1 /* Copyright (c) 2007-2014. The SimGrid Team.
2 * All rights reserved. */
4 /* This program is free software; you can redistribute it and/or modify it
5 * under the terms of the license (GNU LGPL) which comes with this package. */
10 #include "xbt/replay.h"
12 #include "simix/smx_private.h"
13 #include "surf/surf.h"
14 #include "simgrid/sg_config.h"
15 #include "colls/colls.h"
17 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_base, smpi, "Logging specific to SMPI (base)");
20 static int match_recv(void* a, void* b, smx_action_t ignored) {
21 MPI_Request ref = (MPI_Request)a;
22 MPI_Request req = (MPI_Request)b;
23 XBT_DEBUG("Trying to match a recv of src %d against %d, tag %d against %d",ref->src,req->src, ref->tag, req->tag);
25 xbt_assert(ref, "Cannot match recv against null reference");
26 xbt_assert(req, "Cannot match recv against null request");
27 if((ref->src == MPI_ANY_SOURCE || req->src == ref->src)
28 && ((ref->tag == MPI_ANY_TAG && req->tag >=0) || req->tag == ref->tag)){
29 //we match, we can transfer some values
30 // FIXME : move this to the copy function ?
31 if(ref->src == MPI_ANY_SOURCE)ref->real_src = req->src;
32 if(ref->tag == MPI_ANY_TAG)ref->real_tag = req->tag;
33 if(ref->real_size < req->real_size) ref->truncated = 1;
35 ref->detached_sender=req; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
37 XBT_DEBUG("match succeeded");
42 static int match_send(void* a, void* b,smx_action_t ignored) {
43 MPI_Request ref = (MPI_Request)a;
44 MPI_Request req = (MPI_Request)b;
45 XBT_DEBUG("Trying to match a send of src %d against %d, tag %d against %d",ref->src,req->src, ref->tag, req->tag);
46 xbt_assert(ref, "Cannot match send against null reference");
47 xbt_assert(req, "Cannot match send against null request");
49 if((req->src == MPI_ANY_SOURCE || req->src == ref->src)
50 && ((req->tag == MPI_ANY_TAG && ref->tag >=0)|| req->tag == ref->tag))
52 if(req->src == MPI_ANY_SOURCE)req->real_src = ref->src;
53 if(req->tag == MPI_ANY_TAG)req->real_tag = ref->tag;
54 if(req->real_size < ref->real_size) req->truncated = 1;
56 req->detached_sender=ref; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
58 XBT_DEBUG("match succeeded");
64 typedef struct s_smpi_factor *smpi_factor_t;
65 typedef struct s_smpi_factor {
68 double values[4];//arbitrary set to 4
70 xbt_dynar_t smpi_os_values = NULL;
71 xbt_dynar_t smpi_or_values = NULL;
72 xbt_dynar_t smpi_ois_values = NULL;
74 // Methods used to parse and store the values for timing injections in smpi
75 // These are taken from surf/network.c and generalized to have more factors
76 // These methods should be merged with those in surf/network.c (moved somewhere in xbt ?)
78 static int factor_cmp(const void *pa, const void *pb)
80 return (((s_smpi_factor_t*)pa)->factor > ((s_smpi_factor_t*)pb)->factor) ? 1 :
81 (((s_smpi_factor_t*)pa)->factor < ((s_smpi_factor_t*)pb)->factor) ? -1 : 0;
85 static xbt_dynar_t parse_factor(const char *smpi_coef_string)
88 unsigned int iter = 0;
92 xbt_dynar_t smpi_factor, radical_elements, radical_elements2 = NULL;
94 smpi_factor = xbt_dynar_new(sizeof(s_smpi_factor_t), NULL);
95 radical_elements = xbt_str_split(smpi_coef_string, ";");
96 xbt_dynar_foreach(radical_elements, iter, value) {
97 memset(&fact, 0, sizeof(s_smpi_factor_t));
98 radical_elements2 = xbt_str_split(value, ":");
99 if (xbt_dynar_length(radical_elements2) <2 || xbt_dynar_length(radical_elements2) > 5)
100 xbt_die("Malformed radical for smpi factor!");
101 for(i =0; i<xbt_dynar_length(radical_elements2);i++ ){
103 fact.factor = atol(xbt_dynar_get_as(radical_elements2, i, char *));
105 fact.values[fact.nb_values] = atof(xbt_dynar_get_as(radical_elements2, i, char *));
110 xbt_dynar_push_as(smpi_factor, s_smpi_factor_t, fact);
111 XBT_DEBUG("smpi_factor:\t%ld : %d values, first: %f", fact.factor, fact.nb_values ,fact.values[0]);
112 xbt_dynar_free(&radical_elements2);
114 xbt_dynar_free(&radical_elements);
116 xbt_dynar_sort(smpi_factor, &factor_cmp);
117 xbt_dynar_foreach(smpi_factor, iter, fact) {
118 XBT_DEBUG("smpi_factor:\t%ld : %d values, first: %f", fact.factor, fact.nb_values ,fact.values[0]);
123 static double smpi_os(double size)
125 if (!smpi_os_values) {
126 smpi_os_values = parse_factor(sg_cfg_get_string("smpi/os"));
127 smpi_register_static(smpi_os_values, xbt_dynar_free_voidp);
129 unsigned int iter = 0;
130 s_smpi_factor_t fact;
132 xbt_dynar_foreach(smpi_os_values, iter, fact) {
133 if (size <= fact.factor) {
134 XBT_DEBUG("os : %f <= %ld return %f", size, fact.factor, current);
