-
Notifications
You must be signed in to change notification settings - Fork 1
/
domain.c
2723 lines (2262 loc) · 89.7 KB
/
domain.c
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
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
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
953
954
955
956
957
958
959
960
961
962
963
964
965
966
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
#include <mpi.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <math.h>
#include "allvars.h"
#include "proto.h"
/*! \file domain.c
* \brief code for domain decomposition
*
* This file contains the code for the domain decomposition of the
* simulation volume. The domains are constructed from disjoint subsets
* of the leaves of a fiducial top-level tree that covers the full
* simulation volume. Domain boundaries hence run along tree-node
* divisions of a fiducial global BH tree. As a result of this method, the
* tree force are in principle strictly independent of the way the domains
* are cut. The domain decomposition can be carried out for an arbitrary
* number of CPUs. Individual domains are not cubical, but spatially
* coherent since the leaves are traversed in a Peano-Hilbert order and
* individual domains form segments along this order. This also ensures
* that each domain has a small surface to volume ratio, which minimizes
* communication.
*/
/*
* This file was originally part of the GADGET3 code developed by
* Volker Springel. The code has been modified
* somewhat by Phil Hopkins (phopkins@caltech.edu) for GIZMO; these
* modifications do not change the core algorithm, but have optimized it in
* some places, changed relative weighting factors for different levels in the
* domain decomposition, and similar details. Also how some memory issues are
* handled has been updated to reflect the newer more general parallelization
* structures in GIZMO.
*/
/*! toGo[task*NTask + partner] gives the number of particles in task 'task'
* that have to go to task 'partner'
*/
static int *toGo, *toGoGas;
static int *toGet, *toGetGas;
static int *list_NumPart;
static int *list_N_gas;
static int *list_load;
static int *list_loadgas;
static double *list_work;
static double *list_workgas;
extern int old_MaxPart, new_MaxPart;
#ifdef SEPARATE_STELLARDOMAINDECOMP
#define KD_COUNT_GAS_IN_DOMAIN
#define KD_COUNT_STARS_IN_DOMAIN
static int *toGoStars, *toGetStars, *list_N_stars, *list_loadstars;
static double *list_workstars;
#define N_DOMAINDECOMP_QUEUES 4
#else
#define N_DOMAINDECOMP_QUEUES 3
#endif
static struct local_topnode_data
{
peanokey Size; /*!< number of Peano-Hilbert mesh-cells represented by top-level node */
peanokey StartKey; /*!< first Peano-Hilbert key in top-level node */
long long Count; /*!< counts the number of particles in this top-level node */
double Cost;
double GasCost;
int Daughter; /*!< index of first daughter cell (out of 8) of top-level node */
int Leaf; /*!< if the node is a leaf, this gives its number when all leaves are traversed in Peano-Hilbert order */
int Parent;
int PIndex; /*!< first particle in node */
}
*topNodes; /*!< points to the root node of the top-level tree */
static struct peano_hilbert_data
{
peanokey key;
int index;
}
*mp;
static void domain_insertnode(struct local_topnode_data *treeA, struct local_topnode_data *treeB, int noA, int noB);
static void domain_add_cost(struct local_topnode_data *treeA, int noA, long long count, double cost, double gascost);
static float *domainWork; /*!< a table that gives the total "work" due to the particles stored by each processor */
static float *domainWorkGas; /*!< a table that gives the total "work" due to the particles stored by each processor */
static int *domainCount; /*!< a table that gives the total number of particles held by each processor */
static int *domainCountGas; /*!< a table that gives the total number of gas cells held by each processor */
static int domain_allocated_flag = 0;
static int maxLoad, maxLoadgas;
static double totgravcost, gravcost, totgascost, gascost;
#ifdef SEPARATE_STELLARDOMAINDECOMP
static float *domainWorkStars;
static int *domainCountStars;
static int maxLoadstars;
static double totstarcost, starcost;
#endif
static long long totpartcount;
static int UseAllParticles;
/*! This is the main routine for the domain decomposition. It acts as a driver routine that allocates various temporary buffers, maps the
* particles back onto the periodic box if needed, and then does the domain decomposition, and a final Peano-Hilbert order of all particles as a tuning measure. */
void domain_Decomposition(int UseAllTimeBins, int SaveKeys, int do_particle_mergesplit_key)
{
int i, ret, retsum, diff, highest_bin_to_include; size_t bytes, all_bytes; double t0, t1;
/* call first -before- a merge-split, to be sure particles are in the correct order in the tree */
// TO: we don't have to call this before merge_and_split particles()
// Actually we shouldn't because there are tree-walks in merge_and_split_particles().
