mpi_mod.f90 107 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
!**********************************************************************
! Copyright 1998,1999,2000,2001,2002,2005,2007,2008,2009,2010         *
! Andreas Stohl, Petra Seibert, A. Frank, Gerhard Wotawa,             *
! Caroline Forster, Sabine Eckhardt, John Burkhart, Harald Sodemann   *
!                                                                     *
! This file is part of FLEXPART.                                      *
!                                                                     *
! FLEXPART 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 3 of the License, or   *
! (at your option) any later version.                                 *
!                                                                     *
! FLEXPART 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 FLEXPART.  If not, see <http://www.gnu.org/licenses/>.   *
!**********************************************************************

module mpi_mod

!*****************************************************************************
!                                                                            *
!  DESCRIPTION                                                               *
!    This module contains subroutines and common variables used for the      *
!    MPI parallelization of FLEXPART.                                        *
!                                                                            *
!  NOTE                                                                      *
!    Depending on the MPI library installed on your system (e.g. mpich2,     *
!    OpenMPI) you may need to choose below in this file between              *
!      use mpi                                                               *
!    (if the MPI library comes with the file 'mpi.mod'); or                  *
!      include 'mpif.h'                                                      *
!                                                                            *
!                                                                            *
!*****************************************************************************
!                                                                            *
! Variables:                                                                 *
!                                                                            *
! mp_ierr                 MPI error code                                     *
! mp_np                   Number of MPI processes                            *
! mp_pid                  Process ID of each MPI process                     *
! mp_seed                 Parameter for random number seed                   *
Espen Sollum's avatar
Espen Sollum committed
46
47
! read_grp_min            Minimum number of processes at which one will be   *
!                         used as reader                                     *
48
49
50
51
52
53
! numpart_mpi,            Number of particles per node                       *
! maxpart_mpi                                                                *
! mp_partid               MPI process ID for particle calculation            *
! mp_partgroup_           Refers to the subset of processors performing      *
!                         loops over particles. Will be all processes        *
!                         unless a dedicated process runs getfields/readwind *
Espen Sollum's avatar
Espen Sollum committed
54
! lmp_sync                If .false., use asynchronous MPI                   *
55
! mp_cp                   Real precision to use for deposition fields        *
56
57
58
59
60
61
62
!                                                                            *
!                                                                            *
!                                                                            *
!                                                                            *
!*****************************************************************************

  use mpi 
63
  use par_mod, only: dp,sp
64
65
66
67
68
69
70
71
72
73
  use com_mod, only: lroot

  implicit none

!  include 'mpif.h'

  public

! Set aside a process for reading windfields if using at least these many processes
!==================================================
Espen Sollum's avatar
Espen Sollum committed
74
  integer, parameter, private :: read_grp_min=4
75
76
77
78
79
80
81
82
83
84
85
86
!==================================================

! Variables for each MPI process in the world group
  integer :: mp_ierr, mp_np, mp_pid, mp_partid
  integer, private :: world_group_id

! Variables for MPI processes in the 'particle' group
  integer, allocatable, dimension(:) :: mp_partgroup_rank
  integer :: mp_partgroup_comm, mp_partgroup_pid, mp_partgroup_np

  integer :: mp_seed=0
  integer, parameter :: mp_sp=MPI_REAL4, mp_dp=MPI_REAL8
87
  integer :: mp_cp
88
89
90
  integer, parameter :: id_root=0 ! master process

! MPI tags/requests for send/receive operation
Espen Sollum's avatar
Espen Sollum committed
91
  integer :: tm1
92
  integer, parameter :: nvar_async=26
93
!integer, dimension(:), allocatable :: tags
94
95
  integer, dimension(:), allocatable :: reqs

96
97
98
! Status array used for certain MPI operations (MPI_RECV)
  integer, dimension(MPI_STATUS_SIZE) :: mp_status

99
100
101
102
103
104
105
106
107

  integer :: id_read   ! readwind/getfield process
  integer :: numpart_mpi,maxpart_mpi ! number of particles per node
  integer :: tot_numpart=0
  integer :: mp_comm_used ! global or subgroup communicator

  logical :: lmpreader=.false. ! is set to true for reading process(es) only.
  logical :: lmp_use_reader=.false. ! true if separate readwind process is used

108
! .true. if only using synchronous MPI send/recv (default)
109
110
! If setting this to .false., numwfmem must be set to 3
!===============================================================================
111
  logical :: lmp_sync=.true. 
112
113
!===============================================================================

Espen Sollum's avatar
Espen Sollum committed
114
! mp_dbg_mode       Used for debugging MPI.
115
116
117
118
! mp_dev_mode       various checks related to debugging the parallel code
! mp_dbg_out        write some arrays to data file for debugging
! mp_measure_time   Measure cpu/wall time, write out at end of run
! mp_time_barrier   Measure MPI barrier time
119
120
! mp_exact_numpart  Use an extra MPI communication to give the exact number of particles
!                   to standard output (this does *not* otherwise affect the simulation) 
121
122
123
  logical, parameter :: mp_dbg_mode = .false.
  logical, parameter :: mp_dev_mode = .false.
  logical, parameter :: mp_dbg_out = .false.
124
  logical, parameter :: mp_time_barrier=.true.
Espen Sollum's avatar
Espen Sollum committed
125
  logical, parameter :: mp_measure_time=.false.
126
  logical, parameter :: mp_exact_numpart=.true.
127
128

