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;
235 if (flags & SEND) smpi_datatype_unuse(datatype);
241 void smpi_empty_status(MPI_Status * status)
243 if(status != MPI_STATUS_IGNORE) {
244 status->MPI_SOURCE = MPI_ANY_SOURCE;
245 status->MPI_TAG = MPI_ANY_TAG;
246 status->MPI_ERROR = MPI_SUCCESS;
251 void smpi_action_trace_run(char *path)
255 xbt_dict_cursor_t cursor;
259 action_fp = fopen(path, "r");
260 xbt_assert(action_fp != NULL, "Cannot open %s: %s", path,
264 if (!xbt_dict_is_empty(action_queues)) {
266 ("Not all actions got consumed. If the simulation ended successfully (without deadlock), you may want to add new processes to your deployment file.");
269 xbt_dict_foreach(action_queues, cursor, name, todo) {
270 XBT_WARN("Still %lu actions for %s", xbt_dynar_length(todo), name);
276 xbt_dict_free(&action_queues);
277 action_queues = xbt_dict_new_homogeneous(NULL);
280 static void smpi_mpi_request_free_voidp(void* request)
282 MPI_Request req = request;
283 smpi_mpi_request_free(&req);
286 /* MPI Low level calls */
287 MPI_Request smpi_mpi_send_init(void *buf, int count, MPI_Datatype datatype,
288 int dst, int tag, MPI_Comm comm)
290 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
291 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
292 comm, PERSISTENT | SEND | PREPARED);
297 MPI_Request smpi_mpi_ssend_init(void *buf, int count, MPI_Datatype datatype,
298 int dst, int tag, MPI_Comm comm)
300 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
301 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
302 comm, PERSISTENT | SSEND | SEND | PREPARED);
307 MPI_Request smpi_mpi_recv_init(void *buf, int count, MPI_Datatype datatype,
308 int src, int tag, MPI_Comm comm)
310 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
311 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,
312 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 if (request->flags & RECV) {
324 print_request("New recv", request);
325 //FIXME: if receive is posted with a large size, but send is smaller, mailboxes may not match !
326 if (request->size < sg_cfg_get_int("smpi/async_small_thres"))
327 mailbox = smpi_process_mailbox_small();
329 mailbox = smpi_process_mailbox();
330 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
331 request->real_size=request->size;
332 smpi_datatype_use(request->old_type);
333 smpi_comm_use(request->comm);
334 request->action = simcall_comm_irecv(mailbox, request->buf,
335 &request->real_size, &match_recv,
338 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
339 double sleeptime = request->detached ? smpi_or(request->size) : 0.0;
341 simcall_process_sleep(sleeptime);
342 XBT_DEBUG("receiving size of %zu : sleep %f ", request->size, smpi_or(request->size));
348 int receiver = request->dst;//smpi_group_index(smpi_comm_group(request->comm), request->dst);
351 int rank = smpi_process_index();
352 if (TRACE_smpi_view_internals()) {
353 TRACE_smpi_send(rank, rank, receiver,request->size);
356 /* if(receiver == MPI_UNDEFINED) {*/
357 /* XBT_WARN("Trying to send a message to a wrong rank");*/
360 print_request("New send", request);
361 if (request->size < sg_cfg_get_int("smpi/async_small_thres")) { // eager mode
362 mailbox = smpi_process_remote_mailbox_small(receiver);
364 XBT_DEBUG("Send request %p is not in the permanent receive mailbox (buf: %p)",request,request->buf);
365 mailbox = smpi_process_remote_mailbox(receiver);
367 if ( (! (request->flags & SSEND)) && (request->size < sg_cfg_get_int("smpi/send_is_detached_thres"))) {
369 request->detached = 1;
371 if(request->old_type->has_subtype == 0){
372 oldbuf = request->buf;
373 if (!_xbt_replay_is_active() && oldbuf && request->size!=0){
374 request->buf = xbt_malloc(request->size);
375 memcpy(request->buf,oldbuf,request->size);
378 XBT_DEBUG("Send request %p is detached; buf %p copied into %p",request,oldbuf,request->buf);
381 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
382 request->real_size=request->size;
383 smpi_datatype_use(request->old_type);
384 smpi_comm_use(request->comm);
