1 /* Copyright (c) 2007, 2008, 2009, 2010. 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)
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;
223 request->truncated = 0;
224 request->real_size = 0;
225 request->real_tag = 0;
232 if (flags & SEND) smpi_datatype_unuse(datatype);
238 void smpi_empty_status(MPI_Status * status)
240 if(status != MPI_STATUS_IGNORE) {
241 status->MPI_SOURCE = MPI_ANY_SOURCE;
242 status->MPI_TAG = MPI_ANY_TAG;
243 status->MPI_ERROR = MPI_SUCCESS;
248 void smpi_action_trace_run(char *path)
252 xbt_dict_cursor_t cursor;
256 action_fp = fopen(path, "r");
257 xbt_assert(action_fp != NULL, "Cannot open %s: %s", path,
261 if (!xbt_dict_is_empty(action_queues)) {
263 ("Not all actions got consumed. If the simulation ended successfully (without deadlock), you may want to add new processes to your deployment file.");
266 xbt_dict_foreach(action_queues, cursor, name, todo) {
267 XBT_WARN("Still %lu actions for %s", xbt_dynar_length(todo), name);
273 xbt_dict_free(&action_queues);
274 action_queues = xbt_dict_new_homogeneous(NULL);
277 static void smpi_mpi_request_free_voidp(void* request)
279 MPI_Request req = request;
280 smpi_mpi_request_free(&req);
283 /* MPI Low level calls */
284 MPI_Request smpi_mpi_send_init(void *buf, int count, MPI_Datatype datatype,
285 int dst, int tag, MPI_Comm comm)
287 MPI_Request request =
288 build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
289 comm, PERSISTENT | SEND | PREPARED);
294 MPI_Request smpi_mpi_ssend_init(void *buf, int count, MPI_Datatype datatype,
295 int dst, int tag, MPI_Comm comm)
297 MPI_Request request =
298 build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
299 comm, PERSISTENT | SSEND | SEND | PREPARED);
304 MPI_Request smpi_mpi_recv_init(void *buf, int count, MPI_Datatype datatype,
305 int src, int tag, MPI_Comm comm)
307 MPI_Request request =
308 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,
309 comm, PERSISTENT | RECV | PREPARED);
314 void smpi_mpi_start(MPI_Request request)
318 xbt_assert(!request->action,
319 "Cannot (re)start a non-finished communication");
320 if(request->flags & PREPARED)request->flags &= ~PREPARED;
321 if(request->flags & RECV) {
322 print_request("New recv", request);
323 if (request->size < sg_cfg_get_int("smpi/async_small_thres"))
324 mailbox = smpi_process_mailbox_small();
326 mailbox = smpi_process_mailbox();
327 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
328 request->real_size=request->size;
329 smpi_datatype_use(request->old_type);
330 smpi_comm_use(request->comm);
331 request->action = simcall_comm_irecv(mailbox, request->buf, &request->real_size, &match_recv, request);
333 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
334 double sleeptime = request->detached ? smpi_or(request->size) : 0.0;
336 simcall_process_sleep(sleeptime);
337 XBT_DEBUG("receiving size of %zu : sleep %lf ", request->size, smpi_or(request->size));
343 int receiver = request->dst;//smpi_group_index(smpi_comm_group(request->comm), request->dst);
346 int rank = smpi_process_index();
347 if (TRACE_smpi_view_internals()) {
348 TRACE_smpi_send(rank, rank, receiver);
351 /* if(receiver == MPI_UNDEFINED) {*/
352 /* XBT_WARN("Trying to send a message to a wrong rank");*/
355 print_request("New send", request);
356 if (request->size < sg_cfg_get_int("smpi/async_small_thres")) { // eager mode
357 mailbox = smpi_process_remote_mailbox_small(receiver);
359 XBT_DEBUG("Send request %p is not in the permanent receive mailbox (buf: %p)",request,request->buf);
360 mailbox = smpi_process_remote_mailbox(receiver);
362 if ( (! (request->flags & SSEND)) && (request->size < sg_cfg_get_int("smpi/send_is_detached_thres"))) {
364 request->detached = 1;
366 if(request->old_type->has_subtype == 0){
367 oldbuf = request->buf;
368 if (oldbuf && request->size!=0){
369 request->buf = xbt_malloc(request->size);
370 memcpy(request->buf,oldbuf,request->size);
373 XBT_DEBUG("Send request %p is detached; buf %p copied into %p",request,oldbuf,request->buf);
376 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
377 request->real_size=request->size;
378 smpi_datatype_use(request->old_type);
379 smpi_comm_use(request->comm);
