Routines for optimizing load balance between processes in HFX calculations. More...

Functions/Subroutines
subroutine, public	hfx_load_balance (x_data, eps_schwarz, particle_set, max_set, para_env, coeffs_set, coeffs_kind, is_assoc_atomic_block_global, do_periodic, load_balance_parameter, kind_of, basis_parameter, pmax_set, pmax_atom, i_thread, n_threads, cell, do_p_screening, map_atom_to_kind_atom, nkind, eval_type, pmax_block, use_virial)
	Distributes the computation of eri's to all available processes. More...

subroutine, public	hfx_update_load_balance (x_data, para_env, load_balance_parameter, i_thread, n_threads, eval_type)
	Cheap way of redistributing the eri's. More...

integer(kind=int_8) function, public	cost_model (nsa, nsb, nsc, nsd, npgfa, npgfb, npgfc, npgfd, ratio, p1, p2, p3)
	estimates the cost of a set quartet with info available at load balance time i.e. without much info on the primitives primitives More...

subroutine, public	collect_load_balance_info (para_env, x_data, iw, n_threads, i_thread, eval_type)
	... More...

Variables
real(kind=dp), dimension(12), parameter, public	p1_energy = (/2.9461408209700424_dp, 1.0624718662999657_dp, -1.91570128356921242E-002_dp, -1.6668495454436603_dp, 1.7512639006523709_dp, -9.76074323945336081E-002_dp, 2.6230786127311889_dp, -0.31870737623014189_dp, 7.9588203912690973_dp, 1.8331423413134813_dp, -0.15427618665346299_dp, 0.19749436090711650_dp/)

real(kind=dp), dimension(12), parameter, public	p2_energy = (/2.3104682960662593_dp, 1.8744052737304417_dp, -9.36564055598656797E-002_dp, 0.64284973765086939_dp, 1.0137565430060556_dp, -6.80088178288954567E-003_dp, 1.1692629207374552_dp, -2.6314710080507573_dp, 19.237814781880786_dp, 1.0505934173661349_dp, 0.80382371955699250_dp, 0.49903401991818103_dp/)

real(kind=dp), dimension(2), parameter, public	p3_energy = (/7.82336287670072350E-002_dp, 0.38073304105744837_dp/)

Detailed Description

Routines for optimizing load balance between processes in HFX calculations.

History: 04.2008 created [Manuel Guidon]

Author: Manuel Guidon

Function/Subroutine Documentation

◆ hfx_load_balance()

subroutine, public hfx_load_balance_methods::hfx_load_balance	(	type(hfx_type), pointer	x_data,
		real(dp), intent(in)	eps_schwarz,
		type(particle_type), dimension(:), pointer	particle_set,
		integer, intent(in)	max_set,
		type(mp_para_env_type), pointer	para_env,
		type(hfx_screen_coeff_type), dimension(:, :, :, :), pointer	coeffs_set,
		type(hfx_screen_coeff_type), dimension(:, :), pointer	coeffs_kind,
		integer, dimension(:, :)	is_assoc_atomic_block_global,
		logical	do_periodic,
		type(hfx_load_balance_type), pointer	load_balance_parameter,
		integer, dimension(*)	kind_of,
		type(hfx_basis_type), dimension(:), pointer	basis_parameter,
		type(hfx_p_kind), dimension(:), pointer	pmax_set,
		real(dp), dimension(:, :), pointer	pmax_atom,
		integer, intent(in)	i_thread,
		integer, intent(in)	n_threads,
		type(cell_type), pointer	cell,
		logical, intent(in)	do_p_screening,
		integer, dimension(:), pointer	map_atom_to_kind_atom,
		integer, intent(in)	nkind,
		integer, intent(in)	eval_type,
		real(dp), dimension(:, :), pointer	pmax_block,
		logical, intent(in)	use_virial
	)

Distributes the computation of eri's to all available processes.

Parameters

x_data	Object that stores the indices array
eps_schwarz	screening parameter
particle_set,atomic_kind_set,para_env	...
max_set	Maximum number of set to be considered
para_env	para_env
coeffs_set	screening functions
coeffs_kind	screening functions
is_assoc_atomic_block_global	KS-matrix sparsity
do_periodic	flag for periodicity
load_balance_parameter	Parameters for Monte-Carlo routines
kind_of	helper array for mapping
basis_parameter	Basis set parameters
pmax_set	Initial screening matrix
pmax_atom	...
i_thread	Process ID of current Thread
n_threads	Total Number of Threads
cell	cell
do_p_screening	Flag for initial p screening
map_atom_to_kind_atom	...
nkind	...
eval_type	...
pmax_block	...
use_virial	...

