File: tm_kpartitioning.c

package info (click to toggle)
openmpi 5.0.8-4
  • links: PTS, VCS
  • area: main
  • in suites:
  • size: 201,684 kB
  • sloc: ansic: 613,078; makefile: 42,353; sh: 11,194; javascript: 9,244; f90: 7,052; java: 6,404; perl: 5,179; python: 1,859; lex: 740; fortran: 61; cpp: 20; tcl: 12
file content (882 lines) | stat: -rw-r--r-- 25,473 bytes parent folder | download | duplicates (4)
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
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
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
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
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
406
407
408
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
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
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
550
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
643
644
645
646
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
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
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
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
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
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
#include "tm_mapping.h"
#include "tm_mt.h"
#include "tm_kpartitioning.h"
#include "k-partitioning.h"
#include <stdlib.h>
#include <stdio.h>
#include "config.h"
#include "ompi_config.h"

#if defined(HAVE_LIBSCOTCH)
#include <scotch.h>
#endif  /* defined(HAVE_LIBSCOTCH) */


#define USE_KL_KPART 0
#define KL_KPART_GREEDY_TRIALS 0

static int verbose_level = ERROR;

#define MAX_TRIALS 10
#define USE_KL_STRATEGY 1


#define TM_MIN(a,b) ((a)<(b)?(a):(b))


OMPI_HIDDEN int  tm_fill_tab(int **,int *,int,int,int,int);
OMPI_HIDDEN void tm_complete_obj_weight(double **,int,int);

static void allocate_vertex(int,int *,com_mat_t *,int,int *,int);
static double eval_cost(int *, com_mat_t *);
static int *kpartition_greedy(int, com_mat_t *,int,int *,int);
static constraint_t *split_constraints (int *,int,int,tm_topology_t *,int, int);
static com_mat_t **split_com_mat(com_mat_t *,int,int,int *);
static int **split_vertices(int *,int,int,int *);
static void free_tab_com_mat(com_mat_t **,int);
static void free_tab_local_vertices(int **,int);
static void free_const_tab(constraint_t *,int);
static void kpartition_build_level_topology(tm_tree_t *,com_mat_t *,int,int,tm_topology_t *,
				     int *,int *,int,double *,double *);

static int greedy_flag = 0;

void tm_set_greedy_flag(int new_val){
  greedy_flag = new_val;
}

int tm_get_greedy_flag(){
  return greedy_flag;
}


#if defined(HAVE_LIBSCOTCH)

SCOTCH_Graph* com_mat_to_scotch_graph(com_mat_t *com_mat, int n){
  double **mat = com_mat->comm;
  SCOTCH_Num vertnbr = n;	// number of vertices
  SCOTCH_Num edgenbr = vertnbr*vertnbr;			// number of edges
  /* adjacency list */
  SCOTCH_Num *verttab = (SCOTCH_Num *)malloc(sizeof(SCOTCH_Num) * (vertnbr+1));
  /* loads of vertices */
  /* SCOTCH_Num *velotab = (SCOTCH_Num *)malloc(sizeof(SCOTCH_Num) * vertnbr); */
  /* id of the neighbors */
  SCOTCH_Num *edgetab = (SCOTCH_Num *)malloc(sizeof(SCOTCH_Num) * edgenbr);
  /* number of bytes exchanged */
  SCOTCH_Num *edlotab = (SCOTCH_Num *)malloc(sizeof(SCOTCH_Num) * edgenbr);
  SCOTCH_Graph *graphptr =  SCOTCH_graphAlloc();

  int edgeNum = 0;
  int i,j;

  /* Building with the communication matrix */
  for(i = 0; i < com_mat->n ; i++) {
    verttab[i] = edgeNum;
    for(j = 0; j < i; j++) {
      if(mat[i][j]){
	edgetab[edgeNum] = j;
	edlotab[edgeNum] = (SCOTCH_Num)mat[i][j];
	edgeNum++;
      }
    }
    /* ensure i!=j. Hence, avoid to test it...*/
    for(j = i+1 ; j < com_mat->n ; j++) {
      if(mat[i][j]){
	edgetab[edgeNum] = j;
	edlotab[edgeNum] = (SCOTCH_Num)mat[i][j];
	edgeNum++;
      }
    }
  }
  

