioner< MemorySpace, CellPerTileDim, NumSpaceDim >::SparseDimPartitioner	(	MPI_Comm	comm,
		float	max_workload_coeff,
		int	workload_num,
		int	num_step_rebalance,
		const std::array< int, num_space_dim > &	ranks_per_dim,
		const std::array< int, num_space_dim > &	global_cells_per_dim,
		int	max_optimize_iteration = 10 )

inline

Constructor - user-defined ranks_per_dim communicator.

Parameters

comm	MPI communicator to decide the rank nums in each dimension
max_workload_coeff	threshold factor for re-partition
workload_num	total workload(particle/tile) number, used to compute workload_threshold
num_step_rebalance	the simulation step number after which one should check if repartition is needed
ranks_per_dim	3D array, user-defined MPI rank constrains in per dimension
global_cells_per_dim	3D array, global cells in each dimension
max_optimize_iteration	max iteration number to run the optimization

Member Function Documentation

◆ averageRankWorkload() [1/2]

template<typename MemorySpace, unsigned long long CellPerTileDim = 4, std::size_t NumSpaceDim = 3>

int Cabana::Grid::SparseDimPartitioner< MemorySpace, CellPerTileDim, NumSpaceDim >::averageRankWorkload ( )

inline

compute the average workload on each MPI rank

Returns: average workload on each rank

◆ averageRankWorkload() [2/2]

template<typename MemorySpace, unsigned long long CellPerTileDim = 4, std::size_t NumSpaceDim = 3>

template<typename WorkloadViewHost>

int Cabana::Grid::SparseDimPartitioner< MemorySpace, CellPerTileDim, NumSpaceDim >::averageRankWorkload ( WorkloadViewHost & prefix_sum_view )

inline

compute the average workload on each MPI rank

Parameters

prefix_sum_view Host mirror of _workload_prefix_sum

Returns: average workload on each rank

◆ computeFullPrefixSum()

template<typename MemorySpace, unsigned long long CellPerTileDim = 4, std::size_t NumSpaceDim = 3>

void Cabana::Grid::SparseDimPartitioner< MemorySpace, CellPerTileDim, NumSpaceDim >::computeFullPrefixSum ( MPI_Comm comm )

inline

reduce the total workload in all MPI ranks; 2. compute the workload prefix sum matrix for all MPI ranks

Parameters

comm	MPI communicator used for workload reduction

◆ computeImbalanceFactor()

template<typename MemorySpace, unsigned long long CellPerTileDim = 4, std::size_t NumSpaceDim = 3>

float Cabana::Grid::SparseDimPartitioner< MemorySpace, CellPerTileDim, NumSpaceDim >::computeImbalanceFactor ( MPI_Comm cart_comm )

inline

compute the imbalance factor for the current partition

Parameters

cart_comm MPI cartesian communicator

Returns: the imbalance factor = workload on current rank / ave_workload

◆ computeLocalWorkLoad() [1/2]

template<typename MemorySpace, unsigned long long CellPerTileDim = 4, std::size_t NumSpaceDim = 3>

template<class ParticlePosViewType, typename ArrayType, typename CellUnit>

void Cabana::Grid::SparseDimPartitioner< MemorySpace, CellPerTileDim, NumSpaceDim >::computeLocalWorkLoad	(	const ParticlePosViewType &	view,
		int	particle_num,
		const ArrayType &	global_lower_corner,
		const CellUnit	dx )

inline

compute the workload in the current MPI rank from particle positions (each particle count for 1 workload value)

Parameters

view	particle positions view
particle_num	total particle number
global_lower_corner	the coordinate of the domain global lower corner
dx	cell dx size

◆ computeLocalWorkLoad() [2/2]

template<typename MemorySpace, unsigned long long CellPerTileDim = 4, std::size_t NumSpaceDim = 3>

template<class SparseMapType>

void Cabana::Grid::SparseDimPartitioner< MemorySpace, CellPerTileDim, NumSpaceDim >::computeLocalWorkLoad ( const SparseMapType & sparseMap )

inline

compute the workload in the current MPI rank from sparseMap (the workload of a tile is 1 if the tile is occupied, 0 otherwise)

Parameters

sparseMap sparseMap in the current rank

◆ currentRankWorkload() [1/2]

template<typename MemorySpace, unsigned long long CellPerTileDim = 4, std::size_t NumSpaceDim = 3>

int Cabana::Grid::SparseDimPartitioner< MemorySpace, CellPerTileDim, NumSpaceDim >::currentRankWorkload ( MPI_Comm cart_comm )

inline

compute the total workload on the current MPI rank

Parameters

cart_comm MPI cartesian communicator

Returns: total workload on current rank

◆ currentRankWorkload() [2/2]

template<typename MemorySpace, unsigned long long CellPerTileDim = 4, std::size_t NumSpaceDim = 3>

template<typename PartitionViewHost, typename WorkloadViewHost>

int Cabana::Grid::SparseDimPartitioner< MemorySpace, CellPerTileDim, NumSpaceDim >::currentRankWorkload	(	MPI_Comm	cart_comm,
		PartitionViewHost &	rec_view,
		WorkloadViewHost &	prefix_sum_view )

inline

compute the total workload on the current MPI rank

Parameters

cart_comm	MPI cartesian communicator
rec_view	Host mirror of _rec_partition_dev
prefix_sum_view	Host mirror of _workload_prefix_sum

