engine.c 69.8 KB
Newer Older
Pedro Gonnet's avatar
Pedro Gonnet committed
1
/*******************************************************************************
2
 * This file is part of SWIFT.
3
 * Copyright (c) 2012 Pedro Gonnet (pedro.gonnet@durham.ac.uk)
4
 *
Pedro Gonnet's avatar
Pedro Gonnet committed
5
6
7
8
 * This program is free software: you can redistribute it and/or modify
 * it under the terms of the GNU Lesser General Public License as published
 * by the Free Software Foundation, either version 3 of the License, or
 * (at your option) any later version.
9
 *
Pedro Gonnet's avatar
Pedro Gonnet committed
10
11
12
13
 * 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.
14
 *
Pedro Gonnet's avatar
Pedro Gonnet committed
15
16
 * You should have received a copy of the GNU Lesser General Public License
 * along with this program.  If not, see <http://www.gnu.org/licenses/>.
17
 *
Pedro Gonnet's avatar
Pedro Gonnet committed
18
19
20
21
22
23
24
25
26
 ******************************************************************************/

/* Config parameters. */
#include "../config.h"

/* Some standard headers. */
#include <float.h>
#include <limits.h>
#include <sched.h>
27
28
29
30
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <unistd.h>
Angus Lepper's avatar
Angus Lepper committed
31
#include <stdbool.h>
Pedro Gonnet's avatar
Pedro Gonnet committed
32

33
34
/* MPI headers. */
#ifdef WITH_MPI
35
#include <mpi.h>
36
/* METIS headers only used when MPI is also available. */
37
#ifdef HAVE_METIS
38
#include <metis.h>
39
40
41
#endif
#endif

Angus Lepper's avatar
Angus Lepper committed
42
43
44
45
#ifdef HAVE_LIBNUMA
#include <numa.h>
#endif

46
/* This object's header. */
Pedro Gonnet's avatar
Pedro Gonnet committed
47
#include "engine.h"
48
49

/* Local headers. */
50
#include "atomic.h"
51
#include "cell.h"
52
53
#include "cycle.h"
#include "debug.h"
54
#include "error.h"
55
#include "part.h"
56
#include "timers.h"
Pedro Gonnet's avatar
Pedro Gonnet committed
57

58
59
const char *engine_policy_names[10] = {
    "none",   "rand",     "steal",     "keep", "block",
60
    "fix_dt", "cpu_tight", "mpi",  "numa_affinity"};
61

62
63
64
/** The rank of the engine as a global variable (for messages). */
int engine_rank;

65
66
67
68
69
70
71
72
73
74
/**
 * @brief Link a density/force task to a cell.
 *
 * @param e The #engine.
 * @param l The #link.
 * @param t The #task.
 *
 * @return The new #link pointer.
 */

75
struct link *engine_addlink(struct engine *e, struct link *l, struct task *t) {
76

77
78
  const int ind = atomic_inc(&e->nr_links);
  if (ind >= e->size_links) {
79
    error("Link table overflow.");
80
81
  }
  struct link *res = &e->links[ind];
82
83
84
85
  res->next = l;
  res->t = t;
  return res;
}
86
87
88
89
90
91
92
93

/**
 * @brief Generate the ghost and kick tasks for a hierarchy of cells.
 *
 * @param e The #engine.
 * @param c The #cell.
 * @param super The super #cell.
 */
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111

void engine_mkghosts(struct engine *e, struct cell *c, struct cell *super) {

  int k;
  struct scheduler *s = &e->sched;

  /* Am I the super-cell? */
  if (super == NULL && c->nr_tasks > 0) {

    /* Remember me. */
    super = c;

    /* Local tasks only... */
    if (c->nodeID == e->nodeID) {

      /* Generate the ghost task. */
      c->ghost = scheduler_addtask(s, task_type_ghost, task_subtype_none, 0, 0,
                                   c, NULL, 0);
Matthieu Schaller's avatar
Matthieu Schaller committed
112
113
114
115
      /* Add the drift task. */
      c->drift = scheduler_addtask(s, task_type_drift, task_subtype_none, 0, 0,
                                   c, NULL, 0);
      /* Add the init task. */
Matthieu Schaller's avatar
Matthieu Schaller committed
116
117
      c->init = scheduler_addtask(s, task_type_init, task_subtype_none, 0, 0, c,
                                  NULL, 0);
Matthieu Schaller's avatar
Matthieu Schaller committed
118
      /* Add the kick task. */
Matthieu Schaller's avatar
Matthieu Schaller committed
119
120
      c->kick = scheduler_addtask(s, task_type_kick, task_subtype_none, 0, 0, c,
                                  NULL, 0);
121
    }
122
  }
123

124
125
126
127
128
129
130
131
  /* Set the super-cell. */
  c->super = super;

  /* Recurse. */
  if (c->split)
    for (k = 0; k < 8; k++)
      if (c->progeny[k] != NULL) engine_mkghosts(e, c->progeny[k], super);
}
132

133
/**
134
 * @brief Redistribute the particles amongst the nodes according
135
136
137
138
139
 *      to their cell's node IDs.
 *
 * @param e The #engine.
 */

140
void engine_redistribute(struct engine *e) {
141

142
#ifdef WITH_MPI
143

144
145
146
147
148
149
150
151
  int nr_nodes = e->nr_nodes, nodeID = e->nodeID;
  struct space *s = e->s;
  int my_cells = 0;
  int *cdim = s->cdim;
  struct cell *cells = s->cells;
  int nr_cells = s->nr_cells;