381 //if we are giving back the control to the user without waiting for completion, we have to inject timings
382 double sleeptime =0.0;
383 if(request->detached || (request->flags & (ISEND|SSEND))){// issend should be treated as isend
384 //isend and send timings may be different
385 sleeptime = (request->flags & ISEND)? smpi_ois(request->size) : smpi_os(request->size);
389 simcall_process_sleep(sleeptime);
390 XBT_DEBUG("sending size of %zu : sleep %lf ", request->size, smpi_os(request->size));
394 simcall_comm_isend(mailbox, request->size, -1.0,
395 request->buf, request->real_size,
397 &smpi_mpi_request_free_voidp, // how to free the userdata if a detached send fails
399 // detach if msg size < eager/rdv switch limit
403 /* FIXME: detached sends are not traceable (request->action == NULL) */
405 simcall_set_category(request->action, TRACE_internal_smpi_get_category());
413 void smpi_mpi_startall(int count, MPI_Request * requests)
417 for(i = 0; i < count; i++) {
418 smpi_mpi_start(requests[i]);
422 void smpi_mpi_request_free(MPI_Request * request)
424 if((*request) != MPI_REQUEST_NULL){
425 (*request)->refcount--;
426 if((*request)->refcount<0) xbt_die("wrong refcount");
428 if((*request)->refcount==0){
429 print_request("Destroying", (*request));
431 *request = MPI_REQUEST_NULL;
433 print_request("Decrementing", (*request));
437 xbt_die("freeing an already free request");
441 MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype,
442 int dst, int tag, MPI_Comm comm)
444 MPI_Request request =
445 build_request(buf==MPI_BOTTOM ? (void*)0 : buf , count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
446 comm, NON_PERSISTENT | SEND | PREPARED);
451 MPI_Request smpi_mpi_isend(void *buf, int count, MPI_Datatype datatype,
452 int dst, int tag, MPI_Comm comm)
454 MPI_Request request =
455 build_request(buf==MPI_BOTTOM?(void*)0:buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
456 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 =
466 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 =
478 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 =
487 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);
490 smpi_mpi_start(request);
494 void smpi_mpi_recv(void *buf, int count, MPI_Datatype datatype, int src,
495 int tag, MPI_Comm comm, MPI_Status * status)
498 request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm);
499 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 =
508 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);
510 smpi_mpi_start(request);
511 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
515 void smpi_mpi_ssend(void *buf, int count, MPI_Datatype datatype,
516 int dst, int tag, MPI_Comm comm)
518 MPI_Request request =
519 build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
520 comm, NON_PERSISTENT | SSEND | SEND);
522 smpi_mpi_start(request);
523 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
526 void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
527 int dst, int sendtag, void *recvbuf, int recvcount,
528 MPI_Datatype recvtype, int src, int recvtag,
529 MPI_Comm comm, MPI_Status * status)
531 MPI_Request requests[2];
533 int myid=smpi_process_index();
534 if ((smpi_group_index(smpi_comm_group(comm), dst) == myid) && (smpi_group_index(smpi_comm_group(comm), src) == myid)) {
535 smpi_datatype_copy(sendbuf, sendcount, sendtype,
536 recvbuf, recvcount, recvtype);
540 smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
542 smpi_irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
543 smpi_mpi_startall(2, requests);
544 smpi_mpi_waitall(2, requests, stats);
545 if(status != MPI_STATUS_IGNORE) {
546 // Copy receive status
551 int smpi_mpi_get_count(MPI_Status * status, MPI_Datatype datatype)
553 return status->count / smpi_datatype_size(datatype);
556 static void finish_wait(MPI_Request * request, MPI_Status * status)
558 MPI_Request req = *request;
559 if(status != MPI_STATUS_IGNORE)
560 smpi_empty_status(status);
562 if(!(req->detached && req->flags & SEND) && !(req->flags & PREPARED)){
563 if(status != MPI_STATUS_IGNORE) {
564 int src = req->src == MPI_ANY_SOURCE ? req->real_src : req->src;
565 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(req->comm), src);
