Source

ompi-svn-mirror / oshmem / mca / scoll / basic / scoll_basic_barrier.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
/*
 * Copyright (c) 2012      Mellanox Technologies, Inc.
 *                         All rights reserved.
 * $COPYRIGHT$
 * 
 * Additional copyrights may follow
 * 
 * $HEADER$
 */

#include "oshmem_config.h"
#include <stdio.h>
#include <stdlib.h>

#include "orte/mca/grpcomm/grpcomm.h"

#include "oshmem/constants.h"
#include "oshmem/mca/spml/spml.h"
#include "oshmem/mca/scoll/scoll.h"
#include "oshmem/mca/scoll/base/base.h"
#include "oshmem/proc/proc.h"
#include "scoll_basic.h"


static int __algorithm_central_counter(struct oshmem_group_t *group, long *pSync);
static int __algorithm_tournament(struct oshmem_group_t *group, long *pSync);
static int __algorithm_recursive_doubling(struct oshmem_group_t *group, long *pSync);
static int __algorithm_dissemination(struct oshmem_group_t *group, long *pSync);
static int __algorithm_basic(struct oshmem_group_t *group, long *pSync);
static int __algorithm_adaptive(struct oshmem_group_t *group, long *pSync);


int mca_scoll_basic_barrier(struct oshmem_group_t *group, long *pSync, int alg)
{
    int rc = OSHMEM_SUCCESS;

    /* Arguments validation */
    if (!group)
    {
        SCOLL_ERROR("Active set (group) of PE is not defined");
        rc = OSHMEM_ERR_BAD_PARAM;
    }

    if ((rc == OSHMEM_SUCCESS) && oshmem_proc_group_is_member(group))
    {
        if (pSync)
        {
            alg = ( alg == SCOLL_DEFAULT_ALG ? mca_scoll_basic_param_barrier_algorithm : alg);
            switch(alg)
            {
                case SCOLL_ALG_BARRIER_CENTRAL_COUNTER:
                {
                    rc = __algorithm_central_counter(group, pSync);
                    break;
                }
                case SCOLL_ALG_BARRIER_TOURNAMENT:
                {
                    rc = __algorithm_tournament(group, pSync);
                    break;
                }
                case SCOLL_ALG_BARRIER_RECURSIVE_DOUBLING:
                {
                    rc = __algorithm_recursive_doubling(group, pSync);
                    break;
                }
                case SCOLL_ALG_BARRIER_DISSEMINATION:
                {
                    rc = __algorithm_dissemination(group, pSync);
                    break;
                }
                case SCOLL_ALG_BARRIER_BASIC:
                {
                    rc = __algorithm_basic(group, pSync);
                    break;
                }
                case SCOLL_ALG_BARRIER_ADAPTIVE:
                {
                    rc = __algorithm_adaptive(group, pSync);
                    break;
                }
                default:
                {
                    rc = __algorithm_recursive_doubling(group, pSync);
                }
            }
        }
        else
        {
            SCOLL_ERROR("Incorrect argument pSync");
            rc = OSHMEM_ERR_BAD_PARAM;
        }
    }

    return rc;
}


/*
    This algorithm is quite simple and straightforward. But because of it�s obvious simplicity and
    the naive prove for correctness it is implemented quite often. One node asks peers if they are
    achieve barrier state. When all processors are ready it signals to go ahead.
    Outlay:
    NP-1 competing network transfers are needed to implement the counter 
    The memory usage is constant (1 byte) per node.
*/
static int __algorithm_central_counter(struct oshmem_group_t *group, long *pSync)
{
    int rc = OSHMEM_SUCCESS;
    long value = SHMEM_SYNC_INIT;
    int root_id = 0;
    int PE_root = oshmem_proc_pe(group->proc_array[root_id]);
    int i = 0;

    SCOLL_VERBOSE(12, "[#%d] Barrier algorithm: Central Counter", group->my_pe);
    SCOLL_VERBOSE(15, "[#%d] pSync[0] = %ld", group->my_pe, pSync[0]);