! for measuring CPU/Wall time
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
  real(sp),private :: mp_comm_time_beg, mp_comm_time_end, mp_comm_time_total=0.
  real(dp),private :: mp_comm_wtime_beg, mp_comm_wtime_end, mp_comm_wtime_total=0.
  real(sp),private :: mp_root_time_beg, mp_root_time_end, mp_root_time_total=0.
  real(dp),private :: mp_root_wtime_beg, mp_root_wtime_end, mp_root_wtime_total=0.
  real(sp),private :: mp_barrier_time_beg, mp_barrier_time_end, mp_barrier_time_total=0.
  real(dp),private :: mp_barrier_wtime_beg, mp_barrier_wtime_end, mp_barrier_wtime_total=0.
  real(sp),private :: tm_nploop_beg, tm_nploop_end, tm_nploop_total=0.
  real(sp),private :: tm_tot_beg, tm_tot_end, tm_tot_total=0.
  real(dp),private :: mp_getfields_wtime_beg, mp_getfields_wtime_end, mp_getfields_wtime_total=0.
  real(sp),private :: mp_getfields_time_beg, mp_getfields_time_end, mp_getfields_time_total=0.
  real(dp),private :: mp_readwind_wtime_beg, mp_readwind_wtime_end, mp_readwind_wtime_total=0.
  real(sp),private :: mp_readwind_time_beg, mp_readwind_time_end, mp_readwind_time_total=0.
  real(dp),private :: mp_io_wtime_beg, mp_io_wtime_end, mp_io_wtime_total=0.
  real(sp),private :: mp_io_time_beg, mp_io_time_end, mp_io_time_total=0.
  real(dp),private :: mp_wetdepo_wtime_beg, mp_wetdepo_wtime_end, mp_wetdepo_wtime_total=0.
  real(sp),private :: mp_wetdepo_time_beg, mp_wetdepo_time_end, mp_wetdepo_time_total=0.
  real(dp),private :: mp_advance_wtime_beg, mp_advance_wtime_end, mp_advance_wtime_total=0.
  real(dp),private :: mp_conccalc_time_beg, mp_conccalc_time_end, mp_conccalc_time_total=0.
  real(dp),private :: mp_total_wtime_beg, mp_total_wtime_end, mp_total_wtime_total=0.
  real(dp),private :: mp_vt_wtime_beg, mp_vt_wtime_end, mp_vt_wtime_total
  real(sp),private :: mp_vt_time_beg, mp_vt_time_end, mp_vt_time_total
150
151
152
153

! dat_lun           logical unit number for i/o
  integer, private :: dat_lun 

154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
! Temporary arrays for particles (allocated and deallocated as needed)
  integer, allocatable, dimension(:) :: nclass_tmp, npoint_tmp, itra1_tmp, idt_tmp, &
       & itramem_tmp, itrasplit_tmp
  real(kind=dp), allocatable, dimension(:) :: xtra1_tmp, ytra1_tmp
  real, allocatable, dimension(:) :: ztra1_tmp 
  real, allocatable, dimension(:,:) :: xmass1_tmp

! mp_redist_fract        Exchange particles between processes if relative difference
!                        is larger. A good value is between 0.0 and 0.5
! mp_maxpart_factor      Allocate more memory per process than strictly needed
!                        (safety factor). Recommended value between 1.5 and 2.5
! mp_min_redist          Do not redistribute particles if below this limit
  real, parameter :: mp_redist_fract=0.2, mp_maxpart_factor=1.5
  integer,parameter :: mp_min_redist=100000


170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
contains

  subroutine mpif_init
!***********************************************************************
! DESCRIPTION
!   Initialize MPI.
!   
!   Create the global communicator MPI_COMM_WORLD
!   If using a separate MPI process for getfields/readwind, a subgroup
!   is created for the other processes.
!
! VARIABLES
!   mpi_mode    default 0, set to 2/3 if running MPI version
!   mp_np       number of running processes, decided at run-time
!***********************************************************************
185
    use par_mod, only: maxpart, numwfmem, dep_prec
186
    use com_mod, only: mpi_mode, verbosity
187
188
189
190
191

    implicit none

    integer :: i,j,s,addmaxpart=0

192
! Each process gets an ID (mp_pid) in the range 0,..,mp_np-1
193
194
195
196
197
198
199
200
    call MPI_INIT(mp_ierr)
    if (mp_ierr /= 0) goto 100
    call MPI_COMM_RANK(MPI_COMM_WORLD, mp_pid, mp_ierr)
    if (mp_ierr /= 0) goto 100
    call MPI_COMM_SIZE(MPI_COMM_WORLD, mp_np, mp_ierr)
    if (mp_ierr /= 0) goto 100


201
! Variable mpi_mode is used to handle subroutines common to parallel/serial version
202
203
204
205
206
207
208
209
210
211
    if (lmp_sync) then
      mpi_mode=2 ! hold 2 windfields in memory
    else
      mpi_mode=3 ! hold 3 windfields in memory
    end if

    if (mp_pid.ne.0) then
      lroot = .false.
    end if

212
213
214
215
! Set MPI precision to use for transferring deposition fields
!************************************************************
    if (dep_prec==dp) then
      mp_cp = MPI_REAL8
216
! TODO: write info message for serial version as well 
Espen Sollum's avatar
Espen Sollum committed
217
      if (lroot.and.verbosity>0) write(*,*) 'Using double precision for deposition fields'
218
219
    else if (dep_prec==sp) then
      mp_cp = MPI_REAL4
Espen Sollum's avatar
Espen Sollum committed
220
      if (lroot.and.verbosity>0) write(*,*) 'Using single precision for deposition fields'
221
222
223
224
225
    else
      write(*,*) 'ERROR: something went wrong setting MPI real precision'
      stop
    end if

226
227
228
! Check for sensible combination of parameters
!*********************************************

