multicomm.F90

Go to the documentation of this file.

!! Copyright (C) 2002-2015 M. Marques, A. Castro, A. Rubio, G. Bertsch, X. Andrade
!!
!! This program is free software; you can redistribute it and/or modify
!! it under the terms of the GNU General Public License as published by
!! the Free Software Foundation; either version 2, or (at your option)
!! any later version.
!!
!! This program is distributed in the hope that it will be useful,
!! but WITHOUT ANY WARRANTY; without even the implied warranty of
!! MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
!! GNU General Public License for more details.
!!
!! You should have received a copy of the GNU General Public License
!! along with this program; if not, write to the Free Software
!! Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA
!! 02110-1301, USA.
!!
 
#include "global.h"
 
 
module multicomm_oct_m
  use blacs_oct_m
  use calc_mode_par_oct_m
  use debug_oct_m
  use global_oct_m
  use, intrinsic :: iso_fortran_env
  use messages_oct_m
  use mpi_oct_m
  use mpi_lib_oct_m
  use namespace_oct_m
#if defined(HAVE_OPENMP)
  use omp_lib
#endif
  use parser_oct_m
  use profiling_oct_m
  use utils_oct_m
 
  implicit none
 
  private
 
  public ::                          &
    multicomm_divide_range,          &
    multicomm_divide_range_omp,      &
#if defined(HAVE_MPI)
    multicomm_create_all_pairs,      &
#endif
    multicomm_t,                     &
    multicomm_all_pairs_t,           &
    multicomm_init,                  &
    multicomm_end,                   &
    multicomm_all_pairs_copy,        &
    multicomm_strategy_is_parallel,  &
    multicomm_is_slave,              &
    multicomm_have_slaves
 
  integer, public, parameter ::      &
    P_MASTER           = 1,          &
    p_slave            = 2
 
  integer, public, parameter ::      &
    PAR_AUTO            = -1,        &
    par_no              =  0
 
  integer,           parameter :: n_par_types = 4
  character(len=11), parameter :: par_types(0:n_par_types) = &
    (/                              &
    "serial    ",                   &
    "ParDomains",                   &
    "ParStates ",                   &
    "ParKPoints",                   &
    "ParOther  "                    &
    /)
 
  integer, parameter :: MAX_INDEX = 5
 
  type multicomm_t
    private
    integer          :: n_node
 
    integer, public  :: par_strategy
 
    integer, allocatable         :: group_sizes(:)
    integer, allocatable, public :: who_am_i(:)
    type(MPI_Comm), allocatable, public :: group_comm(:)
    type(MPI_Comm), public       :: dom_st_comm     = mpi_comm_undefined      
    type(MPI_Comm), public       :: st_kpt_comm     = mpi_comm_undefined      
    type(MPI_Comm), public       :: dom_st_kpt_comm = mpi_comm_undefined      
    type(mpi_grp_t), public      :: intranode_grp
    type(mpi_grp_t), public      :: internode_grp
 
    integer          :: nthreads
    integer          :: node_type
    logical          :: have_slaves
 
    type(MPI_Comm)          :: full_comm = mpi_comm_undefined       
    integer                 :: full_comm_rank
    type(MPI_Comm), public  :: master_comm = mpi_comm_undefined     
    integer                 :: master_comm_rank
    type(MPI_Comm), public  :: slave_intercomm = mpi_comm_undefined 
 
    logical          :: reorder_ranks
  end type multicomm_t
 
  type multicomm_all_pairs_t
    private
    type(mpi_grp_t)  :: grp
    integer          :: rounds
    integer, allocatable, public :: schedule(:, :)
  end type multicomm_all_pairs_t
 
contains
 
  ! ---------------------------------------------------------
  subroutine multicomm_all_pairs_copy(apout, apin)
    type(multicomm_all_pairs_t), intent(inout) :: apout
    type(multicomm_all_pairs_t), intent(in)    :: apin
 
    push_sub(multicomm_all_pairs_copy)
 
    call mpi_grp_copy(apout%grp, apin%grp)
    apout%rounds = apin%rounds
    if (allocated(apin%schedule)) then
      safe_allocate(apout%schedule(1:size(apin%schedule, 1), 1:size(apin%schedule, 2)))
      apout%schedule = apin%schedule
    end if
 
    pop_sub(multicomm_all_pairs_copy)
  end subroutine multicomm_all_pairs_copy
 
  ! ---------------------------------------------------------
  subroutine multicomm_init(mc, namespace, base_grp, mode_para, n_node, index_range, min_range)
    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),        intent(inout) :: index_range(:)
    integer,               intent(in)    :: min_range(:)
 
    integer :: ii, num_slaves, slave_level, ipar
    integer :: parse(1:P_STRATEGY_MAX), default(1:P_STRATEGY_MAX)
    integer :: parallel_mask, default_mask
 
    push_sub(multicomm_init)
 
    mc%n_node  = n_node
    parallel_mask = mode_para%parallel_mask()
    default_mask = mode_para%default_parallel_mask()
 
    call messages_print_with_emphasis(msg="Parallelization", namespace=namespace)
 