  /* Start by sorting the particles according to their nodes and
152
153
     getting the counts. The counts array is indexed as
     count[from * nr_nodes + to]. */
154
155
156
157
158
159
160
161
162
163
164
165
  int *counts, *dest;
  double ih[3], dim[3];
  ih[0] = s->ih[0];
  ih[1] = s->ih[1];
  ih[2] = s->ih[2];
  dim[0] = s->dim[0];
  dim[1] = s->dim[1];
  dim[2] = s->dim[2];
  if ((counts = (int *)malloc(sizeof(int) *nr_nodes *nr_nodes)) == NULL ||
      (dest = (int *)malloc(sizeof(int) * s->nr_parts)) == NULL)
    error("Failed to allocate count and dest buffers.");
  bzero(counts, sizeof(int) * nr_nodes * nr_nodes);
166
167
168
  struct part *parts = s->parts;
  for (int k = 0; k < s->nr_parts; k++) {
    for (int j = 0; j < 3; j++) {
169
170
171
172
173
      if (parts[k].x[j] < 0.0)
        parts[k].x[j] += dim[j];
      else if (parts[k].x[j] >= dim[j])
        parts[k].x[j] -= dim[j];
    }
Pedro Gonnet's avatar
Pedro Gonnet committed
174
175
    const int cid = cell_getid(cdim, parts[k].x[0] * ih[0],
                               parts[k].x[1] * ih[1], parts[k].x[2] * ih[2]);
176
177
178
    /* if (cid < 0 || cid >= s->nr_cells)
       error("Bad cell id %i for part %i at [%.3e,%.3e,%.3e].",
             cid, k, parts[k].x[0], parts[k].x[1], parts[k].x[2]); */
179
180
181
    dest[k] = cells[cid].nodeID;
    counts[nodeID * nr_nodes + dest[k]] += 1;
  }
182
  space_parts_sort(s, dest, s->nr_parts, 0, nr_nodes - 1);
183
184
185
186
187
188
189
190

  /* Get all the counts from all the nodes. */
  if (MPI_Allreduce(MPI_IN_PLACE, counts, nr_nodes * nr_nodes, MPI_INT, MPI_SUM,
                    MPI_COMM_WORLD) != MPI_SUCCESS)
    error("Failed to allreduce particle transfer counts.");

  /* Get the new number of parts for this node, be generous in allocating. */
  int nr_parts = 0;
191
  for (int k = 0; k < nr_nodes; k++) nr_parts += counts[k * nr_nodes + nodeID];
192
193
  struct part *parts_new = NULL;
  struct xpart *xparts_new = NULL, *xparts = s->xparts;
194
195
196
197
198
199
200
201
202
203
204
  if (posix_memalign((void **)&parts_new, part_align,
                     sizeof(struct part) * nr_parts * 1.2) != 0 ||
      posix_memalign((void **)&xparts_new, part_align,
                     sizeof(struct xpart) * nr_parts * 1.2) != 0)
    error("Failed to allocate new part data.");

  /* Emit the sends and recvs for the particle data. */
  MPI_Request *reqs;
  if ((reqs = (MPI_Request *)malloc(sizeof(MPI_Request) * 4 * nr_nodes)) ==
      NULL)
    error("Failed to allocate MPI request list.");
205
206
207
208
209
210
211
212
213
214
215
216
217
  for (int k = 0; k < 4 * nr_nodes; k++) reqs[k] = MPI_REQUEST_NULL;
  for (int offset_send = 0, offset_recv = 0, k = 0; k < nr_nodes; k++) {
    int ind_send = nodeID * nr_nodes + k;
    int ind_recv = k * nr_nodes + nodeID;
    if (counts[ind_send] > 0) {
      if (k == nodeID) {
        memcpy(&parts_new[offset_recv], &s->parts[offset_send],
               sizeof(struct part) * counts[ind_recv]);
        memcpy(&xparts_new[offset_recv], &s->xparts[offset_send],
               sizeof(struct xpart) * counts[ind_recv]);
        offset_send += counts[ind_send];
        offset_recv += counts[ind_recv];
      } else {
218
        if (MPI_Isend(&s->parts[offset_send], counts[ind_send],
219
                      e->part_mpi_type, k, 2 * ind_send + 0, MPI_COMM_WORLD,
Pedro Gonnet's avatar
Pedro Gonnet committed
220
                      &reqs[4 * k]) != MPI_SUCCESS)
221
          error("Failed to isend parts to node %i.", k);
222
        if (MPI_Isend(&s->xparts[offset_send], counts[ind_send],
223
                      e->xpart_mpi_type, k, 2 * ind_send + 1, MPI_COMM_WORLD,
Pedro Gonnet's avatar
Pedro Gonnet committed
224
                      &reqs[4 * k + 1]) != MPI_SUCCESS)
225
226
227
          error("Failed to isend xparts to node %i.", k);
        offset_send += counts[ind_send];
      }
228
    }
229
    if (k != nodeID && counts[ind_recv] > 0) {
230
231
      if (MPI_Irecv(&parts_new[offset_recv], counts[ind_recv], e->part_mpi_type,
                    k, 2 * ind_recv + 0, MPI_COMM_WORLD,
Pedro Gonnet's avatar
Pedro Gonnet committed
232
                    &reqs[4 * k + 2]) != MPI_SUCCESS)
233
        error("Failed to emit irecv of parts from node %i.", k);
234
      if (MPI_Irecv(&xparts_new[offset_recv], counts[ind_recv],
235
                    e->xpart_mpi_type, k, 2 * ind_recv + 1, MPI_COMM_WORLD,
Pedro Gonnet's avatar
Pedro Gonnet committed
236
                    &reqs[4 * k + 3]) != MPI_SUCCESS)
237
        error("Failed to emit irecv of parts from node %i.", k);
238
      offset_recv += counts[ind_recv];
239
240
    }
  }
241