//rearrange_particle_sequence();
if((All.Ti_Current > All.TimeBegin)&&(do_particle_mergesplit_key==1))
{
merge_and_split_particles(); /* do the particle split/merge operations: only do this on tree-building super-steps */
}
rearrange_particle_sequence(); /* must be called after merge_and_split_particles, and should always be called before new domains are built */
UseAllParticles = UseAllTimeBins;
for(i = 0; i < NumPart; i++) {if(P[i].Ti_current != All.Ti_Current) {drift_particle(i, All.Ti_Current);}}
force_treefree();
domain_free();
if(old_MaxPart) {All.MaxPart = new_MaxPart; old_MaxPart = 0;}
#ifdef BOX_PERIODIC
do_box_wrapping(); /* map the particles back onto the box */
#endif
MPI_Barrier(MPI_COMM_WORLD); CPU_Step[CPU_DRIFT] += measure_time(); // sync everything after merge-split and rearrange //
TreeReconstructFlag = 1; /* ensures that new tree will be constructed */
#ifdef SINGLE_STAR_SINK_DYNAMICS
All.NumForcesSinceLastDomainDecomp = 0;
#endif
/* we take the closest cost factor */
if(UseAllParticles) {highest_bin_to_include = All.HighestOccupiedTimeBin;} else {highest_bin_to_include = All.HighestActiveTimeBin;}
for(i = 1, TakeLevel = 0, diff = abs(All.LevelToTimeBin[0] - highest_bin_to_include); i < GRAVCOSTLEVELS; i++)
{if(diff > abs(All.LevelToTimeBin[i] - highest_bin_to_include)) {TakeLevel = i; diff = abs(All.LevelToTimeBin[i] - highest_bin_to_include);}}
PRINT_STATUS("Domain decomposition building... LevelToTimeBin[TakeLevel=%d]=%d (presently allocated=%g MB)", TakeLevel, All.LevelToTimeBin[TakeLevel], AllocatedBytes / (1024.0 * 1024.0));
t0 = my_second();
do
{
domain_allocate();
all_bytes = 0;
Key = (peanokey *) mymalloc("domain_key", bytes = (sizeof(peanokey) * All.MaxPart));
all_bytes += bytes;
toGo = (int *) mymalloc("toGo", bytes = (sizeof(int) * NTask));
all_bytes += bytes;
toGoGas = (int *) mymalloc("toGoGas", bytes = (sizeof(int) * NTask));
all_bytes += bytes;
toGet = (int *) mymalloc("toGet", bytes = (sizeof(int) * NTask));
all_bytes += bytes;
toGetGas = (int *) mymalloc("toGetGas", bytes = (sizeof(int) * NTask));
all_bytes += bytes;
list_NumPart = (int *) mymalloc("list_NumPart", bytes = (sizeof(int) * NTask));
all_bytes += bytes;
list_N_gas = (int *) mymalloc("list_N_gas", bytes = (sizeof(int) * NTask));
all_bytes += bytes;
list_load = (int *) mymalloc("list_load", bytes = (sizeof(int) * NTask));
all_bytes += bytes;
list_loadgas = (int *) mymalloc("list_loadgas", bytes = (sizeof(int) * NTask));
all_bytes += bytes;
list_work = (double *) mymalloc("list_work", bytes = (sizeof(double) * NTask));
all_bytes += bytes;
list_workgas = (double *) mymalloc("list_workgas", bytes = (sizeof(double) * NTask));
all_bytes += bytes;
domainWork = (float *) mymalloc("domainWork", bytes = (MaxTopNodes * sizeof(float)));
all_bytes += bytes;
domainWorkGas = (float *) mymalloc("domainWorkGas", bytes = (MaxTopNodes * sizeof(float)));
all_bytes += bytes;
domainCount = (int *) mymalloc("domainCount", bytes = (MaxTopNodes * sizeof(int)));
all_bytes += bytes;
domainCountGas = (int *) mymalloc("domainCountGas", bytes = (MaxTopNodes * sizeof(int)));
all_bytes += bytes;
#ifdef SEPARATE_STELLARDOMAINDECOMP
toGoStars = (int *) mymalloc("toGoStars", bytes = (sizeof(int) * NTask)); all_bytes += bytes;
toGetStars = (int *) mymalloc("toGetStars", bytes = (sizeof(int) * NTask)); all_bytes += bytes;
list_N_stars = (int *) mymalloc("list_N_stars", bytes = (sizeof(int) * NTask)); all_bytes += bytes;
list_loadstars = (int *) mymalloc("list_loadstars", bytes = (sizeof(int) * NTask)); all_bytes += bytes;
list_workstars = (double *) mymalloc("list_workstars", bytes = (sizeof(double) * NTask)); all_bytes += bytes;
domainWorkStars = (float *) mymalloc("domainWorkStars", bytes = (MaxTopNodes * sizeof(float))); all_bytes += bytes;
domainCountStars = (int *) mymalloc("domainCountStars", bytes = (MaxTopNodes * sizeof(int))); all_bytes += bytes;
#endif
topNodes = (struct local_topnode_data *) mymalloc("topNodes", bytes =
