mpi_mod.f90 110 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
!**********************************************************************
! 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/>.   *
!**********************************************************************

22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
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
63
64
65
66
!*****************************************************************************
!                                                                            *
! Modification by DJM, 2017-05-09 - added #ifdef USE_MPIINPLACE cpp          *
! directive to mpif_tm_reduce_grid() to insure that MPI_IN_PLACE is          *
! used in the MPI_Reduce() only if specifically compiled with that           *
! directive.                                                                 *
!                                                                            *
67
68
69
!*****************************************************************************

  use mpi 
70
  use par_mod, only: dp,sp
71
72
73
74
75
76
77
78
79
80
  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
81
  integer, parameter, private :: read_grp_min=4
82
83
84
85
86
87
88
89
!==================================================

! 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
90
  integer, allocatable, dimension(:) :: npart_per_process
91
92
93
94
  integer :: mp_partgroup_comm, mp_partgroup_pid, mp_partgroup_np

  integer :: mp_seed=0
  integer, parameter :: mp_sp=MPI_REAL4, mp_dp=MPI_REAL8
95
  integer :: mp_cp
96
97
98
  integer, parameter :: id_root=0 ! master process

! MPI tags/requests for send/receive operation
Espen Sollum's avatar
Espen Sollum committed
99
  integer :: tm1
100
  integer, parameter :: nvar_async=26
101
!integer, dimension(:), allocatable :: tags
102
103
  integer, dimension(:), allocatable :: reqs

104
105
106
! Status array used for certain MPI operations (MPI_RECV)
  integer, dimension(MPI_STATUS_SIZE) :: mp_status

107
108
109
110
111
112
113
114
115

  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

116
! .true. if only using synchronous MPI send/recv (default)
117
118
! If setting this to .false., numwfmem must be set to 3
!===============================================================================
119
  logical :: lmp_sync=.true. 
120
121
!===============================================================================

Espen Sollum's avatar
Espen Sollum committed
122
! mp_dbg_mode       Used for debugging MPI.
123
124
125
126
! 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
127
! mp_exact_numpart  Use an extra MPI communication to give the exact number of particles
128
!                   to standard output (this does not otherwise affect the simulation)
129
130
131
  logical, parameter :: mp_dbg_mode = .false.
  logical, parameter :: mp_dev_mode = .false.
  logical, parameter :: mp_dbg_out = .false.
132
  logical, parameter :: mp_time_barrier=.true.
Espen Sollum's avatar
Espen Sollum committed
133
  logical, parameter :: mp_measure_time=.false.
134
  logical, parameter :: mp_exact_numpart=.true.
135
136

! for measuring CPU/Wall time
137
138
139
140
141
142
143
144
145
146
  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.
147
148
  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.
149
150
151
152
153
154
155
156
157
  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
158
159
160
161

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

162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
! 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


178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
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
!***********************************************************************
193
194
    use par_mod, only: maxpart, numwfmem, dep_prec, maxreceptor, maxspec
    use com_mod, only: mpi_mode, verbosity, creceptor0
195
196
197
198
199

    implicit none

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

200
! Each process gets an ID (mp_pid) in the range 0,..,mp_np-1
201
202
203
204
205
206
207
208
    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


209
! Variable mpi_mode is used to handle subroutines common to parallel/serial version
210
211
212
213
214
215
216
217
218
219
    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

220
221
222
223
! Set MPI precision to use for transferring deposition fields
!************************************************************
    if (dep_prec==dp) then
      mp_cp = MPI_REAL8
224
! TODO: write info message for serial version as well 
Espen Sollum's avatar
Espen Sollum committed
225
      if (lroot.and.verbosity>0) write(*,*) 'Using double precision for deposition fields'
226
227
    else if (dep_prec==sp) then
      mp_cp = MPI_REAL4
Espen Sollum's avatar
Espen Sollum committed
228
      if (lroot.and.verbosity>0) write(*,*) 'Using single precision for deposition fields'
229
230
231
232
233
    else
      write(*,*) 'ERROR: something went wrong setting MPI real precision'
      stop
    end if

234
235
236
! Check for sensible combination of parameters
!*********************************************

237
238
239
240
241
242
243
244
    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)
245
246
247
248
      stop
    else if (lmp_sync.and.numwfmem.ne.2.and.lroot) then
      write(*,FMT='(80("#"))')
      write(*,*) '#### mpi_mod::mpif_init> WARNING: ', &
249
           & 'numwfmem should be set to 2 for syncronous'
250
251
      write(*,*) ' reading. Results will still be valid, but unneccesary memory &
           &is allocated.'
252
      write(*,FMT='(80("#"))')
Espen Sollum's avatar
Espen Sollum committed
253
! Force "syncronized" version if all processes will call getfields
254
    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
255
256
257
258
259
260
      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
261
      lmp_sync=.true.
262
    end if
263
264