229
230
231
232
233
234
235
236
    if (.not.lmp_sync.and.numwfmem.ne.3) then
      if (lroot) then
        write(*,FMT='(80("#"))')
        write(*,*) '#### mpi_mod::mpif_init> ERROR: ', &
             & 'numwfmem must be set to 3 for asyncronous reading ####'
        write(*,FMT='(80("#"))')
      end if
      call MPI_FINALIZE(mp_ierr)
237
238
239
240
      stop
    else if (lmp_sync.and.numwfmem.ne.2.and.lroot) then
      write(*,FMT='(80("#"))')
      write(*,*) '#### mpi_mod::mpif_init> WARNING: ', &
241
           & 'numwfmem should be set to 2 for syncronous'
242
243
      write(*,*) ' reading. Results will still be valid, but unneccesary memory &
           &is allocated.'
244
      write(*,FMT='(80("#"))')
Espen Sollum's avatar
Espen Sollum committed
245
! Force "syncronized" version if all processes will call getfields
246
    else if ((.not.lmp_sync.and.mp_np.lt.read_grp_min).or.(mp_np.eq.1)) then
Espen Sollum's avatar
Espen Sollum committed
247
248
249
250
251
252
      if (lroot) then
        write(*,FMT='(80("#"))')
        write(*,*) '#### mpi_mod::mpif_init> WARNING: ', &
             & 'all procs call getfields. Setting lmp_sync=.true.'
        write(*,FMT='(80("#"))')
      end if
253
      lmp_sync=.true.
254
    end if
255
256

! TODO: Add more warnings for unimplemented flexpart features
257
258
259
260
261
262

! Set ID of process that calls getfield/readwind. 
! Using the last process in the group ensures statistical identical results
! as running with one process less but not using separate read process
!**********************************************************************

263
!    id_read = min(mp_np-1, 1)
264
265
266
267
268
269
    id_read = mp_np-1

    if (mp_pid.eq.id_read) lmpreader=.true.

    call MPI_Comm_group (MPI_COMM_WORLD, world_group_id, mp_ierr)

Espen Sollum's avatar
Espen Sollum committed
270
! Create a MPI group/communicator that will calculate trajectories. 
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
! Skip this step if program is run with only a few processes
!************************************************************************
    allocate(mp_partgroup_rank(0:mp_np-2))

! This allows checking for allocation of arrays when no subgroup is used
    mp_partgroup_pid=0

    if (read_grp_min.lt.2) then
      write(*,*) '#### mpi_mod::mpif_init> ERROR ####', &
           & 'read_grp_min should be at least 2. Exiting'
      stop
    end if

    if (mp_np.ge.read_grp_min) then
      lmp_use_reader = .true.

Espen Sollum's avatar
Espen Sollum committed
287
! Map the subgroup IDs= 0,1,2,...,mp_np-2, skipping reader process
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
      j=-1 
      do i=0, mp_np-2 !loop over all (subgroup) IDs
        j=j+1
        if (i.eq.id_read) then
          j=j+1
        end if
        mp_partgroup_rank(i) = j
      end do

      call MPI_Group_incl (world_group_id, mp_np-1, mp_partgroup_rank, &
           &mp_partgroup_pid, mp_ierr)
      if (mp_ierr /= 0) goto 100
      call MPI_Comm_create (MPI_COMM_WORLD, mp_partgroup_pid, mp_partgroup_comm, mp_ierr)
      if (mp_ierr /= 0) goto 100

      if (mp_pid.ne.id_read) then
        call MPI_Comm_rank (mp_partgroup_comm, mp_partgroup_pid, mp_ierr)
        if (mp_ierr /= 0) goto 100

! Get size of new group (=mp_np-1)
        call MPI_COMM_SIZE(mp_partgroup_comm, mp_partgroup_np, mp_ierr)
        if (mp_ierr /= 0) goto 100
        if (mp_partgroup_np.ne.mp_np-1) then
          write(*,*)  '#### mpi_mod:: mpif_init> ERROR ####&
               & mp_partgroup_np.ne.mp_np-1'
          stop
        endif

      else
        mp_partgroup_pid = -1
      end if
    end if

    if (lmp_use_reader) then
      mp_comm_used = mp_partgroup_comm
      mp_partid = mp_partgroup_pid
      mp_partgroup_np=mp_np-1
    else
      mp_comm_used = MPI_COMM_WORLD
      mp_partgroup_np = mp_np
      mp_partid = mp_pid
    end if

! Set maxpart per process
332
! eso 08/2016: Increase maxpart per process, in case of unbalanced distribution
333
    maxpart_mpi=int(mp_maxpart_factor*real(maxpart)/real(mp_partgroup_np))
334
335
336
337
338
339
340
341
342
343
344

! Do not allocate particle data arrays for readwind process
    if (lmpreader.and.lmp_use_reader) then
      maxpart_mpi=0
    end if

! Set random seed for each non-root process
    if (mp_pid.gt.0) then
      s = 244
      mp_seed = -abs(mod((s*181)*((mp_pid-83)*359), 104729))
    end if
Espen Sollum's avatar
Espen Sollum committed
345
346
347
348
    if (mp_dbg_mode) then
      mp_seed=0
      if (lroot) write(*,*) 'MPI: setting seed=0'
    end if
349
350
351
352