(MaxTopNodes * sizeof(struct local_topnode_data)));
all_bytes += bytes;
PRINT_STATUS(" ..using %g MB of temporary storage for domain decomposition... (presently allocated=%g MB)",all_bytes / (1024.0 * 1024.0), AllocatedBytes / (1024.0 * 1024.0));
maxLoad = (int) (All.MaxPart * REDUC_FAC);
maxLoadgas = (int) (All.MaxPartGas * REDUC_FAC);
#ifdef SEPARATE_STELLARDOMAINDECOMP
maxLoadstars = (int) (All.MaxPart * REDUC_FAC);
#endif
report_memory_usage(&HighMark_domain, "DOMAIN");
ret = domain_decompose();
/* copy what we need for the topnodes */
for(i = 0; i < NTopnodes; i++)
{
TopNodes[i].StartKey = topNodes[i].StartKey;
TopNodes[i].Size = topNodes[i].Size;
TopNodes[i].Daughter = topNodes[i].Daughter;
TopNodes[i].Leaf = topNodes[i].Leaf;
}
myfree(topNodes);
#ifdef SEPARATE_STELLARDOMAINDECOMP
myfree(domainCountStars);
myfree(domainWorkStars);
myfree(list_workstars);
myfree(list_loadstars);
myfree(list_N_stars);
myfree(toGetStars);
myfree(toGoStars);
#endif
myfree(domainCountGas);
myfree(domainCount);
myfree(domainWorkGas);
myfree(domainWork);
myfree(list_workgas);
myfree(list_work);
myfree(list_loadgas);
myfree(list_load);
myfree(list_N_gas);
myfree(list_NumPart);
myfree(toGetGas);
myfree(toGet);
myfree(toGoGas);
myfree(toGo);
MPI_Allreduce(&ret, &retsum, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
if(retsum)
{
myfree(Key);
domain_free();
if(ThisTask == 0) {printf("Increasing TopNodeAllocFactor=%g ", All.TopNodeAllocFactor);}
All.TopNodeAllocFactor *= 1.3;
PRINT_STATUS("..new value=%g", All.TopNodeAllocFactor);
if(All.TopNodeAllocFactor > 1000) {printf("something seems to be going seriously wrong here. Stopping.\n"); fflush(stdout); endrun(781);}
}
}
while(retsum);
t1 = my_second();
PRINT_STATUS(" ..domain decomposition done. (took %g sec)", timediff(t0, t1));
CPU_Step[CPU_DOMAIN] += measure_time();
for(i = 0; i < NumPart; i++) {if(P[i].Type > 5 || P[i].Type < 0) {printf("task=%d: P[i=%d].Type=%d\n", ThisTask, i, P[i].Type); endrun(111111);}}
#ifdef PEANOHILBERT
#ifdef SUBFIND
if(GrNr < 0) /* we don't do it when SUBFIND is executed for a certain group */
#endif
{peano_hilbert_order();}
CPU_Step[CPU_PEANO] += measure_time();
#endif
myfree(Key);
memmove(TopNodes + NTopnodes, DomainTask, NTopnodes * sizeof(int));
TopNodes = (struct topnode_data *) myrealloc(TopNodes, bytes = (NTopnodes * sizeof(struct topnode_data) + NTopnodes * sizeof(int)));
PRINT_STATUS(" ..freed %g MByte in top-level domain structure", (MaxTopNodes - NTopnodes) * sizeof(struct topnode_data) / (1024.0 * 1024.0));
DomainTask = (int *) (TopNodes + NTopnodes);
force_treeallocate((int) (All.TreeAllocFactor * All.MaxPart) + NTopnodes, All.MaxPart);
reconstruct_timebins();
}
/*! This function allocates all the stuff that will be required for the tree-construction/walk later on */
void domain_allocate(void)
{
size_t bytes, all_bytes = 0;
MaxTopNodes = (int) (All.TopNodeAllocFactor * All.MaxPart + 1);
DomainStartList = (int *) mymalloc("DomainStartList", bytes = (NTask * MULTIPLEDOMAINS * sizeof(int)));
all_bytes += bytes;
DomainEndList = (int *) mymalloc("DomainEndList", bytes = (NTask * MULTIPLEDOMAINS * sizeof(int)));
all_bytes += bytes;
TopNodes = (struct topnode_data *) mymalloc("TopNodes", bytes = (MaxTopNodes * sizeof(struct topnode_data) + MaxTopNodes * sizeof(int)));
all_bytes += bytes;
DomainTask = (int *) (TopNodes + MaxTopNodes);
PRINT_STATUS(" ..allocated %g MByte for top-level domain structure", all_bytes / (1024.0 * 1024.0));
domain_allocated_flag = 1;
}
void domain_free(void)
{
if(domain_allocated_flag)
{
myfree(TopNodes);
myfree(DomainEndList);
myfree(DomainStartList);
domain_allocated_flag = 0;
}
}
static struct topnode_data *save_TopNodes;
static int *save_DomainStartList, *save_DomainEndList;
void domain_free_trick(void)
{
if(domain_allocated_flag)
{
save_TopNodes = TopNodes;
save_DomainEndList = DomainEndList;
save_DomainStartList = DomainStartList;
domain_allocated_flag = 0;
}
else
{endrun(131231);}
}
