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);
84 static xbt_dynar_t parse_factor(const char *smpi_coef_string)
87 unsigned int iter = 0;
90 xbt_dynar_t smpi_factor, radical_elements, radical_elements2 = NULL;
92 smpi_factor = xbt_dynar_new(sizeof(s_smpi_factor_t), NULL);
93 radical_elements = xbt_str_split(smpi_coef_string, ";");
94 xbt_dynar_foreach(radical_elements, iter, value) {
96 radical_elements2 = xbt_str_split(value, ":");
97 if (xbt_dynar_length(radical_elements2) <2 || xbt_dynar_length(radical_elements2) > 5)
98 xbt_die("Malformed radical for smpi factor!");
99 for(i =0; i<xbt_dynar_length(radical_elements2);i++ ){
101 fact.factor = atol(xbt_dynar_get_as(radical_elements2, i, char *));
103 fact.values[fact.nb_values] = atof(xbt_dynar_get_as(radical_elements2, i, char *));
108 xbt_dynar_push_as(smpi_factor, s_smpi_factor_t, fact);
109 XBT_DEBUG("smpi_factor:\t%ld : %d values, first: %f", fact.factor, fact.nb_values ,fact.values[0]);
110 xbt_dynar_free(&radical_elements2);
112 xbt_dynar_free(&radical_elements);
114 xbt_dynar_sort(smpi_factor, &factor_cmp);
115 xbt_dynar_foreach(smpi_factor, iter, fact) {
116 XBT_DEBUG("smpi_factor:\t%ld : %d values, first: %f", fact.factor, fact.nb_values ,fact.values[0]);
121 static double smpi_os(double size)
123 if (!smpi_os_values) {
124 smpi_os_values = parse_factor(sg_cfg_get_string("smpi/os"));
125 smpi_register_static(smpi_os_values, xbt_dynar_free_voidp);
127 unsigned int iter = 0;
128 s_smpi_factor_t fact;
130 xbt_dynar_foreach(smpi_os_values, iter, fact) {
131 if (size <= fact.factor) {
132 XBT_DEBUG("os : %lf <= %ld return %f", size, fact.factor, current);
135 current=fact.values[0]+fact.values[1]*size;
138 XBT_DEBUG("os : %lf > %ld return %f", size, fact.factor, current);
143 static double smpi_ois(double size)
145 if (!smpi_ois_values) {
146 smpi_ois_values = parse_factor(sg_cfg_get_string("smpi/ois"));
147 smpi_register_static(smpi_ois_values, xbt_dynar_free_voidp);
149 unsigned int iter = 0;
150 s_smpi_factor_t fact;
152 xbt_dynar_foreach(smpi_ois_values, iter, fact) {
153 if (size <= fact.factor) {
154 XBT_DEBUG("ois : %lf <= %ld return %f", size, fact.factor, current);
157 current=fact.values[0]+fact.values[1]*size;
160 XBT_DEBUG("ois : %lf > %ld return %f", size, fact.factor, current);
165 static double smpi_or(double size)
167 if (!smpi_or_values) {
168 smpi_or_values = parse_factor(sg_cfg_get_string("smpi/or"));
169 smpi_register_static(smpi_or_values, xbt_dynar_free_voidp);
171 unsigned int iter = 0;
172 s_smpi_factor_t fact;
174 xbt_dynar_foreach(smpi_or_values, iter, fact) {
175 if (size <= fact.factor) {
176 XBT_DEBUG("or : %lf <= %ld return %f", size, fact.factor, current);
179 current=fact.values[0]+fact.values[1]*size;
181 XBT_DEBUG("or : %lf > %ld return %f", size, fact.factor, current);
186 static MPI_Request build_request(void *buf, int count,
187 MPI_Datatype datatype, int src, int dst,
188 int tag, MPI_Comm comm, unsigned flags)
190 MPI_Request request = NULL;
192 void *old_buf = NULL;
194 request = xbt_new(s_smpi_mpi_request_t, 1);
196 s_smpi_subtype_t *subtype = datatype->substruct;
198 if(datatype->has_subtype == 1){
199 // This part handles the problem of non-contiguous memory
201 buf = count==0 ? NULL : xbt_malloc(count*smpi_datatype_size(datatype));
203 subtype->serialize(old_buf, buf, count, datatype->substruct);
208 // This part handles the problem of non-contiguous memory (for the
209 // unserialisation at the reception)
210 request->old_buf = old_buf;
211 request->old_type = datatype;
213 request->size = smpi_datatype_size(datatype) * count;
217 request->comm = comm;
218 request->action = NULL;
219 request->flags = flags;
220 request->detached = 0;
221 request->detached_sender = NULL;
222 request->real_src = 0;
224 request->truncated = 0;
225 request->real_size = 0;
226 request->real_tag = 0;
233 if (flags & SEND) smpi_datatype_unuse(datatype);