    /* Set current state as WAIT */
    pSync[0] = SHMEM_SYNC_WAIT;

    /* Root processes synchronization */
    if (PE_root == group->my_pe)
    {
        int pe_cur = 0;
        long wait_pe_count = 0;
        int* wait_pe_array = NULL;

        wait_pe_array = malloc(sizeof(*wait_pe_array) * group->proc_count);
        if (wait_pe_array)
        {
            SCOLL_VERBOSE(14, "[#%d] PE is the root", group->my_pe);

            wait_pe_count = group->proc_count;
            for (i = 0; i < group->proc_count; i++) 
            {
                wait_pe_array[i] = oshmem_proc_pe(group->proc_array[i]);
            }
            wait_pe_array[root_id] = OSHMEM_PE_INVALID;
            wait_pe_count--;

            while (wait_pe_count)
            {
                for (i = 0; (i < group->proc_count) && (rc == OSHMEM_SUCCESS); i++) 
                {
                    pe_cur = wait_pe_array[i];
                    if (pe_cur != OSHMEM_PE_INVALID)
                    {
                        rc = MCA_SPML_CALL(get((void*)pSync, sizeof(value), (void*)&value, pe_cur));
                        if ( (rc == OSHMEM_SUCCESS) && (value == SHMEM_SYNC_WAIT) )
                        {
                            wait_pe_array[i] = OSHMEM_PE_INVALID;
                            wait_pe_count--;
                            SCOLL_VERBOSE(14, "[#%d] PE#%d is ready (wait list counter: %d)", group->my_pe, pe_cur, (int)wait_pe_count);
                        }
                    }
                }
            }

            SCOLL_VERBOSE(14, "[#%d] PE signals to all", group->my_pe);
            value = SHMEM_SYNC_RUN;
            for (i = 0; (i < group->proc_count) && (rc == OSHMEM_SUCCESS); i++) 
            {
                pe_cur = oshmem_proc_pe(group->proc_array[i]);
                if (pe_cur != PE_root)
                {
                    rc = MCA_SPML_CALL(put((void*)pSync, sizeof(value), (void*)&value, pe_cur));
                }
            }

            free(wait_pe_array);
        }
        else
        {
            rc = OSHMEM_ERR_OUT_OF_RESOURCE;
        }

        /* Possibly this is unnecessary...
           But imagine the scenario when you have 2 sequential barriers and the root PE is the fastest one.
           The root could leave the first barrier and in the second barrier it could get SHMEM_SYNC_WAIT value on
           remote node before the remote node receives its SHMEM_SYNC_RUN value in the first barrier
        */
        /* TODO: actually it must be quiet */
        MCA_SPML_CALL(fence());
    }
    /* Wait for RUN signal */
    else
    {
        SCOLL_VERBOSE(14, "[#%d] PE waits for a signal from root", group->my_pe);

        value = SHMEM_SYNC_RUN;
        rc = MCA_SPML_CALL(wait((void*)pSync, SHMEM_CMP_EQ, (void*)&value, SHMEM_LONG));
    }

    /* Restore initial values */
    SCOLL_VERBOSE(12, "[#%d] Restore special synchronization array", group->my_pe);
    for (i = 0; pSync && (i < _SHMEM_BARRIER_SYNC_SIZE); i++) 
    {
        pSync[i] = _SHMEM_SYNC_VALUE;
    }

    SCOLL_VERBOSE(15, "[#%d] pSync[0] = %ld", group->my_pe, pSync[0]);

    return rc;
}


/*
    The Tournament Barrier, proposed by Hengsen, Finkel and Manser is mostly suitable for shared memory
    multiprocessors because it benefits from several caching mechanisms.
    The algorithm is similar to a tournament game. In each round two
    nodes play against each other. The winner is known in advance and waits until the looser arrives. The
    winners play against each other in the next round. The overall winner (the champion) notifies all others
    about the end of the barrier.
    Outlay:
    The game scales with log2(NP) and uses 1 byte of memory.
*/
static int __algorithm_tournament(struct oshmem_group_t *group, long *pSync)
{
    int rc = OSHMEM_SUCCESS;
    int round = 0;
    int exit_flag = group->proc_count - 1;
    long value = SHMEM_SYNC_INIT;
    int my_id = oshmem_proc_group_find_id(group, group->my_pe);
    int peer_id = 0;
    int peer_pe = 0;
    int i = 0;