void domain_allocate_trick(void)
{
domain_allocated_flag = 1;
TopNodes = save_TopNodes;
DomainEndList = save_DomainEndList;
DomainStartList = save_DomainStartList;
}
/* this function determines how particle work-costs are 'weighted' for load-balancing. if you
have additional, expensive physics which only apply to a subset of particles, it may be worth
up-weighting those particles here, so the code knows to try and spread them around. otherwise,
they may end up all bunched onto the same processor */
double domain_particle_cost_multiplier(int i)
{
double multiplier = 0;
if(P[i].Type == 0) /* for gas, weight particles with large neighbor number more, since they require more work */
{
double nngb_reduced = PPP[i].NumNgb; /* remember, in density.c we reduce this by pow(1/NUMDIMS), for use in other routines: need to correct here */
#if (NUMDIMS==3)
multiplier = nngb_reduced*nngb_reduced*nngb_reduced / All.DesNumNgb;
#elif (NUMDIMS==2)
multiplier = nngb_reduced*nngb_reduced / All.DesNumNgb;
#else
multiplier = nngb_reduced / All.DesNumNgb;
#endif
if(multiplier < 0.5) {multiplier = 0.5;} // floor //
} // end gas check
#if defined(GALSF) /* with star formation active, we will up-weight star particles which are active feedback sources */
#ifndef CHIMES /* With CHIMES, the chemistry dominates the cost, so we boost (dense) gas but not stars. */
if(((P[i].Type == 4)||((All.ComovingIntegrationOn==0)&&((P[i].Type == 2)||(P[i].Type==3))))&&(P[i].Mass>0))
{
double star_age = evaluate_stellar_age_Gyr(i);
if(star_age>0.1) {multiplier = 3.125;} else {if(star_age>0.035) {multiplier = 5.;} else {multiplier = 10.;}}
}
#endif
#endif
#ifdef CHIMES
/* With CHIMES, cost is dominated by the chemistry, particularly in dense gas. We therefore boost the cost factor of gas particles with nH >~ 1 cm^-3. */
if(P[i].Type == 0) {double nH_cgs = SphP[i].Density * All.cf_a3inv * UNIT_DENSITY_IN_NHCGS; if(nH_cgs > 1) {multiplier = 10.0;}}
#endif
return multiplier;
}
/* simple function to return costfactor for pure gravity calculation: based just on gravcost calculation, with constant for safety */
double domain_particle_costfactor(int i)
{
return 0.1 + P[i].GravCost[TakeLevel];
}
/*! This function carries out the actual domain decomposition for all
* particle types. It will try to balance the work-load for each domain,
* as estimated based on the P[i]-GravCost values. The decomposition will
* respect the maximum allowed memory-imbalance given by the value of
* PartAllocFactor.
*/
int domain_decompose(void)
{
int i, no, status;
long long sumtogo, sumload, sumloadgas;
int maxload, maxloadgas, multipledomains = MULTIPLEDOMAINS;
double sumwork, maxwork, sumworkgas, maxworkgas;
#ifdef SEPARATE_STELLARDOMAINDECOMP
long long sumloadstars;
int maxloadstars;
double sumworkstars, maxworkstars;
starcost = 0;
#endif
for(i = 0; i < 6; i++) {NtypeLocal[i] = 0;}
for(i = 0, gravcost = gascost = 0; i < NumPart; i++)
{
#ifdef SUBFIND
if(GrNr >= 0 && P[i].GrNr != GrNr) {continue;}
#endif
NtypeLocal[P[i].Type]++;
double wt_0 = domain_particle_costfactor(i);
double wt_mult = domain_particle_cost_multiplier(i);
gravcost += (1 + wt_mult) * wt_0;
//if(P[i].Type == 0) {if(TimeBinActive[P[i].TimeBin] || UseAllParticles) {gascost += wt_mult;}}
if(P[i].Type == 0) {if(TimeBinActive[P[i].TimeBin] || UseAllParticles) {gascost += wt_0;}}
#ifdef SEPARATE_STELLARDOMAINDECOMP
if(P[i].Type == 4 || P[i].Type == 5) {if(TimeBinActive[P[i].TimeBin] || UseAllParticles) {starcost += wt_0;}}
#endif
}
/* because Ntype[] is of type `long long', we cannot do a simple MPI_Allreduce() to sum the total particle numbers */
sumup_large_ints(6, NtypeLocal, Ntype);
for(i = 0, totpartcount = 0; i < 6; i++) {totpartcount += Ntype[i];}
MPI_Allreduce(&gravcost, &totgravcost, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
MPI_Allreduce(&gascost, &totgascost, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
/* determine global dimensions of domain grid */
domain_findExtent();