! Allocate request array for asynchronous MPI
    if (.not.lmp_sync) then
      allocate(reqs(0:nvar_async*mp_np-1))
Espen Sollum's avatar
Espen Sollum committed
353
      reqs(:)=MPI_REQUEST_NULL
354
355
    else
      allocate(reqs(0:1))
Espen Sollum's avatar
Espen Sollum committed
356
      reqs(:)=MPI_REQUEST_NULL
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
    end if

    goto 101

100 write(*,*) '#### mpi_mod::mpif_init> ERROR ####', mp_ierr
    stop

101 end subroutine mpif_init


  subroutine mpif_mpi_barrier
!***********************************************************************
! Set MPI barrier (all processes are synchronized here), with the
! possible exception of a separate 'readwind' process.
! Optionally measure cpu/wall time.
!
!***********************************************************************
    implicit none

    if (mp_time_barrier) then
      call cpu_time(mp_barrier_time_beg)
      mp_barrier_wtime_beg = mpi_wtime()
    endif

    call MPI_BARRIER(mp_comm_used, mp_ierr)
    if (mp_ierr /= 0) goto 100

    if (mp_time_barrier) then
      call cpu_time(mp_barrier_time_end)
      mp_barrier_wtime_end = mpi_wtime()

      mp_barrier_time_total=mp_barrier_time_total+&
           &(mp_barrier_time_end-mp_barrier_time_beg)
      mp_barrier_wtime_total=mp_barrier_wtime_total+&
           &(mp_barrier_wtime_end-mp_barrier_wtime_beg)
    end if

    goto 101

100 write(*,*) '#### mpi_mod::mpif_mpi_barrier> ERROR ####', mp_ierr
    stop

101 end subroutine mpif_mpi_barrier


  subroutine mpif_com_mod_allocate
!*******************************************************************************    
! Dynamic allocation of arrays (in serial code these have size maxpart)
!
406
407
408
! Not in use anymore, moved to com_mod for interoperability with serial version
! 
! TODO: remove 
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
!*******************************************************************************
    use com_mod

    implicit none 

    integer :: array_size

    array_size = maxpart_mpi

    allocate(itra1(array_size),npoint(array_size),&
         & nclass(array_size),idt(array_size),&
         & itramem(array_size),itrasplit(array_size),&
         & xtra1(array_size),ytra1(array_size),&
         & ztra1(array_size),xmass1(array_size, maxspec))

    allocate(uap(array_size),ucp(array_size),&
         & uzp(array_size),us(array_size),&
         & vs(array_size),&
         & ws(array_size),cbt(array_size))


  end subroutine mpif_com_mod_allocate


  subroutine mpif_tm_send_dims
!***********************************************************************
! Distribute array dimensions from pid0 to all processes.
! numpart_mpi/numpart is sent to allow dynamic allocation
!
! Currently not in use.
!
! Variables of similar type (integer, real) are placed in an array
! and sent collectively, to avoid the overhead associated with individual
! MPI send/receives
!
!
!***********************************************************************
    use com_mod

    implicit none

    integer :: add_part=0

Espen Sollum's avatar
Espen Sollum committed
452
    call MPI_Bcast(numpart, 1, MPI_INTEGER, id_root, mp_comm_used, mp_ierr)
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470

! MPI subgroup does the particle-loop 
    if (lmp_use_reader) then
      if (mod(numpart,mp_partgroup_np).ne.0) add_part=1
      numpart_mpi=int(numpart/mp_partgroup_np)+add_part
    else
      if (mod(numpart,mp_np).ne.0) add_part=1
      numpart_mpi=int(numpart/mp_np)+add_part
    end if


! redefine numpart as 'numpart per process' throughout the code
!**************************************************************
    numpart = numpart_mpi

  end subroutine mpif_tm_send_dims


471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
  subroutine mpif_send_part_properties(num_part)
!***********************************************************************
! Distribute particle properties from root to all processes.
!  
! Used by init_domainfill_mpi
!
! Variables:
! 
! num_part        input, number of particles per process (rounded up)
!
!***********************************************************************
    use com_mod

    implicit none

    integer,intent(in) :: num_part
    integer :: i,jj, addone

489
490
491
492
493
494
495
496
497
498
499
500
501
! Exit if too many particles
    if (num_part.gt.maxpart_mpi) then
      write(*,*) '#####################################################'
      write(*,*) '#### ERROR - TOTAL NUMBER OF PARTICLES REQUIRED  ####'
      write(*,*) '#### EXCEEDS THE MAXIMUM ALLOWED NUMBER. REDUCE  ####'
      write(*,*) '#### EITHER NUMBER OF PARTICLES PER RELEASE POINT####'
      write(*,*) '#### OR INCREASE MAXPART.                        ####'
      write(*,*) '#####################################################'
!      call MPI_FINALIZE(mp_ierr)
      stop
    end if


502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
! Time for MPI communications
!****************************
    if (mp_measure_time) call mpif_mtime('commtime',0)

! Distribute variables, send from pid 0 to other processes (including itself)
!****************************************************************************

    call MPI_SCATTER(nclass_tmp,num_part,MPI_INTEGER,nclass,&
         &num_part,MPI_INTEGER,id_root,mp_comm_used,mp_ierr)
    if (mp_ierr /= 0) goto 600 
    call MPI_SCATTER(npoint_tmp,num_part,MPI_INTEGER,npoint,&
         &num_part,MPI_INTEGER,id_root,mp_comm_used,mp_ierr)
    if (mp_ierr /= 0) goto 600 
    call MPI_SCATTER(itra1_tmp,num_part,MPI_INTEGER,itra1,&
         &num_part,MPI_INTEGER,id_root,mp_comm_used,mp_ierr)
    if (mp_ierr /= 0) goto 600 
    call MPI_SCATTER(idt_tmp,num_part,MPI_INTEGER,idt,&
         &num_part,MPI_INTEGER,id_root,mp_comm_used,mp_ierr)
    if (mp_ierr /= 0) goto 600 
    call MPI_SCATTER(itramem_tmp,num_part,MPI_INTEGER,itramem,&
         &num_part,MPI_INTEGER,id_root,mp_comm_used,mp_ierr)
    if (mp_ierr /= 0) goto 600 
    call MPI_SCATTER(itrasplit_tmp,num_part,MPI_INTEGER,itrasplit,&
         &num_part,MPI_INTEGER,id_root,mp_comm_used,mp_ierr)
    if (mp_ierr /= 0) goto 600 
    call MPI_SCATTER(xtra1_tmp,num_part,mp_dp,xtra1,&
         &num_part,mp_dp,id_root,mp_comm_used,mp_ierr)
    if (mp_ierr /= 0) goto 600 
    call MPI_SCATTER(ytra1_tmp,num_part,mp_dp,ytra1,&
         &num_part,mp_dp,id_root,mp_comm_used,mp_ierr)
    if (mp_ierr /= 0) goto 600 
    call MPI_SCATTER(ztra1_tmp,num_part,mp_sp,ztra1,&
         &num_part,mp_sp,id_root,mp_comm_used,mp_ierr)
    if (mp_ierr /= 0) goto 600 
    do i=1,nspec
      call MPI_SCATTER(xmass1_tmp(:,i),num_part,mp_sp,xmass1(:,i),&
           &num_part,mp_sp,id_root,mp_comm_used,mp_ierr)
      if (mp_ierr /= 0) goto 600 
    end do