Returns: total workload on current rank

◆ initializeRecPartition()

template<typename MemorySpace, unsigned long long CellPerTileDim = 4, std::size_t NumSpaceDim = 3>

void Cabana::Grid::SparseDimPartitioner< MemorySpace, CellPerTileDim, NumSpaceDim >::initializeRecPartition	(	std::vector< int > &	rec_partition_i,
		std::vector< int > &	rec_partition_j,
		std::vector< int > &	rec_partition_k )

inline

Initialize the tile partition; partition in each dimension has the form [0, p_1, ..., p_n, total_tile_num], so the partition would be [0, p_1), [p_1, p_2) ... [p_n, total_tile_num].

Parameters

rec_partition_i	partition array in dimension i
rec_partition_j	partition array in dimension j
rec_partition_k	partition array in dimension k

◆ optimizePartition() [1/3]

template<typename MemorySpace, unsigned long long CellPerTileDim = 4, std::size_t NumSpaceDim = 3>

void Cabana::Grid::SparseDimPartitioner< MemorySpace, CellPerTileDim, NumSpaceDim >::optimizePartition	(	bool &	is_changed,
		int	iter_seed )

inline

optimize the partition in three dimensions separately

Parameters

is_changed	label if the partition is changed after the optimization
iter_seed	seed number to choose the starting dimension of the optimization

◆ optimizePartition() [2/3]

template<typename MemorySpace, unsigned long long CellPerTileDim = 4, std::size_t NumSpaceDim = 3>

template<class ParticlePosViewType, typename ArrayType, typename CellUnit>

int Cabana::Grid::SparseDimPartitioner< MemorySpace, CellPerTileDim, NumSpaceDim >::optimizePartition	(	const ParticlePosViewType &	view,
		int	particle_num,
		const ArrayType &	global_lower_corner,
		const CellUnit	dx,
		MPI_Comm	comm )

inline

iteratively optimize the partition

Parameters

view	particle positions view
particle_num	total particle number
global_lower_corner	the coordinate of the domain global lower corner
dx	cell dx size
comm	MPI communicator used for workload reduction

Returns: iteration number

◆ optimizePartition() [3/3]

template<typename MemorySpace, unsigned long long CellPerTileDim = 4, std::size_t NumSpaceDim = 3>

template<class SparseMapType>

int Cabana::Grid::SparseDimPartitioner< MemorySpace, CellPerTileDim, NumSpaceDim >::optimizePartition	(	const SparseMapType &	sparseMap,
		MPI_Comm	comm )

inline

iteratively optimize the partition

Parameters

sparseMap	sparseMap in the current rank
comm	MPI communicator used for workload reduction

Returns: iteration number

◆ ownedCellInfo()

template<typename MemorySpace, unsigned long long CellPerTileDim = 4, std::size_t NumSpaceDim = 3>

void Cabana::Grid::SparseDimPartitioner< MemorySpace, CellPerTileDim, NumSpaceDim >::ownedCellInfo	(	MPI_Comm	cart_comm,
		const std::array< int, num_space_dim > &	,
		std::array< int, num_space_dim > &	owned_num_cell,
		std::array< int, num_space_dim > &	global_cell_offset ) const

inlineoverridevirtual

Get the owned number of cells and the global cell offset of the current MPI rank.

Parameters

cart_comm	The MPI Cartesian communicator for the partitioning.
owned_num_cell	(Return) The owned number of cells of the current MPI rank in each dimension.
global_cell_offset	(Return) The global cell offset of the current MPI rank in each dimension

Implements Cabana::Grid::BlockPartitioner< 3 >.

◆ ownedCellsPerDimension()

template<typename MemorySpace, unsigned long long CellPerTileDim = 4, std::size_t NumSpaceDim = 3>

std::array< int, num_space_dim > Cabana::Grid::SparseDimPartitioner< MemorySpace, CellPerTileDim, NumSpaceDim >::ownedCellsPerDimension ( MPI_Comm cart_comm ) const

inline

Get the cell number in each dimension owned by the current MPI rank.

Parameters

cart_comm MPI cartesian communicator

◆ ownedTileInfo()

template<typename MemorySpace, unsigned long long CellPerTileDim = 4, std::size_t NumSpaceDim = 3>

void Cabana::Grid::SparseDimPartitioner< MemorySpace, CellPerTileDim, NumSpaceDim >::ownedTileInfo	(	MPI_Comm	cart_comm,
		std::array< int, num_space_dim > &	owned_num_tile,
		std::array< int, num_space_dim > &	global_tile_offset ) const

inline

Get the owned number of tiles and the global tile offset of the current MPI rank.