    SCOLL_VERBOSE(12, "[#%d] Barrier algorithm: Tournament", group->my_pe);
    SCOLL_VERBOSE(15, "[#%d] pSync[0] = %ld", group->my_pe, pSync[0]);

    /* Set current state as WAIT */
    pSync[0] = SHMEM_SYNC_WAIT;

    while (exit_flag && (rc == OSHMEM_SUCCESS))
    {
        /* Define a peer for competition */
        peer_id = my_id ^ (1 << round);

        /* Update exit condition and round counter */
        exit_flag >>= 1;
        round++;

        /* Do not have peer for tournament */
        if (peer_id >= group->proc_count) continue;

        if ( my_id < peer_id )
        {
            pSync[0] = peer_id;
            value = my_id;

            SCOLL_VERBOSE(14, "[#%d] round = %d wait", group->my_pe, round);
            rc = MCA_SPML_CALL(wait((void*)pSync, SHMEM_CMP_EQ, (void*)&value, SHMEM_LONG));
        }
        else
        {
            peer_pe = oshmem_proc_pe(group->proc_array[peer_id]);

#if 1 /* It is ugly implementation of compare and swap operation 
         Usage of this hack does not give performance improvement but
         it is expected that shmem_long_cswap() will make it faster.
       */
            do
            {
                MCA_SPML_CALL(get((void*)pSync, sizeof(value), (void*)&value, peer_pe));
            } while (value != my_id);

            SCOLL_VERBOSE(14, "[#%d] round = %d signals to #%d", group->my_pe, round, peer_pe);
            value = peer_id;
            rc = MCA_SPML_CALL(put((void*)pSync, sizeof(value), (void*)&value, peer_pe));
#else
            SCOLL_VERBOSE(14, "[#%d] round = %d signals to #%d", group->my_pe, round, peer_pe);
            do
            {
                rc = MCA_ATOMIC_CALL(cswap((void*)pSync, (void*)&value, (const void*)&my_id, (const void*)&peer_id, sizeof(value), peer_pe));
            } while (value != my_id);
#endif
            SCOLL_VERBOSE(14, "[#%d] round = %d wait", group->my_pe, round);
            value = SHMEM_SYNC_RUN;
            rc = MCA_SPML_CALL(wait((void*)pSync, SHMEM_CMP_EQ, (void*)&value, SHMEM_LONG));

            break;
        }
    }

    /* Restore initial values */
    SCOLL_VERBOSE(12, "[#%d] Restore special synchronization array", group->my_pe);
    for (i = 0; pSync && (i < _SHMEM_BARRIER_SYNC_SIZE); i++) 
    {
        pSync[i] = _SHMEM_SYNC_VALUE;
    }

    /* Send result to all PE in group */
    if ( (my_id == 0) && (rc == OSHMEM_SUCCESS))
    {
        SCOLL_VERBOSE(14, "[#%d] signals to all", group->my_pe);

        value = SHMEM_SYNC_RUN;
        for (peer_id = 1; (peer_id < group->proc_count) && (rc == OSHMEM_SUCCESS); peer_id++) 
        {
            peer_pe = oshmem_proc_pe(group->proc_array[peer_id]);
            rc = MCA_SPML_CALL(put((void*)pSync, sizeof(value), (void*)&value, peer_pe));
        }
    }