  /* for(i = baseval; i < com_mat->n ; i++) { */
  /*   verttab[i] = edgeNum; */
  /*   /\* velotab[i] = (SCOTCH_Num) ceil(ogr->vertices[i].getVertexLoad() * ratio); *\/ */
  /*   for(j = baseval; j < com_mat->n ; j++) { */
  /*     if((mat[i][j] || mat[j][i]) && (i!=j)){ */
  /* 	edgetab[edgeNum] = j; */
  /* 	edlotab[edgeNum] = (SCOTCH_Num) ((mat[i][j] + mat[j][i])/2); */
  /* 	edgeNum++; */
  /*     } */
  /*   } */
  /* } */

  /* adding the dumb vertices: they have no neighbor*/
  for(i = com_mat->n ; i<vertnbr ; i++) {
    verttab[i] = edgeNum;
  }

  verttab[i] = edgeNum;

  if(tm_get_verbose_level() >=DEBUG){
    printf("Graph converted to Scotch format: edgeNum=%d, edgenbr = %lld, vertnbr = %lld\n",edgeNum, (long long int)edgenbr, (long long int)vertnbr);
  }

  assert(edgeNum <= edgenbr);
  edgenbr = edgeNum;

  SCOTCH_graphInit(graphptr);
  SCOTCH_graphBuild(graphptr, 0, vertnbr, verttab, verttab+1, NULL, NULL, edgenbr, edgetab, edlotab); 

  return graphptr;
}



int  check_partition(SCOTCH_Num *parttab, int k, int n){
  int *count = CALLOC(sizeof(int), k);
  int i;
  for(i=0; i<n; i++){
    count[parttab[i]]++;
  }

  int target= n/k;

  for(i = 0; i<k ; i++){
    if(count[i] != target){
      if(tm_get_verbose_level()>=INFO)
	fprintf(stdout, "Error in partition: %d vertices in partition %d while expecting %d vertices\n",count[i], i, target);
      FREE(count);
      return 0;
    }
  }

   FREE(count);
  return 1;
}


/* n is the number of element in teh graoh with dumlb_vertices
   comm_mat->n is the nulber of processes (i.e. the size of teh graph without dumb veritcies*/
int  *kpartition_scotch(int k, com_mat_t *com_mat, int n, int *constraints, int nb_constraints){
  SCOTCH_Num    partnbr = (SCOTCH_Num) k;
  SCOTCH_Graph* graphptr;
  SCOTCH_Strat  strat;
  SCOTCH_Num    straval;
  SCOTCH_Num   *parttab = (SCOTCH_Num *)MALLOC(sizeof(SCOTCH_Num) * n);
  int          *partition = (int *)MALLOC(sizeof(int) * n);
  int          i, j;
  int          *nb_dumb = (int *)MALLOC(sizeof(int) * k); /*number of dumb vertices per partition */ 
  int          dumb_id, min_nb_dumb = n, sum_dumb = 0, p;
  /* if(SCOTCH_graphCheck(graphptr) == 1){ */
  /*   fprintf(stderr,"Bad scotch graph! Exiting program...\n"); */
  /*   exit(-1); */
  /* } */

  /* printf("Correct scotch graph (%d, %d)!\n", SCOTCH_numSizeof(), sizeof(SCOTCH_Num)); */

  for(i=0;i<n;i++)
    parttab[i] = -1;


  /* put "dumb" vertices in the correct partition if there are any*/
  /*constraints are leaves that can be used */
  if (nb_constraints){
    int end, start = 0;
    for( i = 0 ; i < k ; i ++){
      int max_val = (i+1)* (n/k);
      end = start;
      while( end < nb_constraints){
	if(constraints[end] >= max_val)
	  break;
	end++;
      }
      /* now end - start is the number of constraints for the ith subtree
	 hence the number of dumb vertices in partition i is the differences between the
	 number of leaves of the subtree (n/k) and the number of constraints
      */
      nb_dumb[i] = n/k - (end-start);
      sum_dumb += nb_dumb[i];
      if(nb_dumb[i] < min_nb_dumb){
	min_nb_dumb = nb_dumb[i];
     }
       start=end;
    }