137 current=fact.values[0]+fact.values[1]*size;
140 XBT_DEBUG("os : %f > %ld return %f", size, fact.factor, current);
145 static double smpi_ois(double size)
147 if (!smpi_ois_values) {
148 smpi_ois_values = parse_factor(sg_cfg_get_string("smpi/ois"));
149 smpi_register_static(smpi_ois_values, xbt_dynar_free_voidp);
151 unsigned int iter = 0;
152 s_smpi_factor_t fact;
154 xbt_dynar_foreach(smpi_ois_values, iter, fact) {
155 if (size <= fact.factor) {
156 XBT_DEBUG("ois : %f <= %ld return %f", size, fact.factor, current);
159 current=fact.values[0]+fact.values[1]*size;
162 XBT_DEBUG("ois : %f > %ld return %f", size, fact.factor, current);
167 static double smpi_or(double size)
169 if (!smpi_or_values) {
170 smpi_or_values = parse_factor(sg_cfg_get_string("smpi/or"));
171 smpi_register_static(smpi_or_values, xbt_dynar_free_voidp);
173 unsigned int iter = 0;
174 s_smpi_factor_t fact;
176 xbt_dynar_foreach(smpi_or_values, iter, fact) {
177 if (size <= fact.factor) {
178 XBT_DEBUG("or : %f <= %ld return %f", size, fact.factor, current);
181 current=fact.values[0]+fact.values[1]*size;
183 XBT_DEBUG("or : %f > %ld return %f", size, fact.factor, current);
188 static MPI_Request build_request(void *buf, int count,
189 MPI_Datatype datatype, int src, int dst,
190 int tag, MPI_Comm comm, unsigned flags)
192 MPI_Request request = NULL;
194 void *old_buf = NULL;
196 request = xbt_new(s_smpi_mpi_request_t, 1);
198 s_smpi_subtype_t *subtype = datatype->substruct;
200 if(datatype->has_subtype == 1){
201 // This part handles the problem of non-contiguous memory
203 buf = count==0 ? NULL : xbt_malloc(count*smpi_datatype_size(datatype));
205 subtype->serialize(old_buf, buf, count, datatype->substruct);
210 // This part handles the problem of non-contiguous memory (for the
211 // unserialisation at the reception)
212 request->old_buf = old_buf;
213 request->old_type = datatype;
215 request->size = smpi_datatype_size(datatype) * count;
219 request->comm = comm;
220 request->action = NULL;
221 request->flags = flags;
222 request->detached = 0;
223 request->detached_sender = NULL;
224 request->real_src = 0;
226 request->truncated = 0;
227 request->real_size = 0;
228 request->real_tag = 0;
229 if(flags & PERSISTENT)
230 request->refcount = 1;
232 request->refcount = 0;
238 if (flags & SEND) smpi_datatype_unuse(datatype);
244 void smpi_empty_status(MPI_Status * status)
246 if(status != MPI_STATUS_IGNORE) {
247 status->MPI_SOURCE = MPI_ANY_SOURCE;
248 status->MPI_TAG = MPI_ANY_TAG;
249 status->MPI_ERROR = MPI_SUCCESS;
254 void smpi_action_trace_run(char *path)
258 xbt_dict_cursor_t cursor;
262 action_fp = fopen(path, "r");
263 xbt_assert(action_fp != NULL, "Cannot open %s: %s", path,
267 if (!xbt_dict_is_empty(action_queues)) {
269 ("Not all actions got consumed. If the simulation ended successfully (without deadlock), you may want to add new processes to your deployment file.");
272 xbt_dict_foreach(action_queues, cursor, name, todo) {
273 XBT_WARN("Still %lu actions for %s", xbt_dynar_length(todo), name);
279 xbt_dict_free(&action_queues);
280 action_queues = xbt_dict_new_homogeneous(NULL);
283 static void smpi_mpi_request_free_voidp(void* request)
285 MPI_Request req = request;
286 smpi_mpi_request_free(&req);
289 /* MPI Low level calls */
290 MPI_Request smpi_mpi_send_init(void *buf, int count, MPI_Datatype datatype,
291 int dst, int tag, MPI_Comm comm)
293 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
294 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
295 comm, PERSISTENT | SEND | PREPARED);
299 MPI_Request smpi_mpi_ssend_init(void *buf, int count, MPI_Datatype datatype,
300 int dst, int tag, MPI_Comm comm)
302 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
303 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
304 comm, PERSISTENT | SSEND | SEND | PREPARED);
308 MPI_Request smpi_mpi_recv_init(void *buf, int count, MPI_Datatype datatype,
309 int src, int tag, MPI_Comm comm)
311 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
312 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag,
313 comm, PERSISTENT | RECV | PREPARED);
317 void smpi_mpi_start(MPI_Request request)
321 xbt_assert(!request->action, "Cannot (re)start a non-finished communication");
322 request->flags &= ~PREPARED;
323 request->flags &= ~FINISHED;
326 if (request->flags & RECV) {
327 print_request("New recv", request);
328 //FIXME: if receive is posted with a large size, but send is smaller, mailboxes may not match !