    !%Variable ReorderRanks
    !%Default no
    !%Type logical
    !%Section Execution::Parallelization
    !%Description
    !% This variable controls whether the ranks are reorganized to have a more
    !% compact distribution with respect to domain parallelization which needs
    !% to communicate most often. Depending on the system, this can improve
    !% communication speeds.
    !%End
    call parse_variable(namespace, 'ReorderRanks', .false., mc%reorder_ranks)
 
    call messages_obsolete_variable(namespace, 'ParallelizationStrategy')
    call messages_obsolete_variable(namespace, 'ParallelizationGroupRanks')
 
    do ipar = 1, p_strategy_max
      default(ipar) = par_no
      if (bitand(default_mask, ibset(0, ipar - 1)) /= 0) then
        default(ipar) = par_auto
      end if
    end do
 
    !%Variable ParDomains
    !%Type integer
    !%Default auto
    !%Section Execution::Parallelization
    !%Description
    !% This variable controls the number of processors used for the
    !% parallelization in domains.
    !% The special value <tt>auto</tt>, the default, lets Octopus
    !% decide how many processors will be assigned for this
    !% strategy. To disable parallelization in domains, you can use
    !% <tt>ParDomains = no</tt> (or set the number of processors to
    !% 1).
    !%
    !% The total number of processors required is the multiplication
    !% of the processors assigned to each parallelization strategy.
    !%Option auto -1
    !% The number of processors is assigned automatically.
    !%Option no 0
    !% This parallelization strategy is not used.
    !%End
    call parse_variable(namespace, 'ParDomains', default(p_strategy_domains), parse(p_strategy_domains))
 
    !%Variable ParStates
    !%Type integer
    !%Section Execution::Parallelization
    !%Description
    !% This variable controls the number of processors used for the
    !% parallelization in states. The special value <tt>auto</tt> lets
    !% Octopus decide how many processors will be assigned for this
    !% strategy. To disable parallelization in states, you can use
    !% <tt>ParStates = no</tt> (or set the number of processors to 1).
    !%
    !% The default value depends on the <tt>CalculationMode</tt>. For
    !% <tt>CalculationMode = td</tt> the default is <tt>auto</tt>, while
    !% for for other modes the default is <tt>no</tt>.
    !%
    !% The total number of processors required is the multiplication
    !% of the processors assigned to each parallelization strategy.
    !%Option auto -1
    !% The number of processors is assigned automatically.
    !%Option no 0
    !% This parallelization strategy is not used.
    !%End
    call parse_variable(namespace, 'ParStates', default(p_strategy_states), parse(p_strategy_states))
 
    !%Variable ParKPoints
    !%Type integer
    !%Default auto
    !%Section Execution::Parallelization
    !%Description
    !% This variable controls the number of processors used for the
    !% parallelization in K-Points and/or spin.
    !% The special value <tt>auto</tt> lets Octopus decide how many processors will be
    !% assigned for this strategy. To disable parallelization in
    !% KPoints, you can use <tt>ParKPoints = no</tt> (or set the
    !% number of processors to 1).
    !%
    !% The total number of processors required is the multiplication
    !% of the processors assigned to each parallelization strategy.
    !%Option auto -1
    !% The number of processors is assigned automatically.
    !%Option no 0
    !% This parallelization strategy is not used.
    !%End
    call parse_variable(namespace, 'ParKPoints', default(p_strategy_kpoints), parse(p_strategy_kpoints))
 
    !%Variable ParOther
    !%Type integer
    !%Default auto
    !%Section Execution::Parallelization
    !%Description
    !% This variable controls the number of processors used for the
    !% 'other' parallelization mode, that is CalculatioMode
    !% dependent. For <tt>CalculationMode = casida</tt>, it means
    !% parallelization in electron-hole pairs.
    !%
    !% The special value <tt>auto</tt>,
    !% the default, lets Octopus decide how many processors will be
    !% assigned for this strategy. To disable parallelization in
    !% Other, you can use <tt>ParOther = no</tt> (or set the
    !% number of processors to 1).
    !%
    !% The total number of processors required is the multiplication
    !% of the processors assigned to each parallelization strategy.
    !%Option auto -1
    !% The number of processors is assigned automatically.
    !%Option no 0
    !% This parallelization strategy is not used.
    !%End
    call parse_variable(namespace, 'ParOther', default(p_strategy_other), parse(p_strategy_other))
 
    do ipar = 1, p_strategy_max
      if (parse(ipar) == par_no) parse(ipar) = 1
    end do
 
    call strategy()
 
    mc%have_slaves = .false.
 