    /* Imagine we have n=12, k=3, nb_dumb[0] = 3, nb_dumb[1] = 2, nb_dumb[2] = 3, hence min_nb_dumb = 2 and sum_dumb = 8 
       So, we have 8 fix vertices and 12-8 = 4 free vertices
       We want scotch to allocate the 6 free vertices such that the whole partition is balanced (4 vertex in each) : 
       1 in parttion 0,   2 in partition 1 and 1 in partition 2.
       To do so we can fill partab as follows: 
       {-1, -1, -1, -1, 0, 0, 0, 1, 1, 2, 2, 2} and call scotch with a n=12 vertices graph with SCOTCH_STRATBALANCE
       dumb_id = n - sum_dumb;
       for(i = 0;i<k;i++){
         for( j = 0; j < nb_dumb[i]; j ++ ){
           parttab[dumb_id] = i;
           dumb_id++;
         } 
       }
       
       A more efficient solution is to fill partab as follows
       {-1, -1, -1, -1, 0, 2, 0, 0, 1, 1, 2, 2} and call Scotch with 
       a p = 6 (n-sum_dumb+ sum_{i}(nb_dumb[i]-min_dumb) vertices graph. 
       Scotch will then only use the 8 fist element of partab
    */

    dumb_id = n - sum_dumb; /* now dumb_id is the number of free vertices*/
    for(i = 0 ; i < k ; i++){
      for( j = 0; j < nb_dumb[i] - min_nb_dumb; j ++ ){
	parttab[dumb_id] = i;
	dumb_id++;
      } 
    }
    p = dumb_id; 
    for(i = 0 ; i < k ; i++){
      for( j = 0 ; j < min_nb_dumb ; j ++ ){
	parttab[dumb_id] = i;
	dumb_id++;
      } 
    }
  }else{
    p=n; /* if no constraint use n vertices */
  }       


  graphptr = com_mat_to_scotch_graph(com_mat, p);
  
  SCOTCH_stratInit (&strat);
  straval = SCOTCH_STRATBALANCE;
  if(k>4)
    straval = SCOTCH_STRATSPEED;
  SCOTCH_stratGraphMapBuild (&strat, straval, partnbr, 0);


  if(tm_get_verbose_level()>=DEBUG){
    printf("Before Scotch (p=%d, n=%d): \n", p, n);
    for(i = 0 ; i < n; i++){
      printf("%d ",(int)parttab[i]);
    }
    printf("\n");
  }

  if(SCOTCH_graphPartFixed(graphptr, partnbr, &strat, parttab) == 0){
    if(tm_get_verbose_level()>=DEBUG){
      printf("After Scotch: \n");
      for(i = 0 ; i < n; i++){
	printf("%d ",(int)parttab[i]);
      }
      printf("\n");
    }
  }else{
    if(tm_get_verbose_level()>=CRITICAL){
      fprintf(stderr,"Scotch Partitionning failed\n");  
    }
    exit(-1);
  }

  if(!check_partition(parttab, partnbr, n)){
    if(tm_get_verbose_level()>=INFO){
      printf("falling from Scotch to greedy partionning\n");
    }
    FREE(partition);
    partition = kpartition_greedy(k, com_mat, n, constraints, nb_constraints);
  }else{
    for(i=0;i<n;i++)
      partition[i] = parttab [i];
  }

  SCOTCH_stratExit (&strat);
  SCOTCH_graphExit(graphptr);
  SCOTCH_memFree(graphptr);
  FREE(parttab);
  FREE(nb_dumb);
  
  return partition;
}

#endif /* defined(HAVE_LIBSCOTCH) */


static void allocate_vertex(int u, int *res, com_mat_t *com_mat, int n, int *size, int max_size)
{
  int i,best_part=0;
  double cost, best_cost = -1;