239 void smpi_empty_status(MPI_Status * status)
241 if(status != MPI_STATUS_IGNORE) {
242 status->MPI_SOURCE = MPI_ANY_SOURCE;
243 status->MPI_TAG = MPI_ANY_TAG;
244 status->MPI_ERROR = MPI_SUCCESS;
249 void smpi_action_trace_run(char *path)
253 xbt_dict_cursor_t cursor;
257 action_fp = fopen(path, "r");
258 xbt_assert(action_fp != NULL, "Cannot open %s: %s", path,
262 if (!xbt_dict_is_empty(action_queues)) {
264 ("Not all actions got consumed. If the simulation ended successfully (without deadlock), you may want to add new processes to your deployment file.");
267 xbt_dict_foreach(action_queues, cursor, name, todo) {
268 XBT_WARN("Still %lu actions for %s", xbt_dynar_length(todo), name);
274 xbt_dict_free(&action_queues);
275 action_queues = xbt_dict_new_homogeneous(NULL);
278 static void smpi_mpi_request_free_voidp(void* request)
280 MPI_Request req = request;
281 smpi_mpi_request_free(&req);
284 /* MPI Low level calls */
285 MPI_Request smpi_mpi_send_init(void *buf, int count, MPI_Datatype datatype,
286 int dst, int tag, MPI_Comm comm)
288 MPI_Request request = NULL;
289 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
290 comm, PERSISTENT | SEND | PREPARED);
295 MPI_Request smpi_mpi_ssend_init(void *buf, int count, MPI_Datatype datatype,
296 int dst, int tag, MPI_Comm comm)
298 MPI_Request request = NULL;
299 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
300 comm, PERSISTENT | SSEND | SEND | PREPARED);
305 MPI_Request smpi_mpi_recv_init(void *buf, int count, MPI_Datatype datatype,
306 int src, int tag, MPI_Comm comm)
308 MPI_Request request = NULL;
309 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,
310 comm, PERSISTENT | RECV | PREPARED);
315 void smpi_mpi_start(MPI_Request request)
319 xbt_assert(!request->action,
320 "Cannot (re)start a non-finished communication");
321 if(request->flags & PREPARED)request->flags &= ~PREPARED;
322 if(request->flags & RECV) {
323 print_request("New recv", request);
324 //FIXME: if receive is posted with a large size, but send is smaller, mailboxes may not match !
325 if (request->size < sg_cfg_get_int("smpi/async_small_thres"))
326 mailbox = smpi_process_mailbox_small();
328 mailbox = smpi_process_mailbox();
329 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
330 request->real_size=request->size;
331 smpi_datatype_use(request->old_type);
332 smpi_comm_use(request->comm);
333 request->action = simcall_comm_irecv(mailbox, request->buf, &request->real_size, &match_recv, request);
335 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
336 double sleeptime = request->detached ? smpi_or(request->size) : 0.0;
338 simcall_process_sleep(sleeptime);
339 XBT_DEBUG("receiving size of %zu : sleep %lf ", request->size, smpi_or(request->size));
345 int receiver = request->dst;//smpi_group_index(smpi_comm_group(request->comm), request->dst);
348 int rank = smpi_process_index();
349 if (TRACE_smpi_view_internals()) {
350 TRACE_smpi_send(rank, rank, receiver,request->size);
353 /* if(receiver == MPI_UNDEFINED) {*/
354 /* XBT_WARN("Trying to send a message to a wrong rank");*/
357 print_request("New send", request);
358 if (request->size < sg_cfg_get_int("smpi/async_small_thres")) { // eager mode
359 mailbox = smpi_process_remote_mailbox_small(receiver);
361 XBT_DEBUG("Send request %p is not in the permanent receive mailbox (buf: %p)",request,request->buf);
362 mailbox = smpi_process_remote_mailbox(receiver);
364 if ( (! (request->flags & SSEND)) && (request->size < sg_cfg_get_int("smpi/send_is_detached_thres"))) {
366 request->detached = 1;
368 if(request->old_type->has_subtype == 0){
369 oldbuf = request->buf;
370 if (oldbuf && request->size!=0){
371 request->buf = xbt_malloc(request->size);
372 memcpy(request->buf,oldbuf,request->size);
375 XBT_DEBUG("Send request %p is detached; buf %p copied into %p",request,oldbuf,request->buf);
378 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