    if (mp_measure_time) call mpif_mtime('commtime',1)

    goto 601

600 write(*,*) "mpi_mod> mp_ierr \= 0", mp_ierr
    stop

! After the transfer of particles it it possible that the value of
550
! "numpart" is larger than the actual used, so we reduce it if there are
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
! invalid particles at the end of the arrays
601 do i=num_part, 1, -1
      if (itra1(i).eq.-999999999) then
        numpart=numpart-1
      else
        exit
      end if
    end do


!601 end subroutine mpif_send_part_properties
  end subroutine mpif_send_part_properties


  subroutine mpif_calculate_part_redist(itime)
!***********************************************************************
! Calculate number of particles to redistribute between processes. This routine 
! can be called at regular time intervals to keep a level number of
! particles on each process.
!
! First, all processes report their local "numpart" to each other, which is
! stored in array "numpart_mpi(np)". This array is sorted from low to
! high values, along with a corresponding process ID array "idx_arr(np)".
! If the relative difference between the highest and lowest "numpart_mpi" value
! is above a threshold, particles are transferred from process idx_arr(np-1)
! to process idx_arr(0). Repeat for processes idx_arr(np-i) and idx_arr(i)
! for all valid i. 
! Note: If np is an odd number, the process with the median
! number of particles is left unchanged
!
! VARIABLES
! itime        input, current time
!***********************************************************************
    use com_mod

    implicit none
    
    integer, intent(in) :: itime
    real :: pmin,z
    integer :: i,jj,nn, num_part=1,m,imin, num_trans
    logical :: first_iter
    integer,allocatable,dimension(:) :: numparticles_mpi, idx_arr
    real,allocatable,dimension(:) :: sorted ! TODO: we don't really need this

! Exit if running with only 1 process
!************************************************************************
    if (mp_np.eq.1) return

! All processes exchange information on number of particles
!****************************************************************************
    allocate(numparticles_mpi(0:mp_partgroup_np-1), &
         &idx_arr(0:mp_partgroup_np-1), sorted(0:mp_partgroup_np-1))

    call MPI_Allgather(numpart, 1, MPI_INTEGER, numparticles_mpi, &
         & 1, MPI_INTEGER, mp_comm_used, mp_ierr)


! Sort from lowest to highest
! Initial guess: correct order
    sorted(:) = numparticles_mpi(:)
    do i=0, mp_partgroup_np-1
      idx_arr(i) = i
    end do

! For each successive element in index array, see if a lower value exists
    do i=0, mp_partgroup_np-2
      pmin=sorted(i)
      imin=idx_arr(i)
      m=i+1
      do jj=m, mp_partgroup_np-1
        if (pmin.le.sorted(jj)) cycle
        z=pmin
        pmin=sorted(jj)
        sorted(jj)=z

        nn=imin
        imin=idx_arr(jj)
        idx_arr(jj)=nn
      end do
      sorted(i)=pmin
      idx_arr(i)=imin
    end do

! For each pair of processes, transfer particles if the difference is above a
! limit, and numpart at sending process large enough

    m=mp_partgroup_np-1 ! index for last sorted process (most particles)
    do i=0,mp_partgroup_np/2-1
      num_trans = numparticles_mpi(idx_arr(m)) - numparticles_mpi(idx_arr(i))
      if (mp_partid.eq.idx_arr(m).or.mp_partid.eq.idx_arr(i)) then
        if ( numparticles_mpi(idx_arr(m)).gt.mp_min_redist.and.&
             & real(num_trans)/real(numparticles_mpi(idx_arr(m))).gt.mp_redist_fract) then
643
644
645
646
! DBG
          ! write(*,*) 'mp_partid, idx_arr(m), idx_arr(i), mp_min_redist, num_trans, numparticles_mpi', &
          !      &mp_partid, idx_arr(m), idx_arr(i), mp_min_redist, num_trans, numparticles_mpi
! DBG
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
          call mpif_redist_part(itime, idx_arr(m), idx_arr(i), num_trans/2)
        end if
      end if
      m=m-1
    end do

    deallocate(numparticles_mpi, idx_arr, sorted)

  end subroutine mpif_calculate_part_redist


  subroutine mpif_redist_part(itime, src_proc, dest_proc, num_trans)
!***********************************************************************
! Transfer particle properties between two arbitrary processes.
! 
! VARIABLES
!
! itime           input, current time
! src_proc        input, ID of source process
! dest_proc       input, ID of destination process
! num_trans       input, number of particles to transfer
!
!************************************************************************
    use com_mod !TODO:  ,only: nclass etc

    implicit none 

    integer, intent(in) :: itime, src_proc, dest_proc, num_trans
    integer :: ll, ul ! lower and upper indices in arrays
    integer :: arr_sz ! size of temporary arrays
    integer :: mtag   ! MPI message tag
    integer :: i, j, minpart, ipart, maxnumpart
 