242
243
244
245
  /* Wait for all the sends and recvs to tumble in. */
  MPI_Status stats[4 * nr_nodes];
  int res;
  if ((res = MPI_Waitall(4 * nr_nodes, reqs, stats)) != MPI_SUCCESS) {
246
    for (int k = 0; k < 4 * nr_nodes; k++) {
247
248
249
250
      char buff[MPI_MAX_ERROR_STRING];
      int res;
      MPI_Error_string(stats[k].MPI_ERROR, buff, &res);
      message("request %i has error '%s'.", k, buff);
251
    }
252
253
    error("Failed during waitall for part data.");
  }
254

255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
  /* Verify that all parts are in the right place. */
  /* for ( k = 0 ; k < nr_parts ; k++ ) {
      cid = cell_getid( cdim , parts_new[k].x[0]*ih[0] , parts_new[k].x[1]*ih[1]
     , parts_new[k].x[2]*ih[2] );
      if ( cells[ cid ].nodeID != nodeID )
          error( "Received particle (%i) that does not belong here (nodeID=%i)."
     , k , cells[ cid ].nodeID );
      } */

  /* Set the new part data, free the old. */
  free(parts);
  free(xparts);
  s->parts = parts_new;
  s->xparts = xparts_new;
  s->nr_parts = nr_parts;
  s->size_parts = 1.2 * nr_parts;

  /* Be verbose about what just happened. */
273
  for (int k = 0; k < nr_cells; k++)
274
275
276
277
278
279
280
281
282
    if (cells[k].nodeID == nodeID) my_cells += 1;
  message("node %i now has %i parts in %i cells.", nodeID, nr_parts, my_cells);

  /* Clean up other stuff. */
  free(reqs);
  free(counts);
  free(dest);

#else
283
  error("SWIFT was not compiled with MPI support.");
284
285
#endif
}
286

287
/**
288
 * @brief Repartition the cells amongst the nodes.
289
290
291
 *
 * @param e The #engine.
 */
292
293

void engine_repartition(struct engine *e) {
294
295
296

#if defined(WITH_MPI) && defined(HAVE_METIS)

297
298
  int i, j, k, l, cid, cjd, ii, jj, kk, res;
  idx_t *inds, *nodeIDs;
299
  idx_t *weights_v = NULL, *weights_e = NULL;
300
301
302
303
304
305
306
307
308
  struct space *s = e->s;
  int nr_cells = s->nr_cells, my_cells = 0;
  struct cell *cells = s->cells;
  int ind[3], *cdim = s->cdim;
  struct task *t, *tasks = e->sched.tasks;
  struct cell *ci, *cj;
  int nr_nodes = e->nr_nodes, nodeID = e->nodeID;
  float wscale = 1e-3, vscale = 1e-3, wscale_buff;
  idx_t wtot = 0;
309
310
  idx_t wmax = 1e9 / e->nr_nodes;
  idx_t wmin;
311
312
313
314

  /* Clear the repartition flag. */
  e->forcerepart = 0;

315
316
317
  /* Nothing to do if only using a single node. Also avoids METIS
   * bug that doesn't handle this case well. */
  if (nr_nodes == 1) return;
318

319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
  /* Allocate the inds and weights. */
  if ((inds = (idx_t *)malloc(sizeof(idx_t) * 26 *nr_cells)) == NULL ||
      (weights_v = (idx_t *)malloc(sizeof(idx_t) *nr_cells)) == NULL ||
      (weights_e = (idx_t *)malloc(sizeof(idx_t) * 26 *nr_cells)) == NULL ||
      (nodeIDs = (idx_t *)malloc(sizeof(idx_t) * nr_cells)) == NULL)
    error("Failed to allocate inds and weights arrays.");

  /* Fill the inds array. */
  for (cid = 0; cid < nr_cells; cid++) {
    ind[0] = cells[cid].loc[0] / s->cells[cid].h[0] + 0.5;
    ind[1] = cells[cid].loc[1] / s->cells[cid].h[1] + 0.5;
    ind[2] = cells[cid].loc[2] / s->cells[cid].h[2] + 0.5;
    l = 0;
    for (i = -1; i <= 1; i++) {
      ii = ind[0] + i;
      if (ii < 0)
        ii += cdim[0];
      else if (ii >= cdim[0])
        ii -= cdim[0];
      for (j = -1; j <= 1; j++) {
        jj = ind[1] + j;
        if (jj < 0)
          jj += cdim[1];
        else if (jj >= cdim[1])
          jj -= cdim[1];
        for (k = -1; k <= 1; k++) {
          kk = ind[2] + k;
          if (kk < 0)
            kk += cdim[2];
          else if (kk >= cdim[2])
            kk -= cdim[2];
          if (i || j || k) {
            inds[cid * 26 + l] = cell_getid(cdim, ii, jj, kk);
            l += 1;
          }
354
        }
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
      }
    }
  }

  /* Init the weights arrays. */
  bzero(weights_e, sizeof(idx_t) * 26 * nr_cells);
  bzero(weights_v, sizeof(idx_t) * nr_cells);

  /* Loop over the tasks... */
  for (j = 0; j < e->sched.nr_tasks; j++) {

    /* Get a pointer to the kth task. */
    t = &tasks[j];