386 //if we are giving back the control to the user without waiting for completion, we have to inject timings
387 double sleeptime =0.0;
388 if(request->detached || (request->flags & (ISEND|SSEND))){// issend should be treated as isend
389 //isend and send timings may be different
390 sleeptime = (request->flags & ISEND)? smpi_ois(request->size) : smpi_os(request->size);
394 simcall_process_sleep(sleeptime);
395 XBT_DEBUG("sending size of %zu : sleep %f ", request->size, smpi_os(request->size));
399 simcall_comm_isend(mailbox, request->size, -1.0,
400 request->buf, request->real_size,
402 &xbt_free, // how to free the userdata if a detached send fails
404 // detach if msg size < eager/rdv switch limit
408 /* FIXME: detached sends are not traceable (request->action == NULL) */
410 simcall_set_category(request->action, TRACE_internal_smpi_get_category());
418 void smpi_mpi_startall(int count, MPI_Request * requests)
421 if(requests==NULL) return;
423 for(i = 0; i < count; i++) {
424 smpi_mpi_start(requests[i]);
428 void smpi_mpi_request_free(MPI_Request * request)
430 if((*request) != MPI_REQUEST_NULL){
431 (*request)->refcount--;
432 if((*request)->refcount<0) xbt_die("wrong refcount");
434 if((*request)->refcount==0){
435 print_request("Destroying", (*request));
437 *request = MPI_REQUEST_NULL;
439 print_request("Decrementing", (*request));
443 xbt_die("freeing an already free request");
447 MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype,
448 int dst, int tag, MPI_Comm comm)
450 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
451 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf , count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
452 comm, NON_PERSISTENT | ISEND | SEND | PREPARED);
456 MPI_Request smpi_mpi_isend(void *buf, int count, MPI_Datatype datatype,
457 int dst, int tag, MPI_Comm comm)
459 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
460 request = build_request(buf==MPI_BOTTOM?(void*)0:buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
461 comm, NON_PERSISTENT | ISEND | SEND);
462 smpi_mpi_start(request);
466 MPI_Request smpi_mpi_issend(void *buf, int count, MPI_Datatype datatype,
467 int dst, int tag, MPI_Comm comm)
469 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
470 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
471 comm, NON_PERSISTENT | ISEND | SSEND | SEND);
472 smpi_mpi_start(request);
478 MPI_Request smpi_irecv_init(void *buf, int count, MPI_Datatype datatype,
479 int src, int tag, MPI_Comm comm)
481 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
482 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,
483 comm, NON_PERSISTENT | RECV | PREPARED);
487 MPI_Request smpi_mpi_irecv(void *buf, int count, MPI_Datatype datatype,
488 int src, int tag, MPI_Comm comm)
490 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
491 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,
492 comm, NON_PERSISTENT | RECV);
493 smpi_mpi_start(request);
497 void smpi_mpi_recv(void *buf, int count, MPI_Datatype datatype, int src,
498 int tag, MPI_Comm comm, MPI_Status * status)
500 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
501 request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm);
502 smpi_mpi_wait(&request, status);
508 void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst,
509 int tag, MPI_Comm comm)
511 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
512 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
513 comm, NON_PERSISTENT | SEND);
515 smpi_mpi_start(request);
516 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
520 void smpi_mpi_ssend(void *buf, int count, MPI_Datatype datatype,
521 int dst, int tag, MPI_Comm comm)
523 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
524 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
525 comm, NON_PERSISTENT | SSEND | SEND);
527 smpi_mpi_start(request);
528 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
532 void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
533 int dst, int sendtag, void *recvbuf, int recvcount,
534 MPI_Datatype recvtype, int src, int recvtag,
