1 /* Copyright (c) 2007-2013. 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, &request->real_size, &match_recv, request);
336 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
337 double sleeptime = request->detached ? smpi_or(request->size) : 0.0;
339 simcall_process_sleep(sleeptime);
340 XBT_DEBUG("receiving size of %zu : sleep %f ", request->size, smpi_or(request->size));
346 int receiver = request->dst;//smpi_group_index(smpi_comm_group(request->comm), request->dst);
349 int rank = smpi_process_index();
350 if (TRACE_smpi_view_internals()) {
351 TRACE_smpi_send(rank, rank, receiver,request->size);
354 /* if(receiver == MPI_UNDEFINED) {*/
355 /* XBT_WARN("Trying to send a message to a wrong rank");*/
358 print_request("New send", request);
359 if (request->size < sg_cfg_get_int("smpi/async_small_thres")) { // eager mode
360 mailbox = smpi_process_remote_mailbox_small(receiver);
362 XBT_DEBUG("Send request %p is not in the permanent receive mailbox (buf: %p)",request,request->buf);
363 mailbox = smpi_process_remote_mailbox(receiver);
365 if ( (! (request->flags & SSEND)) && (request->size < sg_cfg_get_int("smpi/send_is_detached_thres"))) {
367 request->detached = 1;
369 if(request->old_type->has_subtype == 0){
370 oldbuf = request->buf;
371 if (oldbuf && request->size!=0){
372 request->buf = xbt_malloc(request->size);
373 memcpy(request->buf,oldbuf,request->size);
376 XBT_DEBUG("Send request %p is detached; buf %p copied into %p",request,oldbuf,request->buf);
379 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
380 request->real_size=request->size;
381 smpi_datatype_use(request->old_type);
382 smpi_comm_use(request->comm);
384 //if we are giving back the control to the user without waiting for completion, we have to inject timings
385 double sleeptime =0.0;
386 if(request->detached || (request->flags & (ISEND|SSEND))){// issend should be treated as isend
387 //isend and send timings may be different
388 sleeptime = (request->flags & ISEND)? smpi_ois(request->size) : smpi_os(request->size);
392 simcall_process_sleep(sleeptime);
393 XBT_DEBUG("sending size of %zu : sleep %f ", request->size, smpi_os(request->size));
397 simcall_comm_isend(mailbox, request->size, -1.0,
398 request->buf, request->real_size,
400 &smpi_mpi_request_free_voidp, // how to free the userdata if a detached send fails
402 // detach if msg size < eager/rdv switch limit
406 /* FIXME: detached sends are not traceable (request->action == NULL) */
408 simcall_set_category(request->action, TRACE_internal_smpi_get_category());
416 void smpi_mpi_startall(int count, MPI_Request * requests)
419 if(requests==NULL) return;
421 for(i = 0; i < count; i++) {
422 smpi_mpi_start(requests[i]);
426 void smpi_mpi_request_free(MPI_Request * request)
428 if((*request) != MPI_REQUEST_NULL){
429 (*request)->refcount--;
430 if((*request)->refcount<0) xbt_die("wrong refcount");
432 if((*request)->refcount==0){
433 print_request("Destroying", (*request));
435 *request = MPI_REQUEST_NULL;
437 print_request("Decrementing", (*request));
441 xbt_die("freeing an already free request");
445 MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype,
446 int dst, int tag, MPI_Comm comm)
448 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
449 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf , count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
450 comm, NON_PERSISTENT | ISEND | SEND | PREPARED);
454 MPI_Request smpi_mpi_isend(void *buf, int count, MPI_Datatype datatype,
455 int dst, int tag, MPI_Comm comm)
457 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
458 request = build_request(buf==MPI_BOTTOM?(void*)0:buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
459 comm, NON_PERSISTENT | ISEND | SEND);
460 smpi_mpi_start(request);
464 MPI_Request smpi_mpi_issend(void *buf, int count, MPI_Datatype datatype,
465 int dst, int tag, MPI_Comm comm)
467 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