    /* Skip un-interesting tasks. */
    if (t->type != task_type_self && t->type != task_type_pair &&
        t->type != task_type_sub && t->type != task_type_ghost &&
Matthieu Schaller's avatar
Matthieu Schaller committed
372
        t->type != task_type_drift && t->type != task_type_kick &&
373
        t->type != task_type_init)
374
375
376
377
      continue;

    /* Get the task weight. */
    idx_t w = (t->toc - t->tic) * wscale;
Matthieu Schaller's avatar
Matthieu Schaller committed
378
    if (w < 0) error("Bad task weight (%" SCIDX ").", w);
379
380
381
382
383
384
385
386
387

    /* Do we need to re-scale? */
    wtot += w;
    while (wtot > wmax) {
      wscale /= 2;
      wtot /= 2;
      w /= 2;
      for (k = 0; k < 26 * nr_cells; k++) weights_e[k] *= 0.5;
      for (k = 0; k < nr_cells; k++) weights_v[k] *= 0.5;
388
    }
389
390
391
392
393
394
395
396
397
398
399
400
401
402

    /* Get the top-level cells involved. */
    for (ci = t->ci; ci->parent != NULL; ci = ci->parent)
      ;
    if (t->cj != NULL)
      for (cj = t->cj; cj->parent != NULL; cj = cj->parent)
        ;
    else
      cj = NULL;

    /* Get the cell IDs. */
    cid = ci - cells;

    /* Different weights for different tasks. */
Matthieu Schaller's avatar
Matthieu Schaller committed
403
    if (t->type == task_type_ghost || t->type == task_type_drift ||
404
        t->type == task_type_kick) {
405
406
407
408
409
410
411
412
413
414
415
416
417

      /* Particle updates add only to vertex weight. */
      weights_v[cid] += w;

    }

    /* Self interaction? */
    else if ((t->type == task_type_self && ci->nodeID == nodeID) ||
             (t->type == task_type_sub && cj == NULL && ci->nodeID == nodeID)) {

      /* Self interactions add only to vertex weight. */
      weights_v[cid] += w;

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
452
453
454
455
456
457
458
459
460
461
462
463
464
465

    /* Pair? */
    else if (t->type == task_type_pair ||
             (t->type == task_type_sub && cj != NULL)) {

      /* In-cell pair? */
      if (ci == cj) {

        /* Add weight to vertex for ci. */
        weights_v[cid] += w;

      }

      /* Distinct cells with local ci? */
      else if (ci->nodeID == nodeID) {

        /* Index of the jth cell. */
        cjd = cj - cells;

        /* Add half of weight to each cell. */
        if (ci->nodeID == nodeID) weights_v[cid] += 0.5 * w;
        if (cj->nodeID == nodeID) weights_v[cjd] += 0.5 * w;

        /* Add Weight to edge. */
        for (k = 26 * cid; inds[k] != cjd; k++)
          ;
        weights_e[k] += w;
        for (k = 26 * cjd; inds[k] != cid; k++)
          ;
        weights_e[k] += w;
      }
    }
  }

  /* Get the minimum scaling and re-scale if necessary. */
  if ((res = MPI_Allreduce(&wscale, &wscale_buff, 1, MPI_FLOAT, MPI_MIN,
                           MPI_COMM_WORLD)) != MPI_SUCCESS) {
    char buff[MPI_MAX_ERROR_STRING];
    MPI_Error_string(res, buff, &i);
    error("Failed to allreduce the weight scales (%s).", buff);
  }
  if (wscale_buff != wscale) {
    float scale = wscale_buff / wscale;
    for (k = 0; k < 26 * nr_cells; k++) weights_e[k] *= scale;
    for (k = 0; k < nr_cells; k++) weights_v[k] *= scale;
  }

466
/* Merge the weights arrays across all nodes. */
467
468
469
470
#if IDXTYPEWIDTH == 32
  if ((res = MPI_Reduce((nodeID == 0) ? MPI_IN_PLACE : weights_v, weights_v,
                        nr_cells, MPI_INT, MPI_SUM, 0, MPI_COMM_WORLD)) !=
      MPI_SUCCESS) {
471
#else
472
473
474
  if ((res = MPI_Reduce((nodeID == 0) ? MPI_IN_PLACE : weights_v, weights_v,
                        nr_cells, MPI_LONG_LONG_INT, MPI_SUM, 0,
                        MPI_COMM_WORLD)) != MPI_SUCCESS) {
475
#endif
476
477
478
479
480
481
482
483
    char buff[MPI_MAX_ERROR_STRING];
    MPI_Error_string(res, buff, &i);
    error("Failed to allreduce vertex weights (%s).", buff);
  }
#if IDXTYPEWIDTH == 32
  if (MPI_Reduce((nodeID == 0) ? MPI_IN_PLACE : weights_e, weights_e,
                 26 * nr_cells, MPI_INT, MPI_SUM, 0,
                 MPI_COMM_WORLD) != MPI_SUCCESS)
484
#else
485
486
487
  if (MPI_Reduce((nodeID == 0) ? MPI_IN_PLACE : weights_e, weights_e,
                 26 * nr_cells, MPI_LONG_LONG_INT, MPI_SUM, 0,
                 MPI_COMM_WORLD) != MPI_SUCCESS)
488
#endif
489
490
491
492
493
    error("Failed to allreduce edge weights.");

  /* As of here, only one node needs to compute the partition. */
  if (nodeID == 0) {