Parameters

cart_comm	The MPI Cartesian communicator for the partitioning.
owned_num_tile	(Return) The owned number of tiles of the current MPI rank in each dimension.
global_tile_offset	(Return) The global tile offset of the current MPI rank in each dimension

◆ ownedTilesPerDimension()

template<typename MemorySpace, unsigned long long CellPerTileDim = 4, std::size_t NumSpaceDim = 3>

std::array< int, num_space_dim > Cabana::Grid::SparseDimPartitioner< MemorySpace, CellPerTileDim, NumSpaceDim >::ownedTilesPerDimension ( MPI_Comm cart_comm ) const

inline

Get the tile number in each dimension owned by the current MPI rank.

Parameters

cart_comm MPI cartesian communicator

◆ ranksPerDimension() [1/2]

template<typename MemorySpace, unsigned long long CellPerTileDim = 4, std::size_t NumSpaceDim = 3>

std::array< int, num_space_dim > Cabana::Grid::SparseDimPartitioner< MemorySpace, CellPerTileDim, NumSpaceDim >::ranksPerDimension ( MPI_Comm comm )

inline

Compute the number of MPI ranks in each dimension of the grid from the given MPI communicator.

Parameters

comm	The communicator to use for the partitioning

◆ ranksPerDimension() [2/2]

template<typename MemorySpace, unsigned long long CellPerTileDim = 4, std::size_t NumSpaceDim = 3>

std::array< int, num_space_dim > Cabana::Grid::SparseDimPartitioner< MemorySpace, CellPerTileDim, NumSpaceDim >::ranksPerDimension	(	MPI_Comm	comm,
		const std::array< int, num_space_dim > &	) const

inlineoverridevirtual

Get the number of MPI ranks in each dimension of the grid from the given MPI communicator.

Parameters

comm	The communicator to use for the partitioning

Implements Cabana::Grid::BlockPartitioner< 3 >.

Member Data Documentation

◆ cell_bits_per_tile_dim

template<typename MemorySpace, unsigned long long CellPerTileDim = 4, std::size_t NumSpaceDim = 3>

unsigned long long Cabana::Grid::SparseDimPartitioner< MemorySpace, CellPerTileDim, NumSpaceDim >::cell_bits_per_tile_dim

staticconstexpr

Initial value:

=

bitCount( CellPerTileDim )

Cabana::Grid::bitCount

KOKKOS_INLINE_FUNCTION constexpr Integer bitCount(Integer input_int) noexcept

(Host/Device) Compute the lease bit number needed to index input integer

Definition Cabana_Grid_SparseIndexSpace.hpp:80

Number of bits (per dimension) needed to index the cells inside a tile.

◆ cell_num_per_tile_dim

template<typename MemorySpace, unsigned long long CellPerTileDim = 4, std::size_t NumSpaceDim = 3>

unsigned long long Cabana::Grid::SparseDimPartitioner< MemorySpace, CellPerTileDim, NumSpaceDim >::cell_num_per_tile_dim

staticconstexpr

Initial value:

=

1 << cell_bits_per_tile_dim

Cabana::Grid::SparseDimPartitioner::cell_bits_per_tile_dim

static constexpr unsigned long long cell_bits_per_tile_dim

Number of bits (per dimension) needed to index the cells inside a tile.

Definition Cabana_Grid_SparseDimPartitioner.hpp:70

Number of cells inside each tile (per dimension) Tile size reset to power of 2

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

grid/src/Cabana_Grid_SparseDimPartitioner.hpp

Classes

Public Types

Public Member Functions

Static Public Attributes

Detailed Description

Member Typedef Documentation

◆ partition_view_host

◆ workload_view_host

Constructor & Destructor Documentation

◆ SparseDimPartitioner() [1/2]

◆ SparseDimPartitioner() [2/2]

Member Function Documentation

◆ averageRankWorkload() [1/2]

◆ averageRankWorkload() [2/2]

◆ computeFullPrefixSum()

◆ computeImbalanceFactor()

◆ computeLocalWorkLoad() [1/2]

◆ computeLocalWorkLoad() [2/2]

◆ currentRankWorkload() [1/2]

◆ currentRankWorkload() [2/2]

◆ initializeRecPartition()

◆ optimizePartition() [1/3]

◆ optimizePartition() [2/3]

◆ optimizePartition() [3/3]

◆ ownedCellInfo()

◆ ownedCellsPerDimension()

◆ ownedTileInfo()

◆ ownedTilesPerDimension()

◆ ranksPerDimension() [1/2]

◆ ranksPerDimension() [2/2]

Member Data Documentation

◆ cell_bits_per_tile_dim

◆ cell_num_per_tile_dim