468 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
469 comm, NON_PERSISTENT | ISEND | SSEND | SEND);
470 smpi_mpi_start(request);
476 MPI_Request smpi_irecv_init(void *buf, int count, MPI_Datatype datatype,
477 int src, int tag, MPI_Comm comm)
479 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
480 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,
481 comm, NON_PERSISTENT | RECV | PREPARED);
485 MPI_Request smpi_mpi_irecv(void *buf, int count, MPI_Datatype datatype,
486 int src, int tag, MPI_Comm comm)
488 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
489 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,
490 comm, NON_PERSISTENT | RECV);
491 smpi_mpi_start(request);
495 void smpi_mpi_recv(void *buf, int count, MPI_Datatype datatype, int src,
496 int tag, MPI_Comm comm, MPI_Status * status)
498 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
499 request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm);
500 smpi_mpi_wait(&request, status);
506 void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst,
507 int tag, MPI_Comm comm)
509 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
510 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
511 comm, NON_PERSISTENT | SEND);
513 smpi_mpi_start(request);
514 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
518 void smpi_mpi_ssend(void *buf, int count, MPI_Datatype datatype,
519 int dst, int tag, MPI_Comm comm)
521 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
522 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
523 comm, NON_PERSISTENT | SSEND | SEND);
525 smpi_mpi_start(request);
526 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
530 void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
531 int dst, int sendtag, void *recvbuf, int recvcount,
532 MPI_Datatype recvtype, int src, int recvtag,
533 MPI_Comm comm, MPI_Status * status)
535 MPI_Request requests[2];
537 int myid=smpi_process_index();
538 if ((smpi_group_index(smpi_comm_group(comm), dst) == myid) && (smpi_group_index(smpi_comm_group(comm), src) == myid)) {
539 smpi_datatype_copy(sendbuf, sendcount, sendtype,
540 recvbuf, recvcount, recvtype);
544 smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
546 smpi_irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
547 smpi_mpi_startall(2, requests);
548 smpi_mpi_waitall(2, requests, stats);
549 if(status != MPI_STATUS_IGNORE) {
550 // Copy receive status
555 int smpi_mpi_get_count(MPI_Status * status, MPI_Datatype datatype)
557 return status->count / smpi_datatype_size(datatype);
560 static void finish_wait(MPI_Request * request, MPI_Status * status)
562 MPI_Request req = *request;
563 smpi_empty_status(status);
565 if(!(req->detached && req->flags & SEND) && !(req->flags & PREPARED)){
566 if(status != MPI_STATUS_IGNORE) {
567 int src = req->src == MPI_ANY_SOURCE ? req->real_src : req->src;
568 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(req->comm), src);
569 status->MPI_TAG = req->tag == MPI_ANY_TAG ? req->real_tag : req->tag;
570 status->MPI_ERROR = req->truncated ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
571 // this handles the case were size in receive differs from size in send
572 // FIXME: really this should just contain the count of receive-type blocks,
574 status->count = req->real_size;
577 print_request("Finishing", req);
578 MPI_Datatype datatype = req->old_type;
580 if(datatype->has_subtype == 1){
581 // This part handles the problem of non-contignous memory
582 // the unserialization at the reception
583 s_smpi_subtype_t *subtype = datatype->substruct;
584 if(req->flags & RECV) {
585 subtype->unserialize(req->buf, req->old_buf, req->real_size/smpi_datatype_size(datatype) , datatype->substruct);
587 if(req->detached == 0) free(req->buf);
589 smpi_comm_unuse(req->comm);
590 smpi_datatype_unuse(datatype);
595 if (TRACE_smpi_view_internals()) {
596 if(req->flags & RECV){
597 int rank = smpi_process_index();