494
495
496
497
498
499
500
501
502
503
504
    /* Final rescale of all weights to avoid a large range. Large ranges have
     * been seen to cause an incomplete graph. */
    wmin = wmax;
    wmax = 0.0;
    for (k = 0; k < 26 * nr_cells; k++) {
      wmax = weights_e[k] > wmax ? weights_e[k] : wmax;
      wmin = weights_e[k] < wmin ? weights_e[k] : wmin;
    }
    if ((wmax - wmin) > engine_maxmetisweight) {
      wscale = engine_maxmetisweight / (wmax - wmin);
      for (k = 0; k < 26 * nr_cells; k++) {
505
        weights_e[k] = (weights_e[k] - wmin) * wscale + 1;
506
507
      }
      for (k = 0; k < nr_cells; k++) {
508
        weights_v[k] = (weights_v[k] - wmin) * wscale + 1;
509
510
      }
    }
511

512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
    /* Check that the edge weights are fully symmetric. */
    /* for ( cid = 0 ; cid < nr_cells ; cid++ )
        for ( k = 0 ; k < 26 ; k++ ) {
            cjd = inds[ cid*26 + k ];
            for ( j = 26*cjd ; inds[j] != cid ; j++ );
            if ( weights_e[ cid*26+k ] != weights_e[ j ] )
                error( "Unsymmetric edge weights detected (%i vs %i)." ,
       weights_e[ cid*26+k ] , weights_e[ j ] );
            } */
    /* int w_min = weights_e[0], w_max = weights_e[0], w_tot = weights_e[0];
    for ( k = 1 ; k < 26*nr_cells ; k++ ) {
        w_tot += weights_e[k];
        if ( weights_e[k] < w_min )
            w_min = weights_e[k];
        else if ( weights_e[k] > w_max )
            w_max = weights_e[k];
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
569
    message( "edge weights in [ %i , %i ], tot=%i." , w_min , w_max , w_tot );
    w_min = weights_e[0], w_max = weights_e[0]; w_tot = weights_v[0];
    for ( k = 1 ; k < nr_cells ; k++ ) {
        w_tot += weights_v[k];
        if ( weights_v[k] < w_min )
            w_min = weights_v[k];
        else if ( weights_v[k] > w_max )
            w_max = weights_v[k];
        }
    message( "vertex weights in [ %i , %i ], tot=%i." , w_min , w_max , w_tot );
    */

    /* Make sure there are no zero weights. */
    for (k = 0; k < 26 * nr_cells; k++)
      if (weights_e[k] == 0) weights_e[k] = 1;
    for (k = 0; k < nr_cells; k++)
      if ((weights_v[k] *= vscale) == 0) weights_v[k] = 1;

    /* Allocate and fill the connection array. */
    idx_t *offsets;
    if ((offsets = (idx_t *)malloc(sizeof(idx_t) * (nr_cells + 1))) == NULL)
      error("Failed to allocate offsets buffer.");
    offsets[0] = 0;
    for (k = 0; k < nr_cells; k++) offsets[k + 1] = offsets[k] + 26;

    /* Set the METIS options. +1 to keep the GCC sanitizer happy. */
    idx_t options[METIS_NOPTIONS + 1];
    METIS_SetDefaultOptions(options);
    options[METIS_OPTION_OBJTYPE] = METIS_OBJTYPE_CUT;
    options[METIS_OPTION_NUMBERING] = 0;
    options[METIS_OPTION_CONTIG] = 1;
    options[METIS_OPTION_NCUTS] = 10;
    options[METIS_OPTION_NITER] = 20;
    // options[ METIS_OPTION_UFACTOR ] = 1;

    /* Set the initial partition, although this is probably ignored. */
    for (k = 0; k < nr_cells; k++) nodeIDs[k] = cells[k].nodeID;

    /* Call METIS. */
    idx_t one = 1, idx_nr_cells = nr_cells, idx_nr_nodes = nr_nodes;
    idx_t objval;
570

571
572
573
574
    /* Dump graph in METIS format */
    /*dumpMETISGraph("metis_graph", idx_nr_cells, one, offsets, inds,
                   weights_v, NULL, weights_e);*/

575
576
577
    if (METIS_PartGraphRecursive(&idx_nr_cells, &one, offsets, inds, weights_v,
                                 NULL, weights_e, &idx_nr_nodes, NULL, NULL,
                                 options, &objval, nodeIDs) != METIS_OK)
578
      error("Call to METIS_PartGraphRecursive failed.");
579
580
581
582
583
584

    /* Dump the 3d array of cell IDs. */
    /* printf( "engine_repartition: nodeIDs = reshape( [" );
    for ( i = 0 ; i < cdim[0]*cdim[1]*cdim[2] ; i++ )
        printf( "%i " , (int)nodeIDs[ i ] );
    printf("] ,%i,%i,%i);\n",cdim[0],cdim[1],cdim[2]); */
585
586
587
588

    /* Check that the nodeIDs are ok. */
    for (k = 0; k < nr_cells; k++)
      if (nodeIDs[k] < 0 || nodeIDs[k] >= nr_nodes)
589
        error("Got bad nodeID %" PRIDX " for cell %i.", nodeIDs[k], k);
590
591
592
593
594
595
596
597

    /* Check that the partition is complete and all nodes have some work. */
    int present[nr_nodes];
    int failed = 0;
    for (i = 0; i < nr_nodes; i++) present[i] = 0;
    for (i = 0; i < nr_cells; i++) present[nodeIDs[i]]++;
    for (i = 0; i < nr_nodes; i++) {
      if (!present[i]) {
598
599
        failed = 1;
        message("Node %d is not present after repartition", i);
600
601
602
603
604
605
606
      }
    }