    SCOLL_VERBOSE(15, "[#%d] pSync[0] = %ld", group->my_pe, pSync[0]);

    return rc;
}


/*
    Pairwise Exchange With Recursive Doubling.
    Rinka Gupta, Vinod Tipparaju, Jare Nieplocha, and Dhabaleswar Panda. Efficient Barrier
    using Remote Memory Operations on VIA-Based Clusters. In 2002 IEEE International
    Conference on Cluster Computing (CLUSTER 2002), page 83. IEEE Computer Society, 2002.
    Outlay:
    The algorithm uses a maximum of log2(NP) + 2 network writes and P bytes memory per node.
*/
static int __algorithm_recursive_doubling(struct oshmem_group_t *group, long *pSync)
{
    int rc = OSHMEM_SUCCESS;
    int round = 0;
    int floor2_proc = 0;
    int exit_flag = 0;
    long value = SHMEM_SYNC_INIT;
    int my_id = oshmem_proc_group_find_id(group, group->my_pe);
    int peer_id = 0;
    int peer_pe = 0;
    int i = 0;

    floor2_proc = 1;
    i = group->proc_count;
    i >>= 1;
    while (i)
    {
        i >>= 1;
        floor2_proc <<= 1;
    }

    SCOLL_VERBOSE(12, "[#%d] Barrier algorithm: Recursive Doubling", group->my_pe);
    SCOLL_VERBOSE(15, "[#%d] pSync[0] = %ld floor2_proc = %d", group->my_pe, pSync[0], floor2_proc);

    if (my_id >= floor2_proc)
    {
        /* I am in extra group, my partner is node (my_id-y) in basic group */
        peer_id = my_id - floor2_proc;
        peer_pe = oshmem_proc_pe(group->proc_array[peer_id]);

        SCOLL_VERBOSE(14, "[#%d] is extra and signal to #%d", group->my_pe, peer_pe);
        value = SHMEM_SYNC_WAIT;
        rc = MCA_SPML_CALL(put((void*)pSync, sizeof(value), (void*)&value, peer_pe));

        SCOLL_VERBOSE(14, "[#%d] wait", group->my_pe);
        value = SHMEM_SYNC_RUN;
        rc = MCA_SPML_CALL(wait((void*)pSync, SHMEM_CMP_EQ, (void*)&value, SHMEM_LONG));

        /* Restore initial values */
        SCOLL_VERBOSE(12, "[#%d] Restore special synchronization array", group->my_pe);
        for (i = 0; pSync && (i < _SHMEM_BARRIER_SYNC_SIZE); i++) 
        {
            pSync[i] = _SHMEM_SYNC_VALUE;
        }
    }
    else
    {
        /* Wait for a peer from extra group */
        if ((group->proc_count - floor2_proc) > my_id)
        {
            /* I am in basic group, my partner is node (my_id+y) in extra group */
            peer_id = my_id + floor2_proc;
            peer_pe = oshmem_proc_pe(group->proc_array[peer_id]);

            SCOLL_VERBOSE(14, "[#%d] wait a signal from #%d", group->my_pe, peer_pe);
            value = SHMEM_SYNC_WAIT;
            rc = MCA_SPML_CALL(wait((void*)pSync, SHMEM_CMP_EQ, (void*)&value, SHMEM_LONG));
        }

        /* Pairwise exchange  */
        exit_flag = floor2_proc - 1;
        pSync[0] = round;
        while (exit_flag && (rc == OSHMEM_SUCCESS))
        {
            /* Define a peer for competition */
            peer_id = my_id ^ (1 << round);

            /* Update exit condition and round counter */
            exit_flag >>= 1;
            round++;

            peer_pe = oshmem_proc_pe(group->proc_array[peer_id]);

#if 1 /* It is ugly implementation of compare and swap operation 
         Usage of this hack does not give performance improvement but
         it is expected that shmem_long_cswap() will make it faster.
       */
            do
            {
                MCA_SPML_CALL(get((void*)pSync, sizeof(value), (void*)&value, peer_pe));
            } while (value != (round - 1));