329 if (request->size < sg_cfg_get_int("smpi/async_small_thres"))
330 mailbox = smpi_process_mailbox_small();
332 mailbox = smpi_process_mailbox();
333 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
334 request->real_size=request->size;
335 smpi_datatype_use(request->old_type);
336 smpi_comm_use(request->comm);
337 request->action = simcall_comm_irecv(mailbox, request->buf,
338 &request->real_size, &match_recv,
341 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
342 double sleeptime = request->detached ? smpi_or(request->size) : 0.0;
344 simcall_process_sleep(sleeptime);
345 XBT_DEBUG("receiving size of %zu : sleep %f ", request->size, smpi_or(request->size));
351 int receiver = request->dst;//smpi_group_index(smpi_comm_group(request->comm), request->dst);
354 int rank = smpi_process_index();
355 if (TRACE_smpi_view_internals()) {
356 TRACE_smpi_send(rank, rank, receiver,request->size);
359 /* if(receiver == MPI_UNDEFINED) {*/
360 /* XBT_WARN("Trying to send a message to a wrong rank");*/
363 print_request("New send", request);
364 if (request->size < sg_cfg_get_int("smpi/async_small_thres")) { // eager mode
365 mailbox = smpi_process_remote_mailbox_small(receiver);
367 XBT_DEBUG("Send request %p is not in the permanent receive mailbox (buf: %p)",request,request->buf);
368 mailbox = smpi_process_remote_mailbox(receiver);
371 void* buf = request->buf;
372 if ( (! (request->flags & SSEND)) && (request->size < sg_cfg_get_int("smpi/send_is_detached_thres"))) {
374 request->detached = 1;
376 if(request->old_type->has_subtype == 0){
377 oldbuf = request->buf;
378 if (!_xbt_replay_is_active() && oldbuf && request->size!=0){
379 buf = xbt_malloc(request->size);
380 memcpy(buf,oldbuf,request->size);
383 XBT_DEBUG("Send request %p is detached; buf %p copied into %p",request,oldbuf,request->buf);
386 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
387 request->real_size=request->size;
388 smpi_datatype_use(request->old_type);
389 smpi_comm_use(request->comm);
391 //if we are giving back the control to the user without waiting for completion, we have to inject timings
392 double sleeptime = 0.0;
393 if(request->detached || (request->flags & (ISEND|SSEND))){// issend should be treated as isend
394 //isend and send timings may be different
395 sleeptime = (request->flags & ISEND)? smpi_ois(request->size) : smpi_os(request->size);
398 if(sleeptime != 0.0){
399 simcall_process_sleep(sleeptime);
400 XBT_DEBUG("sending size of %zu : sleep %f ", request->size, smpi_os(request->size));
404 simcall_comm_isend(mailbox, request->size, -1.0,
405 buf, request->real_size,
407 &xbt_free, // how to free the userdata if a detached send fails
409 // detach if msg size < eager/rdv switch limit
413 /* FIXME: detached sends are not traceable (request->action == NULL) */
415 simcall_set_category(request->action, TRACE_internal_smpi_get_category());
423 void smpi_mpi_startall(int count, MPI_Request * requests)
426 if(requests==NULL) return;
428 for(i = 0; i < count; i++) {
429 smpi_mpi_start(requests[i]);
433 void smpi_mpi_request_free(MPI_Request * request)
435 if((*request) != MPI_REQUEST_NULL){
436 (*request)->refcount--;
437 if((*request)->refcount<0) xbt_die("wrong refcount");
439 if((*request)->refcount==0){
440 print_request("Destroying", (*request));
442 *request = MPI_REQUEST_NULL;
444 print_request("Decrementing", (*request));
447 xbt_die("freeing an already free request");
451 MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype,
452 int dst, int tag, MPI_Comm comm)
454 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
455 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf , count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
456 comm, PERSISTENT | ISEND | SEND | PREPARED);
460 MPI_Request smpi_mpi_isend(void *buf, int count, MPI_Datatype datatype,
461 int dst, int tag, MPI_Comm comm)
463 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
464 request = build_request(buf==MPI_BOTTOM?(void*)0:buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
465 comm, NON_PERSISTENT | ISEND | SEND);
466 smpi_mpi_start(request);
470 MPI_Request smpi_mpi_issend(void *buf, int count, MPI_Datatype datatype,
471 int dst, int tag, MPI_Comm comm)
473 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
474 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
475 comm, NON_PERSISTENT | ISEND | SSEND | SEND);
476 smpi_mpi_start(request);
482 MPI_Request smpi_irecv_init(void *buf, int count, MPI_Datatype datatype,
483 int src, int tag, MPI_Comm comm)
485 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
486 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag,
487 comm, PERSISTENT | RECV | PREPARED);
491 MPI_Request smpi_mpi_irecv(void *buf, int count, MPI_Datatype datatype,
492 int src, int tag, MPI_Comm comm)
494 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
495 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag,
496 comm, NON_PERSISTENT | RECV);
497 smpi_mpi_start(request);
501 void smpi_mpi_recv(void *buf, int count, MPI_Datatype datatype, int src,
502 int tag, MPI_Comm comm, MPI_Status * status)
504 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