if(domain_determineTopTree()) {return 1;}
/* find the split of the domain grid */
domain_findSplit_work_balanced(multipledomains * NTask, NTopleaves);
domain_assign_load_or_work_balanced(1,multipledomains);
status = domain_check_memory_bound(multipledomains);
if(status != 0) /* the optimum balanced solution violates memory constraint, let's try something different */
{
if(ThisTask == 0) {printf("Note: the domain decomposition is suboptimum because the ceiling for memory-imbalance is reached\n");}
domain_findSplit_load_balanced(multipledomains * NTask, NTopleaves);
domain_assign_load_or_work_balanced(0,multipledomains);
status = domain_check_memory_bound(multipledomains);
if(status != 0)
{
if(ThisTask == 0) {printf("No domain decomposition that stays within memory bounds is possible.\n");}
endrun(0);
}
}
if(ThisTask == 0)
{
sumload = maxload = sumloadgas = maxloadgas = 0;
sumwork = sumworkgas = maxwork = maxworkgas = 0;
#ifdef SEPARATE_STELLARDOMAINDECOMP
sumloadstars = maxloadstars = 0;
sumworkstars = maxworkstars = 0;
#endif
for(i = 0; i < NTask; i++)
{
sumload += list_load[i];
sumloadgas += list_loadgas[i];
sumwork += list_work[i];
sumworkgas += list_workgas[i];
#ifdef SEPARATE_STELLARDOMAINDECOMP
sumloadstars += list_loadstars[i];
sumworkstars += list_workstars[i];
#endif
if(list_load[i] > maxload) {maxload = list_load[i];}
if(list_loadgas[i] > maxloadgas) {maxloadgas = list_loadgas[i];}
#ifdef SEPARATE_STELLARDOMAINDECOMP
if(list_loadstars[i] > maxloadstars) {maxloadstars = list_loadstars[i];}
#endif
if(list_work[i] > maxwork) {maxwork = list_work[i];}
if(list_workgas[i] > maxworkgas) {maxworkgas = list_workgas[i];}
#ifdef SEPARATE_STELLARDOMAINDECOMP
if(list_workstars[i] > maxworkstars) {maxworkstars = list_workstars[i];}
#endif
}
printf("Balance: gravity work-load balance=%g memory-balance=%g hydro work-load balance=%g\n",
maxwork / (sumwork / NTask), maxload / (((double) sumload) / NTask), maxworkgas / ((sumworkgas + 1.0e-30) / NTask));
}
/* flag the particles that need to be exported */
for(i = 0; i < NumPart; i++)
{
#ifdef SUBFIND
if(GrNr >= 0 && P[i].GrNr != GrNr) {continue;}
#endif
no = 0;
while(topNodes[no].Daughter >= 0) {no = topNodes[no].Daughter + (Key[i] - topNodes[no].StartKey) / (topNodes[no].Size / 8);}
no = topNodes[no].Leaf;
int task = DomainTask[no];
if(task != ThisTask) {P[i].Type |= 32;}
}
int iter = 0, ret;
size_t exchange_limit;
do
{
exchange_limit = FreeBytes - NTask * (24 * sizeof(int) + 16 * sizeof(MPI_Request));
if(exchange_limit <= 0)
{
printf("task=%d: exchange_limit=%d\n", ThisTask, (int) exchange_limit);
endrun(1223);
}
/* determine for each cpu how many particles have to be shifted to other cpus */
ret = domain_countToGo(exchange_limit);
for(i = 0, sumtogo = 0; i < NTask; i++) {sumtogo += toGo[i];}
sumup_longs(1, &sumtogo, &sumtogo);
PRINT_STATUS(" ..iter=%d exchange of %d%09d particles (ret=%d)", iter, (int) (sumtogo / 1000000000), (int) (sumtogo % 1000000000), ret);
domain_exchange();
iter++;
}
while(ret > 0);
return 0;
}
int domain_check_memory_bound(int multipledomains)
{
int ta, m, i;
int load, gasload, max_load, max_gasload;
double work, workgas;
#ifdef SEPARATE_STELLARDOMAINDECOMP
int starsload, max_starsload;
double workstars;
#endif
max_load = max_gasload = 0;
#ifdef SEPARATE_STELLARDOMAINDECOMP
max_starsload = 0;
#endif
for(ta = 0; ta < NTask; ta++)
{
load = gasload = 0;
work = workgas = 0;
#ifdef SEPARATE_STELLARDOMAINDECOMP
starsload = 0;
workstars = 0;
#endif
for(m = 0; m < multipledomains; m++)
for(i = DomainStartList[ta * multipledomains + m]; i <= DomainEndList[ta * multipledomains + m]; i++)
{
load += domainCount[i];
gasload += domainCountGas[i];
work += domainWork[i];
workgas += domainWorkGas[i];
#ifdef SEPARATE_STELLARDOMAINDECOMP
starsload += domainCountStars[i];
workstars += domainWorkStars[i];
#endif
}
list_load[ta] = load;
list_loadgas[ta] = gasload;
list_work[ta] = work;
list_workgas[ta] = workgas;
#ifdef SEPARATE_STELLARDOMAINDECOMP
list_loadstars[ta] = starsload;