    /* If partition failed continue with the current one, but make this
     * clear. */
    if (failed) {
      message(
607
608
          "WARNING: METIS repartition has failed, continuing with "
          "the current partition, load balance will not be optimal");
609
610
      for (k = 0; k < nr_cells; k++) nodeIDs[k] = cells[k].nodeID;
    }
611
612
613
614
615
616
  }

/* Broadcast the result of the partition. */
#if IDXTYPEWIDTH == 32
  if (MPI_Bcast(nodeIDs, nr_cells, MPI_INT, 0, MPI_COMM_WORLD) != MPI_SUCCESS)
    error("Failed to bcast the node IDs.");
617
#else
618
619
620
  if (MPI_Bcast(nodeIDs, nr_cells, MPI_LONG_LONG_INT, 0, MPI_COMM_WORLD) !=
      MPI_SUCCESS)
    error("Failed to bcast the node IDs.");
621
#endif
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637

  /* Set the cell nodeIDs and clear any non-local parts. */
  for (k = 0; k < nr_cells; k++) {
    cells[k].nodeID = nodeIDs[k];
    if (nodeIDs[k] == nodeID) my_cells += 1;
  }

  /* Clean up. */
  free(inds);
  free(weights_v);
  free(weights_e);
  free(nodeIDs);

  /* Now comes the tricky part: Exchange particles between all nodes.
     This is done in two steps, first allreducing a matrix of
     how many particles go from where to where, then re-allocating
638
     the parts array, and emitting the sends and receives.
639
640
641
642
643
644
645
646
647
648
649
     Finally, the space, tasks, and proxies need to be rebuilt. */

  /* Redistribute the particles between the nodes. */
  engine_redistribute(e);

  /* Make the proxies. */
  engine_makeproxies(e);

  /* Tell the engine it should re-build whenever possible */
  e->forcerebuild = 1;

650
#else
651
  error("SWIFT was not compiled with MPI and METIS support.");
652
#endif
653
}
654

655
656
657
658
659
660
661
/**
 * @brief Add up/down gravity tasks to a cell hierarchy.
 *
 * @param e The #engine.
 * @param c The #cell
 * @param up The upward gravity #task.
 * @param down The downward gravity #task.
Matthieu Schaller's avatar
Matthieu Schaller committed
662
 */
663
664
665

void engine_addtasks_grav(struct engine *e, struct cell *c, struct task *up,
                          struct task *down) {
Matthieu Schaller's avatar
Matthieu Schaller committed
666

667
668
669
670
671
672
673
674
675
676
  /* Link the tasks to this cell. */
  c->grav_up = up;
  c->grav_down = down;

  /* Recurse? */
  if (c->split)
    for (int k = 0; k < 8; k++)
      if (c->progeny[k] != NULL)
        engine_addtasks_grav(e, c->progeny[k], up, down);
}
677

678
679
680
681
/**
 * @brief Add send tasks to a hierarchy of cells.
 *
 * @param e The #engine.
682
683
 * @param ci The sending #cell.
 * @param cj The receiving #cell
684
685
 */

686
void engine_addtasks_send(struct engine *e, struct cell *ci, struct cell *cj) {
687

Matthieu Schaller's avatar
Matthieu Schaller committed
688
#ifdef WITH_MPI
689
690
691
  int k;
  struct link *l = NULL;
  struct scheduler *s = &e->sched;
692

693
694
695
696
697
  /* Check if any of the density tasks are for the target node. */
  for (l = ci->density; l != NULL; l = l->next)
    if (l->t->ci->nodeID == cj->nodeID ||
        (l->t->cj != NULL && l->t->cj->nodeID == cj->nodeID))
      break;
698

699
700
  /* If so, attach send tasks. */
  if (l != NULL) {
701

702
703
704
705
706
707
708
    /* Create the tasks. */
    struct task *t_xv =
        scheduler_addtask(&e->sched, task_type_send, task_subtype_none,
                          2 * ci->tag, 0, ci, cj, 0);
    struct task *t_rho =
        scheduler_addtask(&e->sched, task_type_send, task_subtype_none,
                          2 * ci->tag + 1, 0, ci, cj, 0);
709

710
711
    /* The send_rho task depends on the cell's ghost task. */
    scheduler_addunlock(s, ci->super->ghost, t_rho);
712

Matthieu Schaller's avatar
Matthieu Schaller committed
713
714
    /* The send_rho task should unlock the super-cell's kick task. */
    scheduler_addunlock(s, t_rho, ci->super->kick);
715

716
717
    /* The send_xv task should unlock the super-cell's ghost task. */
    scheduler_addunlock(s, t_xv, ci->super->ghost);
718

719
  }
720

721
722
723
724
  /* Recurse? */
  else if (ci->split)
    for (k = 0; k < 8; k++)
      if (ci->progeny[k] != NULL) engine_addtasks_send(e, ci->progeny[k], cj);
Matthieu Schaller's avatar
Matthieu Schaller committed
725
726
727
728

#else
  error("SWIFT was not compiled with MPI support.");
#endif
729
}
730
731
732
733
734
735
736
737
738
739

/**
 * @brief Add recv tasks to a hierarchy of cells.
 *
 * @param e The #engine.
 * @param c The #cell.
 * @param t_xv The recv_xv #task, if it has already been created.
 * @param t_rho The recv_rho #task, if it has already been created.
 */

740
741
void engine_addtasks_recv(struct engine *e, struct cell *c, struct task *t_xv,
                          struct task *t_rho) {
742

Matthieu Schaller's avatar
Matthieu Schaller committed
743
#ifdef WITH_MPI
744
745
  int k;
  struct scheduler *s = &e->sched;
746

747
748
  /* Do we need to construct a recv task? */
  if (t_xv == NULL && c->nr_density > 0) {
749