505 request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm);
506 smpi_mpi_wait(&request, status);
512 void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst,
513 int tag, MPI_Comm comm)
515 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
516 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
517 comm, NON_PERSISTENT | SEND);
519 smpi_mpi_start(request);
520 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
524 void smpi_mpi_ssend(void *buf, int count, MPI_Datatype datatype,
525 int dst, int tag, MPI_Comm comm)
527 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
528 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
529 comm, NON_PERSISTENT | SSEND | SEND);
531 smpi_mpi_start(request);
532 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
536 void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
537 int dst, int sendtag, void *recvbuf, int recvcount,
538 MPI_Datatype recvtype, int src, int recvtag,
539 MPI_Comm comm, MPI_Status * status)
541 MPI_Request requests[2];
543 int myid=smpi_process_index();
544 if ((smpi_group_index(smpi_comm_group(comm), dst) == myid) && (smpi_group_index(smpi_comm_group(comm), src) == myid)) {
545 smpi_datatype_copy(sendbuf, sendcount, sendtype,
546 recvbuf, recvcount, recvtype);
550 smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
552 smpi_irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
553 smpi_mpi_startall(2, requests);
554 smpi_mpi_waitall(2, requests, stats);
555 smpi_mpi_request_free(&requests[0]);
556 smpi_mpi_request_free(&requests[1]);
557 if(status != MPI_STATUS_IGNORE) {
558 // Copy receive status
563 int smpi_mpi_get_count(MPI_Status * status, MPI_Datatype datatype)
565 return status->count / smpi_datatype_size(datatype);
568 static void finish_wait(MPI_Request * request, MPI_Status * status)
570 MPI_Request req = *request;
571 smpi_empty_status(status);
573 if(!(req->detached && req->flags & SEND)
574 && !(req->flags & PREPARED)){
575 if(status != MPI_STATUS_IGNORE) {
576 int src = req->src == MPI_ANY_SOURCE ? req->real_src : req->src;
577 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(req->comm), src);
578 status->MPI_TAG = req->tag == MPI_ANY_TAG ? req->real_tag : req->tag;
579 status->MPI_ERROR = req->truncated ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
580 // this handles the case were size in receive differs from size in send
581 // FIXME: really this should just contain the count of receive-type blocks,
583 status->count = req->real_size;
586 print_request("Finishing", req);
587 MPI_Datatype datatype = req->old_type;
589 if(datatype->has_subtype == 1){
590 // This part handles the problem of non-contignous memory
591 // the unserialization at the reception
592 s_smpi_subtype_t *subtype = datatype->substruct;
593 if(req->flags & RECV) {
594 subtype->unserialize(req->buf, req->old_buf, req->real_size/smpi_datatype_size(datatype) , datatype->substruct);
596 if(req->detached == 0) free(req->buf);
598 smpi_comm_unuse(req->comm);
599 smpi_datatype_unuse(datatype);
604 if (TRACE_smpi_view_internals()) {
605 if(req->flags & RECV){
606 int rank = smpi_process_index();
607 int src_traced = (req->src == MPI_ANY_SOURCE ? req->real_src : req->src);
608 TRACE_smpi_recv(rank, src_traced, rank);
613 if(req->detached_sender!=NULL){
614 smpi_mpi_request_free(&(req->detached_sender));
616 if(req->flags & PERSISTENT)
618 req->flags |= FINISHED;
620 smpi_mpi_request_free(request);
624 int smpi_mpi_test(MPI_Request * request, MPI_Status * status) {
627 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before)
628 smpi_empty_status(status);
630 if (!((*request)->flags & PREPARED)) {
631 if ((*request)->action != NULL)
632 flag = simcall_comm_test((*request)->action);
634 finish_wait(request, status);
635 if (*request != MPI_REQUEST_NULL && !((*request)->flags & PERSISTENT))
636 *request = MPI_REQUEST_NULL;
642 int smpi_mpi_testany(int count, MPI_Request requests[], int *index,
649 *index = MPI_UNDEFINED;
651 comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
652 map = xbt_new(int, count);
654 for(i = 0; i < count; i++) {
655 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action &&
656 !(requests[i]->flags & PREPARED)) {
657 xbt_dynar_push(comms, &requests[i]->action);
663 i = simcall_comm_testany(comms);
664 // not MPI_UNDEFINED, as this is a simix return code
667 finish_wait(&requests[*index], status);
668 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags & NON_PERSISTENT))
669 requests[*index] = MPI_REQUEST_NULL;
673 //all requests are null or inactive, return true
675 smpi_empty_status(status);
678 xbt_dynar_free(&comms);
684 int smpi_mpi_testall(int count, MPI_Request requests[],
688 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
691 for(i=0; i<count; i++){
692 if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags & PREPARED)) {
693 if (smpi_mpi_test(&requests[i], pstat)!=1){
696 requests[i]=MPI_REQUEST_NULL;
699 smpi_empty_status(pstat);
701 if(status != MPI_STATUSES_IGNORE) {
708 void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
710 //FIXME find another wait to avoid busy waiting ?