379 request->real_size=request->size;
380 smpi_datatype_use(request->old_type);
381 smpi_comm_use(request->comm);
383 //if we are giving back the control to the user without waiting for completion, we have to inject timings
384 double sleeptime =0.0;
385 if(request->detached || (request->flags & (ISEND|SSEND))){// issend should be treated as isend
386 //isend and send timings may be different
387 sleeptime = (request->flags & ISEND)? smpi_ois(request->size) : smpi_os(request->size);
391 simcall_process_sleep(sleeptime);
392 XBT_DEBUG("sending size of %zu : sleep %lf ", request->size, smpi_os(request->size));
396 simcall_comm_isend(mailbox, request->size, -1.0,
397 request->buf, request->real_size,
399 &smpi_mpi_request_free_voidp, // how to free the userdata if a detached send fails
401 // detach if msg size < eager/rdv switch limit
405 /* FIXME: detached sends are not traceable (request->action == NULL) */
407 simcall_set_category(request->action, TRACE_internal_smpi_get_category());
415 void smpi_mpi_startall(int count, MPI_Request * requests)
419 for(i = 0; i < count; i++) {
420 smpi_mpi_start(requests[i]);
424 void smpi_mpi_request_free(MPI_Request * request)
426 if((*request) != MPI_REQUEST_NULL){
427 (*request)->refcount--;
428 if((*request)->refcount<0) xbt_die("wrong refcount");
430 if((*request)->refcount==0){
431 print_request("Destroying", (*request));
433 *request = MPI_REQUEST_NULL;
435 print_request("Decrementing", (*request));
439 xbt_die("freeing an already free request");
443 MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype,
444 int dst, int tag, MPI_Comm comm)
446 MPI_Request request = NULL;
447 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf , count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
448 comm, NON_PERSISTENT | ISEND | SEND | PREPARED);
452 MPI_Request smpi_mpi_isend(void *buf, int count, MPI_Datatype datatype,
453 int dst, int tag, MPI_Comm comm)
455 MPI_Request request = NULL;
456 request = build_request(buf==MPI_BOTTOM?(void*)0:buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
457 comm, NON_PERSISTENT | ISEND | SEND);
458 smpi_mpi_start(request);
462 MPI_Request smpi_mpi_issend(void *buf, int count, MPI_Datatype datatype,
463 int dst, int tag, MPI_Comm comm)
465 MPI_Request request = NULL;
466 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
467 comm, NON_PERSISTENT | ISEND | SSEND | SEND);
468 smpi_mpi_start(request);
474 MPI_Request smpi_irecv_init(void *buf, int count, MPI_Datatype datatype,
475 int src, int tag, MPI_Comm comm)
477 MPI_Request request = NULL;
478 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,
479 comm, NON_PERSISTENT | RECV | PREPARED);
483 MPI_Request smpi_mpi_irecv(void *buf, int count, MPI_Datatype datatype,
484 int src, int tag, MPI_Comm comm)
486 MPI_Request request = NULL;
487 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,
488 comm, NON_PERSISTENT | RECV);
489 smpi_mpi_start(request);
493 void smpi_mpi_recv(void *buf, int count, MPI_Datatype datatype, int src,
494 int tag, MPI_Comm comm, MPI_Status * status)
496 MPI_Request request = NULL;
497 request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm);
498 smpi_mpi_wait(&request, status);
504 void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst,
505 int tag, MPI_Comm comm)
507 MPI_Request request = NULL;
508 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
509 comm, NON_PERSISTENT | SEND);
511 smpi_mpi_start(request);
512 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
516 void smpi_mpi_ssend(void *buf, int count, MPI_Datatype datatype,
517 int dst, int tag, MPI_Comm comm)
519 MPI_Request request = NULL;
520 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
521 comm, NON_PERSISTENT | SSEND | SEND);
523 smpi_mpi_start(request);
524 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
528 void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