    if (mc%par_strategy /= p_strategy_serial) then
      safe_allocate(mc%group_sizes(1:p_strategy_max))
 
      mc%group_sizes = 1
 
      do ipar = 1, p_strategy_max
        if (multicomm_strategy_is_parallel(mc, ipar)) then
          mc%group_sizes(ipar) = parse(ipar)
        else if (parse(ipar) /= 1) then
          call messages_write('Ignoring specification for ' // par_types(ipar))
          call messages_new_line()
          call messages_write('This parallelization strategy is not available.')
          call messages_warning()
        end if
      end do
 
      call assign_nodes()
 
 
      !%Variable ParallelizationNumberSlaves
      !%Type integer
      !%Default 0
      !%Section Execution::Parallelization
      !%Description
      !% Slaves are nodes used for task parallelization. The number of
      !% such nodes is given by this variable multiplied by the number
      !% of domains used in domain parallelization.
      !%End
      call parse_variable(namespace, 'ParallelizationNumberSlaves', 0, num_slaves)
 
      ! the slaves must be defined at a certain parallelization level, for the moment this is state parallelization.
      slave_level = p_strategy_states
      mc%have_slaves = (num_slaves > 0)
 
      if (mc%have_slaves) then
        call messages_experimental('Task parallelization')
      end if
 
      ! clear parallel strategies that were available but will not be used
      do ii = 1, p_strategy_max
        if (mc%group_sizes(ii) == 1) mc%par_strategy = ibclr(mc%par_strategy, ii - 1)
      end do
 
      ! reset
      call sanity_check()
    end if
 
    call group_comm_create()
 
    call messages_print_with_emphasis(namespace=namespace)
 
    pop_sub(multicomm_init)
 
  contains
 
    ! ---------------------------------------------------------
    subroutine strategy()
      integer :: jj, ipar
 
      push_sub(multicomm_init.strategy)
 
      if (base_grp%size > 1) then
 
        mc%par_strategy = 0
 
        do ipar = 1, p_strategy_max
          if (parse(ipar) == par_auto .or. parse(ipar) > 1) then
            mc%par_strategy = ibset(mc%par_strategy, ipar - 1)
          end if
        end do
 
        if (mc%par_strategy /= bitand(mc%par_strategy, parallel_mask)) then
          call messages_write('Parallelization strategies unavailable for this run mode are being discarded.')
          call messages_warning()
        end if
 
        mc%par_strategy = bitand(mc%par_strategy, parallel_mask)
 
        if (mc%par_strategy == p_strategy_serial) then
          message(1) = "More than one node is available, but this run mode cannot run with the requested parallelization."
          message(2) = "Please select a parallelization strategy compatible with"
          jj = 2
          do ii = 1, n_par_types
            if (bitand(parallel_mask, 2**(ii - 1)) /= 0) then
              jj = jj + 1
              write(message(jj), '(2a)') "  -> ", par_types(ii)
            end if
          end do
          jj=jj+1
          write(message(jj),'(a,i6)') "mc%par_strategy is : ",mc%par_strategy
          call messages_fatal(jj, only_root_writes = .true.)
        end if
      else
        mc%par_strategy = p_strategy_serial
      end if
 
      mc%nthreads = 1
#if defined(HAVE_OPENMP)
      !$omp parallel
      !$omp master
      mc%nthreads = omp_get_num_threads()
      !$omp end master
      !$omp end parallel
#endif
 
      if (mc%par_strategy == p_strategy_serial .and. mc%nthreads == 1) then
        message(1) = "Info: Octopus will run in *serial*"
        call messages_info(1, namespace=namespace)
      else
        write(message(1),'(a)')     'Info: Octopus will run in *parallel*'
        write(message(2),'(a)')     ''
        write(message(3),'(a, i8)') '      Number of processes           :', base_grp%size
        write(message(4),'(a, i8)') '      Number of threads per process :', mc%nthreads
        write(message(5),'(a)')     ''
        call messages_info(5, namespace=namespace)
      end if
 
      pop_sub(multicomm_init.strategy)
    end subroutine strategy
 
    ! ---------------------------------------------------------
 
    subroutine assign_nodes()
      integer :: ii, nn, kk, n_divisors, divisors(1:50)
      integer(int64) :: n_group_max(1:p_strategy_max)
 
      push_sub(multicomm_init.assign_nodes)
 