566 status->MPI_TAG = req->tag == MPI_ANY_TAG ? req->real_tag : req->tag;
567 status->MPI_ERROR = req->truncated ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
568 // this handles the case were size in receive differs from size in send
569 // FIXME: really this should just contain the count of receive-type blocks,
571 status->count = req->real_size;
574 print_request("Finishing", req);
575 MPI_Datatype datatype = req->old_type;
577 if(datatype->has_subtype == 1){
578 // This part handles the problem of non-contignous memory
579 // the unserialization at the reception
580 s_smpi_subtype_t *subtype = datatype->substruct;
581 if(req->flags & RECV) {
582 subtype->unserialize(req->buf, req->old_buf, req->real_size/smpi_datatype_size(datatype) , datatype->substruct);
584 if(req->detached == 0) free(req->buf);
586 smpi_comm_unuse(req->comm);
587 smpi_datatype_unuse(datatype);
592 if (TRACE_smpi_view_internals()) {
593 if(req->flags & RECV){
594 int rank = smpi_process_index();
595 int src_traced = (req->src == MPI_ANY_SOURCE ? req->real_src : req->src);
596 TRACE_smpi_recv(rank, src_traced, rank);
601 if(req->detached_sender!=NULL){
602 smpi_mpi_request_free(&(req->detached_sender));
605 if(req->flags & NON_PERSISTENT) {
606 smpi_mpi_request_free(request);
612 int smpi_mpi_test(MPI_Request * request, MPI_Status * status) {
615 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before)
616 if ((*request)->action == NULL)
619 flag = simcall_comm_test((*request)->action);
621 finish_wait(request, status);
622 request=MPI_REQUEST_NULL;
624 smpi_empty_status(status);
629 int smpi_mpi_testany(int count, MPI_Request requests[], int *index,
636 *index = MPI_UNDEFINED;
638 comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
639 map = xbt_new(int, count);
641 for(i = 0; i < count; i++) {
642 if((requests[i]!=MPI_REQUEST_NULL) && requests[i]->action) {
643 xbt_dynar_push(comms, &requests[i]->action);
649 i = simcall_comm_testany(comms);
650 // not MPI_UNDEFINED, as this is a simix return code
653 finish_wait(&requests[*index], status);
657 //all requests are null or inactive, return true
659 smpi_empty_status(status);
662 xbt_dynar_free(&comms);
668 int smpi_mpi_testall(int count, MPI_Request requests[],
672 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
675 for(i=0; i<count; i++){
676 if(requests[i]!= MPI_REQUEST_NULL){
677 if (smpi_mpi_test(&requests[i], pstat)!=1){
680 requests[i]=MPI_REQUEST_NULL;
683 smpi_empty_status(pstat);
685 if(status != MPI_STATUSES_IGNORE) {
692 void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
694 //FIXME find another wait to avoid busy waiting ?
695 // the issue here is that we have to wait on a nonexistent comm
697 smpi_mpi_iprobe(source, tag, comm, &flag, status);
698 XBT_DEBUG("Busy Waiting on probing : %d", flag);
702 void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
704 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,
705 comm, NON_PERSISTENT | RECV);
707 //to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
708 double sleeptime= sg_cfg_get_double("smpi/iprobe");
709 //multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
710 static int nsleeps = 1;
712 simcall_process_sleep(sleeptime);
714 // behave like a receive, but don't do it
717 print_request("New iprobe", request);
718 // We have to test both mailboxes as we don't know if we will receive one one or another
719 if (sg_cfg_get_int("smpi/async_small_thres")>0){
720 mailbox = smpi_process_mailbox_small();
721 XBT_DEBUG("trying to probe the perm recv mailbox");
722 request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
724 if (request->action==NULL){
725 mailbox = smpi_process_mailbox();
726 XBT_DEBUG("trying to probe the other mailbox");
727 request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
731 MPI_Request req = (MPI_Request)SIMIX_comm_get_src_data(request->action);
733 if(status != MPI_STATUS_IGNORE && !(req->flags & PREPARED)) {
734 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(comm), req->src);
735 status->MPI_TAG = req->tag;
736 status->MPI_ERROR = MPI_SUCCESS;
737 status->count = req->real_size;
739 nsleeps=1;//reset the number of sleeps we will do next time
745 smpi_mpi_request_free(&request);