711 // the issue here is that we have to wait on a nonexistent comm
713 smpi_mpi_iprobe(source, tag, comm, &flag, status);
714 XBT_DEBUG("Busy Waiting on probing : %d", flag);
718 void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
720 MPI_Request request =build_request(NULL, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), source), smpi_comm_rank(comm), tag,
721 comm, PERSISTENT | RECV);
723 //to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
724 double sleeptime= sg_cfg_get_double("smpi/iprobe");
725 //multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
726 static int nsleeps = 1;
728 simcall_process_sleep(sleeptime);
730 // behave like a receive, but don't do it
733 print_request("New iprobe", request);
734 // We have to test both mailboxes as we don't know if we will receive one one or another
735 if (sg_cfg_get_int("smpi/async_small_thres")>0){
736 mailbox = smpi_process_mailbox_small();
737 XBT_DEBUG("trying to probe the perm recv mailbox");
738 request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
740 if (request->action==NULL){
741 mailbox = smpi_process_mailbox();
742 XBT_DEBUG("trying to probe the other mailbox");
743 request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
747 MPI_Request req = (MPI_Request)SIMIX_comm_get_src_data(request->action);
749 if(status != MPI_STATUS_IGNORE && !(req->flags & PREPARED)) {
750 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(comm), req->src);
751 status->MPI_TAG = req->tag;
752 status->MPI_ERROR = MPI_SUCCESS;
753 status->count = req->real_size;
755 nsleeps=1;//reset the number of sleeps we will do next time
761 smpi_mpi_request_free(&request);
766 void smpi_mpi_wait(MPI_Request * request, MPI_Status * status)
768 print_request("Waiting", *request);
769 if ((*request)->flags & PREPARED) {
770 smpi_empty_status(status);
774 if ((*request)->action != NULL) { // this is not a detached send
775 simcall_comm_wait((*request)->action, -1.0);
778 (*request)->action->comm.dst_data = NULL; // dangling pointer : dst_data is freed with a wait, need to set it to NULL for system state comparison
782 finish_wait(request, status);
783 if (*request != MPI_REQUEST_NULL && ((*request)->flags & NON_PERSISTENT))
784 *request = MPI_REQUEST_NULL;
785 // FIXME for a detached send, finish_wait is not called:
788 int smpi_mpi_waitany(int count, MPI_Request requests[],
795 index = MPI_UNDEFINED;
797 // Wait for a request to complete
798 comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
799 map = xbt_new(int, count);
801 XBT_DEBUG("Wait for one of %d", count);
802 for(i = 0; i < count; i++) {
803 if (requests[i] != MPI_REQUEST_NULL
804 && !(requests[i]->flags & PREPARED)
805 && !(requests[i]->flags & FINISHED)) {
806 if (requests[i]->action != NULL) {
807 XBT_DEBUG("Waiting any %p ", requests[i]);
808 xbt_dynar_push(comms, &requests[i]->action);
812 //This is a finished detached request, let's return this one
813 size=0;//so we free the dynar but don't do the waitany call
815 finish_wait(&requests[i], status);//cleanup if refcount = 0
816 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
817 requests[i]=MPI_REQUEST_NULL;//set to null
823 i = simcall_comm_waitany(comms);
825 // not MPI_UNDEFINED, as this is a simix return code
828 finish_wait(&requests[index], status);
829 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
830 requests[index] = MPI_REQUEST_NULL;
834 xbt_dynar_free(&comms);
837 if (index==MPI_UNDEFINED)
838 smpi_empty_status(status);
843 int smpi_mpi_waitall(int count, MPI_Request requests[],
848 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
849 int retvalue = MPI_SUCCESS;
850 //tag invalid requests in the set
851 if (status != MPI_STATUSES_IGNORE) {
852 for (c = 0; c < count; c++) {
853 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL ||
854 (requests[c]->flags & PREPARED)) {
855 smpi_empty_status(&status[c]);
856 } else if (requests[c]->src == MPI_PROC_NULL) {
857 smpi_empty_status(&status[c]);
858 status[c].MPI_SOURCE = MPI_PROC_NULL;
862 for(c = 0; c < count; c++) {
864 if (MC_is_active()) {
865 smpi_mpi_wait(&requests[c], pstat);
868 index = smpi_mpi_waitany(count, requests, pstat);
869 if (index == MPI_UNDEFINED)
871 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
872 requests[index]=MPI_REQUEST_NULL;
874 if (status != MPI_STATUSES_IGNORE) {
875 status[index] = *pstat;
876 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
877 retvalue = MPI_ERR_IN_STATUS;
884 int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices,
889 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
892 for(i = 0; i < incount; i++)
894 index=smpi_mpi_waitany(incount, requests, pstat);
895 if(index!=MPI_UNDEFINED){
896 indices[count] = index;
898 if(status != MPI_STATUSES_IGNORE) {
899 status[index] = *pstat;