535 MPI_Comm comm, MPI_Status * status)
537 MPI_Request requests[2];
539 int myid=smpi_process_index();
540 if ((smpi_group_index(smpi_comm_group(comm), dst) == myid) && (smpi_group_index(smpi_comm_group(comm), src) == myid)) {
541 smpi_datatype_copy(sendbuf, sendcount, sendtype,
542 recvbuf, recvcount, recvtype);
546 smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
548 smpi_irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
549 smpi_mpi_startall(2, requests);
550 smpi_mpi_waitall(2, requests, stats);
551 if(status != MPI_STATUS_IGNORE) {
552 // Copy receive status
557 int smpi_mpi_get_count(MPI_Status * status, MPI_Datatype datatype)
559 return status->count / smpi_datatype_size(datatype);
562 static void finish_wait(MPI_Request * request, MPI_Status * status)
564 MPI_Request req = *request;
565 smpi_empty_status(status);
567 if(!(req->detached && req->flags & SEND) && !(req->flags & PREPARED)){
568 if(status != MPI_STATUS_IGNORE) {
569 int src = req->src == MPI_ANY_SOURCE ? req->real_src : req->src;
570 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(req->comm), src);
571 status->MPI_TAG = req->tag == MPI_ANY_TAG ? req->real_tag : req->tag;
572 status->MPI_ERROR = req->truncated ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
573 // this handles the case were size in receive differs from size in send
574 // FIXME: really this should just contain the count of receive-type blocks,
576 status->count = req->real_size;
579 print_request("Finishing", req);
580 MPI_Datatype datatype = req->old_type;
582 if(datatype->has_subtype == 1){
583 // This part handles the problem of non-contignous memory
584 // the unserialization at the reception
585 s_smpi_subtype_t *subtype = datatype->substruct;
586 if(req->flags & RECV) {
587 subtype->unserialize(req->buf, req->old_buf, req->real_size/smpi_datatype_size(datatype) , datatype->substruct);
589 if(req->detached == 0) free(req->buf);
591 smpi_comm_unuse(req->comm);
592 smpi_datatype_unuse(datatype);
597 if (TRACE_smpi_view_internals()) {
598 if(req->flags & RECV){
599 int rank = smpi_process_index();
600 int src_traced = (req->src == MPI_ANY_SOURCE ? req->real_src : req->src);
601 TRACE_smpi_recv(rank, src_traced, rank);
606 if(req->detached_sender!=NULL){
607 smpi_mpi_request_free(&(req->detached_sender));
610 if(req->flags & NON_PERSISTENT) {
611 smpi_mpi_request_free(request);
617 int smpi_mpi_test(MPI_Request * request, MPI_Status * status) {
620 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before)
621 smpi_empty_status(status);
623 if (!((*request)->flags & PREPARED)) {
624 if ((*request)->action != NULL)
625 flag = simcall_comm_test((*request)->action);
627 finish_wait(request, status);
628 if (*request != MPI_REQUEST_NULL && !((*request)->flags & PERSISTENT))
629 *request = MPI_REQUEST_NULL;
635 int smpi_mpi_testany(int count, MPI_Request requests[], int *index,
642 *index = MPI_UNDEFINED;
644 comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
645 map = xbt_new(int, count);
647 for(i = 0; i < count; i++) {
648 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action &&
649 !(requests[i]->flags & PREPARED)) {
650 xbt_dynar_push(comms, &requests[i]->action);
656 i = simcall_comm_testany(comms);
657 // not MPI_UNDEFINED, as this is a simix return code
660 finish_wait(&requests[*index], status);
661 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags & NON_PERSISTENT))
662 requests[*index] = MPI_REQUEST_NULL;
666 //all requests are null or inactive, return true
668 smpi_empty_status(status);
671 xbt_dynar_free(&comms);
677 int smpi_mpi_testall(int count, MPI_Request requests[],
681 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
684 for(i=0; i<count; i++){
685 if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags & PREPARED)) {
686 if (smpi_mpi_test(&requests[i], pstat)!=1){
689 requests[i]=MPI_REQUEST_NULL;
692 smpi_empty_status(pstat);
694 if(status != MPI_STATUSES_IGNORE) {
701 void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
703 //FIXME find another wait to avoid busy waiting ?