750 void smpi_mpi_wait(MPI_Request * request, MPI_Status * status)
752 print_request("Waiting", *request);
753 if ((*request)->action != NULL) { // this is not a detached send
754 simcall_comm_wait((*request)->action, -1.0);
756 finish_wait(request, status);
757 request=MPI_REQUEST_NULL;
758 // FIXME for a detached send, finish_wait is not called:
761 int smpi_mpi_waitany(int count, MPI_Request requests[],
768 index = MPI_UNDEFINED;
770 // Wait for a request to complete
771 comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
772 map = xbt_new(int, count);
774 XBT_DEBUG("Wait for one of %d", count);
775 for(i = 0; i < count; i++) {
776 if(requests[i] != MPI_REQUEST_NULL) {
777 if (requests[i]->action != NULL) {
778 XBT_DEBUG("Waiting any %p ", requests[i]);
779 xbt_dynar_push(comms, &requests[i]->action);
783 //This is a finished detached request, let's return this one
784 size=0;//so we free the dynar but don't do the waitany call
786 finish_wait(&requests[i], status);//cleanup if refcount = 0
787 requests[i]=MPI_REQUEST_NULL;//set to null
793 i = simcall_comm_waitany(comms);
795 // not MPI_UNDEFINED, as this is a simix return code
798 finish_wait(&requests[index], status);
802 xbt_dynar_free(&comms);
805 if (index==MPI_UNDEFINED)
806 smpi_empty_status(status);
811 int smpi_mpi_waitall(int count, MPI_Request requests[],
816 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
817 int retvalue = MPI_SUCCESS;
818 //tag invalid requests in the set
819 if (status != MPI_STATUSES_IGNORE) {
820 for (c = 0; c < count; c++) {
821 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL) {
822 smpi_empty_status(&status[c]);
823 } else if (requests[c]->src == MPI_PROC_NULL) {
824 smpi_empty_status(&status[c]);
825 status[c].MPI_SOURCE = MPI_PROC_NULL;
829 for(c = 0; c < count; c++) {
830 if (MC_is_active()) {
831 smpi_mpi_wait(&requests[c], pstat);
834 index = smpi_mpi_waitany(count, requests, pstat);
835 if (index == MPI_UNDEFINED)
837 requests[index]=MPI_REQUEST_NULL;
839 if (status != MPI_STATUSES_IGNORE) {
840 status[index] = *pstat;
841 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
842 retvalue = MPI_ERR_IN_STATUS;
849 int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices,
854 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
857 for(i = 0; i < incount; i++)
859 index=smpi_mpi_waitany(incount, requests, pstat);
860 if(index!=MPI_UNDEFINED){
861 indices[count] = index;
863 if(status != MPI_STATUSES_IGNORE) {
864 status[index] = *pstat;
866 requests[index]=MPI_REQUEST_NULL;
868 return MPI_UNDEFINED;
874 int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices,
877 int i, count, count_dead;
879 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
883 for(i = 0; i < incount; i++) {
884 if((requests[i] != MPI_REQUEST_NULL)) {
885 if(smpi_mpi_test(&requests[i], pstat)) {
888 if(status != MPI_STATUSES_IGNORE) {
891 requests[i]=MPI_REQUEST_NULL;
898 if(count_dead==incount)return MPI_UNDEFINED;
902 void smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root,
905 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
906 nary_tree_bcast(buf, count, datatype, root, comm, 4);
909 void smpi_mpi_barrier(MPI_Comm comm)
911 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
912 nary_tree_barrier(comm, 4);
915 void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
916 void *recvbuf, int recvcount, MPI_Datatype recvtype,
917 int root, MPI_Comm comm)
919 int system_tag = COLL_TAG_GATHER;
920 int rank, size, src, index;
921 MPI_Aint lb = 0, recvext = 0;
922 MPI_Request *requests;
924 rank = smpi_comm_rank(comm);
925 size = smpi_comm_size(comm);
927 // Send buffer to root
928 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
930 // FIXME: check for errors
931 smpi_datatype_extent(recvtype, &lb, &recvext);
932 // Local copy from root
933 smpi_datatype_copy(sendbuf, sendcount, sendtype,
934 (char *)recvbuf + root * recvcount * recvext, recvcount, recvtype);
935 // Receive buffers from senders
936 requests = xbt_new(MPI_Request, size - 1);
938 for(src = 0; src < size; src++) {
940 requests[index] = smpi_irecv_init((char *)recvbuf + src * recvcount * recvext,
942 src, system_tag, comm);
946 // Wait for completion of irecv's.