      ! this is the maximum number of processors in each group
      n_group_max(1:p_strategy_max) = max(index_range(1:p_strategy_max), 1_int64)
      do kk = 1, p_strategy_max
        if (.not. multicomm_strategy_is_parallel(mc, kk)) n_group_max(kk) = 1
      end do
 
      call messages_write('Debug info: Allowable group ranks:', new_line = .true.)
      do kk = 1, p_strategy_max
        call messages_write(par_types(kk), fmt = '2x,a12,":",1x')
        call messages_write(n_group_max(kk), new_line = .true.)
      end do
      call messages_info(debug_only=.true.)
 
      nn = mc%n_node
 
      ! first loop, check the processors assigned by the user
      do ipar = p_strategy_max, 1, -1
 
        if (mc%group_sizes(ipar) == par_auto) cycle
 
        if (mc%group_sizes(ipar) > n_group_max(ipar)) then
          call messages_write('The number of processors specified for '//par_types(ipar)//'(')
          call messages_write(mc%group_sizes(ipar))
          call messages_write(')', new_line = .true.)
          call messages_write('is larger than the degrees of freedom for that level (')
          call messages_write(n_group_max(ipar))
          call messages_write(').')
          call messages_warning()
        end if
 
        if (mod(nn, mc%group_sizes(ipar)) /= 0) then
          call messages_write('The number of processors specified for '//par_types(ipar)//'(')
          call messages_write(mc%group_sizes(ipar))
          call messages_write(')', new_line = .true.)
          call messages_write('is not a divisor of the number of processors (')
          call messages_write(mc%n_node)
          call messages_write(').')
          call messages_fatal()
        end if
 
        nn = nn/mc%group_sizes(ipar)
 
      end do
 
      ! second loop, now assign the rest automatically
      do ipar = p_strategy_max, 1, -1
 
        if (mc%group_sizes(ipar) /= par_auto) cycle
 
        n_divisors = ubound(divisors, dim = 1)
        call get_divisors(nn, n_divisors, divisors)
 
        mc%group_sizes(ipar) = nn
        do ii = 2, n_divisors
          if (divisors(ii) > n_group_max(ipar)) then
            mc%group_sizes(ipar) = divisors(ii - 1)
            exit
          end if
        end do
 
        nn = nn/mc%group_sizes(ipar)
 
      end do
 
      pop_sub(multicomm_init.assign_nodes)
    end subroutine assign_nodes
 
 
    ! ---------------------------------------------------------
    subroutine sanity_check()
      real(real64) :: frac
      integer :: ii, kk
      integer(int64) :: jj, n_max
      integer :: real_group_sizes(1:MAX_INDEX)
 
      push_sub(multicomm_init.sanity_check)
 
      if (num_slaves > 0) then
 
        if (mc%group_sizes(slave_level) < num_slaves + 1) then
          message(1) = 'Too many nodes assigned to task parallelization.'
          call messages_fatal(1)
        end if
 
        write(message(1),'(a,i6)') 'Info: Number of slaves nodes              :', &
          num_slaves*product(mc%group_sizes(1:slave_level - 1))
        call messages_info(1)
 
      end if
 
      ! print out some info
      ii = 0
      do kk = p_strategy_max, 1, -1
        real_group_sizes(kk) = mc%group_sizes(kk)
        if (.not. multicomm_strategy_is_parallel(mc, kk)) cycle
        ii = ii + 1
        if (kk == slave_level) real_group_sizes(kk) = real_group_sizes(kk) - num_slaves
        write(message(ii),'(3a,i6,a,i12,a)') 'Info: Number of nodes in ', &
          par_types(kk), ' group:', real_group_sizes(kk), ' (', index_range(kk), ')'
      end do
      call messages_info(ii)
 
      ! do we have the correct number of processors
      if (product(mc%group_sizes(1:p_strategy_max)) /= base_grp%size) then
        write(message(1),'(a)') 'Inconsistent number of processors:'
        write(message(2),'(a,i6)') '  MPI processes      = ', base_grp%size
        write(message(3),'(a,i6)') '  Required processes = ', product(mc%group_sizes(1:p_strategy_max))
        message(4) = ''
        message(5) = 'You probably have a problem in the ParDomains, ParStates, ParKPoints or ParOther.'
        call messages_fatal(5, only_root_writes = .true.)
      end if
 
      if (any(real_group_sizes(1:p_strategy_max) > index_range(1:p_strategy_max))) then
        message(1) = "Could not distribute nodes in parallel job. Most likely you are trying to"
        message(2) = "use too many nodes for the job."
        call messages_fatal(2, only_root_writes = .true.)
      end if
 
      if (any(index_range(1:p_strategy_max) / real_group_sizes(1:p_strategy_max) < min_range(1:p_strategy_max) .and. &
        real_group_sizes(1:p_strategy_max) >  1)) then
        message(1) = "I have fewer elements in a parallel group than recommended."
        message(2) = "Maybe you should reduce the number of nodes."
        call messages_warning(2)
      end if
 