529 int dst, int sendtag, void *recvbuf, int recvcount,
530 MPI_Datatype recvtype, int src, int recvtag,
531 MPI_Comm comm, MPI_Status * status)
533 MPI_Request requests[2];
535 int myid=smpi_process_index();
536 if ((smpi_group_index(smpi_comm_group(comm), dst) == myid) && (smpi_group_index(smpi_comm_group(comm), src) == myid)) {
537 smpi_datatype_copy(sendbuf, sendcount, sendtype,
538 recvbuf, recvcount, recvtype);
542 smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
544 smpi_irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
545 smpi_mpi_startall(2, requests);
546 smpi_mpi_waitall(2, requests, stats);
547 if(status != MPI_STATUS_IGNORE) {
548 // Copy receive status
553 int smpi_mpi_get_count(MPI_Status * status, MPI_Datatype datatype)
555 return status->count / smpi_datatype_size(datatype);
558 static void finish_wait(MPI_Request * request, MPI_Status * status)
560 MPI_Request req = *request;
561 smpi_empty_status(status);
563 if(!(req->detached && req->flags & SEND) && !(req->flags & PREPARED)){
564 if(status != MPI_STATUS_IGNORE) {
565 int src = req->src == MPI_ANY_SOURCE ? req->real_src : req->src;
566 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(req->comm), src);
567 status->MPI_TAG = req->tag == MPI_ANY_TAG ? req->real_tag : req->tag;
568 status->MPI_ERROR = req->truncated ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
569 // this handles the case were size in receive differs from size in send
570 // FIXME: really this should just contain the count of receive-type blocks,
572 status->count = req->real_size;
575 print_request("Finishing", req);
576 MPI_Datatype datatype = req->old_type;
578 if(datatype->has_subtype == 1){
579 // This part handles the problem of non-contignous memory
580 // the unserialization at the reception
581 s_smpi_subtype_t *subtype = datatype->substruct;
582 if(req->flags & RECV) {
583 subtype->unserialize(req->buf, req->old_buf, req->real_size/smpi_datatype_size(datatype) , datatype->substruct);
585 if(req->detached == 0) free(req->buf);
587 smpi_comm_unuse(req->comm);
588 smpi_datatype_unuse(datatype);
593 if (TRACE_smpi_view_internals()) {
594 if(req->flags & RECV){
595 int rank = smpi_process_index();
596 int src_traced = (req->src == MPI_ANY_SOURCE ? req->real_src : req->src);
597 TRACE_smpi_recv(rank, src_traced, rank);
602 if(req->detached_sender!=NULL){
603 smpi_mpi_request_free(&(req->detached_sender));
606 if(req->flags & NON_PERSISTENT) {
607 smpi_mpi_request_free(request);
613 int smpi_mpi_test(MPI_Request * request, MPI_Status * status) {
616 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before)
617 if ((*request)->action == NULL)
620 flag = simcall_comm_test((*request)->action);
622 finish_wait(request, status);
623 *request = MPI_REQUEST_NULL;
625 smpi_empty_status(status);
630 int smpi_mpi_testany(int count, MPI_Request requests[], int *index,
637 *index = MPI_UNDEFINED;
639 comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
640 map = xbt_new(int, count);
642 for(i = 0; i < count; i++) {
643 if((requests[i]!=MPI_REQUEST_NULL) && requests[i]->action) {
644 xbt_dynar_push(comms, &requests[i]->action);
650 i = simcall_comm_testany(comms);
651 // not MPI_UNDEFINED, as this is a simix return code
654 finish_wait(&requests[*index], status);
655 requests[*index] = MPI_REQUEST_NULL;
659 //all requests are null or inactive, return true
661 smpi_empty_status(status);
664 xbt_dynar_free(&comms);
670 int smpi_mpi_testall(int count, MPI_Request requests[],
674 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
677 for(i=0; i<count; i++){
678 if(requests[i]!= MPI_REQUEST_NULL){
679 if (smpi_mpi_test(&requests[i], pstat)!=1){
682 requests[i]=MPI_REQUEST_NULL;
685 smpi_empty_status(pstat);
687 if(status != MPI_STATUSES_IGNORE) {
694 void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
696 //FIXME find another wait to avoid busy waiting ?