750
751
752
753
754
755
756
757
    /* Create the tasks. */
    t_xv = c->recv_xv =
        scheduler_addtask(&e->sched, task_type_recv, task_subtype_none,
                          2 * c->tag, 0, c, NULL, 0);
    t_rho = c->recv_rho =
        scheduler_addtask(&e->sched, task_type_recv, task_subtype_none,
                          2 * c->tag + 1, 0, c, NULL, 0);
  }
758

759
760
761
762
763
764
765
766
767
768
769
770
771
772
  /* Add dependencies. */
  for (struct link *l = c->density; l != NULL; l = l->next) {
    scheduler_addunlock(s, t_xv, l->t);
    scheduler_addunlock(s, l->t, t_rho);
  }
  for (struct link *l = c->force; l != NULL; l = l->next)
    scheduler_addunlock(s, t_rho, l->t);
  if (c->sorts != NULL) scheduler_addunlock(s, t_xv, c->sorts);

  /* Recurse? */
  if (c->split)
    for (k = 0; k < 8; k++)
      if (c->progeny[k] != NULL)
        engine_addtasks_recv(e, c->progeny[k], t_xv, t_rho);
Matthieu Schaller's avatar
Matthieu Schaller committed
773
774
775
776

#else
  error("SWIFT was not compiled with MPI support.");
#endif
777
}
778
779
780
781
782
783

/**
 * @brief Exchange cell structures with other nodes.
 *
 * @param e The #engine.
 */
784
785

void engine_exchange_cells(struct engine *e) {
786
787
788

#ifdef WITH_MPI

789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
  int j, k, pid, count = 0;
  struct pcell *pcells;
  struct space *s = e->s;
  struct cell *cells = s->cells;
  int nr_cells = s->nr_cells;
  int nr_proxies = e->nr_proxies;
  int offset[nr_cells];
  MPI_Request reqs_in[engine_maxproxies];
  MPI_Request reqs_out[engine_maxproxies];
  MPI_Status status;

  /* Run through the cells and get the size of the ones that will be sent off.
   */
  for (k = 0; k < nr_cells; k++) {
    offset[k] = count;
    if (cells[k].sendto)
      count += (cells[k].pcell_size = cell_getsize(&cells[k]));
  }
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
833
834
835
  /* Allocate the pcells. */
  if ((pcells = (struct pcell *)malloc(sizeof(struct pcell) * count)) == NULL)
    error("Failed to allocate pcell buffer.");

  /* Pack the cells. */
  cell_next_tag = 0;
  for (k = 0; k < nr_cells; k++)
    if (cells[k].sendto) {
      cell_pack(&cells[k], &pcells[offset[k]]);
      cells[k].pcell = &pcells[offset[k]];
    }

  /* Launch the proxies. */
  for (k = 0; k < nr_proxies; k++) {
    proxy_cells_exch1(&e->proxies[k]);
    reqs_in[k] = e->proxies[k].req_cells_count_in;
    reqs_out[k] = e->proxies[k].req_cells_count_out;
  }

  /* Wait for each count to come in and start the recv. */
  for (k = 0; k < nr_proxies; k++) {
    if (MPI_Waitany(nr_proxies, reqs_in, &pid, &status) != MPI_SUCCESS ||
        pid == MPI_UNDEFINED)
      error("MPI_Waitany failed.");
    // message( "request from proxy %i has arrived." , pid );
    proxy_cells_exch2(&e->proxies[pid]);
  }

836
  /* Wait for all the sends to have finished too. */
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
  if (MPI_Waitall(nr_proxies, reqs_out, MPI_STATUSES_IGNORE) != MPI_SUCCESS)
    error("MPI_Waitall on sends failed.");

  /* Set the requests for the cells. */
  for (k = 0; k < nr_proxies; k++) {
    reqs_in[k] = e->proxies[k].req_cells_in;
    reqs_out[k] = e->proxies[k].req_cells_out;
  }

  /* Wait for each pcell array to come in from the proxies. */
  for (k = 0; k < nr_proxies; k++) {
    if (MPI_Waitany(nr_proxies, reqs_in, &pid, &status) != MPI_SUCCESS ||
        pid == MPI_UNDEFINED)
      error("MPI_Waitany failed.");
    // message( "cell data from proxy %i has arrived." , pid );
    for (count = 0, j = 0; j < e->proxies[pid].nr_cells_in; j++)
      count += cell_unpack(&e->proxies[pid].pcells_in[count],
                           e->proxies[pid].cells_in[j], e->s);
  }

857
  /* Wait for all the sends to have finished too. */
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
  if (MPI_Waitall(nr_proxies, reqs_out, MPI_STATUSES_IGNORE) != MPI_SUCCESS)
    error("MPI_Waitall on sends failed.");

  /* Count the number of particles we need to import and re-allocate
     the buffer if needed. */
  for (count = 0, k = 0; k < nr_proxies; k++)
    for (j = 0; j < e->proxies[k].nr_cells_in; j++)
      count += e->proxies[k].cells_in[j]->count;
  if (count > s->size_parts_foreign) {
    if (s->parts_foreign != NULL) free(s->parts_foreign);
    s->size_parts_foreign = 1.1 * count;
    if (posix_memalign((void **)&s->parts_foreign, part_align,
                       sizeof(struct part) * s->size_parts_foreign) != 0)
      error("Failed to allocate foreign part data.");
  }
873