680
681
682
683
684
685
686
687
! Check for invalid input arguments
!**********************************
 if (src_proc.eq.dest_proc) then
   write(*,*) '<mpi_mod::mpif_redist_part>: Error: &
        &src_proc.eq.dest_proc' 
   stop
 end if

688
689
690
691
692
693
694
695
696
697
698
699
700
! Measure time for MPI communications
!************************************
    if (mp_measure_time) call mpif_mtime('commtime',0)

! Send particles
!***************
    if (mp_partid.eq.src_proc) then
      mtag = 2000

! Array indices for the transferred particles
      ll=numpart-num_trans+1
      ul=numpart

701
702
703
704
705
      if (mp_dev_mode) then
        write(*,FMT='(72("#"))')
        write(*,*) "Sending ", num_trans, "particles (from/to)", src_proc, dest_proc
        write(*,*) "numpart before: ", numpart
      end if
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743

      call MPI_SEND(nclass(ll:ul),num_trans,&
           & MPI_INTEGER,dest_proc,mtag+1,mp_comm_used,mp_ierr)

      call MPI_SEND(npoint(ll:ul),num_trans,&
           & MPI_INTEGER,dest_proc,mtag+2,mp_comm_used,mp_ierr)

      call MPI_SEND(itra1(ll:ul),num_trans, &
           & MPI_INTEGER,dest_proc,mtag+3,mp_comm_used,mp_ierr)

      call MPI_SEND(idt(ll:ul),num_trans, &
           & MPI_INTEGER,dest_proc,mtag+4,mp_comm_used,mp_ierr)

      call MPI_SEND(itramem(ll:ul),num_trans, &
           & MPI_INTEGER,dest_proc,mtag+5,mp_comm_used,mp_ierr)

      call MPI_SEND(itrasplit(ll:ul),num_trans,&
           & MPI_INTEGER,dest_proc,mtag+6,mp_comm_used,mp_ierr)

      call MPI_SEND(xtra1(ll:ul),num_trans, &
           & mp_dp,dest_proc,mtag+7,mp_comm_used,mp_ierr)

      call MPI_SEND(ytra1(ll:ul),num_trans,&
           & mp_dp,dest_proc,mtag+8,mp_comm_used,mp_ierr)

      call MPI_SEND(ztra1(ll:ul),num_trans,&
           & mp_sp,dest_proc,mtag+9,mp_comm_used,mp_ierr)

      do j=1,nspec
        call MPI_SEND(xmass1(ll:ul,j),num_trans,&
             & mp_sp,dest_proc,mtag+(9+j),mp_comm_used,mp_ierr)
      end do

! Terminate transferred particles, update numpart
      itra1(ll:ul) = -999999999

      numpart = numpart-num_trans

744
745
746
747
      if (mp_dev_mode) then
        write(*,*) "numpart after: ", numpart
        write(*,FMT='(72("#"))')
      end if
748
749
750
751
752
753
754
755
756
757

    else if (mp_partid.eq.dest_proc) then

! Receive particles
!******************
      mtag = 2000
      ! Array indices for the transferred particles
      ll=numpart+1
      ul=numpart+num_trans

758
759
760
761
762
      if (mp_dev_mode) then
        write(*,FMT='(72("#"))')
        write(*,*) "Receiving ", num_trans, "particles (from/to)", src_proc, dest_proc
        write(*,*) "numpart before: ", numpart
      end if
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811

! We could receive the data directly at nclass(ll:ul) etc., but this leaves
! vacant spaces at lower indices. Using temporary arrays instead.
      arr_sz = ul-ll+1
      allocate(itra1_tmp(arr_sz),npoint_tmp(arr_sz),nclass_tmp(arr_sz),&
           & idt_tmp(arr_sz),itramem_tmp(arr_sz),itrasplit_tmp(arr_sz),&
           & xtra1_tmp(arr_sz),ytra1_tmp(arr_sz),ztra1_tmp(arr_sz),&
           & xmass1_tmp(arr_sz, maxspec))
      
      call MPI_RECV(nclass_tmp,num_trans,&
           & MPI_INTEGER,src_proc,mtag+1,mp_comm_used,mp_status,mp_ierr)

      call MPI_RECV(npoint_tmp,num_trans,&
           & MPI_INTEGER,src_proc,mtag+2,mp_comm_used,mp_status,mp_ierr)

      call MPI_RECV(itra1_tmp,num_trans, &
           & MPI_INTEGER,src_proc,mtag+3,mp_comm_used,mp_status,mp_ierr)

      call MPI_RECV(idt_tmp,num_trans, &
           & MPI_INTEGER,src_proc,mtag+4,mp_comm_used,mp_status,mp_ierr)

      call MPI_RECV(itramem_tmp,num_trans, &
           & MPI_INTEGER,src_proc,mtag+5,mp_comm_used,mp_status,mp_ierr)

      call MPI_RECV(itrasplit_tmp,num_trans,&
           & MPI_INTEGER,src_proc,mtag+6,mp_comm_used,mp_status,mp_ierr)

      call MPI_RECV(xtra1_tmp,num_trans, &
           & mp_dp,src_proc,mtag+7,mp_comm_used,mp_status,mp_ierr)

      call MPI_RECV(ytra1_tmp,num_trans,&
           & mp_dp,src_proc,mtag+8,mp_comm_used,mp_status,mp_ierr)

      call MPI_RECV(ztra1_tmp,num_trans,&
           & mp_sp,src_proc,mtag+9,mp_comm_used,mp_status,mp_ierr)

      do j=1,nspec
        call MPI_RECV(xmass1_tmp(:,j),num_trans,&
             & mp_sp,src_proc,mtag+(9+j),mp_comm_used,mp_status,mp_ierr)
      end do