list_workstars[ta] = workstars;
#endif
if(load > max_load) {max_load = load;}
if(gasload > max_gasload) {max_gasload = gasload;}
#ifdef SEPARATE_STELLARDOMAINDECOMP
if(starsload > max_starsload) {max_starsload = starsload;}
#endif
}
#ifdef SUBFIND
if(GrNr >= 0)
{
load = max_load;
gasload = max_gasload;
#ifdef SEPARATE_STELLARDOMAINDECOMP
starsload = max_starsload;
#endif
for(i = 0; i < NumPart; i++)
{
if(P[i].GrNr != GrNr)
{
load++;
if(P[i].Type == 0) {gasload++;}
#ifdef SEPARATE_STELLARDOMAINDECOMP
if(P[i].Type == 4 || P[i].Type == 5) {starsload++;}
#endif
}
}
MPI_Allreduce(&load, &max_load, 1, MPI_INT, MPI_MAX, MPI_COMM_WORLD);
MPI_Allreduce(&gasload, &max_gasload, 1, MPI_INT, MPI_MAX, MPI_COMM_WORLD);
#ifdef SEPARATE_STELLARDOMAINDECOMP
MPI_Allreduce(&starsload, &max_starsload, 1, MPI_INT, MPI_MAX, MPI_COMM_WORLD);
#endif
}
#endif
if(max_load > maxLoad)
{
if(ThisTask == 0) {printf("desired memory imbalance=%g (limit=%d, needed=%d)\n", (max_load * All.PartAllocFactor) / maxLoad, maxLoad, max_load);}
return 1;
}
if(max_gasload > maxLoadgas)
{
if(ThisTask == 0) {printf("desired memory imbalance=%g (GAS/FLUID) (limit=%d, needed=%d)\n", (max_gasload * All.PartAllocFactor) / maxLoadgas, maxLoadgas, max_gasload);}
return 1;
}
#ifdef SEPARATE_STELLARDOMAINDECOMP
if(max_starsload > maxLoadstars)
{
if(ThisTask == 0) {printf("desired memory imbalance=%g (STARS/SINKS) (limit=%d, needed=%d)\n", (max_starsload * All.PartAllocFactor) / maxLoadstars, maxLoadstars, max_starsload);}
return 1;
}
#endif
return 0;
}
void domain_exchange(void)
{
long count_togo = 0, count_togo_gas = 0, count_get = 0, count_get_gas = 0;
long *count, *count_gas, *offset, *offset_gas;
long *count_recv, *count_recv_gas, *offset_recv, *offset_recv_gas;
long i, n, ngrp, no, target;
struct particle_data *partBuf;
struct gas_cell_data *gasBuf;
peanokey *keyBuf;
count = (long *) mymalloc("count", NTask * sizeof(long));
count_gas = (long *) mymalloc("count_gas", NTask * sizeof(long));
offset = (long *) mymalloc("offset", NTask * sizeof(long));
offset_gas = (long *) mymalloc("offset_gas", NTask * sizeof(long));
count_recv = (long *) mymalloc("count_recv", NTask * sizeof(long));
count_recv_gas = (long *) mymalloc("count_recv_gas", NTask * sizeof(long));
offset_recv = (long *) mymalloc("offset_recv", NTask * sizeof(long));
offset_recv_gas = (long *) mymalloc("offset_recv_gas", NTask * sizeof(long));
#ifdef SEPARATE_STELLARDOMAINDECOMP
int count_togo_stars = 0, count_get_stars = 0;
int *count_stars, *offset_stars;
int *count_recv_stars, *offset_recv_stars;
count_stars = (int *) mymalloc("count_stars", NTask * sizeof(int));
offset_stars = (int *) mymalloc("offset_stars", NTask * sizeof(int));
count_recv_stars = (int *) mymalloc("count_recv_stars", NTask * sizeof(int));
offset_recv_stars = (int *) mymalloc("offset_recv_stars", NTask * sizeof(int));
#endif
long prec_offset, prec_count;
long *decrease;
decrease = (long *) mymalloc("decrease", NTask * sizeof(long));
for(i = 1, offset_gas[0] = 0, decrease[0] = 0; i < NTask; i++)
{
offset_gas[i] = offset_gas[i - 1] + toGoGas[i - 1];
decrease[i] = toGoGas[i - 1];
}
prec_offset = offset_gas[NTask - 1] + toGoGas[NTask - 1];
#ifdef SEPARATE_STELLARDOMAINDECOMP
offset_stars[0] = prec_offset;
for(i = 1; i < NTask; i++)
{
offset_stars[i] = offset_stars[i - 1] + toGoStars[i - 1];
decrease[i] += toGoStars[i - 1];
}
prec_offset = offset_stars[NTask - 1] + toGoStars[NTask - 1];
#endif
offset[0] = prec_offset;
for(i = 1; i < NTask; i++) {offset[i] = offset[i - 1] + (toGo[i - 1] - decrease[i]);}
myfree(decrease);
for(i = 0; i < NTask; i++)
{
count_togo += toGo[i];
count_togo_gas += toGoGas[i];
count_get += toGet[i];
count_get_gas += toGetGas[i];
#ifdef SEPARATE_STELLARDOMAINDECOMP
count_togo_stars += toGoStars[i];
count_get_stars += toGetStars[i];
#endif
}
partBuf = (struct particle_data *) mymalloc("partBuf", count_togo * sizeof(struct particle_data));