704 // the issue here is that we have to wait on a nonexistent comm
706 smpi_mpi_iprobe(source, tag, comm, &flag, status);
707 XBT_DEBUG("Busy Waiting on probing : %d", flag);
711 void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
713 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,
714 comm, NON_PERSISTENT | RECV);
716 //to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
717 double sleeptime= sg_cfg_get_double("smpi/iprobe");
718 //multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
719 static int nsleeps = 1;
721 simcall_process_sleep(sleeptime);
723 // behave like a receive, but don't do it
726 print_request("New iprobe", request);
727 // We have to test both mailboxes as we don't know if we will receive one one or another
728 if (sg_cfg_get_int("smpi/async_small_thres")>0){
729 mailbox = smpi_process_mailbox_small();
730 XBT_DEBUG("trying to probe the perm recv mailbox");
731 request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
733 if (request->action==NULL){
734 mailbox = smpi_process_mailbox();
735 XBT_DEBUG("trying to probe the other mailbox");
736 request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
740 MPI_Request req = (MPI_Request)SIMIX_comm_get_src_data(request->action);
742 if(status != MPI_STATUS_IGNORE && !(req->flags & PREPARED)) {
743 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(comm), req->src);
744 status->MPI_TAG = req->tag;
745 status->MPI_ERROR = MPI_SUCCESS;
746 status->count = req->real_size;
748 nsleeps=1;//reset the number of sleeps we will do next time
754 smpi_mpi_request_free(&request);
759 void smpi_mpi_wait(MPI_Request * request, MPI_Status * status)
761 print_request("Waiting", *request);
762 if ((*request)->flags & PREPARED) {
763 smpi_empty_status(status);
767 if ((*request)->action != NULL) { // this is not a detached send
768 simcall_comm_wait((*request)->action, -1.0);
771 (*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
775 finish_wait(request, status);
776 if (*request != MPI_REQUEST_NULL && ((*request)->flags & NON_PERSISTENT))
777 *request = MPI_REQUEST_NULL;
778 // FIXME for a detached send, finish_wait is not called:
781 int smpi_mpi_waitany(int count, MPI_Request requests[],
788 index = MPI_UNDEFINED;
790 // Wait for a request to complete
791 comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
792 map = xbt_new(int, count);
794 XBT_DEBUG("Wait for one of %d", count);
795 for(i = 0; i < count; i++) {
796 if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags & PREPARED)) {
797 if (requests[i]->action != NULL) {
798 XBT_DEBUG("Waiting any %p ", requests[i]);
799 xbt_dynar_push(comms, &requests[i]->action);
803 //This is a finished detached request, let's return this one
804 size=0;//so we free the dynar but don't do the waitany call
806 finish_wait(&requests[i], status);//cleanup if refcount = 0
807 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
808 requests[i]=MPI_REQUEST_NULL;//set to null
814 i = simcall_comm_waitany(comms);
816 // not MPI_UNDEFINED, as this is a simix return code
819 finish_wait(&requests[index], status);
820 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
821 requests[index] = MPI_REQUEST_NULL;