  /*printf("\n");
    print_1D_tab(res,n);*/
  if(u>=com_mat->n){
    for( i = 0 ; i < n ; i++)
      if (( res[i] != -1 ) && ( size[res[i]] < max_size )){
	best_part = res[i];
	break;
      }

  }else{
    for( i = 0 ; i < n ; i++){
      if (( res[i] != -1 ) && ( size[res[i]] < max_size )){
	cost = (((i)<com_mat->n)) ?com_mat->comm[u][i]:0;
	/* if((n<=16) && (u==8)){ */
	/*   printf("u=%d, i=%d: %f\n",u, i, cost); */
	/* } */
	if (( cost > best_cost)){
	  best_cost = cost;
	  best_part = res[i];
	}
      }
    }
  }
  /* if(n<=16){ */
  /*   printf("size[%d]: %d\n",best_part, size[best_part]); */
  /*   printf("putting(%.2f): %d -> %d\n",best_cost, u, best_part); */
  /* } */

  res[u] = best_part;
  size[best_part]++;
}

static double eval_cost(int *partition, com_mat_t *com_mat)
{
  double cost = 0;
  int i,j;

  for( i = 0 ; i < com_mat->n ; i++ )
    for( j = i+1 ; j < com_mat->n ; j++ )
      if(partition[i] != partition[j])
	cost += com_mat->comm[i][j];

  return cost;
}

static int  *kpartition_greedy(int k, com_mat_t *com_mat, int n, int *constraints, int nb_constraints)
{
  int *partition = NULL, *best_partition=NULL, *size = NULL;
  int i,j,nb_trials;
  int max_size, max_val;
  double cost, best_cost = -1;
  int start, end;
  int dumb_id, nb_dumb;
  int vl = tm_get_verbose_level();


  if(nb_constraints > n){
    if(vl >= ERROR){
      fprintf(stderr,"Error more constraints (%d) than the problem size (%d)!\n",nb_constraints, n);
    }
    return NULL;
  }

  max_size = n/k;

  if(vl >= DEBUG){
    printf("max_size = %d (n=%d,k=%d)\ncom_mat->n-1=%d\n",max_size,n,k,com_mat->n-1);
    printf("nb_constraints = %d\n",nb_constraints);

    if(n<=16){
      printf("Constraints: ");
      tm_print_1D_tab(constraints,nb_constraints);
    }
  }
  /* if(com_mat->n){ */
  /*   printf ("val [n-1][0]= %f\n",com_mat->comm[com_mat->n-1][0]); */
  /* } */


  for( nb_trials = 0 ; nb_trials < MAX_TRIALS ; nb_trials++ ){
    partition = (int *)MALLOC(sizeof(int)*n);
    for ( i = 0 ; i < n ; i ++ )
      partition[i] = -1;

    size = (int *)CALLOC(k,sizeof(int));



    /* put "dumb" vertices in the correct partition if there are any*/
    /*constraints are leaves that can be used */
    if (nb_constraints){
      start = 0;
      dumb_id = n-1;
      for( i = 0 ; i < k ; i ++){
	max_val = (i+1)* (n/k);
	end = start;
	while( end < nb_constraints){
	  if(constraints[end] >= max_val)
	    break;
	  end++;
	}
	/* now end - start is the number of constraints for the ith subtree
	   hence the number of dumb vertices is the differences between the
	   number of leaves of the subtree (n/k) and the number of constraints
	*/
	nb_dumb = n/k - (end-start);
	/* if(n<=16){ */
	/*   printf("max_val: %d, nb_dumb=%d, start=%d, end=%d, size=%d\n",max_val, nb_dumb, start, end, n/k); */
	/* } */
	/* dumb vertices are the one with highest indices:
	   put them in the ith partitions*/
	for( j = 0; j < nb_dumb; j ++ ){
	  partition[dumb_id] = i;
	  dumb_id--;
	}
	/* increase the size of the ith partition accordingly*/
	size[i] += nb_dumb;
	start=end;
      }
    }
    /* if(n<=16){ */
    /*   printf("After dumb vertices mapping: ");print_1D_tab(partition,n); */
    /* } */