gasBuf = (struct gas_cell_data *) mymalloc("gasBuf", count_togo_gas * sizeof(struct gas_cell_data));
#ifdef CHIMES
struct gasVariables *gasChimesBuf;
ChimesFloat *gasAbundancesBuf, *gasAbundancesRecvBuf, *tempAbundanceArray;
int abunIndex;
gasChimesBuf = (struct gasVariables *) mymalloc("chiBuf", count_togo_gas * sizeof(struct gasVariables));
gasAbundancesBuf = (ChimesFloat *) mymalloc("abunBuf", count_togo_gas * ChimesGlobalVars.totalNumberOfSpecies * sizeof(ChimesFloat));
gasAbundancesRecvBuf = (ChimesFloat *) mymalloc("xRecBuf", count_get_gas * ChimesGlobalVars.totalNumberOfSpecies * sizeof(ChimesFloat));
tempAbundanceArray = (ChimesFloat *) malloc(ChimesGlobalVars.totalNumberOfSpecies * sizeof(ChimesFloat));
#endif
keyBuf = (peanokey *) mymalloc("keyBuf", count_togo * sizeof(peanokey));
for(i = 0; i < NTask; i++) {count[i] = count_gas[i] = 0;}
#ifdef SEPARATE_STELLARDOMAINDECOMP
for(i = 0; i < NTask; i++) {count_stars[i] = 0;}
#endif
for(n = 0; n < NumPart; n++)
{
if((P[n].Type & (32 + 16)) == (32 + 16)) /* flagged with both 16 and 32 */
{
P[n].Type &= 15; /* clear 16 and 32 */
no = 0;
/* new code - not clear if strictly necessary or just optimization */
/*
peanokey mask = ((peanokey)7) << (3 * (BITS_PER_DIMENSION - 1));
int shift = 3 * (BITS_PER_DIMENSION - 1);
while(topNodes[no].Daughter >= 0)
{
no = topNodes[no].Daughter + (int)((Key[n] & mask) >> shift);
mask >>= 3;
shift -= 3;
}
*/
/* old code */
while(topNodes[no].Daughter >= 0) {no = topNodes[no].Daughter + (Key[n] - topNodes[no].StartKey) / (topNodes[no].Size / 8);}
no = topNodes[no].Leaf;
target = DomainTask[no];
if(P[n].Type == 0)
{
partBuf[offset_gas[target] + count_gas[target]] = P[n];
keyBuf[offset_gas[target] + count_gas[target]] = Key[n];
#ifdef CHIMES
for(i = 0; i < ChimesGlobalVars.totalNumberOfSpecies; i++) {gasAbundancesBuf[((offset_gas[target] + count_gas[target]) * ChimesGlobalVars.totalNumberOfSpecies) + i] = ChimesGasVars[n].abundances[i];}
free_gas_abundances_memory(&(ChimesGasVars[n]), &ChimesGlobalVars);
ChimesGasVars[n].abundances = NULL;
ChimesGasVars[n].isotropic_photon_density = NULL;
ChimesGasVars[n].G0_parameter = NULL;
ChimesGasVars[n].H2_dissocJ = NULL;
gasChimesBuf[offset_gas[target] + count_gas[target]] = ChimesGasVars[n];
#endif
gasBuf[offset_gas[target] + count_gas[target]] = SphP[n];
count_gas[target]++;
}
#ifdef SEPARATE_STELLARDOMAINDECOMP
else if(P[n].Type == 4 || P[n].Type == 5)
{
partBuf[offset_stars[target] + count_stars[target]] = P[n];
keyBuf[offset_stars[target] + count_stars[target]] = Key[n];
count_stars[target]++;
}
#endif
else
{
partBuf[offset[target] + count[target]] = P[n];
keyBuf[offset[target] + count[target]] = Key[n];
count[target]++;
}
if(P[n].Type == 0)
{
P[n] = P[N_gas - 1];
SphP[n] = SphP[N_gas - 1];
Key[n] = Key[N_gas - 1];
#ifdef CHIMES
if (n < N_gas - 1)
{
for(abunIndex = 0; abunIndex < ChimesGlobalVars.totalNumberOfSpecies; abunIndex++)
{tempAbundanceArray[abunIndex] = ChimesGasVars[N_gas - 1].abundances[abunIndex];}
free_gas_abundances_memory(&(ChimesGasVars[N_gas - 1]), &ChimesGlobalVars);
ChimesGasVars[N_gas - 1].abundances = NULL;
ChimesGasVars[N_gas - 1].isotropic_photon_density = NULL;
ChimesGasVars[N_gas - 1].G0_parameter = NULL;
ChimesGasVars[N_gas - 1].H2_dissocJ = NULL;
ChimesGasVars[n] = ChimesGasVars[N_gas - 1];
allocate_gas_abundances_memory(&(ChimesGasVars[n]), &ChimesGlobalVars);
for (abunIndex = 0; abunIndex < ChimesGlobalVars.totalNumberOfSpecies; abunIndex++)
{ChimesGasVars[n].abundances[abunIndex] = tempAbundanceArray[abunIndex];}
}
#endif
P[N_gas - 1] = P[NumPart - 1];
Key[N_gas - 1] = Key[NumPart - 1];
NumPart--;
N_gas--;
n--;
}
#ifdef SEPARATE_STELLARDOMAINDECOMP
else if(P[n].Type == 4 || P[n].Type == 5)
{
if(n < NumPart - 1)
{
P[n] = P[NumPart - 1];
Key[n] = Key[NumPart - 1];
}
NumPart--;
N_stars--;
n--;
}
#endif
else
{
P[n] = P[NumPart - 1];
Key[n] = Key[NumPart - 1];
NumPart--;
n--;
}
}
}
#ifdef CHIMES
free(tempAbundanceArray);
#endif
long count_totget;
count_totget = count_get_gas;