! This is the maximum value numpart can possibly have after the transfer
      maxnumpart=numpart+num_trans

! Search for vacant space and copy from temporary storage
!********************************************************
      minpart=1
      do i=1, num_trans
! Take into acount that we may have transferred invalid particles
812
        if (itra1_tmp(i).ne.itime) cycle
813
814
        do ipart=minpart,maxnumpart
          if (itra1(ipart).ne.itime) then
815
816
817
818
819
820
821
822
823
824
            itra1(ipart) = itra1_tmp(i)
            npoint(ipart) = npoint_tmp(i)
            nclass(ipart) = nclass_tmp(i)
            idt(ipart) = idt_tmp(i)
            itramem(ipart) = itramem_tmp(i)
            itrasplit(ipart) = itrasplit_tmp(i)
            xtra1(ipart) = xtra1_tmp(i)
            ytra1(ipart) = ytra1_tmp(i)
            ztra1(ipart) = ztra1_tmp(i)
            xmass1(ipart,:) = xmass1_tmp(i,:)
825
826
            goto 200 ! Storage space has been found, stop searching
          end if
827
! :TODO: add check for if too many particles requiried
828
        end do
829
200     minpart=ipart+1
830
      end do
831
832
! Increase numpart, if necessary
      numpart=max(numpart,ipart)
833

834
835
      deallocate(itra1_tmp,npoint_tmp,nclass_tmp,idt_tmp,itramem_tmp,&
           &itrasplit_tmp,xtra1_tmp,ytra1_tmp,ztra1_tmp,xmass1_tmp)
836

837
838
839
840
      if (mp_dev_mode) then
        write(*,*) "numpart after: ", numpart
        write(*,FMT='(72("#"))')
      end if
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855

    else
! This routine should only be called by the two participating processes
      write(*,*) "ERROR: wrong process has entered mpi_mod::mpif_redist_part"
      stop
      return
    end if

! Measure time for MPI communications
!************************************
    if (mp_measure_time) call mpif_mtime('commtime',1)

  end subroutine mpif_redist_part


856
857
858
859
  subroutine mpif_tm_send_vars
!***********************************************************************
! Distribute particle variables from pid0 to all processes.
! Called from timemanager
860
! *NOT IN USE* at the moment, but can be useful for debugging
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
!
!***********************************************************************
    use com_mod

    implicit none

    integer :: i

! Time for MPI communications
!****************************
    if (mp_measure_time) call mpif_mtime('commtime',0)

! Distribute variables, send from pid 0 to other processes
!**********************************************************************

! integers
    if (lroot) then
      call MPI_SCATTER(npoint,numpart_mpi,MPI_INTEGER,MPI_IN_PLACE,&
879
           & numpart_mpi,MPI_INTEGER,id_root,mp_comm_used,mp_ierr)
880
881
      if (mp_ierr /= 0) goto 600 
      call MPI_SCATTER(idt,numpart_mpi,MPI_INTEGER,MPI_IN_PLACE,&
882
           & numpart_mpi,MPI_INTEGER,id_root,mp_comm_used,mp_ierr)
883
884
      if (mp_ierr /= 0) goto 600 
      call MPI_SCATTER(itra1,numpart_mpi,MPI_INTEGER,MPI_IN_PLACE,&
885
           & numpart_mpi,MPI_INTEGER,id_root,mp_comm_used,mp_ierr)
886
887
      if (mp_ierr /= 0) goto 600 
      call MPI_SCATTER(nclass,numpart_mpi,MPI_INTEGER,MPI_IN_PLACE,&
888
           & numpart_mpi,MPI_INTEGER,id_root,mp_comm_used,mp_ierr)
889
890
      if (mp_ierr /= 0) goto 600 
      call MPI_SCATTER(itramem,numpart_mpi,MPI_INTEGER,MPI_IN_PLACE,&
891
           & numpart_mpi,MPI_INTEGER,id_root,mp_comm_used,mp_ierr)
892
893
894
895
      if (mp_ierr /= 0) goto 600 

! int2
      call MPI_SCATTER(cbt,numpart_mpi,MPI_INTEGER2,MPI_IN_PLACE,&
896
           & numpart_mpi,MPI_INTEGER2,id_root,mp_comm_used,mp_ierr)
897
898
899
      if (mp_ierr /= 0) goto 600

! real
900
      call MPI_SCATTER(uap,numpart_mpi,mp_sp,MPI_IN_PLACE,&
901
           & numpart_mpi,mp_sp,id_root,mp_comm_used,mp_ierr)
902
      if (mp_ierr /= 0) goto 600
903
      call MPI_SCATTER(ucp,numpart_mpi,mp_sp,MPI_IN_PLACE,&
904
           & numpart_mpi,mp_sp,id_root,mp_comm_used,mp_ierr)
905
      if (mp_ierr /= 0) goto 600
906
      call MPI_SCATTER(uzp,numpart_mpi,mp_sp,MPI_IN_PLACE,&
907
           & numpart_mpi,mp_sp,id_root,mp_comm_used,mp_ierr)
908
909
      if (mp_ierr /= 0) goto 600