    /* choose k initial "true" vertices at random and put them in a different partition */
    for ( i = 0 ; i < k ; i ++ ){
      /* if the partition is full of dumb vertices go to next partition*/
      if(size[i] >= max_size)
	continue;
      /* find a vertex not allready partitionned*/
      do{
	/* call the mersenne twister PRNG of tm_mt.c*/
	j =  tm_genrand_int32() % n;
      } while ( partition[j] != -1 );
      /* allocate and update size of partition*/
      partition[j] = i;
      /* if(n<=16){ */
      /* 	printf("random: %d -> %d\n",j,i); */
      /* } */
      size[i]++;
    }

    /* allocate each unaloacted vertices in the partition that maximize the communication*/
    for( i = 0 ;  i < n ; i ++)
      if( partition[i] == -1)
	allocate_vertex(i, partition, com_mat, n, size, max_size);

    cost = eval_cost(partition,com_mat);
    /* if(n<=16){ */
    /*   print_1D_tab(partition,n); */
    /*   printf("cost=%.2f\n",cost); */
    /* } */
    if((cost<best_cost) || (best_cost == -1)){
      best_cost=cost;
      FREE(best_partition);
      best_partition=partition;
    }else
      FREE(partition);

    FREE(size);
  }

  /*print_1D_tab(best_partition,n);
  printf("best_cost=%.2f\n",best_cost);
  */
  return best_partition;
}

int *tm_kpartition(int k, com_mat_t *com_mat, int n, int *constraints, int nb_constraints)
{
  int *res= NULL;

  if( n%k != 0){
    if(verbose_level >= ERROR)
      fprintf(stderr,"Error: Cannot partition %d elements in %d parts\n",n,k);
    return NULL;
  }

  /* if(USE_KL_KPART) */
  /*   res = kPartitioning(comm, n, k, constraints, nb_constraints, KL_KPART_GREEDY_TRIALS); */
  /* else */


#if defined(HAVE_LIBSCOTCH)
  if(!greedy_flag){
    if(verbose_level >= DEBUG)
      printf("Using Scotch\n");
    res = kpartition_scotch(k, com_mat, n, constraints, nb_constraints);
  }else{
    if(verbose_level >= DEBUG)
      printf("Using greedy partitionning\n");
    res = kpartition_greedy(k, com_mat, n, constraints, nb_constraints);
  }
#else  /* defined(HAVE_LIBSCOTCH) */
  if(verbose_level >= DEBUG)
    printf("Using greedy partitionning\n");
  res = kpartition_greedy(k, com_mat, n, constraints, nb_constraints);
#endif  /* defined(HAVE_LIBSCOTCH) */
  return res;
}

static constraint_t *split_constraints (int *constraints, int nb_constraints, int k, tm_topology_t *topology, int depth, int N)
{
  constraint_t *const_tab = NULL;
  int nb_leaves, start, end;
  int i;
  int vl = tm_get_verbose_level();

  const_tab = (constraint_t *)CALLOC(k,sizeof(constraint_t));

  /* nb_leaves is the number of leaves of the current subtree
     this will help to determine where to split constraints and how to shift values
  */
  nb_leaves = tm_compute_nb_leaves_from_level( depth + 1, topology );

/* split the constraints into k sub-constraints
     each sub-contraints 'i' contains constraints of value in [i*nb_leaves,(i+1)*nb_leaves[
   */
  start = 0;

  for( i = 0; i < k; i++ ){
    /*returns the indice in constraints that contains the smallest value not copied
      end is used to compute the number of copied elements (end-size) and is used as the next staring indices*/
    end = tm_fill_tab(&(const_tab[i].constraints), constraints, nb_constraints,start, (i+1) * nb_leaves, i * nb_leaves);
    const_tab[i].length = end-start;
    if(vl>=DEBUG){
      printf("Step %d\n",i);
      printf("\tConstraint: "); tm_print_1D_tab(constraints, nb_constraints);
      printf("\tSub constraint: "); tm_print_1D_tab(const_tab[i].constraints, end-start);
    }