598 int src_traced = (req->src == MPI_ANY_SOURCE ? req->real_src : req->src);
599 TRACE_smpi_recv(rank, src_traced, rank);
604 if(req->detached_sender!=NULL){
605 smpi_mpi_request_free(&(req->detached_sender));
608 if(req->flags & NON_PERSISTENT) {
609 smpi_mpi_request_free(request);
615 int smpi_mpi_test(MPI_Request * request, MPI_Status * status) {
618 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before)
619 smpi_empty_status(status);
621 if (!((*request)->flags & PREPARED)) {
622 if ((*request)->action != NULL)
623 flag = simcall_comm_test((*request)->action);
625 finish_wait(request, status);
626 *request = MPI_REQUEST_NULL;
632 int smpi_mpi_testany(int count, MPI_Request requests[], int *index,
639 *index = MPI_UNDEFINED;
641 comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
642 map = xbt_new(int, count);
644 for(i = 0; i < count; i++) {
645 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action &&
646 !(requests[i]->flags & PREPARED)) {
647 xbt_dynar_push(comms, &requests[i]->action);
653 i = simcall_comm_testany(comms);
654 // not MPI_UNDEFINED, as this is a simix return code
657 finish_wait(&requests[*index], status);
658 requests[*index] = MPI_REQUEST_NULL;
662 //all requests are null or inactive, return true
664 smpi_empty_status(status);
667 xbt_dynar_free(&comms);
673 int smpi_mpi_testall(int count, MPI_Request requests[],
677 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
680 for(i=0; i<count; i++){
681 if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags & PREPARED)) {
682 if (smpi_mpi_test(&requests[i], pstat)!=1){
685 requests[i]=MPI_REQUEST_NULL;
688 smpi_empty_status(pstat);
690 if(status != MPI_STATUSES_IGNORE) {
697 void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
699 //FIXME find another wait to avoid busy waiting ?
700 // the issue here is that we have to wait on a nonexistent comm
702 smpi_mpi_iprobe(source, tag, comm, &flag, status);
703 XBT_DEBUG("Busy Waiting on probing : %d", flag);
707 void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
709 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,
710 comm, NON_PERSISTENT | RECV);
712 //to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
713 double sleeptime= sg_cfg_get_double("smpi/iprobe");
714 //multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
715 static int nsleeps = 1;
717 simcall_process_sleep(sleeptime);
719 // behave like a receive, but don't do it
722 print_request("New iprobe", request);
723 // We have to test both mailboxes as we don't know if we will receive one one or another
724 if (sg_cfg_get_int("smpi/async_small_thres")>0){
725 mailbox = smpi_process_mailbox_small();
726 XBT_DEBUG("trying to probe the perm recv mailbox");
727 request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
729 if (request->action==NULL){
730 mailbox = smpi_process_mailbox();
731 XBT_DEBUG("trying to probe the other mailbox");
732 request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
736 MPI_Request req = (MPI_Request)SIMIX_comm_get_src_data(request->action);
738 if(status != MPI_STATUS_IGNORE && !(req->flags & PREPARED)) {
739 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(comm), req->src);
740 status->MPI_TAG = req->tag;
741 status->MPI_ERROR = MPI_SUCCESS;
742 status->count = req->real_size;
744 nsleeps=1;//reset the number of sleeps we will do next time
750 smpi_mpi_request_free(&request);
755 void smpi_mpi_wait(MPI_Request * request, MPI_Status * status)
757 print_request("Waiting", *request);
758 if ((*request)->flags & PREPARED) {
759 smpi_empty_status(status);
763 if ((*request)->action != NULL) { // this is not a detached send
764 simcall_comm_wait((*request)->action, -1.0);
769 (*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
772 finish_wait(request, status);
773 *request = MPI_REQUEST_NULL;
774 // FIXME for a detached send, finish_wait is not called:
777 int smpi_mpi_waitany(int count, MPI_Request requests[],