825 xbt_dynar_free(&comms);
828 if (index==MPI_UNDEFINED)
829 smpi_empty_status(status);
834 int smpi_mpi_waitall(int count, MPI_Request requests[],
839 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
840 int retvalue = MPI_SUCCESS;
841 //tag invalid requests in the set
842 if (status != MPI_STATUSES_IGNORE) {
843 for (c = 0; c < count; c++) {
844 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL ||
845 (requests[c]->flags & PREPARED)) {
846 smpi_empty_status(&status[c]);
847 } else if (requests[c]->src == MPI_PROC_NULL) {
848 smpi_empty_status(&status[c]);
849 status[c].MPI_SOURCE = MPI_PROC_NULL;
853 for(c = 0; c < count; c++) {
854 if (MC_is_active()) {
855 smpi_mpi_wait(&requests[c], pstat);
858 index = smpi_mpi_waitany(count, requests, pstat);
859 if (index == MPI_UNDEFINED)
861 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
862 requests[index]=MPI_REQUEST_NULL;
864 if (status != MPI_STATUSES_IGNORE) {
865 status[index] = *pstat;
866 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
867 retvalue = MPI_ERR_IN_STATUS;
874 int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices,
879 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
882 for(i = 0; i < incount; i++)
884 index=smpi_mpi_waitany(incount, requests, pstat);
885 if(index!=MPI_UNDEFINED){
886 indices[count] = index;
888 if(status != MPI_STATUSES_IGNORE) {
889 status[index] = *pstat;
891 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
892 requests[index]=MPI_REQUEST_NULL;
894 return MPI_UNDEFINED;
900 int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices,
903 int i, count, count_dead;
905 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
909 for(i = 0; i < incount; i++) {
910 if((requests[i] != MPI_REQUEST_NULL)) {
911 if(smpi_mpi_test(&requests[i], pstat)) {
914 if(status != MPI_STATUSES_IGNORE) {
917 if (requests[i]->flags & NON_PERSISTENT)
918 requests[i]=MPI_REQUEST_NULL;
924 if(count_dead==incount)return MPI_UNDEFINED;
928 void smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root,
931 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
932 nary_tree_bcast(buf, count, datatype, root, comm, 4);
935 void smpi_mpi_barrier(MPI_Comm comm)
937 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
938 nary_tree_barrier(comm, 4);
941 void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
942 void *recvbuf, int recvcount, MPI_Datatype recvtype,
943 int root, MPI_Comm comm)
945 int system_tag = COLL_TAG_GATHER;
946 int rank, size, src, index;
947 MPI_Aint lb = 0, recvext = 0;
948 MPI_Request *requests;
950 rank = smpi_comm_rank(comm);
951 size = smpi_comm_size(comm);
953 // Send buffer to root
954 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
956 // FIXME: check for errors
957 smpi_datatype_extent(recvtype, &lb, &recvext);
958 // Local copy from root
959 smpi_datatype_copy(sendbuf, sendcount, sendtype,
960 (char *)recvbuf + root * recvcount * recvext, recvcount, recvtype);
961 // Receive buffers from senders
962 requests = xbt_new(MPI_Request, size - 1);
964 for(src = 0; src < size; src++) {
966 requests[index] = smpi_irecv_init((char *)recvbuf + src * recvcount * recvext,
968 src, system_tag, comm);
972 // Wait for completion of irecv's.