697 // the issue here is that we have to wait on a nonexistent comm
699 smpi_mpi_iprobe(source, tag, comm, &flag, status);
700 XBT_DEBUG("Busy Waiting on probing : %d", flag);
704 void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
706 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,
707 comm, NON_PERSISTENT | RECV);
709 //to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
710 double sleeptime= sg_cfg_get_double("smpi/iprobe");
711 //multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
712 static int nsleeps = 1;
714 simcall_process_sleep(sleeptime);
716 // behave like a receive, but don't do it
719 print_request("New iprobe", request);
720 // We have to test both mailboxes as we don't know if we will receive one one or another
721 if (sg_cfg_get_int("smpi/async_small_thres")>0){
722 mailbox = smpi_process_mailbox_small();
723 XBT_DEBUG("trying to probe the perm recv mailbox");
724 request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
726 if (request->action==NULL){
727 mailbox = smpi_process_mailbox();
728 XBT_DEBUG("trying to probe the other mailbox");
729 request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
733 MPI_Request req = (MPI_Request)SIMIX_comm_get_src_data(request->action);
735 if(status != MPI_STATUS_IGNORE && !(req->flags & PREPARED)) {
736 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(comm), req->src);
737 status->MPI_TAG = req->tag;
738 status->MPI_ERROR = MPI_SUCCESS;
739 status->count = req->real_size;
741 nsleeps=1;//reset the number of sleeps we will do next time
747 smpi_mpi_request_free(&request);
752 void smpi_mpi_wait(MPI_Request * request, MPI_Status * status)
754 print_request("Waiting", *request);
755 if ((*request)->action != NULL) { // this is not a detached send
756 simcall_comm_wait((*request)->action, -1.0);
761 (*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
764 finish_wait(request, status);
765 *request = MPI_REQUEST_NULL;
766 // FIXME for a detached send, finish_wait is not called:
769 int smpi_mpi_waitany(int count, MPI_Request requests[],
776 index = MPI_UNDEFINED;
778 // Wait for a request to complete
779 comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
780 map = xbt_new(int, count);
782 XBT_DEBUG("Wait for one of %d", count);
783 for(i = 0; i < count; i++) {
784 if(requests[i] != MPI_REQUEST_NULL) {
785 if (requests[i]->action != NULL) {
786 XBT_DEBUG("Waiting any %p ", requests[i]);
787 xbt_dynar_push(comms, &requests[i]->action);
791 //This is a finished detached request, let's return this one
792 size=0;//so we free the dynar but don't do the waitany call
794 finish_wait(&requests[i], status);//cleanup if refcount = 0
795 requests[i]=MPI_REQUEST_NULL;//set to null
801 i = simcall_comm_waitany(comms);
803 // not MPI_UNDEFINED, as this is a simix return code
806 finish_wait(&requests[index], status);
807 requests[index] = MPI_REQUEST_NULL;
811 xbt_dynar_free(&comms);
814 if (index==MPI_UNDEFINED)
815 smpi_empty_status(status);
820 int smpi_mpi_waitall(int count, MPI_Request requests[],
825 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
826 int retvalue = MPI_SUCCESS;
827 //tag invalid requests in the set
828 if (status != MPI_STATUSES_IGNORE) {
829 for (c = 0; c < count; c++) {
830 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL) {
831 smpi_empty_status(&status[c]);
832 } else if (requests[c]->src == MPI_PROC_NULL) {
833 smpi_empty_status(&status[c]);
834 status[c].MPI_SOURCE = MPI_PROC_NULL;
838 for(c = 0; c < count; c++) {
839 if (MC_is_active()) {
840 smpi_mpi_wait(&requests[c], pstat);
843 index = smpi_mpi_waitany(count, requests, pstat);
844 if (index == MPI_UNDEFINED)
846 requests[index]=MPI_REQUEST_NULL;
848 if (status != MPI_STATUSES_IGNORE) {
849 status[index] = *pstat;
850 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
851 retvalue = MPI_ERR_IN_STATUS;
858 int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices,
863 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
866 for(i = 0; i < incount; i++)
868 index=smpi_mpi_waitany(incount, requests, pstat);
869 if(index!=MPI_UNDEFINED){
870 indices[count] = index;
872 if(status != MPI_STATUSES_IGNORE) {
873 status[index] = *pstat;
875 requests[index]=MPI_REQUEST_NULL;
877 return MPI_UNDEFINED;
883 int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices,
886 int i, count, count_dead;
888 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
892 for(i = 0; i < incount; i++) {
893 if((requests[i] != MPI_REQUEST_NULL)) {
894 if(smpi_mpi_test(&requests[i], pstat)) {
897 if(status != MPI_STATUSES_IGNORE) {
900 requests[i]=MPI_REQUEST_NULL;
907 if(count_dead==incount)return MPI_UNDEFINED;
911 void smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root,
914 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
915 nary_tree_bcast(buf, count, datatype, root, comm, 4);
918 void smpi_mpi_barrier(MPI_Comm comm)
920 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
921 nary_tree_barrier(comm, 4);
924 void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
925 void *recvbuf, int recvcount, MPI_Datatype recvtype,
926 int root, MPI_Comm comm)
928 int system_tag = COLL_TAG_GATHER;
929 int rank, size, src, index;
930 MPI_Aint lb = 0, recvext = 0;
931 MPI_Request *requests;
933 rank = smpi_comm_rank(comm);
934 size = smpi_comm_size(comm);
936 // Send buffer to root
937 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
939 // FIXME: check for errors
940 smpi_datatype_extent(recvtype, &lb, &recvext);
941 // Local copy from root
942 smpi_datatype_copy(sendbuf, sendcount, sendtype,
943 (char *)recvbuf + root * recvcount * recvext, recvcount, recvtype);
944 // Receive buffers from senders
945 requests = xbt_new(MPI_Request, size - 1);
947 for(src = 0; src < size; src++) {
949 requests[index] = smpi_irecv_init((char *)recvbuf + src * recvcount * recvext,
951 src, system_tag, comm);
955 // Wait for completion of irecv's.