            SCOLL_VERBOSE(14, "[#%d] round = %d signals to #%d", group->my_pe, round, peer_pe);
            value = round;
            rc = MCA_SPML_CALL(put((void*)pSync, sizeof(value), (void*)&value, peer_pe));
#else
            SCOLL_VERBOSE(14, "[#%d] round = %d signals to #%d", group->my_pe, round, peer_pe);
            {
                long cond = round - 1;
                do
                {
                    rc = MCA_ATOMIC_CALL(cswap((void*)pSync, (void*)&value, (const void*)&cond, (const void*)&round, sizeof(value), peer_pe));
                } while (value != (round-1));
            }
#endif

            SCOLL_VERBOSE(14, "[#%d] round = %d wait", group->my_pe, round);
            value = round;
            rc = MCA_SPML_CALL(wait((void*)pSync, SHMEM_CMP_GE, (void*)&value, SHMEM_LONG));
        }

        /* Restore initial values */
        SCOLL_VERBOSE(12, "[#%d] Restore special synchronization array", group->my_pe);
        for (i = 0; pSync && (i < _SHMEM_BARRIER_SYNC_SIZE); i++) 
        {
            pSync[i] = _SHMEM_SYNC_VALUE;
        }

        /* Notify a peer from extra group */
        if ((group->proc_count - floor2_proc) > my_id)
        {
            /* I am in basic group, my partner is node (my_id+y) in extra group */
            peer_id = my_id + floor2_proc;
            peer_pe = oshmem_proc_pe(group->proc_array[peer_id]);

            SCOLL_VERBOSE(14, "[#%d] signals to #%d", group->my_pe, peer_pe);
            value = SHMEM_SYNC_RUN;
            rc = MCA_SPML_CALL(put((void*)pSync, sizeof(value), (void*)&value, peer_pe));
        }
    }

    SCOLL_VERBOSE(15, "[#%d] pSync[0] = %ld", group->my_pe, pSync[0]);

    return rc;
}


/*
    The Dissemination Barrier, introduced by Hengsen, Finkel and Manser in 1998.
    The algorithm is mostly an improvement of the Butterfly Barrier for non power of two processor counts. 
    It uses the same pairwise synchronization but with other partners.
    Outlay:
    The game scales with log2(NP) and uses 1 byte of memory.
*/
static int __algorithm_dissemination(struct oshmem_group_t *group, long *pSync)
{
    int rc = OSHMEM_SUCCESS;
    int round = 0;
    int log2_proc = 0;
    long value = SHMEM_SYNC_INIT;
    int my_id = oshmem_proc_group_find_id(group, group->my_pe);
    int peer_id = 0;
    int peer_pe = 0;
    int i = 0;

    log2_proc = scoll_log2((unsigned long)group->proc_count);

    SCOLL_VERBOSE(12, "[#%d] Barrier algorithm: Dissemination", group->my_pe);
    SCOLL_VERBOSE(15, "[#%d] pSync[0] = %ld floor2_proc = %d", group->my_pe, pSync[0], log2_proc);

    pSync[0] = round;
    for (round = 0; (round <= log2_proc) && (rc == OSHMEM_SUCCESS); round++)
    {
        /* Define a peer to send signal */
        peer_id = (my_id + (1 << round)) % group->proc_count;

        peer_pe = oshmem_proc_pe(group->proc_array[peer_id]);

#if 1 /* It is ugly implementation of compare and swap operation 
         Usage of this hack does not give performance improvement but
         it is expected that shmem_long_cswap() will make it faster.
       */
        do
        {
            MCA_SPML_CALL(get((void*)pSync, sizeof(value), (void*)&value, peer_pe));
        } while (value != round);

        SCOLL_VERBOSE(14, "[#%d] round = %d signals to #%d", group->my_pe, round, peer_pe);
        value = round + 1;
        rc = MCA_SPML_CALL(put((void*)pSync, sizeof(value), (void*)&value, peer_pe));
#endif

        SCOLL_VERBOSE(14, "[#%d] round = %d wait", group->my_pe, round);
        value = round + 1;
        rc = MCA_SPML_CALL(wait((void*)pSync, SHMEM_CMP_GE, (void*)&value, SHMEM_LONG));
    }