874
  /* Unpack the cells and link to the particle data. */
875
  struct part *parts = s->parts_foreign;
876
877
878
879
  for (k = 0; k < nr_proxies; k++) {
    for (count = 0, j = 0; j < e->proxies[k].nr_cells_in; j++) {
      count += cell_link(e->proxies[k].cells_in[j], parts);
      parts = &parts[e->proxies[k].cells_in[j]->count];
880
    }
881
882
883
884
885
886
887
888
889
  }
  s->nr_parts_foreign = parts - s->parts_foreign;

  /* Is the parts buffer large enough? */
  if (s->nr_parts_foreign > s->size_parts_foreign)
    error("Foreign parts buffer too small.");

  /* Free the pcell buffer. */
  free(pcells);
890

891
892
893
894
#else
  error("SWIFT was not compiled with MPI support.");
#endif
}
895
896
897
898
899

/**
 * @brief Exchange straying parts with other nodes.
 *
 * @param e The #engine.
900
901
 * @param offset The index in the parts array as of which the foreign parts
 *reside.
902
903
904
905
906
 * @param ind The ID of the foreign #cell.
 * @param N The number of stray parts.
 *
 * @return The number of arrived parts copied to parts and xparts.
 */
907
908

int engine_exchange_strays(struct engine *e, int offset, int *ind, int N) {
909
910
911

#ifdef WITH_MPI

912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
  int k, pid, count = 0, nr_in = 0, nr_out = 0;
  MPI_Request reqs_in[2 * engine_maxproxies];
  MPI_Request reqs_out[2 * engine_maxproxies];
  MPI_Status status;
  struct proxy *p;
  struct space *s = e->s;

  /* Re-set the proxies. */
  for (k = 0; k < e->nr_proxies; k++) e->proxies[k].nr_parts_out = 0;

  /* Put the parts into the corresponding proxies. */
  for (k = 0; k < N; k++) {
    int node_id = e->s->cells[ind[k]].nodeID;
    if (node_id < 0 || node_id >= e->nr_nodes)
      error("Bad node ID %i.", node_id);
    pid = e->proxy_ind[node_id];
    if (pid < 0)
      error(
          "Do not have a proxy for the requested nodeID %i for part with "
          "id=%llu, x=[%e,%e,%e].",
          node_id, s->parts[offset + k].id, s->parts[offset + k].x[0],
          s->parts[offset + k].x[1], s->parts[offset + k].x[2]);
    proxy_parts_load(&e->proxies[pid], &s->parts[offset + k],
                     &s->xparts[offset + k], 1);
  }

  /* Launch the proxies. */
  for (k = 0; k < e->nr_proxies; k++) {
    proxy_parts_exch1(&e->proxies[k]);
    reqs_in[k] = e->proxies[k].req_parts_count_in;
    reqs_out[k] = e->proxies[k].req_parts_count_out;
  }

  /* Wait for each count to come in and start the recv. */
  for (k = 0; k < e->nr_proxies; k++) {
    if (MPI_Waitany(e->nr_proxies, reqs_in, &pid, &status) != MPI_SUCCESS ||
        pid == MPI_UNDEFINED)
      error("MPI_Waitany failed.");
    // message( "request from proxy %i has arrived." , pid );
    proxy_parts_exch2(&e->proxies[pid]);
  }
953

954
  /* Wait for all the sends to have finished too. */
955
956
957
  if (MPI_Waitall(e->nr_proxies, reqs_out, MPI_STATUSES_IGNORE) != MPI_SUCCESS)
    error("MPI_Waitall on sends failed.");

958
  /* Count the total number of incoming particles and make sure we have
959
960
961
962
963
964
     enough space to accommodate them. */
  int count_in = 0;
  for (k = 0; k < e->nr_proxies; k++) count_in += e->proxies[k].nr_parts_in;
  message("sent out %i particles, got %i back.", N, count_in);
  if (offset + count_in > s->size_parts) {
    s->size_parts = (offset + count_in) * 1.05;
965
966
    struct part *parts_new = NULL;
    struct xpart *xparts_new = NULL;
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
    if (posix_memalign((void **)&parts_new, part_align,
                       sizeof(struct part) * s->size_parts) != 0 ||
        posix_memalign((void **)&xparts_new, part_align,
                       sizeof(struct xpart) * s->size_parts) != 0)
      error("Failed to allocate new part data.");
    memcpy(parts_new, s->parts, sizeof(struct part) * offset);
    memcpy(xparts_new, s->xparts, sizeof(struct xpart) * offset);
    free(s->parts);
    free(s->xparts);
    s->parts = parts_new;
    s->xparts = xparts_new;
  }

  /* Collect the requests for the particle data from the proxies. */
  for (k = 0; k < e->nr_proxies; k++) {
    if (e->proxies[k].nr_parts_in > 0) {
      reqs_in[2 * k] = e->proxies[k].req_parts_in;
      reqs_in[2 * k + 1] = e->proxies[k].req_xparts_in;
      nr_in += 1;
    } else
      reqs_in[2 * k] = reqs_in[2 * k + 1] = MPI_REQUEST_NULL;
    if (e->proxies[k].nr_parts_out > 0) {
      reqs_out[2 * k] = e->proxies[k].req_parts_out;
      reqs_out[2 * k + 1] = e->proxies[k].req_xparts_out;
      nr_out += 1;
    } else
      reqs_out[2 * k] = reqs_out[2 * k + 1] = MPI_REQUEST_NULL;
  }

  /* Wait for each part array to come in and collect the new
     parts from the proxies. */
  for (k = 0; k < 2 * (nr_in + nr_out); k++) {
    int err;
    if ((err = MPI_Waitany(2 * e->nr_proxies, reqs_in, &pid, &status)) !=