This module handles the communicators for the various parallelization strategies. More...

Detailed Description

This module handles the communicators for the various parallelization strategies.

Given a number n_index of indices and their ranges index_range(1:n_index), we divide the n_nodes in groups and create communicators for each group. Each group is associated with one index. The min_range indicates the minimum number of elements in each processor. For example, given min_range(1) = 25000, the algorithm will try to put at least 25000 points in each processor.

Example:

 Given 3 indices with ranges
 index_range(1) = 100000, (number of points in the mesh)
 index_range(2) = 15, (number of states)
 index_range(3) = 2, (number of k-points)

and 12 processors, we could get

 mc%group_sizes = (2, 3, 2)

which means that

space is divided in 2 domains per state,
the states are divided in 3 groups, i.e. 5 states per processor, and
the whole setting is duplicated because of the 2 k-points.

To perform collective operations (like a reduce), you can use the communicators provided in mcgroup_comm(:). For example, to sum over states, the communicator to use is mcgroup_comm(P_STRATEGY_STATES)

You can use the routine multicomm_strategy_is_parallel to know if a certain index is parallelized.

Data Types
type	multicomm_all_pairs_t
	An all-pairs communication schedule for a given group. More...

type	multicomm_t
	Stores all communicators and groups. More...

Functions/Subroutines
subroutine, public	multicomm_all_pairs_copy (apout, apin)

subroutine, public	multicomm_init (mc, namespace, base_grp, mode_para, n_node, index_range, min_range)
	create index and domain communicators More...

subroutine, public	multicomm_end (mc)

logical pure function, public	multicomm_strategy_is_parallel (mc, level)

subroutine, public	multicomm_create_all_pairs (mpi_grp, ap)
	This routine uses the one-factorization (or near-one-factorization of a complete graph to construct an all-pair communication schedule (cf. Wang, X., Blum, E. K., Parker, D. S., and Massey, D. 1997. The dance-party problem and its application to collective communication in computer networks. Parallel Comput. 23, 8 (Aug. 1997), 1141-1156. More...

subroutine, public	multicomm_divide_range (nobjs, nprocs, istart, ifinal, lsize, scalapack_compat)
	Divide the range of numbers [1, nobjs] between nprocs processors. More...

subroutine, public	multicomm_divide_range_omp (nobjs, ini, nobjs_loc)
	Function to divide the range of numbers from 1 to nobjs between all available threads with OpenMP. More...

logical pure function, public	multicomm_is_slave (this)

logical pure function, public	multicomm_have_slaves (this)

Variables
integer, parameter, public	p_master = 1

integer, parameter, public	p_slave = 2

integer, parameter, public	par_auto = -1

integer, parameter, public	par_no = 0

integer, parameter	n_par_types = 4

character(len=11), dimension(0:n_par_types), parameter	par_types = (/ "serial ", "ParDomains", "ParStates ", "ParKPoints", "ParOther " /)

integer, parameter	max_index = 5

Function/Subroutine Documentation

◆ multicomm_all_pairs_copy()

subroutine, public multicomm_oct_m::multicomm_all_pairs_copy	(	type(multicomm_all_pairs_t), intent(inout)	apout,
		type(multicomm_all_pairs_t), intent(in)	apin
	)

Definition at line 245 of file multicomm.F90.

◆ multicomm_init()

subroutine, public multicomm_oct_m::multicomm_init	(	type(multicomm_t), intent(out)	mc,
		type(namespace_t), intent(in)	namespace,
		type(mpi_grp_t), intent(in)	base_grp,
		type(calc_mode_par_t), intent(in)	mode_para,
		integer, intent(in)	n_node,
		integer(int64), dimension(:), intent(inout)	index_range,
		integer, dimension(:), intent(in)	min_range
	)

create index and domain communicators

Definition at line 263 of file multicomm.F90.

◆ multicomm_end()

subroutine, public multicomm_oct_m::multicomm_end ( type(multicomm_t), intent(inout) mc )

Definition at line 890 of file multicomm.F90.