947 smpi_mpi_startall(size - 1, requests);
948 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
954 void smpi_mpi_reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts,
955 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
959 int rank = smpi_process_index();
962 /* arbitrarily choose root as rank 0 */
963 size = smpi_comm_size(comm);
965 displs = xbt_new(int, size);
966 for (i = 0; i < size; i++) {
968 count += recvcounts[i];
970 tmpbuf=(void*)xbt_malloc(count*smpi_datatype_get_extent(datatype));
971 mpi_coll_reduce_fun(sendbuf, tmpbuf, count, datatype, op, 0, comm);
972 smpi_mpi_scatterv(tmpbuf, recvcounts, displs, datatype, recvbuf,
973 recvcounts[rank], datatype, 0, comm);
978 void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
979 void *recvbuf, int *recvcounts, int *displs,
980 MPI_Datatype recvtype, int root, MPI_Comm comm)
982 int system_tag = COLL_TAG_GATHERV;
983 int rank, size, src, index;
984 MPI_Aint lb = 0, recvext = 0;
985 MPI_Request *requests;
987 rank = smpi_comm_rank(comm);
988 size = smpi_comm_size(comm);
990 // Send buffer to root
991 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
993 // FIXME: check for errors
994 smpi_datatype_extent(recvtype, &lb, &recvext);
995 // Local copy from root
996 smpi_datatype_copy(sendbuf, sendcount, sendtype,
997 (char *)recvbuf + displs[root] * recvext,
998 recvcounts[root], recvtype);
999 // Receive buffers from senders
1000 requests = xbt_new(MPI_Request, size - 1);
1002 for(src = 0; src < size; src++) {
1005 smpi_irecv_init((char *)recvbuf + displs[src] * recvext,
1006 recvcounts[src], recvtype, src, system_tag, comm);
1010 // Wait for completion of irecv's.
1011 smpi_mpi_startall(size - 1, requests);
1012 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1017 void smpi_mpi_allgather(void *sendbuf, int sendcount,
1018 MPI_Datatype sendtype, void *recvbuf,
1019 int recvcount, MPI_Datatype recvtype,
1022 int system_tag = COLL_TAG_ALLGATHER;
1023 int rank, size, other, index;
1024 MPI_Aint lb = 0, recvext = 0;
1025 MPI_Request *requests;
1027 rank = smpi_comm_rank(comm);
1028 size = smpi_comm_size(comm);
1029 // FIXME: check for errors
1030 smpi_datatype_extent(recvtype, &lb, &recvext);
1031 // Local copy from self
1032 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1033 (char *)recvbuf + rank * recvcount * recvext, recvcount,
1035 // Send/Recv buffers to/from others;
1036 requests = xbt_new(MPI_Request, 2 * (size - 1));
1038 for(other = 0; other < size; other++) {
1041 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1044 requests[index] = smpi_irecv_init((char *)recvbuf + other * recvcount * recvext,
1045 recvcount, recvtype, other,
1050 // Wait for completion of all comms.
1051 smpi_mpi_startall(2 * (size - 1), requests);
1052 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1056 void smpi_mpi_allgatherv(void *sendbuf, int sendcount,
1057 MPI_Datatype sendtype, void *recvbuf,
1058 int *recvcounts, int *displs,
1059 MPI_Datatype recvtype, MPI_Comm comm)
1061 int system_tag = COLL_TAG_ALLGATHERV;
1062 int rank, size, other, index;
1063 MPI_Aint lb = 0, recvext = 0;
1064 MPI_Request *requests;
1066 rank = smpi_comm_rank(comm);
1067 size = smpi_comm_size(comm);
1068 // FIXME: check for errors
1069 smpi_datatype_extent(recvtype, &lb, &recvext);
1070 // Local copy from self
1071 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1072 (char *)recvbuf + displs[rank] * recvext,
1073 recvcounts[rank], recvtype);
1074 // Send buffers to others;
1075 requests = xbt_new(MPI_Request, 2 * (size - 1));
1077 for(other = 0; other < size; other++) {
1080 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1084 smpi_irecv_init((char *)recvbuf + displs[other] * recvext, recvcounts[other],
1085 recvtype, other, system_tag, comm);
1089 // Wait for completion of all comms.