      ! calculate fraction of idle time
      frac = m_one
      do ii = 1, p_strategy_max
        n_max = ceiling(real(index_range(ii), real64) / real(real_group_sizes(ii)), real64)
        jj = n_max*real_group_sizes(ii)
        frac = frac*(m_one - real(jj - index_range(ii), real64)  / real(jj, real64) )
      end do
 
      write(message(1), '(a,f5.2,a)') "Info: Octopus will waste at least ", &
        (m_one - frac)*100.0_real64, "% of computer time."
      if (frac < 0.8_real64) then
        message(2) = "Usually a number of processors which is a multiple of small primes is best."
        call messages_warning(2)
      else
        call messages_info(1)
      end if
 
      pop_sub(multicomm_init.sanity_check)
    end subroutine sanity_check
 
    ! ---------------------------------------------------------
    subroutine group_comm_create()
#if defined(HAVE_MPI)
      logical :: dim_mask(MAX_INDEX)
      integer :: i_strategy, irank
      logical :: reorder, periodic_mask(MAX_INDEX)
      integer :: coords(MAX_INDEX)
      type(mpi_comm) :: new_comm
      integer :: new_comm_size
      character(len=6) :: node_type
      type(mpi_grp_t) :: reorder_grp
      type(mpi_group) :: base_group, reorder_group
      integer :: ranks(base_grp%size)
      integer :: ii, jj, kk, ll, nn
      type(mpi_comm) :: reorder_comm
#endif
 
      push_sub(multicomm_init.group_comm_create)
 
      mc%node_type = p_master
 
      safe_allocate(mc%group_comm(1:p_strategy_max))
      safe_allocate(mc%who_am_i(1:p_strategy_max))
 
      mc%group_comm = mpi_comm_undefined
      mc%who_am_i   = 0
 
#if defined(HAVE_MPI)
      mc%full_comm = mpi_comm_null
      mc%slave_intercomm = mpi_comm_null
      if (mc%par_strategy /= p_strategy_serial) then
        if (mc%reorder_ranks) then
          ! first, reorder the ranks
          ! this is done to get a column-major ordering of the ranks in the
          ! Cartesian communicator, since they a ordered row-major otherwise
          call mpi_comm_group(base_grp%comm, base_group, mpi_err)
          if (mpi_err /= mpi_success) then
            message(1) = "Error in getting MPI group!"
            call messages_fatal(1)
          end if
          ! now transpose the hypercube => get rank numbers in column-major order
          nn = 1
          do ii = 1, mc%group_sizes(1)
            do jj = 1, mc%group_sizes(2)
              do kk = 1, mc%group_sizes(3)
                do ll = 1, mc%group_sizes(4)
                  ranks(nn) = (ll-1)*mc%group_sizes(3)*mc%group_sizes(2)*mc%group_sizes(1) &
                    + (kk-1)*mc%group_sizes(2)*mc%group_sizes(1) &
                    + (jj-1)*mc%group_sizes(1) + ii - 1
                  nn = nn + 1
                end do
              end do
            end do
          end do
          call mpi_group_incl(base_group, base_grp%size, ranks, reorder_group, mpi_err)
          if (mpi_err /= mpi_success) then
            message(1) = "Error in creating MPI group!"
            call messages_fatal(1)
          end if
          ! now get the reordered communicator
          call mpi_comm_create(base_grp%comm, reorder_group, reorder_comm, mpi_err)
          if (mpi_err /= mpi_success) then
            message(1) = "Error in creating reordered communicator!"
            call messages_fatal(1)
          end if
          call mpi_grp_init(reorder_grp, reorder_comm)
        else
          call mpi_grp_copy(reorder_grp, base_grp)
        end if
 
        ! Multilevel parallelization is organized in a hypercube. We
        ! use an MPI Cartesian topology to generate the communicators
        ! that correspond to each level.
 
        ! create the topology
        periodic_mask = .false.
        reorder = .true.
 
        ! The domain and states dimensions have to be periodic (2D torus)
        ! in order to circulate matrix blocks.
        periodic_mask(p_strategy_domains) = multicomm_strategy_is_parallel(mc, p_strategy_domains)
        periodic_mask(p_strategy_states)  = multicomm_strategy_is_parallel(mc, p_strategy_states)
 
        ! We allow reordering of ranks.
        call mpi_cart_create(reorder_grp%comm, p_strategy_max, mc%group_sizes, periodic_mask, reorder, mc%full_comm, mpi_err)
 
        call mpi_comm_rank(mc%full_comm, mc%full_comm_rank, mpi_err)
 
        ! get the coordinates of the current processor
        call mpi_cart_coords(mc%full_comm, mc%full_comm_rank, p_strategy_max, coords, mpi_err)
 