    if(end-start > N/k){
      if(vl >= ERROR){
	fprintf(stderr, "Error in spliting constraint at step %d. N=%d k= %d, length = %d\n", i, N, k, end-start);
      }
      FREE(const_tab);
      return NULL;
    }
    const_tab[i].id = i;
    start = end;
  }

  return const_tab;
}


/* split the com_mat of order n in k partiton according to parmutition table*/
static com_mat_t **split_com_mat(com_mat_t *com_mat, int n, int k, int *partition)
{
  com_mat_t **res = NULL, *sub_com_mat;
  double **sub_mat = NULL;
  int *perm = NULL;
  int cur_part, i, ii, j, jj, m = n/k, s;

  res = (com_mat_t**)MALLOC(k*sizeof(com_mat_t *));


  if(verbose_level >= DEBUG){
    printf("Partition: "); tm_print_1D_tab(partition,n);
    tm_display_tab(com_mat->comm,com_mat->n);
    printf("m=%d,n=%d,k=%d\n",m,n,k);
    printf("perm=%p\n", (void *)perm);
  }

  perm  = (int*)MALLOC(sizeof(int)*m);
  for( cur_part = 0 ; cur_part < k ; cur_part ++ ){

    /* build perm such that submat[i][j] correspond to com_mat[perm[i]][perm[j]] according to the partition*/
    s = 0;
    /* The partition is of size n. n can be larger than the communication matrix order
       as only the input problem are in the communication matrix while n is of the size
       of all the element (including the added one where it is possible to map computation) :
       we can have more compute units than processes*/
    for( j = 0; j < com_mat->n; j ++)
      if ( partition[j] == cur_part )
	perm[s++] = j;

    if(s>m){
      if(verbose_level >= CRITICAL){
	fprintf(stderr,"Partition: "); tm_print_1D_tab(partition,n);
	tm_display_tab(com_mat->comm,com_mat->n);
	fprintf(stderr,"too many elements of the partition for the permuation (s=%d>%d=m). n=%d, k=%d, cur_part= %d\n",s,m,n,k, cur_part);
      }
      exit(-1);
    }
    /* s is now the size of the non zero sub matrix for this partition*/
    /* built a sub-matrix for partition cur_part*/
    sub_mat = (double **) MALLOC(sizeof(double *) * s);
    for( i = 0 ; i < s ; i++)
      sub_mat[i] = (double *) MALLOC(sizeof(double ) * s);

    /* build the sub_mat corresponding to the partiion cur_part*/
    for ( i = 0 ; i < s ; i ++){
      ii = perm[i];
      for( j = i ; j < s ; j ++){
	jj = perm[j];
	sub_mat[i][j] = com_mat->comm[ii][jj];
	sub_mat[j][i] = sub_mat[i][j];
      }
    }

    sub_com_mat = (com_mat_t *)MALLOC(sizeof(com_mat_t));
    sub_com_mat -> n = s;
    sub_com_mat -> comm = sub_mat;


    /*  printf("\n\npartition:%d\n",cur_part);display_tab(sub_mat,m);*/

    /* assign the sub_mat to the result*/
    res[cur_part] = sub_com_mat;
  }

 FREE(perm);

  return res;
}

static int **split_vertices( int *vertices, int n, int k, int *partition)
{
  int **res = NULL, *sub_vertices = NULL;
  int m = n/k;
  int i, j, cur_part;

  /*allocate resuts*/
  res = (int**) MALLOC(sizeof(int*) * k);


  if(verbose_level >= DEBUG){
    printf("Partition: ");tm_print_1D_tab(partition,n);
    printf("Vertices id: ");tm_print_1D_tab(vertices,n);
  }