! TODO: Add more warnings for unimplemented flexpart features
265
266
267
268
269
270

! 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
!**********************************************************************

271
!    id_read = min(mp_np-1, 1)
272
273
274
275
276
277
    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
278
! Create a MPI group/communicator that will calculate trajectories. 
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
! 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
295
! Map the subgroup IDs= 0,1,2,...,mp_np-2, skipping reader process
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
332
333
334
335
336
337
338
339
      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
340
! ESO 08/2016: Increase maxpart per process, in case of unbalanced distribution
341
    maxpart_mpi=int(mp_maxpart_factor*real(maxpart)/real(mp_partgroup_np))
342
    if (mp_np == 1) maxpart_mpi = maxpart
343
344
345
346
347
348
349
350
351
352
353

! 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
354
355
356
357
    if (mp_dbg_mode) then
      mp_seed=0
      if (lroot) write(*,*) 'MPI: setting seed=0'
    end if
358
359
360
361

! 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
362
      reqs(:)=MPI_REQUEST_NULL
363
364
    else
      allocate(reqs(0:1))
Espen Sollum's avatar
Espen Sollum committed
365
      reqs(:)=MPI_REQUEST_NULL
366
367
    end if

368
369
370
! Allocate array for number of particles per process    
    allocate(npart_per_process(0:mp_partgroup_np-1))

371
372
373
374
375
376
377
! Write whether MPI_IN_PLACE is used or not
#ifdef USE_MPIINPLACE
    if (lroot) write(*,*) 'Using MPI_IN_PLACE operations'
#else
    if (lroot) allocate(creceptor0(maxreceptor,maxspec))
    if (lroot) write(*,*) 'Not using MPI_IN_PLACE operations'
#endif
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
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
    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)
!
425
426
427
! Not in use anymore, moved to com_mod for interoperability with serial version
! 
! TODO: remove 
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
!*******************************************************************************
    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
471
    call MPI_Bcast(numpart, 1, MPI_INTEGER, id_root, mp_comm_used, mp_ierr)
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489

! 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


490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
  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

508
509
510
511
512
513
514
515
516
517
518
519
520
! 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


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
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
! 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
569
! "numpart" is larger than the actual used, so we reduce it if there are
570
! invalid particles at the end of the arrays
571

572
601 do i=numpart, 1, -1
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
      if (itra1(i).eq.-999999999) then
        numpart=numpart-1
      else
        exit
      end if
    end do


  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
611
    integer,allocatable,dimension(:) :: idx_arr
612
613
614
615
616
617
618
619
    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
!****************************************************************************
620
    allocate( idx_arr(0:mp_partgroup_np-1), sorted(0:mp_partgroup_np-1))
621

622
    call MPI_Allgather(numpart, 1, MPI_INTEGER, npart_per_process, &
623
624
625
626
627
         & 1, MPI_INTEGER, mp_comm_used, mp_ierr)


! Sort from lowest to highest
! Initial guess: correct order
628
    sorted(:) = npart_per_process(:)
629
630
631
632
    do i=0, mp_partgroup_np-1
      idx_arr(i) = i
    end do

633
634
635
! Do not rebalance particles for ipout=3    
    if (ipout.eq.3) return

636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
! 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
660
      num_trans = npart_per_process(idx_arr(m)) - npart_per_process(idx_arr(i))
661
      if (mp_partid.eq.idx_arr(m).or.mp_partid.eq.idx_arr(i)) then
662
663
        if ( npart_per_process(idx_arr(m)).gt.mp_min_redist.and.&
             & real(num_trans)/real(npart_per_process(idx_arr(m))).gt.mp_redist_fract) then
664
! DBG
665
666
          ! write(*,*) 'mp_partid, idx_arr(m), idx_arr(i), mp_min_redist, num_trans, npart_per_process', &
          !      &mp_partid, idx_arr(m), idx_arr(i), mp_min_redist, num_trans, npart_per_process
667
! DBG
668
669
670
671
672
673
          call mpif_redist_part(itime, idx_arr(m), idx_arr(i), num_trans/2)
        end if
      end if
      m=m-1
    end do

674
    deallocate(idx_arr, sorted)
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700

  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
 
701
702
703
704
705
706
707
708
! 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