901 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
902 requests[index]=MPI_REQUEST_NULL;
904 return MPI_UNDEFINED;
910 int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices,
913 int i, count, count_dead;
915 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
919 for(i = 0; i < incount; i++) {
920 if((requests[i] != MPI_REQUEST_NULL)) {
921 if(smpi_mpi_test(&requests[i], pstat)) {
924 if(status != MPI_STATUSES_IGNORE) {
927 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->flags & NON_PERSISTENT)
928 requests[i]=MPI_REQUEST_NULL;
934 if(count_dead==incount)return MPI_UNDEFINED;
938 void smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root,
941 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
942 nary_tree_bcast(buf, count, datatype, root, comm, 4);
945 void smpi_mpi_barrier(MPI_Comm comm)
947 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
948 nary_tree_barrier(comm, 4);
951 void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
952 void *recvbuf, int recvcount, MPI_Datatype recvtype,
953 int root, MPI_Comm comm)
955 int system_tag = COLL_TAG_GATHER;
956 int rank, size, src, index;
957 MPI_Aint lb = 0, recvext = 0;
958 MPI_Request *requests;
960 rank = smpi_comm_rank(comm);
961 size = smpi_comm_size(comm);
963 // Send buffer to root
964 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
966 // FIXME: check for errors
967 smpi_datatype_extent(recvtype, &lb, &recvext);
968 // Local copy from root
969 smpi_datatype_copy(sendbuf, sendcount, sendtype,
970 (char *)recvbuf + root * recvcount * recvext, recvcount, recvtype);
971 // Receive buffers from senders
972 requests = xbt_new(MPI_Request, size - 1);
974 for(src = 0; src < size; src++) {
976 requests[index] = smpi_irecv_init((char *)recvbuf + src * recvcount * recvext,
978 src, system_tag, comm);
982 // Wait for completion of irecv's.
983 smpi_mpi_startall(size - 1, requests);
984 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
985 for(src = 0; src < size-1; src++) {
986 smpi_mpi_request_free(&requests[src]);
993 void smpi_mpi_reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts,
994 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
998 int rank = smpi_process_index();
1001 /* arbitrarily choose root as rank 0 */
1002 size = smpi_comm_size(comm);
1004 displs = xbt_new(int, size);
1005 for (i = 0; i < size; i++) {
1007 count += recvcounts[i];
1009 tmpbuf=(void*)xbt_malloc(count*smpi_datatype_get_extent(datatype));
1010 mpi_coll_reduce_fun(sendbuf, tmpbuf, count, datatype, op, 0, comm);
1011 smpi_mpi_scatterv(tmpbuf, recvcounts, displs, datatype, recvbuf,
1012 recvcounts[rank], datatype, 0, comm);
1017 void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1018 void *recvbuf, int *recvcounts, int *displs,
1019 MPI_Datatype recvtype, int root, MPI_Comm comm)
1021 int system_tag = COLL_TAG_GATHERV;
1022 int rank, size, src, index;
1023 MPI_Aint lb = 0, recvext = 0;
1024 MPI_Request *requests;
1026 rank = smpi_comm_rank(comm);
1027 size = smpi_comm_size(comm);
1029 // Send buffer to root
1030 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
1032 // FIXME: check for errors
1033 smpi_datatype_extent(recvtype, &lb, &recvext);
1034 // Local copy from root
1035 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1036 (char *)recvbuf + displs[root] * recvext,
1037 recvcounts[root], recvtype);
1038 // Receive buffers from senders
1039 requests = xbt_new(MPI_Request, size - 1);
1041 for(src = 0; src < size; src++) {
1044 smpi_irecv_init((char *)recvbuf + displs[src] * recvext,
1045 recvcounts[src], recvtype, src, system_tag, comm);
1049 // Wait for completion of irecv's.
1050 smpi_mpi_startall(size - 1, requests);
1051 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1052 for(src = 0; src < size-1; src++) {
1053 smpi_mpi_request_free(&requests[src]);
1059 void smpi_mpi_allgather(void *sendbuf, int sendcount,
1060 MPI_Datatype sendtype, void *recvbuf,
1061 int recvcount, MPI_Datatype recvtype,
1064 int system_tag = COLL_TAG_ALLGATHER;
1065 int rank, size, other, index;
1066 MPI_Aint lb = 0, recvext = 0;
1067 MPI_Request *requests;
1069 rank = smpi_comm_rank(comm);
1070 size = smpi_comm_size(comm);
1071 // FIXME: check for errors
1072 smpi_datatype_extent(recvtype, &lb, &recvext);
1073 // Local copy from self
1074 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1075 (char *)recvbuf + rank * recvcount * recvext, recvcount,
1077 // Send/Recv buffers to/from others;
1078 requests = xbt_new(MPI_Request, 2 * (size - 1));
1080 for(other = 0; other < size; other++) {
1083 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1086 requests[index] = smpi_irecv_init((char *)recvbuf + other * recvcount * recvext,
1087 recvcount, recvtype, other,
1092 // Wait for completion of all comms.