  /*split the vertices tab of the partition cur_part  to the sub_vertices tab*/
  for( cur_part = 0; cur_part < k ; cur_part ++){
    sub_vertices = (int*) MALLOC(sizeof(int) * m);
    i = 0;
    for( j = 0; j < n; j ++)
      if ( partition[j] == cur_part )
	sub_vertices[i++] = vertices[j];
    res[cur_part] = sub_vertices;
    if(verbose_level >= DEBUG){
      printf("partition %d: ",cur_part);tm_print_1D_tab(sub_vertices,m);
    }
  }
  /*exit(-1);*/
  return res;
}

static void free_tab_com_mat(com_mat_t **mat,int k)
{
  int i,j;
  if( !mat )
    return;

  for ( i = 0 ; i < k ; i ++){
    for ( j = 0 ; j < mat[i]->n ; j ++)
      FREE( mat[i]->comm[j] );
    FREE( mat[i]->comm );
    FREE(mat[i]);

  }
  FREE(mat);
}

static void free_tab_local_vertices(int **mat, int k)
{
  int i; /* m=n/k; */
  if( !mat )
    return;

  for ( i = 0 ; i < k ; i ++){
    FREE( mat[i] );
  }
  FREE(mat);
}


static void free_const_tab(constraint_t *const_tab, int k)
{
  int i;

  if( !const_tab )
    return;

  for(i = 0; i < k; i++){
    if(const_tab[i].length)
      FREE(const_tab[i].constraints);
  }

  FREE(const_tab);
}

#if 0
static void check_com_mat(com_mat_t *com_mat){
  int i,j;

  for( i = 0 ; i < com_mat->n ; i++ )
    for( j = 0 ; j < com_mat->n ; j++ )
      if(com_mat->comm[i][j]<0){
	printf("com_mat->comm[%d][%d]= %f\n",i,j,com_mat->comm[i][j]);
	exit(-1);
      }
}
#endif

static void print_tab(int n){
  for(;n;n--)
    fprintf(stdout,"\t");
}

static void display_partition(int *partition, int *local_vertices, int n, int depth, int k){
  int cur_part, j;
  print_tab(depth);fprintf(stdout,"Partitions at depth=%d\n",depth);
  for( cur_part = 0; cur_part < k ; cur_part ++){
    print_tab(depth); fprintf(stdout,"%d :",cur_part);
    for( j = 0; j < n; j ++){
      if ( partition[j] == cur_part ){
	if(local_vertices[j]!=-1)
	  fprintf(stdout,"%d ",local_vertices[j]);
      }
    }
    fprintf(stdout,"\n");
  }	
}

static void kpartition_build_level_topology(tm_tree_t *cur_node, com_mat_t *com_mat, int N, int depth,
				     tm_topology_t *topology, int *local_vertices,
				     int *constraints, int nb_constraints,
				     double *obj_weight, double *comm_speed)
{
  com_mat_t **tab_com_mat = NULL; /* table of comunication matrix. We will have k of such comunication matrix, one for each subtree */
  int k = topology->arity[depth];
  tm_tree_t **tab_child = NULL;
  int *partition = NULL;
  int **tab_local_vertices = NULL;
  constraint_t *const_tab = NULL;
  int i;
  verbose_level = tm_get_verbose_level();

  /* if we are at the bottom of the tree set cur_node
   and return*/
  if ( depth == topology->nb_levels - 1 ){
    if(verbose_level>=DEBUG)
      printf("id : %d, com_mat= %p\n",local_vertices[0], (void *)com_mat->comm);
    tm_set_node(cur_node,NULL, 0, NULL, local_vertices[0], 0, NULL, depth);
    return;
  }


  if(verbose_level >= DEBUG){
    printf("Partitionning Matrix of size %d (problem size= %d) in %d partitions\n", com_mat->n, N, k);
  }

  /* check_com_mat(com_mat); */

  /* partition the com_matrix in k partitions*/
  partition = tm_kpartition(k, com_mat, N, constraints, nb_constraints);

  if(verbose_level>=INFO)
    display_partition(partition, local_vertices, N, depth, k);

  /* exit(-1); */
  /* split the communication matrix in k parts according to the partition just found above */
  tab_com_mat = split_com_mat( com_mat, N, k, partition);

  /* split the local vertices in k parts according to the partition just found above */
  tab_local_vertices = split_vertices( local_vertices, N, k, partition);