History: 06.2007 created [Manuel Guidon] 08.2007 new parallel scheme [Manuel Guidon] 09.2007 new 'modulo' parellel scheme and Monte Carlo step [Manuel Guidon] 11.2007 parallelize load balance on box_idx1 [Manuel Guidon] 02.2009 completely refactored [Manuel Guidon]

Author: Manuel Guidon

Note: The optimization is done via a binning procedure followed by simple Monte Carlo procedure: In a first step the total amount of integrals in the system is calculated, taking into account the sparsity of the KS-matrix , the screening based on near/farfield approximations and if desired the screening on an initial density matrix. In a second step, bins are generate that contain approximately the same number of integrals, and a cost for these bins is estimated (currently the number of integrals) In a third step, a Monte Carlo procedure optimizes the assignment of the different loads to each process At the end each process owns an unique array of atomic indices-ranges that are used to decide whether a process has to calculate a certain bunch of integrals or not

Definition at line 123 of file hfx_load_balance_methods.F.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ hfx_update_load_balance()

subroutine, public hfx_load_balance_methods::hfx_update_load_balance	(	type(hfx_type), pointer	x_data,
		type(mp_para_env_type), intent(in)	para_env,
		type(hfx_load_balance_type)	load_balance_parameter,
		integer, intent(in)	i_thread,
		integer, intent(in)	n_threads,
		integer, intent(in)	eval_type
	)

Cheap way of redistributing the eri's.

Parameters

x_data	Object that stores the indices array
para_env	para_env
load_balance_parameter	contains parmameter for Monte-Carlo routines
i_thread	current thread ID
n_threads	Total Number of threads
eval_type	...

History: 12.2007 created [Manuel Guidon] 02.2009 optimize Memory Usage [Manuel Guidon]

Author: Manuel Guidon

Note: The cost matrix is given by the walltime for each bin that is measured during the calculation

Definition at line 1480 of file hfx_load_balance_methods.F.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ cost_model()

integer(kind=int_8) function, public hfx_load_balance_methods::cost_model	(	integer	nsa,
		integer	nsb,
		integer	nsc,
		integer	nsd,
		integer	npgfa,
		integer	npgfb,
		integer	npgfc,
		integer	npgfd,
		real(kind=dp)	ratio,
		real(kind=dp), dimension(12), intent(in)	p1,
		real(kind=dp), dimension(12), intent(in)	p2,
		real(kind=dp), dimension(2), intent(in)	p3
	)

estimates the cost of a set quartet with info available at load balance time i.e. without much info on the primitives primitives

Parameters

nsa	...
nsb	...
nsc	...
nsd	...
npgfa	...
npgfb	...
npgfc	...
npgfd	...
ratio	...
p1	...
p2	...
p3	...

Returns: ...

History: 08.2009 created Joost VandeVondele

Author: Joost VandeVondele

Definition at line 1765 of file hfx_load_balance_methods.F.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ collect_load_balance_info()

subroutine, public hfx_load_balance_methods::collect_load_balance_info	(	type(mp_para_env_type), intent(in)	para_env,
		type(hfx_type), pointer	x_data,
		integer, intent(in)	iw,
		integer, intent(in)	n_threads,
		integer, intent(in)	i_thread,
		integer, intent(in)	eval_type
	)

...

Parameters

para_env	...
x_data	...
iw	...
n_threads	...
i_thread	...
eval_type	...

Definition at line 2327 of file hfx_load_balance_methods.F.

Here is the caller graph for this function:

Variable Documentation

◆ p1_energy

real(kind=dp), dimension(12), parameter, public hfx_load_balance_methods::p1_energy = (/2.9461408209700424_dp, 1.0624718662999657_dp, -1.91570128356921242E-002_dp, -1.6668495454436603_dp, 1.7512639006523709_dp, -9.76074323945336081E-002_dp, 2.6230786127311889_dp, -0.31870737623014189_dp, 7.9588203912690973_dp, 1.8331423413134813_dp, -0.15427618665346299_dp, 0.19749436090711650_dp/)

Definition at line 45 of file hfx_load_balance_methods.F.

◆ p2_energy

real(kind=dp), dimension(12), parameter, public hfx_load_balance_methods::p2_energy = (/2.3104682960662593_dp, 1.8744052737304417_dp, -9.36564055598656797E-002_dp, 0.64284973765086939_dp, 1.0137565430060556_dp, -6.80088178288954567E-003_dp, 1.1692629207374552_dp, -2.6314710080507573_dp, 19.237814781880786_dp, 1.0505934173661349_dp, 0.80382371955699250_dp, 0.49903401991818103_dp/)

Definition at line 51 of file hfx_load_balance_methods.F.

◆ p3_energy

real(kind=dp), dimension(2), parameter, public hfx_load_balance_methods::p3_energy = (/7.82336287670072350E-002_dp, 0.38073304105744837_dp/)

Definition at line 57 of file hfx_load_balance_methods.F.

Functions/Subroutines

Variables

Detailed Description

Function/Subroutine Documentation

◆ hfx_load_balance()

◆ hfx_update_load_balance()

◆ cost_model()

◆ collect_load_balance_info()

Variable Documentation

◆ p1_energy

◆ p2_energy

◆ p3_energy