1093 smpi_mpi_startall(2 * (size - 1), requests);
1094 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1095 for(other = 0; other < 2*(size-1); other++) {
1096 smpi_mpi_request_free(&requests[other]);
1101 void smpi_mpi_allgatherv(void *sendbuf, int sendcount,
1102 MPI_Datatype sendtype, void *recvbuf,
1103 int *recvcounts, int *displs,
1104 MPI_Datatype recvtype, MPI_Comm comm)
1106 int system_tag = COLL_TAG_ALLGATHERV;
1107 int rank, size, other, index;
1108 MPI_Aint lb = 0, recvext = 0;
1109 MPI_Request *requests;
1111 rank = smpi_comm_rank(comm);
1112 size = smpi_comm_size(comm);
1113 // FIXME: check for errors
1114 smpi_datatype_extent(recvtype, &lb, &recvext);
1115 // Local copy from self
1116 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1117 (char *)recvbuf + displs[rank] * recvext,
1118 recvcounts[rank], recvtype);
1119 // Send buffers to others;
1120 requests = xbt_new(MPI_Request, 2 * (size - 1));
1122 for(other = 0; other < size; other++) {
1125 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1129 smpi_irecv_init((char *)recvbuf + displs[other] * recvext, recvcounts[other],
1130 recvtype, other, system_tag, comm);
1134 // Wait for completion of all comms.
1135 smpi_mpi_startall(2 * (size - 1), requests);
1136 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1137 for(other = 0; other < 2*(size-1); other++) {
1138 smpi_mpi_request_free(&requests[other]);
1143 void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1144 void *recvbuf, int recvcount, MPI_Datatype recvtype,
1145 int root, MPI_Comm comm)
1147 int system_tag = COLL_TAG_SCATTER;
1148 int rank, size, dst, index;
1149 MPI_Aint lb = 0, sendext = 0;
1150 MPI_Request *requests;
1152 rank = smpi_comm_rank(comm);
1153 size = smpi_comm_size(comm);
1155 // Recv buffer from root
1156 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1159 // FIXME: check for errors
1160 smpi_datatype_extent(sendtype, &lb, &sendext);
1161 // Local copy from root
1162 if(recvbuf!=MPI_IN_PLACE){
1163 smpi_datatype_copy((char *)sendbuf + root * sendcount * sendext,
1164 sendcount, sendtype, recvbuf, recvcount, recvtype);
1166 // Send buffers to receivers
1167 requests = xbt_new(MPI_Request, size - 1);
1169 for(dst = 0; dst < size; dst++) {
1171 requests[index] = smpi_isend_init((char *)sendbuf + dst * sendcount * sendext,
1172 sendcount, sendtype, dst,
1177 // Wait for completion of isend's.
1178 smpi_mpi_startall(size - 1, requests);
1179 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1180 for(dst = 0; dst < size-1; dst++) {
1181 smpi_mpi_request_free(&requests[dst]);
1187 void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs,
1188 MPI_Datatype sendtype, void *recvbuf, int recvcount,
1189 MPI_Datatype recvtype, int root, MPI_Comm comm)
1191 int system_tag = COLL_TAG_SCATTERV;
1192 int rank, size, dst, index;
1193 MPI_Aint lb = 0, sendext = 0;
1194 MPI_Request *requests;
1196 rank = smpi_comm_rank(comm);
1197 size = smpi_comm_size(comm);
1199 // Recv buffer from root
1200 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1203 // FIXME: check for errors
1204 smpi_datatype_extent(sendtype, &lb, &sendext);
1205 // Local copy from root
1206 if(recvbuf!=MPI_IN_PLACE){
1207 smpi_datatype_copy((char *)sendbuf + displs[root] * sendext, sendcounts[root],
1208 sendtype, recvbuf, recvcount, recvtype);
1210 // Send buffers to receivers
1211 requests = xbt_new(MPI_Request, size - 1);
1213 for(dst = 0; dst < size; dst++) {
1216 smpi_isend_init((char *)sendbuf + displs[dst] * sendext, sendcounts[dst],
1217 sendtype, dst, system_tag, comm);
1221 // Wait for completion of isend's.
1222 smpi_mpi_startall(size - 1, requests);
1223 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1224 for(dst = 0; dst < size-1; dst++) {
1225 smpi_mpi_request_free(&requests[dst]);
1231 void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count,
1232 MPI_Datatype datatype, MPI_Op op, int root,
1235 int system_tag = COLL_TAG_REDUCE;
1236 int rank, size, src, index;
1237 MPI_Aint lb = 0, dataext = 0;
1238 MPI_Request *requests;
1242 char* sendtmpbuf = (char*) sendbuf;
1243 if( sendbuf == MPI_IN_PLACE ) {
1244 sendtmpbuf = (char *)xbt_malloc(count*smpi_datatype_get_extent(datatype));
1245 smpi_datatype_copy(recvbuf, count, datatype,sendtmpbuf, count, datatype);
1248 rank = smpi_comm_rank(comm);
1249 size = smpi_comm_size(comm);
1250 //non commutative case, use a working algo from openmpi
1251 if(!smpi_op_is_commute(op)){
1252 smpi_coll_tuned_reduce_ompi_basic_linear(sendtmpbuf, recvbuf, count,
1253 datatype, op, root, comm);
1258 // Send buffer to root
1259 smpi_mpi_send(sendtmpbuf, count, datatype, root, system_tag, comm);
1261 // FIXME: check for errors
1262 smpi_datatype_extent(datatype, &lb, &dataext);
1263 // Local copy from root
1264 if (sendtmpbuf && recvbuf)
1265 smpi_datatype_copy(sendtmpbuf, count, datatype, recvbuf, count, datatype);
1266 // Receive buffers from senders
1267 //TODO: make a MPI_barrier here ?