  /* construct a tab of constraints of  size k: one for each partitions*/
  const_tab = split_constraints (constraints, nb_constraints, k, topology, depth, N);

  /* create the table of k nodes of the resulting sub-tree */
  tab_child = (tm_tree_t **) CALLOC (k,sizeof(tm_tree_t*));
  for( i = 0 ; i < k ; i++){
    tab_child[i] = (tm_tree_t *) MALLOC(sizeof(tm_tree_t));
  }

  /* for each child, proceeed recursively*/
  for( i = 0 ; i < k ; i++){
    tab_child[i]->id = i;
    kpartition_build_level_topology ( tab_child[i], tab_com_mat[i], N/k, depth + 1,
				      topology, tab_local_vertices[i],
				      const_tab[i].constraints, const_tab[i].length,
				      obj_weight, comm_speed);
    tab_child[i]->parent = cur_node;
  }

  /* link the node with its child */
  tm_set_node( cur_node, tab_child, k, NULL, cur_node->id, 0, NULL, depth);

  /* free local data*/
  FREE(partition);
  free_tab_com_mat(tab_com_mat,k);
  free_tab_local_vertices(tab_local_vertices,k);
  free_const_tab(const_tab,k);
}


tm_tree_t *tm_kpartition_build_tree_from_topology(tm_topology_t *topology,double **comm,int N, int *constraints, int nb_constraints, double *obj_weight, double *com_speed)
{
  int depth,i, K;
  tm_tree_t *root = NULL;
  int *local_vertices = NULL;
  int nb_cores;
  com_mat_t com_mat;

  verbose_level = tm_get_verbose_level();


  nb_cores=tm_nb_processing_units(topology)*topology->oversub_fact;


  if(verbose_level>=INFO)
    printf("Number of constraints: %d, N=%d, nb_cores = %d, K=%d\n", nb_constraints, N, nb_cores, nb_cores-N);

  if((constraints == NULL) && (nb_constraints != 0)){
    if(verbose_level>=ERROR)
      fprintf(stderr,"size of constraint table not zero while constraint tab is NULL\n");
    return NULL;
  }

  if((constraints != NULL) && (nb_constraints > nb_cores)){
    if(verbose_level>=ERROR)
      fprintf(stderr,"size of constraint table (%d) is greater than the number of cores (%d)\n", nb_constraints, nb_cores);
    return NULL;
  }

  depth = 0;

  /* if we have more cores than processes add new dumb process to the com matrix*/
  if((K=nb_cores - N)>0){
    /* add K element to the object weight*/
    tm_complete_obj_weight(&obj_weight,N,K);
  } else if( K < 0){
    if(verbose_level>=ERROR)
      fprintf(stderr,"Not enough cores!\n");
    return NULL;
  }

  com_mat.comm = comm;
  com_mat.n    = N;

  /*
     local_vertices is the array of vertices that can be used
     the min(N,nb_contraints) 1st element are number from 0 to N
     the last ones have value -1
     the value of this array will be used to number the leaves of the tm_tree_t tree
     that start at "root"

     min(N,nb_contraints) is used to tackle the case where there is less processes than constraints

   */

  local_vertices = (int*) MALLOC (sizeof(int) * (K+N));

  for( i = 0 ; i < TM_MIN(N,nb_constraints) ; i++)
    local_vertices[i] = i;
  for( i = TM_MIN(N,nb_constraints) ;i < N + K ; i++)
    local_vertices[i] = -1;

  /* we assume all objects have the same arity*/
  /* assign the root of the tree*/
  root = (tm_tree_t*) MALLOC (sizeof(tm_tree_t));
  root -> id = 0;


  /*build the tree downward from the root*/
  kpartition_build_level_topology(root, &com_mat, N+K,  depth, topology, local_vertices,
					constraints, nb_constraints, obj_weight, com_speed);

  /*print_1D_tab(local_vertices,K+N);*/
  if(verbose_level>=INFO)
    printf("Build (bottom-up) tree done!\n");



  FREE(local_vertices);


  /* tell the system it is a constraint tree, this is usefull for freeing pointers*/
  root->constraint = 1;

  return root;
}