956 smpi_mpi_startall(size - 1, requests);
957 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
963 void smpi_mpi_reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts,
964 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
968 int rank = smpi_process_index();
971 /* arbitrarily choose root as rank 0 */
972 size = smpi_comm_size(comm);
974 displs = xbt_new(int, size);
975 for (i = 0; i < size; i++) {
977 count += recvcounts[i];
979 tmpbuf=(void*)xbt_malloc(count*smpi_datatype_get_extent(datatype));
980 mpi_coll_reduce_fun(sendbuf, tmpbuf, count, datatype, op, 0, comm);
981 smpi_mpi_scatterv(tmpbuf, recvcounts, displs, datatype, recvbuf,
982 recvcounts[rank], datatype, 0, comm);
987 void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
988 void *recvbuf, int *recvcounts, int *displs,
989 MPI_Datatype recvtype, int root, MPI_Comm comm)
991 int system_tag = COLL_TAG_GATHERV;
992 int rank, size, src, index;
993 MPI_Aint lb = 0, recvext = 0;
994 MPI_Request *requests;
996 rank = smpi_comm_rank(comm);
997 size = smpi_comm_size(comm);
999 // Send buffer to root
1000 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
1002 // FIXME: check for errors
1003 smpi_datatype_extent(recvtype, &lb, &recvext);
1004 // Local copy from root
1005 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1006 (char *)recvbuf + displs[root] * recvext,
1007 recvcounts[root], recvtype);
1008 // Receive buffers from senders
1009 requests = xbt_new(MPI_Request, size - 1);
1011 for(src = 0; src < size; src++) {
1014 smpi_irecv_init((char *)recvbuf + displs[src] * recvext,
1015 recvcounts[src], recvtype, src, system_tag, comm);
1019 // Wait for completion of irecv's.
1020 smpi_mpi_startall(size - 1, requests);
1021 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1026 void smpi_mpi_allgather(void *sendbuf, int sendcount,
1027 MPI_Datatype sendtype, void *recvbuf,
1028 int recvcount, MPI_Datatype recvtype,
1031 int system_tag = COLL_TAG_ALLGATHER;
1032 int rank, size, other, index;
1033 MPI_Aint lb = 0, recvext = 0;
1034 MPI_Request *requests;
1036 rank = smpi_comm_rank(comm);
1037 size = smpi_comm_size(comm);
1038 // FIXME: check for errors
1039 smpi_datatype_extent(recvtype, &lb, &recvext);
1040 // Local copy from self
1041 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1042 (char *)recvbuf + rank * recvcount * recvext, recvcount,
1044 // Send/Recv buffers to/from others;
1045 requests = xbt_new(MPI_Request, 2 * (size - 1));
1047 for(other = 0; other < size; other++) {
1050 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1053 requests[index] = smpi_irecv_init((char *)recvbuf + other * recvcount * recvext,
1054 recvcount, recvtype, other,
1059 // Wait for completion of all comms.