784 index = MPI_UNDEFINED;
786 // Wait for a request to complete
787 comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
788 map = xbt_new(int, count);
790 XBT_DEBUG("Wait for one of %d", count);
791 for(i = 0; i < count; i++) {
792 if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags & PREPARED)) {
793 if (requests[i]->action != NULL) {
794 XBT_DEBUG("Waiting any %p ", requests[i]);
795 xbt_dynar_push(comms, &requests[i]->action);
799 //This is a finished detached request, let's return this one
800 size=0;//so we free the dynar but don't do the waitany call
802 finish_wait(&requests[i], status);//cleanup if refcount = 0
803 requests[i]=MPI_REQUEST_NULL;//set to null
809 i = simcall_comm_waitany(comms);
811 // not MPI_UNDEFINED, as this is a simix return code
814 finish_wait(&requests[index], status);
815 requests[index] = MPI_REQUEST_NULL;
819 xbt_dynar_free(&comms);
822 if (index==MPI_UNDEFINED)
823 smpi_empty_status(status);
828 int smpi_mpi_waitall(int count, MPI_Request requests[],
833 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
834 int retvalue = MPI_SUCCESS;
835 //tag invalid requests in the set
836 if (status != MPI_STATUSES_IGNORE) {
837 for (c = 0; c < count; c++) {
838 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL ||
839 (requests[c]->flags & PREPARED)) {
840 smpi_empty_status(&status[c]);
841 } else if (requests[c]->src == MPI_PROC_NULL) {
842 smpi_empty_status(&status[c]);
843 status[c].MPI_SOURCE = MPI_PROC_NULL;
847 for(c = 0; c < count; c++) {
848 if (MC_is_active()) {
849 smpi_mpi_wait(&requests[c], pstat);
852 index = smpi_mpi_waitany(count, requests, pstat);
853 if (index == MPI_UNDEFINED)
855 requests[index]=MPI_REQUEST_NULL;
857 if (status != MPI_STATUSES_IGNORE) {
858 status[index] = *pstat;
859 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
860 retvalue = MPI_ERR_IN_STATUS;
867 int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices,
872 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
875 for(i = 0; i < incount; i++)
877 index=smpi_mpi_waitany(incount, requests, pstat);
878 if(index!=MPI_UNDEFINED){
879 indices[count] = index;
881 if(status != MPI_STATUSES_IGNORE) {
882 status[index] = *pstat;
884 requests[index]=MPI_REQUEST_NULL;
886 return MPI_UNDEFINED;
892 int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices,
895 int i, count, count_dead;
897 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
901 for(i = 0; i < incount; i++) {
902 if((requests[i] != MPI_REQUEST_NULL)) {
903 if(smpi_mpi_test(&requests[i], pstat)) {
906 if(status != MPI_STATUSES_IGNORE) {
909 requests[i]=MPI_REQUEST_NULL;
916 if(count_dead==incount)return MPI_UNDEFINED;
920 void smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root,
923 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
924 nary_tree_bcast(buf, count, datatype, root, comm, 4);
927 void smpi_mpi_barrier(MPI_Comm comm)
929 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
930 nary_tree_barrier(comm, 4);
933 void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
934 void *recvbuf, int recvcount, MPI_Datatype recvtype,
935 int root, MPI_Comm comm)
937 int system_tag = COLL_TAG_GATHER;
938 int rank, size, src, index;
939 MPI_Aint lb = 0, recvext = 0;
940 MPI_Request *requests;
942 rank = smpi_comm_rank(comm);
943 size = smpi_comm_size(comm);
945 // Send buffer to root
946 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
948 // FIXME: check for errors
949 smpi_datatype_extent(recvtype, &lb, &recvext);
950 // Local copy from root
951 smpi_datatype_copy(sendbuf, sendcount, sendtype,
952 (char *)recvbuf + root * recvcount * recvext, recvcount, recvtype);
953 // Receive buffers from senders
954 requests = xbt_new(MPI_Request, size - 1);
956 for(src = 0; src < size; src++) {
958 requests[index] = smpi_irecv_init((char *)recvbuf + src * recvcount * recvext,
960 src, system_tag, comm);
964 // Wait for completion of irecv's.