        ! find out what type of node this is
        if (coords(slave_level) >= mc%group_sizes(slave_level) - num_slaves) then
          mc%node_type = p_slave
        end if
 
        if (mc%node_type == p_master) then
          mc%group_sizes(slave_level) = mc%group_sizes(slave_level) - num_slaves
        else
          mc%group_sizes(slave_level) = num_slaves
        end if
 
        call mpi_comm_split(mc%full_comm, mc%node_type, mc%full_comm_rank, new_comm, mpi_err)
        assert(new_comm /= mpi_comm_null)
        call mpi_comm_size(new_comm, new_comm_size, mpi_err)
 
        reorder = .false.
        if (product(mc%group_sizes(:)) /= new_comm_size) then
          write(stderr,*) 'node ', mpi_world%rank, ': mc%group_sizes = ', mc%group_sizes, ' new_comm_size = ', new_comm_size
          call mpi_world%barrier()
          assert(product(mc%group_sizes(:)) == new_comm_size)
        end if
        call mpi_cart_create(new_comm, p_strategy_max, mc%group_sizes, periodic_mask, reorder, mc%master_comm, mpi_err)
        assert(mc%master_comm /= mpi_comm_null)
 
        call mpi_comm_free(new_comm, mpi_err)
 
        call mpi_comm_rank(mc%master_comm, mc%master_comm_rank, mpi_err)
 
        ! The "lines" of the Cartesian grid.
        ! Initialize all the communicators, even if they are not parallelized
        do i_strategy = 1, p_strategy_max
          dim_mask             = .false.
          dim_mask(i_strategy) = .true.
          call mpi_cart_sub(mc%master_comm, dim_mask, mc%group_comm(i_strategy), mpi_err)
          call mpi_comm_rank(mc%group_comm(i_strategy), mc%who_am_i(i_strategy), mpi_err)
        end do
 
        ! The domain-state "planes" of the grid (the ones with periodic dimensions).
        dim_mask                     = .false.
        dim_mask(p_strategy_domains) = .true.
        dim_mask(p_strategy_states)  = .true.
        call mpi_cart_sub(mc%master_comm, dim_mask, mc%dom_st_comm, mpi_err)
 
        ! The state-kpoints "planes" of the grid
        dim_mask                     = .false.
        dim_mask(p_strategy_states)  = .true.
        dim_mask(p_strategy_kpoints) = .true.
        call mpi_cart_sub(mc%master_comm, dim_mask, mc%st_kpt_comm, mpi_err)
 
        ! The domains-states-kpoints "cubes" of the grid
        dim_mask                     = .false.
        dim_mask(p_strategy_domains) = .true.
        dim_mask(p_strategy_states)  = .true.
        dim_mask(p_strategy_kpoints) = .true.
        call mpi_cart_sub(mc%master_comm, dim_mask, mc%dom_st_kpt_comm, mpi_err)
 
        if (num_slaves > 0) call create_slave_intercommunicators()
 
        call create_intranode_communicator(base_grp, mc%intranode_grp, mc%internode_grp)
      else
        ! we initialize these communicators so we can use them even in serial
        mc%group_comm = base_grp%comm
        mc%who_am_i   = 0
        mc%master_comm = base_grp%comm
        mc%dom_st_comm = base_grp%comm
        mc%st_kpt_comm = base_grp%comm
        mc%dom_st_kpt_comm = base_grp%comm
        call mpi_grp_copy(mc%intranode_grp, base_grp)
        call mpi_grp_copy(mc%internode_grp, base_grp)
      end if
 
      ! This is temporary debugging information.
      if (debug%info .and. mc%par_strategy /= p_strategy_serial) then
        write(message(1),'(a)') 'Debug: MPI Task Assignment to MPI Groups'
        write(message(2),'(5a10)') 'World', 'Domains', 'States', 'K-Points', 'Other'
        call messages_info(1)
 
        if (mc%node_type == p_slave) then
          node_type = "slave"
        else
          node_type = "master"
        end if
        do irank = 0, mpi_world%size - 1
          if (mpi_world%rank == irank) then
            write(message(1),'(5i10,5x,a)') mpi_world%rank, mc%who_am_i(p_strategy_domains), mc%who_am_i(p_strategy_states), &
              mc%who_am_i(p_strategy_kpoints), mc%who_am_i(p_strategy_other), trim(node_type)
            call messages_info(1, all_nodes = .true.)
          end if
          call mpi_world%barrier()
        end do
      end if
#endif
 
      pop_sub(multicomm_init.group_comm_create)
    end subroutine group_comm_create
 
    ! -----------------------------------------------------
#ifdef HAVE_MPI
    subroutine create_slave_intercommunicators()
      integer :: remote_leader
      integer :: tag
      integer :: coords(MAX_INDEX)
 
      push_sub(multicomm_init.create_slave_intercommunicators)
 