1060 smpi_mpi_startall(2 * (size - 1), requests);
1061 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1065 void smpi_mpi_allgatherv(void *sendbuf, int sendcount,
1066 MPI_Datatype sendtype, void *recvbuf,
1067 int *recvcounts, int *displs,
1068 MPI_Datatype recvtype, MPI_Comm comm)
1070 int system_tag = COLL_TAG_ALLGATHERV;
1071 int rank, size, other, index;
1072 MPI_Aint lb = 0, recvext = 0;
1073 MPI_Request *requests;
1075 rank = smpi_comm_rank(comm);
1076 size = smpi_comm_size(comm);
1077 // FIXME: check for errors
1078 smpi_datatype_extent(recvtype, &lb, &recvext);
1079 // Local copy from self
1080 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1081 (char *)recvbuf + displs[rank] * recvext,
1082 recvcounts[rank], recvtype);
1083 // Send buffers to others;
1084 requests = xbt_new(MPI_Request, 2 * (size - 1));
1086 for(other = 0; other < size; other++) {
1089 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1093 smpi_irecv_init((char *)recvbuf + displs[other] * recvext, recvcounts[other],
1094 recvtype, other, system_tag, comm);
1098 // Wait for completion of all comms.
1099 smpi_mpi_startall(2 * (size - 1), requests);
1100 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1104 void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1105 void *recvbuf, int recvcount, MPI_Datatype recvtype,
1106 int root, MPI_Comm comm)
1108 int system_tag = COLL_TAG_SCATTER;
1109 int rank, size, dst, index;
1110 MPI_Aint lb = 0, sendext = 0;
1111 MPI_Request *requests;
1113 rank = smpi_comm_rank(comm);
1114 size = smpi_comm_size(comm);
1116 // Recv buffer from root
1117 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1120 // FIXME: check for errors
1121 smpi_datatype_extent(sendtype, &lb, &sendext);
1122 // Local copy from root
1123 if(recvbuf!=MPI_IN_PLACE){
1124 smpi_datatype_copy((char *)sendbuf + root * sendcount * sendext,
1125 sendcount, sendtype, recvbuf, recvcount, recvtype);
1127 // Send buffers to receivers
1128 requests = xbt_new(MPI_Request, size - 1);
1130 for(dst = 0; dst < size; dst++) {
1132 requests[index] = smpi_isend_init((char *)sendbuf + dst * sendcount * sendext,
1133 sendcount, sendtype, dst,
1138 // Wait for completion of isend's.
1139 smpi_mpi_startall(size - 1, requests);
1140 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1145 void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs,
1146 MPI_Datatype sendtype, void *recvbuf, int recvcount,
1147 MPI_Datatype recvtype, int root, MPI_Comm comm)
1149 int system_tag = COLL_TAG_SCATTERV;
1150 int rank, size, dst, index;
1151 MPI_Aint lb = 0, sendext = 0;
1152 MPI_Request *requests;
1154 rank = smpi_comm_rank(comm);
1155 size = smpi_comm_size(comm);
1157 // Recv buffer from root
1158 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1161 // FIXME: check for errors
1162 smpi_datatype_extent(sendtype, &lb, &sendext);
1163 // Local copy from root
1164 if(recvbuf!=MPI_IN_PLACE){
1165 smpi_datatype_copy((char *)sendbuf + displs[root] * sendext, sendcounts[root],
1166 sendtype, recvbuf, recvcount, recvtype);
1168 // Send buffers to receivers
1169 requests = xbt_new(MPI_Request, size - 1);
1171 for(dst = 0; dst < size; dst++) {
1174 smpi_isend_init((char *)sendbuf + displs[dst] * sendext, sendcounts[dst],
1175 sendtype, dst, system_tag, comm);
1179 // Wait for completion of isend's.
1180 smpi_mpi_startall(size - 1, requests);
1181 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1186 void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count,
1187 MPI_Datatype datatype, MPI_Op op, int root,
1190 int system_tag = COLL_TAG_REDUCE;
1191 int rank, size, src, index;
1192 MPI_Aint lb = 0, dataext = 0;
1193 MPI_Request *requests;
1197 char* sendtmpbuf = (char*) sendbuf;
1198 if( sendbuf == MPI_IN_PLACE ) {
1199 sendtmpbuf = (char *)xbt_malloc(count*smpi_datatype_get_extent(datatype));
1200 smpi_datatype_copy(recvbuf, count, datatype,sendtmpbuf, count, datatype);
1203 rank = smpi_comm_rank(comm);
1204 size = smpi_comm_size(comm);
1205 //non commutative case, use a working algo from openmpi
1206 if(!smpi_op_is_commute(op)){
1207 smpi_coll_tuned_reduce_ompi_basic_linear(sendtmpbuf, recvbuf, count,
1208 datatype, op, root, comm);
1213 // Send buffer to root
1214 smpi_mpi_send(sendtmpbuf, count, datatype, root, system_tag, comm);
1216 // FIXME: check for errors
1217 smpi_datatype_extent(datatype, &lb, &dataext);
1218 // Local copy from root
1219 if (sendtmpbuf && recvbuf)
1220 smpi_datatype_copy(sendtmpbuf, count, datatype, recvbuf, count, datatype);
1221 // Receive buffers from senders
1222 //TODO: make a MPI_barrier here ?