973 smpi_mpi_startall(size - 1, requests);
974 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
980 void smpi_mpi_reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts,
981 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
985 int rank = smpi_process_index();
988 /* arbitrarily choose root as rank 0 */
989 size = smpi_comm_size(comm);
991 displs = xbt_new(int, size);
992 for (i = 0; i < size; i++) {
994 count += recvcounts[i];
996 tmpbuf=(void*)xbt_malloc(count*smpi_datatype_get_extent(datatype));
997 mpi_coll_reduce_fun(sendbuf, tmpbuf, count, datatype, op, 0, comm);
998 smpi_mpi_scatterv(tmpbuf, recvcounts, displs, datatype, recvbuf,
999 recvcounts[rank], datatype, 0, comm);
1004 void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1005 void *recvbuf, int *recvcounts, int *displs,
1006 MPI_Datatype recvtype, int root, MPI_Comm comm)
1008 int system_tag = COLL_TAG_GATHERV;
1009 int rank, size, src, index;
1010 MPI_Aint lb = 0, recvext = 0;
1011 MPI_Request *requests;
1013 rank = smpi_comm_rank(comm);
1014 size = smpi_comm_size(comm);
1016 // Send buffer to root
1017 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
1019 // FIXME: check for errors
1020 smpi_datatype_extent(recvtype, &lb, &recvext);
1021 // Local copy from root
1022 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1023 (char *)recvbuf + displs[root] * recvext,
1024 recvcounts[root], recvtype);
1025 // Receive buffers from senders
1026 requests = xbt_new(MPI_Request, size - 1);
1028 for(src = 0; src < size; src++) {
1031 smpi_irecv_init((char *)recvbuf + displs[src] * recvext,
1032 recvcounts[src], recvtype, src, system_tag, comm);
1036 // Wait for completion of irecv's.
1037 smpi_mpi_startall(size - 1, requests);
1038 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1043 void smpi_mpi_allgather(void *sendbuf, int sendcount,
1044 MPI_Datatype sendtype, void *recvbuf,
1045 int recvcount, MPI_Datatype recvtype,
1048 int system_tag = COLL_TAG_ALLGATHER;
1049 int rank, size, other, index;
1050 MPI_Aint lb = 0, recvext = 0;
1051 MPI_Request *requests;
1053 rank = smpi_comm_rank(comm);
1054 size = smpi_comm_size(comm);
1055 // FIXME: check for errors
1056 smpi_datatype_extent(recvtype, &lb, &recvext);
1057 // Local copy from self
1058 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1059 (char *)recvbuf + rank * recvcount * recvext, recvcount,
1061 // Send/Recv buffers to/from others;
1062 requests = xbt_new(MPI_Request, 2 * (size - 1));
1064 for(other = 0; other < size; other++) {
1067 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1070 requests[index] = smpi_irecv_init((char *)recvbuf + other * recvcount * recvext,
1071 recvcount, recvtype, other,
1076 // Wait for completion of all comms.
1077 smpi_mpi_startall(2 * (size - 1), requests);
1078 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1082 void smpi_mpi_allgatherv(void *sendbuf, int sendcount,
1083 MPI_Datatype sendtype, void *recvbuf,
1084 int *recvcounts, int *displs,
1085 MPI_Datatype recvtype, MPI_Comm comm)
1087 int system_tag = COLL_TAG_ALLGATHERV;
1088 int rank, size, other, index;
1089 MPI_Aint lb = 0, recvext = 0;
1090 MPI_Request *requests;
1092 rank = smpi_comm_rank(comm);
1093 size = smpi_comm_size(comm);
1094 // FIXME: check for errors
1095 smpi_datatype_extent(recvtype, &lb, &recvext);
1096 // Local copy from self
1097 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1098 (char *)recvbuf + displs[rank] * recvext,
1099 recvcounts[rank], recvtype);
1100 // Send buffers to others;
1101 requests = xbt_new(MPI_Request, 2 * (size - 1));
1103 for(other = 0; other < size; other++) {
1106 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1110 smpi_irecv_init((char *)recvbuf + displs[other] * recvext, recvcounts[other],
1111 recvtype, other, system_tag, comm);
1115 // Wait for completion of all comms.