910
      call MPI_SCATTER(us,numpart_mpi,mp_sp,MPI_IN_PLACE,&
911
           & numpart_mpi,mp_sp,id_root,mp_comm_used,mp_ierr)
912
      if (mp_ierr /= 0) goto 600
913
      call MPI_SCATTER(vs,numpart_mpi,mp_sp,MPI_IN_PLACE,&
914
           & numpart_mpi,mp_sp,id_root,mp_comm_used,mp_ierr)
915
      if (mp_ierr /= 0) goto 600
916
      call MPI_SCATTER(ws,numpart_mpi,mp_sp,MPI_IN_PLACE,&
917
           & numpart_mpi,mp_sp,id_root,mp_comm_used,mp_ierr)
918
919
920
      if (mp_ierr /= 0) goto 600

      call MPI_SCATTER(xtra1,numpart_mpi,mp_dp,MPI_IN_PLACE,&
921
           & numpart_mpi,mp_dp,id_root,mp_comm_used,mp_ierr)
922
923
      if (mp_ierr /= 0) goto 600 
      call MPI_SCATTER(ytra1,numpart_mpi,mp_dp,MPI_IN_PLACE,&
924
           & numpart_mpi,mp_dp,id_root,mp_comm_used,mp_ierr)
925
      if (mp_ierr /= 0) goto 600 
926
      call MPI_SCATTER(ztra1,numpart_mpi,mp_sp,MPI_IN_PLACE,&
927
           & numpart_mpi,mp_sp,id_root,mp_comm_used,mp_ierr)
928
929
930
      if (mp_ierr /= 0) goto 600 

      do i=1,nspec
931
        call MPI_SCATTER(xmass1(:,i),numpart_mpi,mp_sp,MPI_IN_PLACE,&
932
             & numpart_mpi,mp_sp,id_root,mp_comm_used,mp_ierr)
933
934
935
936
937
938
        if (mp_ierr /= 0) goto 600 
      end do

    else ! (mp_pid >= 1)
! integers
      call MPI_SCATTER(npoint,numpart_mpi,MPI_INTEGER,npoint,&
939
           & numpart_mpi,MPI_INTEGER,id_root,mp_comm_used,mp_ierr)
940
941
      if (mp_ierr /= 0) goto 600 
      call MPI_SCATTER(idt,numpart_mpi,MPI_INTEGER,idt,&
942
           & numpart_mpi,MPI_INTEGER,id_root,mp_comm_used,mp_ierr)
943
944
      if (mp_ierr /= 0) goto 600 
      call MPI_SCATTER(itra1,numpart_mpi,MPI_INTEGER,itra1,&
945
           & numpart_mpi,MPI_INTEGER,id_root,mp_comm_used,mp_ierr)
946
947
      if (mp_ierr /= 0) goto 600 
      call MPI_SCATTER(nclass,numpart_mpi,MPI_INTEGER,nclass,&
948
           & numpart_mpi,MPI_INTEGER,id_root,mp_comm_used,mp_ierr)
949
950
      if (mp_ierr /= 0) goto 600 
      call MPI_SCATTER(itramem,numpart_mpi,MPI_INTEGER,itramem,&
951
           & numpart_mpi,MPI_INTEGER,id_root,mp_comm_used,mp_ierr)
952
953
954
955
      if (mp_ierr /= 0) goto 600 

! int2
      call MPI_SCATTER(cbt,numpart_mpi,MPI_INTEGER2,cbt,&
956
           & numpart_mpi,MPI_INTEGER2,id_root,mp_comm_used,mp_ierr)
957
958
959
      if (mp_ierr /= 0) goto 600

! reals
960
      call MPI_SCATTER(uap,numpart_mpi,mp_sp,uap,&
961
           & numpart_mpi,mp_sp,id_root,mp_comm_used,mp_ierr)
962
      if (mp_ierr /= 0) goto 600
963
      call MPI_SCATTER(ucp,numpart_mpi,mp_sp,ucp,&
964
           & numpart_mpi,mp_sp,id_root,mp_comm_used,mp_ierr)
965
      if (mp_ierr /= 0) goto 600
966
      call MPI_SCATTER(uzp,numpart_mpi,mp_sp,uzp,&
967
           & numpart_mpi,mp_sp,id_root,mp_comm_used,mp_ierr)
968
969
      if (mp_ierr /= 0) goto 600

970
      call MPI_SCATTER(us,numpart_mpi,mp_sp,us,&
971
           & numpart_mpi,mp_sp,id_root,mp_comm_used,mp_ierr)
972
      if (mp_ierr /= 0) goto 600
973
      call MPI_SCATTER(vs,numpart_mpi,mp_sp,vs,&
974
           & numpart_mpi,mp_sp,id_root,mp_comm_used,mp_ierr)
975
      if (mp_ierr /= 0) goto 600
976
      call MPI_SCATTER(ws,numpart_mpi,mp_sp,ws,&
977
           & numpart_mpi,mp_sp,id_root,mp_comm_used,mp_ierr)
978
979
980
      if (mp_ierr /= 0) goto 600

      call MPI_SCATTER(xtra1,numpart_mpi,mp_dp,xtra1,&
981
           & numpart_mpi,mp_dp,id_root,mp_comm_used,mp_ierr)
982
983
      if (mp_ierr /= 0) goto 600 
      call MPI_SCATTER(ytra1,numpart_mpi,mp_dp,ytra1,&
984
           & numpart_mpi,mp_dp,id_root,mp_comm_used,mp_ierr)
985
      if (mp_ierr /= 0) goto 600 
986
      call MPI_SCATTER(ztra1,numpart_mpi,mp_sp,ztra1,&
987
           & numpart_mpi,mp_sp,id_root,mp_comm_used,mp_ierr)
988
989
990
      if (mp_ierr /= 0) goto 600 

      do i=1,nspec
991
        call MPI_SCATTER(xmass1(:,i),numpart_mpi,mp_sp,xmass1(:,i),&
992
             & numpart_mpi,mp_sp,id_root,mp_comm_used,mp_ierr)
993
994
995
996
997
998
999
1000
        if (mp_ierr /= 0) goto 600 
      end do

    end if !(lroot)

    if (mp_measure_time) call mpif_mtime('commtime',1)

    goto 601