#ifdef SEPARATE_STELLARDOMAINDECOMP
count_totget += count_get_stars;
#endif
if(count_totget)
{
memmove(P + N_gas + count_totget, P + N_gas, (NumPart - N_gas) * sizeof(struct particle_data));
memmove(Key + N_gas + count_totget, Key + N_gas, (NumPart - N_gas) * sizeof(peanokey));
}
for(i = 0; i < NTask; i++)
{
count_recv_gas[i] = toGetGas[i];
count_recv[i] = toGet[i] - toGetGas[i];
#ifdef SEPARATE_STELLARDOMAINDECOMP
count_recv_stars[i] = toGetStars[i];
count_recv[i] -= toGetStars[i];
#endif
}
for(i = 1, offset_recv_gas[0] = N_gas; i < NTask; i++)
{offset_recv_gas[i] = offset_recv_gas[i - 1] + count_recv_gas[i - 1];}
prec_count = N_gas + count_get_gas;
#ifdef SEPARATE_STELLARDOMAINDECOMP
offset_recv_stars[0] = prec_count;
for(i = 1; i < NTask; i++)
{offset_recv_stars[i] = offset_recv_stars[i - 1] + count_recv_stars[i - 1];}
prec_count += count_get_stars;
#endif
offset_recv[0] = NumPart - N_gas + prec_count;
for(i = 1; i < NTask; i++)
{offset_recv[i] = offset_recv[i - 1] + count_recv[i - 1];}
#ifndef NO_ISEND_IRECV_IN_DOMAIN
int n_requests = 0, max_requests = 10;
MPI_Request *requests;
#ifdef SEPARATE_STELLARDOMAINDECOMP
max_requests += 6; // check this, see what its hardwired to //
#endif
#ifdef CHIMES
max_requests += 4;
#endif
requests = (MPI_Request *) mymalloc("requests", max_requests * NTask * sizeof(MPI_Request));
for(ngrp = 1; ngrp < (1 << PTask); ngrp++)
{
target = ThisTask ^ ngrp;
if(target < NTask)
{
if(count_recv_gas[target] > 0)
{
MPI_Irecv(P + offset_recv_gas[target], count_recv_gas[target] * sizeof(struct particle_data),
MPI_BYTE, target, TAG_PDATA_GAS, MPI_COMM_WORLD, &requests[n_requests++]);
MPI_Irecv(Key + offset_recv_gas[target], count_recv_gas[target] * sizeof(peanokey),
MPI_BYTE, target, TAG_KEY_GAS, MPI_COMM_WORLD, &requests[n_requests++]);
MPI_Irecv(SphP + offset_recv_gas[target],
count_recv_gas[target] * sizeof(struct gas_cell_data), MPI_BYTE, target,
TAG_GASDATA, MPI_COMM_WORLD, &requests[n_requests++]);
#ifdef CHIMES
MPI_Irecv(ChimesGasVars + offset_recv_gas[target],
count_recv_gas[target] * sizeof(struct gasVariables), MPI_BYTE, target,
TAG_CHIMESDATA, MPI_COMM_WORLD, &requests[n_requests++]);
#ifdef CHIMES_USE_DOUBLE_PRECISION
MPI_Irecv(gasAbundancesRecvBuf + ((offset_recv_gas[target] - offset_recv_gas[0]) * ChimesGlobalVars.totalNumberOfSpecies),
count_recv_gas[target] * ChimesGlobalVars.totalNumberOfSpecies, MPI_DOUBLE, target, TAG_ABUNDATA,
MPI_COMM_WORLD, &requests[n_requests++]);
#else
MPI_Irecv(gasAbundancesRecvBuf + ((offset_recv_gas[target] - offset_recv_gas[0]) * ChimesGlobalVars.totalNumberOfSpecies),
count_recv_gas[target] * ChimesGlobalVars.totalNumberOfSpecies, MPI_FLOAT, target, TAG_ABUNDATA,
MPI_COMM_WORLD, &requests[n_requests++]);
#endif
#endif
}
#ifdef SEPARATE_STELLARDOMAINDECOMP
if(count_recv_stars[target] > 0)
{
MPI_Irecv(P + offset_recv_stars[target],
count_recv_stars[target] * sizeof(struct particle_data), MPI_BYTE, target,
TAG_PDATA_STARS, MPI_COMM_WORLD, &requests[n_requests++]);
MPI_Irecv(Key + offset_recv_stars[target], count_recv_stars[target] * sizeof(peanokey),
MPI_BYTE, target, TAG_KEY_STARS, MPI_COMM_WORLD, &requests[n_requests++]);
}
#endif
if(count_recv[target] > 0)
{
MPI_Irecv(P + offset_recv[target], count_recv[target] * sizeof(struct particle_data),
MPI_BYTE, target, TAG_PDATA, MPI_COMM_WORLD, &requests[n_requests++]);
MPI_Irecv(Key + offset_recv[target], count_recv[target] * sizeof(peanokey),
MPI_BYTE, target, TAG_KEY, MPI_COMM_WORLD, &requests[n_requests++]);
}
}
}
MPI_Barrier(MPI_COMM_WORLD); /* not really necessary, but this will guarantee that all receives are
posted before the sends, which helps the stability of MPI on
bluegene, and perhaps some mpich1-clusters */
for(ngrp = 1; ngrp < (1 << PTask); ngrp++)
{
target = ThisTask ^ ngrp;
if(target < NTask)
{
if(count_gas[target] > 0)
{
MPI_Isend(partBuf + offset_gas[target], count_gas[target] * sizeof(struct particle_data),
MPI_BYTE, target, TAG_PDATA_GAS, MPI_COMM_WORLD, &requests[n_requests++]);