1090 smpi_mpi_startall(2 * (size - 1), requests);
1091 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1095 void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1096 void *recvbuf, int recvcount, MPI_Datatype recvtype,
1097 int root, MPI_Comm comm)
1099 int system_tag = COLL_TAG_SCATTER;
1100 int rank, size, dst, index;
1101 MPI_Aint lb = 0, sendext = 0;
1102 MPI_Request *requests;
1104 rank = smpi_comm_rank(comm);
1105 size = smpi_comm_size(comm);
1107 // Recv buffer from root
1108 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1111 // FIXME: check for errors
1112 smpi_datatype_extent(sendtype, &lb, &sendext);
1113 // Local copy from root
1114 if(recvbuf!=MPI_IN_PLACE){
1115 smpi_datatype_copy((char *)sendbuf + root * sendcount * sendext,
1116 sendcount, sendtype, recvbuf, recvcount, recvtype);
1118 // Send buffers to receivers
1119 requests = xbt_new(MPI_Request, size - 1);
1121 for(dst = 0; dst < size; dst++) {
1123 requests[index] = smpi_isend_init((char *)sendbuf + dst * sendcount * sendext,
1124 sendcount, sendtype, dst,
1129 // Wait for completion of isend's.
1130 smpi_mpi_startall(size - 1, requests);
1131 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1136 void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs,
1137 MPI_Datatype sendtype, void *recvbuf, int recvcount,
1138 MPI_Datatype recvtype, int root, MPI_Comm comm)
1140 int system_tag = COLL_TAG_SCATTERV;
1141 int rank, size, dst, index;
1142 MPI_Aint lb = 0, sendext = 0;
1143 MPI_Request *requests;
1145 rank = smpi_comm_rank(comm);
1146 size = smpi_comm_size(comm);
1148 // Recv buffer from root
1149 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1152 // FIXME: check for errors
1153 smpi_datatype_extent(sendtype, &lb, &sendext);
1154 // Local copy from root
1155 if(recvbuf!=MPI_IN_PLACE){
1156 smpi_datatype_copy((char *)sendbuf + displs[root] * sendext, sendcounts[root],
1157 sendtype, recvbuf, recvcount, recvtype);
1159 // Send buffers to receivers
1160 requests = xbt_new(MPI_Request, size - 1);
1162 for(dst = 0; dst < size; dst++) {
1165 smpi_isend_init((char *)sendbuf + displs[dst] * sendext, sendcounts[dst],
1166 sendtype, dst, system_tag, comm);
1170 // Wait for completion of isend's.
1171 smpi_mpi_startall(size - 1, requests);
1172 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1177 void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count,
1178 MPI_Datatype datatype, MPI_Op op, int root,
1181 int system_tag = COLL_TAG_REDUCE;
1182 int rank, size, src, index;
1183 MPI_Aint lb = 0, dataext = 0;
1184 MPI_Request *requests;
1188 char* sendtmpbuf = (char*) sendbuf;
1189 if( sendbuf == MPI_IN_PLACE ) {
1190 sendtmpbuf = (char *)xbt_malloc(count*smpi_datatype_get_extent(datatype));
1191 smpi_datatype_copy(recvbuf, count, datatype,sendtmpbuf, count, datatype);
1194 rank = smpi_comm_rank(comm);
1195 size = smpi_comm_size(comm);
1196 //non commutative case, use a working algo from openmpi
1197 if(!smpi_op_is_commute(op)){
1198 smpi_coll_tuned_reduce_ompi_basic_linear(sendtmpbuf, recvbuf, count,
1199 datatype, op, root, comm);
1204 // Send buffer to root
1205 smpi_mpi_send(sendtmpbuf, count, datatype, root, system_tag, comm);
1207 // FIXME: check for errors
1208 smpi_datatype_extent(datatype, &lb, &dataext);
1209 // Local copy from root
1210 if (sendtmpbuf && recvbuf)
1211 smpi_datatype_copy(sendtmpbuf, count, datatype, recvbuf, count, datatype);
1212 // Receive buffers from senders
1213 //TODO: make a MPI_barrier here ?