709
710
711
712
713
714
715
716
717
718
719
720
721
! 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

722
723
724
725
726
      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
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764

      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

765
766
767
768
      if (mp_dev_mode) then
        write(*,*) "numpart after: ", numpart
        write(*,FMT='(72("#"))')
      end if
769
770
771
772
773
774
775
776
777
778

    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

779
780
781
782
783
      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
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
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832

! 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
833
        if (itra1_tmp(i).ne.itime) cycle
834
835
        do ipart=minpart,maxnumpart
          if (itra1(ipart).ne.itime) then
836
837
838
839
840
841
842
843
844
845
            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,:)
846
847
            goto 200 ! Storage space has been found, stop searching
          end if
848
! :TODO: add check for if too many particles requiried
849
        end do
850
200     minpart=ipart+1
851
      end do
852
853
! Increase numpart, if necessary
      numpart=max(numpart,ipart)
854

855
856
      deallocate(itra1_tmp,npoint_tmp,nclass_tmp,idt_tmp,itramem_tmp,&
           &itrasplit_tmp,xtra1_tmp,ytra1_tmp,ztra1_tmp,xmass1_tmp)
857

858
859
860
861
      if (mp_dev_mode) then
        write(*,*) "numpart after: ", numpart
        write(*,FMT='(72("#"))')
      end if
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876

    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


877
878
879
880
  subroutine mpif_tm_send_vars
!***********************************************************************
! Distribute particle variables from pid0 to all processes.
! Called from timemanager
881
! *NOT IN USE* at the moment, but can be useful for debugging
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
!
!***********************************************************************
    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,&
900
           & numpart_mpi,MPI_INTEGER,id_root,mp_comm_used,mp_ierr)
901
902
      if (mp_ierr /= 0) goto 600 
      call MPI_SCATTER(idt,numpart_mpi,MPI_INTEGER,MPI_IN_PLACE,&
903
           & numpart_mpi,MPI_INTEGER,id_root,mp_comm_used,mp_ierr)
904
905
      if (mp_ierr /= 0) goto 600 
      call MPI_SCATTER(itra1,numpart_mpi,MPI_INTEGER,MPI_IN_PLACE,&
906
           & numpart_mpi,MPI_INTEGER,id_root,mp_comm_used,mp_ierr)
907
908
      if (mp_ierr /= 0) goto 600 
      call MPI_SCATTER(nclass,numpart_mpi,MPI_INTEGER,MPI_IN_PLACE,&
909
           & numpart_mpi,MPI_INTEGER,id_root,mp_comm_used,mp_ierr)
910
911
      if (mp_ierr /= 0) goto 600 
      call MPI_SCATTER(itramem,numpart_mpi,MPI_INTEGER,MPI_IN_PLACE,&
912
           & numpart_mpi,MPI_INTEGER,id_root,mp_comm_used,mp_ierr)
913
914
915
916
      if (mp_ierr /= 0) goto 600 

! int2
      call MPI_SCATTER(cbt,numpart_mpi,MPI_INTEGER2,MPI_IN_PLACE,&
917
           & numpart_mpi,MPI_INTEGER2,id_root,mp_comm_used,mp_ierr)
918
919
920
      if (mp_ierr /= 0) goto 600

! real
921
      call MPI_SCATTER(uap,numpart_mpi,mp_sp,MPI_IN_PLACE,&
922
           & numpart_mpi,mp_sp,id_root,mp_comm_used,mp_ierr)
923
      if (mp_ierr /= 0) goto 600
924
      call MPI_SCATTER(ucp,numpart_mpi,mp_sp,MPI_IN_PLACE,&
925
           & numpart_mpi,mp_sp,id_root,mp_comm_used,mp_ierr)
926
      if (mp_ierr /= 0) goto 600
927
      call MPI_SCATTER(uzp,numpart_mpi,mp_sp,MPI_IN_PLACE,&
928
           & numpart_mpi,mp_sp,id_root,mp_comm_used,mp_ierr)
929
930
      if (mp_ierr /= 0) goto 600