1268 requests = xbt_new(MPI_Request, size - 1);
1269 tmpbufs = xbt_new(void *, size - 1);
1271 for(src = 0; src < size; src++) {
1273 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1275 tmpbufs[index] = xbt_malloc(count * dataext);
1277 smpi_irecv_init(tmpbufs[index], count, datatype, src,
1282 // Wait for completion of irecv's.
1283 smpi_mpi_startall(size - 1, requests);
1284 for(src = 0; src < size - 1; src++) {
1285 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1286 XBT_DEBUG("finished waiting any request with index %d", index);
1287 if(index == MPI_UNDEFINED) {
1290 smpi_mpi_request_free(&requests[index]);
1292 if(op) /* op can be MPI_OP_NULL that does nothing */
1293 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1295 for(index = 0; index < size - 1; index++) {
1296 xbt_free(tmpbufs[index]);
1301 if( sendbuf == MPI_IN_PLACE ) {
1302 xbt_free(sendtmpbuf);
1307 void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count,
1308 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1310 smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, 0, comm);
1311 smpi_mpi_bcast(recvbuf, count, datatype, 0, comm);
1314 void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count,
1315 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1317 int system_tag = -888;
1318 int rank, size, other, index;
1319 MPI_Aint lb = 0, dataext = 0;
1320 MPI_Request *requests;
1323 rank = smpi_comm_rank(comm);
1324 size = smpi_comm_size(comm);
1326 // FIXME: check for errors
1327 smpi_datatype_extent(datatype, &lb, &dataext);
1329 // Local copy from self
1330 smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
1332 // Send/Recv buffers to/from others;
1333 requests = xbt_new(MPI_Request, size - 1);
1334 tmpbufs = xbt_new(void *, rank);
1336 for(other = 0; other < rank; other++) {
1337 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1339 tmpbufs[index] = xbt_malloc(count * dataext);
1341 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1345 for(other = rank + 1; other < size; other++) {
1347 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1350 // Wait for completion of all comms.
1351 smpi_mpi_startall(size - 1, requests);
1353 if(smpi_op_is_commute(op)){
1354 for(other = 0; other < size - 1; other++) {
1355 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1356 if(index == MPI_UNDEFINED) {
1360 // #Request is below rank: it's a irecv
1361 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1365 //non commutative case, wait in order
1366 for(other = 0; other < size - 1; other++) {
1367 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1369 smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1373 for(index = 0; index < rank; index++) {
1374 xbt_free(tmpbufs[index]);
1376 for(index = 0; index < size-1; index++) {
1377 smpi_mpi_request_free(&requests[index]);
1383 void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count,
1384 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1386 int system_tag = -888;
1387 int rank, size, other, index;
1388 MPI_Aint lb = 0, dataext = 0;
1389 MPI_Request *requests;
1391 int recvbuf_is_empty=1;
1392 rank = smpi_comm_rank(comm);
1393 size = smpi_comm_size(comm);
1395 // FIXME: check for errors
1396 smpi_datatype_extent(datatype, &lb, &dataext);
1398 // Send/Recv buffers to/from others;
1399 requests = xbt_new(MPI_Request, size - 1);
1400 tmpbufs = xbt_new(void *, rank);
1402 for(other = 0; other < rank; other++) {
1403 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1405 tmpbufs[index] = xbt_malloc(count * dataext);
1407 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1411 for(other = rank + 1; other < size; other++) {
1413 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1416 // Wait for completion of all comms.
1417 smpi_mpi_startall(size - 1, requests);
1418 if(smpi_op_is_commute(op)){
1419 for(other = 0; other < size - 1; other++) {
1420 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1421 if(index == MPI_UNDEFINED) {
1425 if(recvbuf_is_empty){
1426 smpi_datatype_copy(tmpbufs[index], count, datatype, recvbuf, count, datatype);
1429 // #Request is below rank: it's a irecv
1430 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1434 //non commutative case, wait in order
1435 for(other = 0; other < size - 1; other++) {
1436 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1438 if(recvbuf_is_empty){
1439 smpi_datatype_copy(tmpbufs[other], count, datatype, recvbuf, count, datatype);
1441 }else smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1445 for(index = 0; index < rank; index++) {
1446 xbt_free(tmpbufs[index]);
1448 for(index = 0; index < size-1; index++) {
1449 smpi_mpi_request_free(&requests[index]);