1223 requests = xbt_new(MPI_Request, size - 1);
1224 tmpbufs = xbt_new(void *, size - 1);
1226 for(src = 0; src < size; src++) {
1228 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1230 tmpbufs[index] = xbt_malloc(count * dataext);
1232 smpi_irecv_init(tmpbufs[index], count, datatype, src,
1237 // Wait for completion of irecv's.
1238 smpi_mpi_startall(size - 1, requests);
1239 for(src = 0; src < size - 1; src++) {
1240 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1241 XBT_DEBUG("finished waiting any request with index %d", index);
1242 if(index == MPI_UNDEFINED) {
1245 if(op) /* op can be MPI_OP_NULL that does nothing */
1246 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1248 for(index = 0; index < size - 1; index++) {
1249 xbt_free(tmpbufs[index]);
1254 if( sendbuf == MPI_IN_PLACE ) {
1255 xbt_free(sendtmpbuf);
1260 void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count,
1261 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1263 smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, 0, comm);
1264 smpi_mpi_bcast(recvbuf, count, datatype, 0, comm);
1267 void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count,
1268 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1270 int system_tag = -888;
1271 int rank, size, other, index;
1272 MPI_Aint lb = 0, dataext = 0;
1273 MPI_Request *requests;
1276 rank = smpi_comm_rank(comm);
1277 size = smpi_comm_size(comm);
1279 // FIXME: check for errors
1280 smpi_datatype_extent(datatype, &lb, &dataext);
1282 // Local copy from self
1283 smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
1285 // Send/Recv buffers to/from others;
1286 requests = xbt_new(MPI_Request, size - 1);
1287 tmpbufs = xbt_new(void *, rank);
1289 for(other = 0; other < rank; other++) {
1290 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1292 tmpbufs[index] = xbt_malloc(count * dataext);
1294 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1298 for(other = rank + 1; other < size; other++) {
1300 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1303 // Wait for completion of all comms.
1304 smpi_mpi_startall(size - 1, requests);
1306 if(smpi_op_is_commute(op)){
1307 for(other = 0; other < size - 1; other++) {
1308 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1309 if(index == MPI_UNDEFINED) {
1313 // #Request is below rank: it's a irecv
1314 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1318 //non commutative case, wait in order
1319 for(other = 0; other < size - 1; other++) {
1320 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1322 smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1326 for(index = 0; index < rank; index++) {
1327 xbt_free(tmpbufs[index]);
1333 void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count,
1334 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1336 int system_tag = -888;
1337 int rank, size, other, index;
1338 MPI_Aint lb = 0, dataext = 0;
1339 MPI_Request *requests;
1341 int recvbuf_is_empty=1;
1342 rank = smpi_comm_rank(comm);
1343 size = smpi_comm_size(comm);
1345 // FIXME: check for errors
1346 smpi_datatype_extent(datatype, &lb, &dataext);
1348 // Send/Recv buffers to/from others;
1349 requests = xbt_new(MPI_Request, size - 1);
1350 tmpbufs = xbt_new(void *, rank);
1352 for(other = 0; other < rank; other++) {
1353 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1355 tmpbufs[index] = xbt_malloc(count * dataext);
1357 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1361 for(other = rank + 1; other < size; other++) {
1363 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1366 // Wait for completion of all comms.
1367 smpi_mpi_startall(size - 1, requests);
1368 if(smpi_op_is_commute(op)){
1369 for(other = 0; other < size - 1; other++) {
1370 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1371 if(index == MPI_UNDEFINED) {
1375 if(recvbuf_is_empty){
1376 smpi_datatype_copy(tmpbufs[index], count, datatype, recvbuf, count, datatype);
1379 // #Request is below rank: it's a irecv
1380 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1384 //non commutative case, wait in order
1385 for(other = 0; other < size - 1; other++) {
1386 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1388 if(recvbuf_is_empty){
1389 smpi_datatype_copy(tmpbufs[other], count, datatype, recvbuf, count, datatype);
1391 }else smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1395 for(index = 0; index < rank; index++) {
1396 xbt_free(tmpbufs[index]);