    /* Restore initial values */
    SCOLL_VERBOSE(12, "[#%d] Restore special synchronization array", group->my_pe);
    for (i = 0; pSync && (i < _SHMEM_BARRIER_SYNC_SIZE); i++) 
    {
        pSync[i] = _SHMEM_SYNC_VALUE;
    }

    SCOLL_VERBOSE(15, "[#%d] pSync[0] = %ld", group->my_pe, pSync[0]);

    return rc;
}


static int __algorithm_basic(struct oshmem_group_t *group, long *pSync)
{
    int rc = OSHMEM_SUCCESS;
    int root_id = 0;
    int PE_root = oshmem_proc_pe(group->proc_array[root_id]);
    int i = 0;

    SCOLL_VERBOSE(12, "[#%d] Barrier algorithm: Basic", group->my_pe);

    if (PE_root != group->my_pe)
    {
        rc = MCA_SPML_CALL(send(NULL, 0, PE_root, MCA_SPML_BASE_PUT_STANDARD));
        if (OSHMEM_SUCCESS != rc) {
            return rc;
        }

        rc = MCA_SPML_CALL(recv(NULL, 0, PE_root));
        if (OSHMEM_SUCCESS != rc) {
            return rc;
        }
    }

    /* The root collects and broadcasts the messages. */

    else 
    {
        int pe_cur = 0;

        for (i = 0; (i < group->proc_count) && (rc == OSHMEM_SUCCESS); i++) 
        {
            pe_cur = oshmem_proc_pe(group->proc_array[i]);
            if (pe_cur != PE_root)
            {
                rc = MCA_SPML_CALL(recv(NULL, 0, SHMEM_ANY_SOURCE));
            }
            if (OSHMEM_SUCCESS != rc) {
                return rc;
            }
        }

        for (i = 0; (i < group->proc_count) && (rc == OSHMEM_SUCCESS); i++) 
        {
            pe_cur = oshmem_proc_pe(group->proc_array[i]);
            if (pe_cur != PE_root)
            {
                rc = MCA_SPML_CALL(send(NULL, 0, pe_cur, MCA_SPML_BASE_PUT_STANDARD));
            }
            if (OSHMEM_SUCCESS != rc) {
                return rc;
            }
        }
    }

    return rc;
}

static int __algorithm_adaptive(struct oshmem_group_t *group, long *pSync)
{
    int rc = OSHMEM_SUCCESS;
    bool local_peers_only = true;

    SCOLL_VERBOSE(12, "[#%d] Barrier algorithm: Adaptive", group->my_pe);

    /* check if we have only local peers */
    {
        int i = 0;

        for (i = 0; i < group->proc_count; i++)
        {
            if (i == group->id) continue;

            if ( !OPAL_PROC_ON_LOCAL_NODE(group->proc_array[i]->proc_flags) )
            {
                local_peers_only = false;
                break;
            }
        }
    }

    /* Select algorithm we use:
     * use send/recv way for group in the same node and for np < 32
     * otherwise use put/get way
     */
    if (local_peers_only || (group->proc_count < 32) )
    {
        rc = __algorithm_basic(group, pSync);
    }
    else
    {
        rc = __algorithm_recursive_doubling(group, pSync);
    }

    return rc;
}
Tip: Filter by directory path e.g. /media app.js to search for public/media/app.js.
Tip: Use camelCasing e.g. ProjME to search for ProjectModifiedEvent.java.
Tip: Filter by extension type e.g. /repo .js to search for all .js files in the /repo directory.
Tip: Separate your search with spaces e.g. /ssh pom.xml to search for src/ssh/pom.xml.
Tip: Use ↑ and ↓ arrow keys to navigate and return to view the file.
Tip: You can also navigate files with Ctrl+j (next) and Ctrl+k (previous) and view the file with Ctrl+o.
Tip: You can also navigate files with Alt+j (next) and Alt+k (previous) and view the file with Alt+o.