1116 smpi_mpi_startall(2 * (size - 1), requests);
1117 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1121 void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1122 void *recvbuf, int recvcount, MPI_Datatype recvtype,
1123 int root, MPI_Comm comm)
1125 int system_tag = COLL_TAG_SCATTER;
1126 int rank, size, dst, index;
1127 MPI_Aint lb = 0, sendext = 0;
1128 MPI_Request *requests;
1130 rank = smpi_comm_rank(comm);
1131 size = smpi_comm_size(comm);
1133 // Recv buffer from root
1134 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1137 // FIXME: check for errors
1138 smpi_datatype_extent(sendtype, &lb, &sendext);
1139 // Local copy from root
1140 if(recvbuf!=MPI_IN_PLACE){
1141 smpi_datatype_copy((char *)sendbuf + root * sendcount * sendext,
1142 sendcount, sendtype, recvbuf, recvcount, recvtype);
1144 // Send buffers to receivers
1145 requests = xbt_new(MPI_Request, size - 1);
1147 for(dst = 0; dst < size; dst++) {
1149 requests[index] = smpi_isend_init((char *)sendbuf + dst * sendcount * sendext,
1150 sendcount, sendtype, dst,
1155 // Wait for completion of isend's.
1156 smpi_mpi_startall(size - 1, requests);
1157 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1162 void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs,
1163 MPI_Datatype sendtype, void *recvbuf, int recvcount,
1164 MPI_Datatype recvtype, int root, MPI_Comm comm)
1166 int system_tag = COLL_TAG_SCATTERV;
1167 int rank, size, dst, index;
1168 MPI_Aint lb = 0, sendext = 0;
1169 MPI_Request *requests;
1171 rank = smpi_comm_rank(comm);
1172 size = smpi_comm_size(comm);
1174 // Recv buffer from root
1175 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1178 // FIXME: check for errors
1179 smpi_datatype_extent(sendtype, &lb, &sendext);
1180 // Local copy from root
1181 if(recvbuf!=MPI_IN_PLACE){
1182 smpi_datatype_copy((char *)sendbuf + displs[root] * sendext, sendcounts[root],
1183 sendtype, recvbuf, recvcount, recvtype);
1185 // Send buffers to receivers
1186 requests = xbt_new(MPI_Request, size - 1);
1188 for(dst = 0; dst < size; dst++) {
1191 smpi_isend_init((char *)sendbuf + displs[dst] * sendext, sendcounts[dst],
1192 sendtype, dst, system_tag, comm);
1196 // Wait for completion of isend's.
1197 smpi_mpi_startall(size - 1, requests);
1198 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1203 void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count,
1204 MPI_Datatype datatype, MPI_Op op, int root,
1207 int system_tag = COLL_TAG_REDUCE;
1208 int rank, size, src, index;
1209 MPI_Aint lb = 0, dataext = 0;
1210 MPI_Request *requests;
1214 char* sendtmpbuf = (char*) sendbuf;
1215 if( sendbuf == MPI_IN_PLACE ) {
1216 sendtmpbuf = (char *)xbt_malloc(count*smpi_datatype_get_extent(datatype));
1217 smpi_datatype_copy(recvbuf, count, datatype,sendtmpbuf, count, datatype);
1220 rank = smpi_comm_rank(comm);
1221 size = smpi_comm_size(comm);
1222 //non commutative case, use a working algo from openmpi
1223 if(!smpi_op_is_commute(op)){
1224 smpi_coll_tuned_reduce_ompi_basic_linear(sendtmpbuf, recvbuf, count,
1225 datatype, op, root, comm);
1230 // Send buffer to root
1231 smpi_mpi_send(sendtmpbuf, count, datatype, root, system_tag, comm);
1233 // FIXME: check for errors
1234 smpi_datatype_extent(datatype, &lb, &dataext);
1235 // Local copy from root
1236 if (sendtmpbuf && recvbuf)
1237 smpi_datatype_copy(sendtmpbuf, count, datatype, recvbuf, count, datatype);
1238 // Receive buffers from senders
1239 //TODO: make a MPI_barrier here ?
1240 requests = xbt_new(MPI_Request, size - 1);
1241 tmpbufs = xbt_new(void *, size - 1);
1243 for(src = 0; src < size; src++) {
1245 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1247 tmpbufs[index] = xbt_malloc(count * dataext);
1249 smpi_irecv_init(tmpbufs[index], count, datatype, src,
1254 // Wait for completion of irecv's.