931
      call MPI_SCATTER(us,numpart_mpi,mp_sp,MPI_IN_PLACE,&
932
           & numpart_mpi,mp_sp,id_root,mp_comm_used,mp_ierr)
933
      if (mp_ierr /= 0) goto 600
934
      call MPI_SCATTER(vs,numpart_mpi,mp_sp,MPI_IN_PLACE,&
935
           & numpart_mpi,mp_sp,id_root,mp_comm_used,mp_ierr)
936
      if (mp_ierr /= 0) goto 600
937
      call MPI_SCATTER(ws,numpart_mpi,mp_sp,MPI_IN_PLACE,&
938
           & numpart_mpi,mp_sp,id_root,mp_comm_used,mp_ierr)
939
940
941
      if (mp_ierr /= 0) goto 600

      call MPI_SCATTER(xtra1,numpart_mpi,mp_dp,MPI_IN_PLACE,&
942
           & numpart_mpi,mp_dp,id_root,mp_comm_used,mp_ierr)
943
944
      if (mp_ierr /= 0) goto 600 
      call MPI_SCATTER(ytra1,numpart_mpi,mp_dp,MPI_IN_PLACE,&
945
           & numpart_mpi,mp_dp,id_root,mp_comm_used,mp_ierr)
946
      if (mp_ierr /= 0) goto 600 
947
      call MPI_SCATTER(ztra1,numpart_mpi,mp_sp,MPI_IN_PLACE,&
948
           & numpart_mpi,mp_sp,id_root,mp_comm_used,mp_ierr)
949
950
951
      if (mp_ierr /= 0) goto 600 

      do i=1,nspec
952
        call MPI_SCATTER(xmass1(:,i),numpart_mpi,mp_sp,MPI_IN_PLACE,&
953
             & numpart_mpi,mp_sp,id_root,mp_comm_used,mp_ierr)
954
955
956
957
958
959
        if (mp_ierr /= 0) goto 600 
      end do

    else ! (mp_pid >= 1)
! integers
      call MPI_SCATTER(npoint,numpart_mpi,MPI_INTEGER,npoint,&
960
           & numpart_mpi,MPI_INTEGER,id_root,mp_comm_used,mp_ierr)
961
962
      if (mp_ierr /= 0) goto 600 
      call MPI_SCATTER(idt,numpart_mpi,MPI_INTEGER,idt,&
963
           & numpart_mpi,MPI_INTEGER,id_root,mp_comm_used,mp_ierr)
964
965
      if (mp_ierr /= 0) goto 600 
      call MPI_SCATTER(itra1,numpart_mpi,MPI_INTEGER,itra1,&
966
           & numpart_mpi,MPI_INTEGER,id_root,mp_comm_used,mp_ierr)
967
968
      if (mp_ierr /= 0) goto 600 
      call MPI_SCATTER(nclass,numpart_mpi,MPI_INTEGER,nclass,&
969
           & numpart_mpi,MPI_INTEGER,id_root,mp_comm_used,mp_ierr)
970
971
      if (mp_ierr /= 0) goto 600 
      call MPI_SCATTER(itramem,numpart_mpi,MPI_INTEGER,itramem,&
972
           & numpart_mpi,MPI_INTEGER,id_root,mp_comm_used,mp_ierr)
973
974
975
976
      if (mp_ierr /= 0) goto 600 

! int2
      call MPI_SCATTER(cbt,numpart_mpi,MPI_INTEGER2,cbt,&
977
           & numpart_mpi,MPI_INTEGER2,id_root,mp_comm_used,mp_ierr)
978
979
980
      if (mp_ierr /= 0) goto 600

! reals
981
      call MPI_SCATTER(uap,numpart_mpi,mp_sp,uap,&
982
           & numpart_mpi,mp_sp,id_root,mp_comm_used,mp_ierr)
983
      if (mp_ierr /= 0) goto 600
984
      call MPI_SCATTER(ucp,numpart_mpi,mp_sp,ucp,&
985
           & numpart_mpi,mp_sp,id_root,mp_comm_used,mp_ierr)
986
      if (mp_ierr /= 0) goto 600
987
      call MPI_SCATTER(uzp,numpart_mpi,mp_sp,uzp,&
988
           & numpart_mpi,mp_sp,id_root,mp_comm_used,mp_ierr)
989
990
      if (mp_ierr /= 0) goto 600

991
      call MPI_SCATTER(us,numpart_mpi,mp_sp,us,&
992
           & numpart_mpi,mp_sp,id_root,mp_comm_used,mp_ierr)
993
      if (mp_ierr /= 0) goto 600
994
      call MPI_SCATTER(vs,numpart_mpi,mp_sp,vs,&
995
           & numpart_mpi,mp_sp,id_root,mp_comm_used,mp_ierr)
996
      if (mp_ierr /= 0) goto 600
997
      call MPI_SCATTER(ws,numpart_mpi,mp_sp,ws,&
998
           & numpart_mpi,mp_sp,id_root,mp_comm_used,mp_ierr)
999
1000
      if (mp_ierr /= 0) goto 600