965 smpi_mpi_startall(size - 1, requests);
966 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
972 void smpi_mpi_reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts,
973 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
977 int rank = smpi_process_index();
980 /* arbitrarily choose root as rank 0 */
981 size = smpi_comm_size(comm);
983 displs = xbt_new(int, size);
984 for (i = 0; i < size; i++) {
986 count += recvcounts[i];
988 tmpbuf=(void*)xbt_malloc(count*smpi_datatype_get_extent(datatype));
989 mpi_coll_reduce_fun(sendbuf, tmpbuf, count, datatype, op, 0, comm);
990 smpi_mpi_scatterv(tmpbuf, recvcounts, displs, datatype, recvbuf,
991 recvcounts[rank], datatype, 0, comm);
996 void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
997 void *recvbuf, int *recvcounts, int *displs,
998 MPI_Datatype recvtype, int root, MPI_Comm comm)
1000 int system_tag = COLL_TAG_GATHERV;
1001 int rank, size, src, index;
1002 MPI_Aint lb = 0, recvext = 0;
1003 MPI_Request *requests;
1005 rank = smpi_comm_rank(comm);
1006 size = smpi_comm_size(comm);
1008 // Send buffer to root
1009 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
1011 // FIXME: check for errors
1012 smpi_datatype_extent(recvtype, &lb, &recvext);
1013 // Local copy from root
1014 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1015 (char *)recvbuf + displs[root] * recvext,
1016 recvcounts[root], recvtype);
1017 // Receive buffers from senders
1018 requests = xbt_new(MPI_Request, size - 1);
1020 for(src = 0; src < size; src++) {
1023 smpi_irecv_init((char *)recvbuf + displs[src] * recvext,
1024 recvcounts[src], recvtype, src, system_tag, comm);
1028 // Wait for completion of irecv's.
1029 smpi_mpi_startall(size - 1, requests);
1030 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1035 void smpi_mpi_allgather(void *sendbuf, int sendcount,
1036 MPI_Datatype sendtype, void *recvbuf,
1037 int recvcount, MPI_Datatype recvtype,
1040 int system_tag = COLL_TAG_ALLGATHER;
1041 int rank, size, other, index;
1042 MPI_Aint lb = 0, recvext = 0;
1043 MPI_Request *requests;
1045 rank = smpi_comm_rank(comm);
1046 size = smpi_comm_size(comm);
1047 // FIXME: check for errors
1048 smpi_datatype_extent(recvtype, &lb, &recvext);
1049 // Local copy from self
1050 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1051 (char *)recvbuf + rank * recvcount * recvext, recvcount,
1053 // Send/Recv buffers to/from others;
1054 requests = xbt_new(MPI_Request, 2 * (size - 1));
1056 for(other = 0; other < size; other++) {
1059 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1062 requests[index] = smpi_irecv_init((char *)recvbuf + other * recvcount * recvext,
1063 recvcount, recvtype, other,
1068 // Wait for completion of all comms.
1069 smpi_mpi_startall(2 * (size - 1), requests);
1070 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1074 void smpi_mpi_allgatherv(void *sendbuf, int sendcount,
1075 MPI_Datatype sendtype, void *recvbuf,
1076 int *recvcounts, int *displs,
1077 MPI_Datatype recvtype, MPI_Comm comm)
1079 int system_tag = COLL_TAG_ALLGATHERV;
1080 int rank, size, other, index;
1081 MPI_Aint lb = 0, recvext = 0;
1082 MPI_Request *requests;
1084 rank = smpi_comm_rank(comm);
1085 size = smpi_comm_size(comm);
1086 // FIXME: check for errors
1087 smpi_datatype_extent(recvtype, &lb, &recvext);
1088 // Local copy from self
1089 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1090 (char *)recvbuf + displs[rank] * recvext,
1091 recvcounts[rank], recvtype);
1092 // Send buffers to others;
1093 requests = xbt_new(MPI_Request, 2 * (size - 1));
1095 for(other = 0; other < size; other++) {
1098 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1102 smpi_irecv_init((char *)recvbuf + displs[other] * recvext, recvcounts[other],
1103 recvtype, other, system_tag, comm);
1107 // Wait for completion of all comms.