1255 smpi_mpi_startall(size - 1, requests);
1256 for(src = 0; src < size - 1; src++) {
1257 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1258 XBT_DEBUG("finished waiting any request with index %d", index);
1259 if(index == MPI_UNDEFINED) {
1262 if(op) /* op can be MPI_OP_NULL that does nothing */
1263 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1265 for(index = 0; index < size - 1; index++) {
1266 xbt_free(tmpbufs[index]);
1271 if( sendbuf == MPI_IN_PLACE ) {
1272 xbt_free(sendtmpbuf);
1277 void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count,
1278 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1280 smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, 0, comm);
1281 smpi_mpi_bcast(recvbuf, count, datatype, 0, comm);
1284 void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count,
1285 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1287 int system_tag = -888;
1288 int rank, size, other, index;
1289 MPI_Aint lb = 0, dataext = 0;
1290 MPI_Request *requests;
1293 rank = smpi_comm_rank(comm);
1294 size = smpi_comm_size(comm);
1296 // FIXME: check for errors
1297 smpi_datatype_extent(datatype, &lb, &dataext);
1299 // Local copy from self
1300 smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
1302 // Send/Recv buffers to/from others;
1303 requests = xbt_new(MPI_Request, size - 1);
1304 tmpbufs = xbt_new(void *, rank);
1306 for(other = 0; other < rank; other++) {
1307 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1309 tmpbufs[index] = xbt_malloc(count * dataext);
1311 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1315 for(other = rank + 1; other < size; other++) {
1317 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1320 // Wait for completion of all comms.
1321 smpi_mpi_startall(size - 1, requests);
1323 if(smpi_op_is_commute(op)){
1324 for(other = 0; other < size - 1; other++) {
1325 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1326 if(index == MPI_UNDEFINED) {
1330 // #Request is below rank: it's a irecv
1331 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1335 //non commutative case, wait in order
1336 for(other = 0; other < size - 1; other++) {
1337 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1339 smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1343 for(index = 0; index < rank; index++) {
1344 xbt_free(tmpbufs[index]);
1350 void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count,
1351 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1353 int system_tag = -888;
1354 int rank, size, other, index;
1355 MPI_Aint lb = 0, dataext = 0;
1356 MPI_Request *requests;
1358 int recvbuf_is_empty=1;
1359 rank = smpi_comm_rank(comm);
1360 size = smpi_comm_size(comm);
1362 // FIXME: check for errors
1363 smpi_datatype_extent(datatype, &lb, &dataext);
1365 // Send/Recv buffers to/from others;
1366 requests = xbt_new(MPI_Request, size - 1);
1367 tmpbufs = xbt_new(void *, rank);
1369 for(other = 0; other < rank; other++) {
1370 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1372 tmpbufs[index] = xbt_malloc(count * dataext);
1374 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1378 for(other = rank + 1; other < size; other++) {
1380 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1383 // Wait for completion of all comms.
1384 smpi_mpi_startall(size - 1, requests);
1385 if(smpi_op_is_commute(op)){
1386 for(other = 0; other < size - 1; other++) {
1387 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1388 if(index == MPI_UNDEFINED) {
1392 if(recvbuf_is_empty){
1393 smpi_datatype_copy(tmpbufs[index], count, datatype, recvbuf, count, datatype);
1396 // #Request is below rank: it's a irecv
1397 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1401 //non commutative case, wait in order
1402 for(other = 0; other < size - 1; other++) {
1403 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1405 if(recvbuf_is_empty){
1406 smpi_datatype_copy(tmpbufs[other], count, datatype, recvbuf, count, datatype);
1408 }else smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1412 for(index = 0; index < rank; index++) {
1413 xbt_free(tmpbufs[index]);