◆ multicomm_strategy_is_parallel()

logical pure function, public multicomm_oct_m::multicomm_strategy_is_parallel	(	type(multicomm_t), intent(in)	mc,
		integer, intent(in)	level
	)

Definition at line 926 of file multicomm.F90.

◆ multicomm_create_all_pairs()

subroutine, public multicomm_oct_m::multicomm_create_all_pairs	(	type(mpi_grp_t), intent(in)	mpi_grp,
		type(multicomm_all_pairs_t), intent(out)	ap
	)

This routine uses the one-factorization (or near-one-factorization of a complete graph to construct an all-pair communication schedule (cf. Wang, X., Blum, E. K., Parker, D. S., and Massey, D. 1997. The dance-party problem and its application to collective communication in computer networks. Parallel Comput. 23, 8 (Aug. 1997), 1141-1156.

Definition at line 944 of file multicomm.F90.

◆ multicomm_divide_range()

subroutine, public multicomm_oct_m::multicomm_divide_range	(	integer, intent(in)	nobjs,
		integer, intent(in)	nprocs,
		integer, dimension(:), intent(out)	istart,
		integer, dimension(:), intent(out)	ifinal,
		integer, dimension(:), intent(out), optional	lsize,
		logical, intent(in), optional	scalapack_compat
	)

Divide the range of numbers [1, nobjs] between nprocs processors.

Routine is "thread safe" but cannot be made pure due to the use of ASSERT macro

Parameters

[in]	nobjs	Number of points to divide
[in]	nprocs	Number of processors
[out]	lsize	Number of objects in each partition

Definition at line 1036 of file multicomm.F90.

◆ multicomm_divide_range_omp()

subroutine, public multicomm_oct_m::multicomm_divide_range_omp	(	integer, intent(in)	nobjs,
		integer, intent(out)	ini,
		integer, intent(out)	nobjs_loc
	)

Function to divide the range of numbers from 1 to nobjs between all available threads with OpenMP.

Parameters

[in]	nobjs	Number of points to divide
[out]	ini	Start point of the partition
[out]	nobjs_loc	Number of objects in each partition

Definition at line 1112 of file multicomm.F90.

◆ multicomm_is_slave()

logical pure function, public multicomm_oct_m::multicomm_is_slave ( type(multicomm_t), intent(in) this )

Definition at line 1134 of file multicomm.F90.

◆ multicomm_have_slaves()

logical pure function, public multicomm_oct_m::multicomm_have_slaves ( type(multicomm_t), intent(in) this )

Definition at line 1142 of file multicomm.F90.

Variable Documentation

◆ p_master

integer, parameter, public multicomm_oct_m::p_master = 1

Definition at line 181 of file multicomm.F90.

◆ p_slave

integer, parameter, public multicomm_oct_m::p_slave = 2

Definition at line 181 of file multicomm.F90.

◆ par_auto

integer, parameter, public multicomm_oct_m::par_auto = -1

Definition at line 185 of file multicomm.F90.

◆ par_no

integer, parameter, public multicomm_oct_m::par_no = 0

Definition at line 185 of file multicomm.F90.

◆ n_par_types

integer, parameter multicomm_oct_m::n_par_types = 4

private

Definition at line 189 of file multicomm.F90.

◆ par_types

character(len=11), dimension(0:n_par_types), parameter multicomm_oct_m::par_types = (/ "serial ", "ParDomains", "ParStates ", "ParKPoints", "ParOther " /)

private

Definition at line 190 of file multicomm.F90.

◆ max_index

integer, parameter multicomm_oct_m::max_index = 5

private

Definition at line 199 of file multicomm.F90.

Detailed Description

Data Types

Functions/Subroutines

Variables

Function/Subroutine Documentation

◆ multicomm_all_pairs_copy()

◆ multicomm_init()

◆ multicomm_end()

◆ multicomm_strategy_is_parallel()

◆ multicomm_create_all_pairs()

◆ multicomm_divide_range()

◆ multicomm_divide_range_omp()

◆ multicomm_is_slave()

◆ multicomm_have_slaves()

Variable Documentation

◆ p_master

◆ p_slave

◆ par_auto

◆ par_no

◆ n_par_types

◆ par_types

◆ max_index