1108 smpi_mpi_startall(2 * (size - 1), requests);
1109 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1113 void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1114 void *recvbuf, int recvcount, MPI_Datatype recvtype,
1115 int root, MPI_Comm comm)
1117 int system_tag = COLL_TAG_SCATTER;
1118 int rank, size, dst, index;
1119 MPI_Aint lb = 0, sendext = 0;
1120 MPI_Request *requests;
1122 rank = smpi_comm_rank(comm);
1123 size = smpi_comm_size(comm);
1125 // Recv buffer from root
1126 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1129 // FIXME: check for errors
1130 smpi_datatype_extent(sendtype, &lb, &sendext);
1131 // Local copy from root
1132 if(recvbuf!=MPI_IN_PLACE){
1133 smpi_datatype_copy((char *)sendbuf + root * sendcount * sendext,
1134 sendcount, sendtype, recvbuf, recvcount, recvtype);
1136 // Send buffers to receivers
1137 requests = xbt_new(MPI_Request, size - 1);
1139 for(dst = 0; dst < size; dst++) {
1141 requests[index] = smpi_isend_init((char *)sendbuf + dst * sendcount * sendext,
1142 sendcount, sendtype, dst,
1147 // Wait for completion of isend's.
1148 smpi_mpi_startall(size - 1, requests);
1149 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1154 void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs,
1155 MPI_Datatype sendtype, void *recvbuf, int recvcount,
1156 MPI_Datatype recvtype, int root, MPI_Comm comm)
1158 int system_tag = COLL_TAG_SCATTERV;
1159 int rank, size, dst, index;
1160 MPI_Aint lb = 0, sendext = 0;
1161 MPI_Request *requests;
1163 rank = smpi_comm_rank(comm);
1164 size = smpi_comm_size(comm);
1166 // Recv buffer from root
1167 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1170 // FIXME: check for errors
1171 smpi_datatype_extent(sendtype, &lb, &sendext);
1172 // Local copy from root
1173 if(recvbuf!=MPI_IN_PLACE){
1174 smpi_datatype_copy((char *)sendbuf + displs[root] * sendext, sendcounts[root],
1175 sendtype, recvbuf, recvcount, recvtype);
1177 // Send buffers to receivers
1178 requests = xbt_new(MPI_Request, size - 1);
1180 for(dst = 0; dst < size; dst++) {
1183 smpi_isend_init((char *)sendbuf + displs[dst] * sendext, sendcounts[dst],
1184 sendtype, dst, system_tag, comm);
1188 // Wait for completion of isend's.
1189 smpi_mpi_startall(size - 1, requests);
1190 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1195 void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count,
1196 MPI_Datatype datatype, MPI_Op op, int root,
1199 int system_tag = COLL_TAG_REDUCE;
1200 int rank, size, src, index;
1201 MPI_Aint lb = 0, dataext = 0;
1202 MPI_Request *requests;
1206 char* sendtmpbuf = (char*) sendbuf;
1207 if( sendbuf == MPI_IN_PLACE ) {
1208 sendtmpbuf = (char *)xbt_malloc(count*smpi_datatype_get_extent(datatype));
1209 smpi_datatype_copy(recvbuf, count, datatype,sendtmpbuf, count, datatype);
1212 rank = smpi_comm_rank(comm);
1213 size = smpi_comm_size(comm);
1214 //non commutative case, use a working algo from openmpi
1215 if(!smpi_op_is_commute(op)){
1216 smpi_coll_tuned_reduce_ompi_basic_linear(sendtmpbuf, recvbuf, count,
1217 datatype, op, root, comm);
1222 // Send buffer to root
1223 smpi_mpi_send(sendtmpbuf, count, datatype, root, system_tag, comm);
1225 // FIXME: check for errors
1226 smpi_datatype_extent(datatype, &lb, &dataext);
1227 // Local copy from root
1228 if (sendtmpbuf && recvbuf)
1229 smpi_datatype_copy(sendtmpbuf, count, datatype, recvbuf, count, datatype);
1230 // Receive buffers from senders
1231 //TODO: make a MPI_barrier here ?