      ! create the intercommunicators to communicate with slaves
 
      ! get the coordinates of the current processor
      call mpi_cart_coords(mc%full_comm, mc%full_comm_rank, p_strategy_max, coords, mpi_err)
 
      !now get the rank of the remote_leader
      if (mc%node_type == p_slave) then
        coords(slave_level) = 0
      else
        coords(slave_level) = mc%group_sizes(slave_level)
      end if
      call mpi_cart_rank(mc%full_comm, coords, remote_leader, mpi_err)
 
      ! now create the intercommunicator
      tag = coords(p_strategy_domains)
      call mpi_intercomm_create(mc%group_comm(slave_level), 0, base_grp%comm, remote_leader, tag, mc%slave_intercomm, mpi_err)
 
      pop_sub(multicomm_init.create_slave_intercommunicators)
    end subroutine create_slave_intercommunicators
#endif
  end subroutine multicomm_init
 
 
  ! ---------------------------------------------------------
  subroutine multicomm_end(mc)
    type(multicomm_t), intent(inout) :: mc
 
#if defined(HAVE_MPI)
    integer :: ii
#endif
 
    push_sub(multicomm_end)
 
    if (mc%par_strategy /= p_strategy_serial) then
#if defined(HAVE_MPI)
      ! Delete communicators.
      do ii = 1, p_strategy_max
        ! initialized even if not parallelized
        call mpi_comm_free(mc%group_comm(ii), mpi_err)
      end do
      call mpi_comm_free(mc%dom_st_comm, mpi_err)
      call mpi_comm_free(mc%st_kpt_comm, mpi_err)
      call mpi_comm_free(mc%dom_st_kpt_comm, mpi_err)
      call mpi_comm_free(mc%full_comm, mpi_err)
      call mpi_comm_free(mc%master_comm, mpi_err)
 
      if (multicomm_have_slaves(mc)) call mpi_comm_free(mc%slave_intercomm, mpi_err)
 
#endif
    end if
 
    safe_deallocate_a(mc%group_sizes)
    safe_deallocate_a(mc%group_comm)
    safe_deallocate_a(mc%who_am_i)
 
    pop_sub(multicomm_end)
  end subroutine multicomm_end
 
 
  ! ---------------------------------------------------------
  logical pure function multicomm_strategy_is_parallel(mc, level) result(rr)
    type(multicomm_t), intent(in) :: mc
    integer,           intent(in) :: level
 
    rr = bitand(mc%par_strategy, 2**(level - 1)) /= 0
 
  end function multicomm_strategy_is_parallel
 
 
  ! ---------------------------------------------------------
#if defined(HAVE_MPI)
 
  subroutine multicomm_create_all_pairs(mpi_grp, ap)
    type(mpi_grp_t),             intent(in)  :: mpi_grp
    type(multicomm_all_pairs_t), intent(out) :: ap
 
    integer :: grp_size, rounds, ir, in
 
    push_sub(create_all_pairs)
 
    ap%grp = mpi_grp
    grp_size = mpi_grp%size
 
    ! Number of rounds.
    if (mod(grp_size, 2) == 0) then
      rounds = grp_size - 1
    else
      rounds = grp_size
    end if
    ap%rounds = rounds
 
    ! Calculate schedule.
    safe_allocate(ap%schedule(0:grp_size - 1, 1:rounds))
    do ir = 1, rounds
      do in = 0, grp_size - 1
        ap%schedule(in, ir) = get_partner(in + 1, ir) - 1
      end do
    end do
 
    pop_sub(create_all_pairs)
 
  contains
 
    ! ---------------------------------------------------------
    integer pure function get_partner(in, ir)
      integer, intent(in) :: in, ir
 
      ! No PUSH SUB, called too often.
 
      if (mod(grp_size, 2) == 0) then
        get_partner = get_partner_even(grp_size, in - 1, ir - 1) + 1
      else
        get_partner = get_partner_odd(grp_size, in - 1, ir - 1) + 1
      end if
 
    end function get_partner
 
    ! ---------------------------------------------------------
    integer pure function get_partner_even(grp_size, ii, rr) result(pp)
      integer, intent(in) :: grp_size, ii, rr
 
      integer :: mm
 
      ! No PUSH SUB, called too often.
 
      mm = grp_size / 2
 
      if (ii == 0) then
        pp = rr + 1
      elseif (ii == rr + 1) then
        pp = 0
      else
        ! I never know when to use which remainder function, but here
        ! it has to be the modulo one. Do not change that!
        pp = modulo(2 * rr - ii + 1, 2 * mm - 1) + 1
      end if
 
    end function get_partner_even
 
    ! ---------------------------------------------------------
    integer pure function get_partner_odd(grp_size, ii, rr) result(pp)
      integer, intent(in) :: grp_size, ii, rr
 
      integer :: mm
 
      ! No PUSH SUB, called too often.
 