1214 requests = xbt_new(MPI_Request, size - 1);
1215 tmpbufs = xbt_new(void *, size - 1);
1217 for(src = 0; src < size; src++) {
1219 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1221 tmpbufs[index] = xbt_malloc(count * dataext);
1223 smpi_irecv_init(tmpbufs[index], count, datatype, src,
1228 // Wait for completion of irecv's.
1229 smpi_mpi_startall(size - 1, requests);
1230 for(src = 0; src < size - 1; src++) {
1231 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1232 XBT_DEBUG("finished waiting any request with index %d", index);
1233 if(index == MPI_UNDEFINED) {
1236 if(op) /* op can be MPI_OP_NULL that does nothing */
1237 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1239 for(index = 0; index < size - 1; index++) {
1240 xbt_free(tmpbufs[index]);
1245 if( sendbuf == MPI_IN_PLACE ) {
1246 xbt_free(sendtmpbuf);
1251 void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count,
1252 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1254 smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, 0, comm);
1255 smpi_mpi_bcast(recvbuf, count, datatype, 0, comm);
1258 void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count,
1259 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1261 int system_tag = -888;
1262 int rank, size, other, index;
1263 MPI_Aint lb = 0, dataext = 0;
1264 MPI_Request *requests;
1267 rank = smpi_comm_rank(comm);
1268 size = smpi_comm_size(comm);
1270 // FIXME: check for errors
1271 smpi_datatype_extent(datatype, &lb, &dataext);
1273 // Local copy from self
1274 smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
1276 // Send/Recv buffers to/from others;
1277 requests = xbt_new(MPI_Request, size - 1);
1278 tmpbufs = xbt_new(void *, rank);
1280 for(other = 0; other < rank; other++) {
1281 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1283 tmpbufs[index] = xbt_malloc(count * dataext);
1285 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1289 for(other = rank + 1; other < size; other++) {
1291 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1294 // Wait for completion of all comms.
1295 smpi_mpi_startall(size - 1, requests);
1297 if(smpi_op_is_commute(op)){
1298 for(other = 0; other < size - 1; other++) {
1299 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1300 if(index == MPI_UNDEFINED) {
1304 // #Request is below rank: it's a irecv
1305 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1309 //non commutative case, wait in order
1310 for(other = 0; other < size - 1; other++) {
1311 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1313 smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1317 for(index = 0; index < rank; index++) {
1318 xbt_free(tmpbufs[index]);
1324 void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count,
1325 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1327 int system_tag = -888;
1328 int rank, size, other, index;
1329 MPI_Aint lb = 0, dataext = 0;
1330 MPI_Request *requests;
1332 int recvbuf_is_empty=1;
1333 rank = smpi_comm_rank(comm);
1334 size = smpi_comm_size(comm);
1336 // FIXME: check for errors
1337 smpi_datatype_extent(datatype, &lb, &dataext);
1339 // Send/Recv buffers to/from others;
1340 requests = xbt_new(MPI_Request, size - 1);
1341 tmpbufs = xbt_new(void *, rank);
1343 for(other = 0; other < rank; other++) {
1344 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1346 tmpbufs[index] = xbt_malloc(count * dataext);
1348 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1352 for(other = rank + 1; other < size; other++) {
1354 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1357 // Wait for completion of all comms.
1358 smpi_mpi_startall(size - 1, requests);
1359 if(smpi_op_is_commute(op)){
1360 for(other = 0; other < size - 1; other++) {
1361 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1362 if(index == MPI_UNDEFINED) {
1366 if(recvbuf_is_empty){
1367 smpi_datatype_copy(tmpbufs[index], count, datatype, recvbuf, count, datatype);
1370 // #Request is below rank: it's a irecv
1371 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1375 //non commutative case, wait in order
1376 for(other = 0; other < size - 1; other++) {
1377 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1379 if(recvbuf_is_empty){
1380 smpi_datatype_copy(tmpbufs[other], count, datatype, recvbuf, count, datatype);
1382 }else smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1386 for(index = 0; index < rank; index++) {
1387 xbt_free(tmpbufs[index]);