1232 requests = xbt_new(MPI_Request, size - 1);
1233 tmpbufs = xbt_new(void *, size - 1);
1235 for(src = 0; src < size; src++) {
1237 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1239 tmpbufs[index] = xbt_malloc(count * dataext);
1241 smpi_irecv_init(tmpbufs[index], count, datatype, src,
1246 // Wait for completion of irecv's.
1247 smpi_mpi_startall(size - 1, requests);
1248 for(src = 0; src < size - 1; src++) {
1249 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1250 XBT_DEBUG("finished waiting any request with index %d", index);
1251 if(index == MPI_UNDEFINED) {
1254 if(op) /* op can be MPI_OP_NULL that does nothing */
1255 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1257 for(index = 0; index < size - 1; index++) {
1258 xbt_free(tmpbufs[index]);
1263 if( sendbuf == MPI_IN_PLACE ) {
1264 xbt_free(sendtmpbuf);
1269 void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count,
1270 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1272 smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, 0, comm);
1273 smpi_mpi_bcast(recvbuf, count, datatype, 0, comm);
1276 void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count,
1277 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1279 int system_tag = -888;
1280 int rank, size, other, index;
1281 MPI_Aint lb = 0, dataext = 0;
1282 MPI_Request *requests;
1285 rank = smpi_comm_rank(comm);
1286 size = smpi_comm_size(comm);
1288 // FIXME: check for errors
1289 smpi_datatype_extent(datatype, &lb, &dataext);
1291 // Local copy from self
1292 smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
1294 // Send/Recv buffers to/from others;
1295 requests = xbt_new(MPI_Request, size - 1);
1296 tmpbufs = xbt_new(void *, rank);
1298 for(other = 0; other < rank; other++) {
1299 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1301 tmpbufs[index] = xbt_malloc(count * dataext);
1303 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1307 for(other = rank + 1; other < size; other++) {
1309 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1312 // Wait for completion of all comms.
1313 smpi_mpi_startall(size - 1, requests);
1315 if(smpi_op_is_commute(op)){
1316 for(other = 0; other < size - 1; other++) {
1317 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1318 if(index == MPI_UNDEFINED) {
1322 // #Request is below rank: it's a irecv
1323 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1327 //non commutative case, wait in order
1328 for(other = 0; other < size - 1; other++) {
1329 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1331 smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1335 for(index = 0; index < rank; index++) {
1336 xbt_free(tmpbufs[index]);
1342 void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count,
1343 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1345 int system_tag = -888;
1346 int rank, size, other, index;
1347 MPI_Aint lb = 0, dataext = 0;
1348 MPI_Request *requests;
1350 int recvbuf_is_empty=1;
1351 rank = smpi_comm_rank(comm);
1352 size = smpi_comm_size(comm);
1354 // FIXME: check for errors
1355 smpi_datatype_extent(datatype, &lb, &dataext);
1357 // Send/Recv buffers to/from others;
1358 requests = xbt_new(MPI_Request, size - 1);
1359 tmpbufs = xbt_new(void *, rank);
1361 for(other = 0; other < rank; other++) {
1362 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1364 tmpbufs[index] = xbt_malloc(count * dataext);
1366 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1370 for(other = rank + 1; other < size; other++) {
1372 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1375 // Wait for completion of all comms.
1376 smpi_mpi_startall(size - 1, requests);
1377 if(smpi_op_is_commute(op)){
1378 for(other = 0; other < size - 1; other++) {
1379 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1380 if(index == MPI_UNDEFINED) {
1384 if(recvbuf_is_empty){
1385 smpi_datatype_copy(tmpbufs[index], count, datatype, recvbuf, count, datatype);
1388 // #Request is below rank: it's a irecv
1389 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1393 //non commutative case, wait in order
1394 for(other = 0; other < size - 1; other++) {
1395 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1397 if(recvbuf_is_empty){
1398 smpi_datatype_copy(tmpbufs[other], count, datatype, recvbuf, count, datatype);
1400 }else smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1404 for(index = 0; index < rank; index++) {
1405 xbt_free(tmpbufs[index]);