      mm = (grp_size + 1) / 2
 
      pp = get_partner_even(grp_size + 1, ii, rr)
 
      if (pp == 2 * mm - 1) then
        pp = ii
      end if
 
    end function get_partner_odd
 
  end subroutine multicomm_create_all_pairs
#endif
 
  subroutine multicomm_divide_range(nobjs, nprocs, istart, ifinal, lsize, scalapack_compat)
    integer,           intent(in)  :: nobjs
    integer,           intent(in)  :: nprocs
    integer,           intent(out) :: istart(:)
    integer,           intent(out) :: ifinal(:)
    integer, optional, intent(out) :: lsize(:)
    logical, optional, intent(in)  :: scalapack_compat
 
    integer :: ii, jj, rank
    logical :: scalapack_compat_
    integer :: nbl, size
 
    scalapack_compat_ = optional_default(scalapack_compat, .false.)
#ifndef HAVE_SCALAPACK
    scalapack_compat_ = .false.
#endif
 
    if (scalapack_compat_) then
      nbl = nobjs/nprocs
      if (mod(nobjs, nprocs) /= 0) nbl = nbl + 1
 
      istart(1) = 1
      do rank = 1, nprocs
#ifdef HAVE_SCALAPACK
        size = numroc(nobjs, nbl, rank - 1, 0, nprocs)
#endif
        if (size > 0) then
          if (rank > 1) istart(rank) = ifinal(rank - 1) + 1
          ifinal(rank) = istart(rank) + size - 1
        else
          istart(rank) = 1
          ifinal(rank) = 0
        end if
      end do
    else
 
      if (nprocs <= nobjs) then
 
        ! procs are assigned to groups by round robin
        do rank = 0, nprocs - 1
          jj = nobjs / nprocs
          ii = nobjs - jj*nprocs
          if (ii > 0 .and. rank < ii) then
            jj = jj + 1
            istart(rank + 1) = rank*jj + 1
            ifinal(rank + 1) = istart(rank + 1) + jj - 1
          else
            ifinal(rank + 1) = nobjs - (nprocs - rank - 1)*jj
            istart(rank + 1) = ifinal(rank + 1) - jj + 1
          end if
        end do
 
      else
        do ii = 1, nprocs
          if (ii <= nobjs) then
            istart(ii) = ii
            ifinal(ii) = ii
          else
            istart(ii) = 1
            ifinal(ii) = 0
          end if
        end do
      end if
    end if
 
    if (present(lsize)) then
      lsize(1:nprocs) = ifinal(1:nprocs) - istart(1:nprocs) + 1
      assert(sum(lsize(1:nprocs)) == nobjs)
    end if
 
  end subroutine multicomm_divide_range
 
  ! ---------------------------------------------------------
  ! THREADSAFE
  subroutine multicomm_divide_range_omp(nobjs, ini, nobjs_loc)
    integer, intent(in)    :: nobjs
    integer, intent(out)   :: ini
    integer, intent(out)   :: nobjs_loc
 
    integer :: rank
#ifdef HAVE_OPENMP
    integer, allocatable :: istart(:), ifinal(:), lsize(:)
    integer :: nthreads
#endif
 
    ! no push_sub, threadsafe
    rank = 1
#ifdef HAVE_OPENMP
    nthreads = omp_get_num_threads()
    allocate(istart(1:nthreads))
    allocate(ifinal(1:nthreads))
    allocate(lsize(1:nthreads))
    call multicomm_divide_range(nobjs, nthreads, istart, ifinal, lsize)
    rank   = 1 + omp_get_thread_num()
    ini    = istart(rank)
    nobjs_loc = lsize(rank)
    deallocate(istart)
    deallocate(ifinal)
    deallocate(lsize)
#else
    ini = 1
    nobjs_loc = nobjs
#endif
 
  end subroutine multicomm_divide_range_omp
 
  ! ---------------------------------------------------------
 
  logical pure function multicomm_is_slave(this) result(slave)
    type(multicomm_t), intent(in) :: this
 
    slave = this%node_type == p_slave
  end function multicomm_is_slave
 
  ! ---------------------------------------------------------
 
  logical pure function multicomm_have_slaves(this) result(have_slaves)
    type(multicomm_t), intent(in) :: this
 
    have_slaves = this%have_slaves
  end function multicomm_have_slaves
 
end module multicomm_oct_m
 
 
!! Local Variables:
!! mode: f90
!! coding: utf-8
!! End: