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 //FIXME: if receive is posted with a large size, but send is smaller, mailboxes may not match !
324 if (request->size < sg_cfg_get_int("smpi/async_small_thres"))
325 mailbox = smpi_process_mailbox_small();
327 mailbox = smpi_process_mailbox();
328 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
329 request->real_size=request->size;
330 smpi_datatype_use(request->old_type);
331 smpi_comm_use(request->comm);
332 request->action = simcall_comm_irecv(mailbox, request->buf, &request->real_size, &match_recv, request);
334 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
335 double sleeptime = request->detached ? smpi_or(request->size) : 0.0;
337 simcall_process_sleep(sleeptime);
338 XBT_DEBUG("receiving size of %zu : sleep %lf ", request->size, smpi_or(request->size));
344 int receiver = request->dst;//smpi_group_index(smpi_comm_group(request->comm), request->dst);
347 int rank = smpi_process_index();
348 if (TRACE_smpi_view_internals()) {
349 TRACE_smpi_send(rank, rank, receiver,request->size);
352 /* if(receiver == MPI_UNDEFINED) {*/
353 /* XBT_WARN("Trying to send a message to a wrong rank");*/
356 print_request("New send", request);
357 if (request->size < sg_cfg_get_int("smpi/async_small_thres")) { // eager mode
358 mailbox = smpi_process_remote_mailbox_small(receiver);
360 XBT_DEBUG("Send request %p is not in the permanent receive mailbox (buf: %p)",request,request->buf);
361 mailbox = smpi_process_remote_mailbox(receiver);
363 if ( (! (request->flags & SSEND)) && (request->size < sg_cfg_get_int("smpi/send_is_detached_thres"))) {
365 request->detached = 1;
367 if(request->old_type->has_subtype == 0){
368 oldbuf = request->buf;
369 if (oldbuf && request->size!=0){
370 request->buf = xbt_malloc(request->size);
371 memcpy(request->buf,oldbuf,request->size);
374 XBT_DEBUG("Send request %p is detached; buf %p copied into %p",request,oldbuf,request->buf);
377 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
378 request->real_size=request->size;
379 smpi_datatype_use(request->old_type);
380 smpi_comm_use(request->comm);
382 //if we are giving back the control to the user without waiting for completion, we have to inject timings
383 double sleeptime =0.0;
384 if(request->detached || (request->flags & (ISEND|SSEND))){// issend should be treated as isend
385 //isend and send timings may be different
386 sleeptime = (request->flags & ISEND)? smpi_ois(request->size) : smpi_os(request->size);
390 simcall_process_sleep(sleeptime);
391 XBT_DEBUG("sending size of %zu : sleep %lf ", request->size, smpi_os(request->size));
395 simcall_comm_isend(mailbox, request->size, -1.0,
396 request->buf, request->real_size,
398 &smpi_mpi_request_free_voidp, // how to free the userdata if a detached send fails
400 // detach if msg size < eager/rdv switch limit
404 /* FIXME: detached sends are not traceable (request->action == NULL) */
406 simcall_set_category(request->action, TRACE_internal_smpi_get_category());
414 void smpi_mpi_startall(int count, MPI_Request * requests)
418 for(i = 0; i < count; i++) {
419 smpi_mpi_start(requests[i]);
423 void smpi_mpi_request_free(MPI_Request * request)
425 if((*request) != MPI_REQUEST_NULL){
426 (*request)->refcount--;
427 if((*request)->refcount<0) xbt_die("wrong refcount");
429 if((*request)->refcount==0){
430 print_request("Destroying", (*request));
432 *request = MPI_REQUEST_NULL;
434 print_request("Decrementing", (*request));
438 xbt_die("freeing an already free request");
442 MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype,
443 int dst, int tag, MPI_Comm comm)
445 MPI_Request request =
446 build_request(buf==MPI_BOTTOM ? (void*)0 : buf , count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
447 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 =
456 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);
459 smpi_mpi_start(request);
463 MPI_Request smpi_mpi_issend(void *buf, int count, MPI_Datatype datatype,
464 int dst, int tag, MPI_Comm comm)
466 MPI_Request request =
467 build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
468 comm, NON_PERSISTENT | ISEND | SSEND | SEND);
469 smpi_mpi_start(request);
475 MPI_Request smpi_irecv_init(void *buf, int count, MPI_Datatype datatype,
476 int src, int tag, MPI_Comm comm)
478 MPI_Request request =
479 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,
480 comm, NON_PERSISTENT | RECV | PREPARED);
484 MPI_Request smpi_mpi_irecv(void *buf, int count, MPI_Datatype datatype,
485 int src, int tag, MPI_Comm comm)
487 MPI_Request request =
488 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,
489 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)
499 request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm);
500 smpi_mpi_wait(&request, status);
505 void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst,
506 int tag, MPI_Comm comm)
508 MPI_Request request =
509 build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
510 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 =
520 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);
527 void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
528 int dst, int sendtag, void *recvbuf, int recvcount,
529 MPI_Datatype recvtype, int src, int recvtag,
530 MPI_Comm comm, MPI_Status * status)
532 MPI_Request requests[2];
534 int myid=smpi_process_index();
535 if ((smpi_group_index(smpi_comm_group(comm), dst) == myid) && (smpi_group_index(smpi_comm_group(comm), src) == myid)) {
536 smpi_datatype_copy(sendbuf, sendcount, sendtype,
537 recvbuf, recvcount, recvtype);
541 smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
543 smpi_irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
544 smpi_mpi_startall(2, requests);
545 smpi_mpi_waitall(2, requests, stats);
546 if(status != MPI_STATUS_IGNORE) {
547 // Copy receive status
552 int smpi_mpi_get_count(MPI_Status * status, MPI_Datatype datatype)
554 return status->count / smpi_datatype_size(datatype);
557 static void finish_wait(MPI_Request * request, MPI_Status * status)
559 MPI_Request req = *request;
560 if(status != MPI_STATUS_IGNORE)
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);
658 //all requests are null or inactive, return true
660 smpi_empty_status(status);
663 xbt_dynar_free(&comms);
669 int smpi_mpi_testall(int count, MPI_Request requests[],
673 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
676 for(i=0; i<count; i++){
677 if(requests[i]!= MPI_REQUEST_NULL){
678 if (smpi_mpi_test(&requests[i], pstat)!=1){
681 requests[i]=MPI_REQUEST_NULL;
684 smpi_empty_status(pstat);
686 if(status != MPI_STATUSES_IGNORE) {
693 void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
695 //FIXME find another wait to avoid busy waiting ?
696 // the issue here is that we have to wait on a nonexistent comm
698 smpi_mpi_iprobe(source, tag, comm, &flag, status);
699 XBT_DEBUG("Busy Waiting on probing : %d", flag);
703 void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
705 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,
706 comm, NON_PERSISTENT | RECV);
708 //to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
709 double sleeptime= sg_cfg_get_double("smpi/iprobe");
710 //multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
711 static int nsleeps = 1;
713 simcall_process_sleep(sleeptime);
715 // behave like a receive, but don't do it
718 print_request("New iprobe", request);
719 // We have to test both mailboxes as we don't know if we will receive one one or another
720 if (sg_cfg_get_int("smpi/async_small_thres")>0){
721 mailbox = smpi_process_mailbox_small();
722 XBT_DEBUG("trying to probe the perm recv mailbox");
723 request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
725 if (request->action==NULL){
726 mailbox = smpi_process_mailbox();
727 XBT_DEBUG("trying to probe the other mailbox");
728 request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
732 MPI_Request req = (MPI_Request)SIMIX_comm_get_src_data(request->action);
734 if(status != MPI_STATUS_IGNORE && !(req->flags & PREPARED)) {
735 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(comm), req->src);
736 status->MPI_TAG = req->tag;
737 status->MPI_ERROR = MPI_SUCCESS;
738 status->count = req->real_size;
740 nsleeps=1;//reset the number of sleeps we will do next time
746 smpi_mpi_request_free(&request);
751 void smpi_mpi_wait(MPI_Request * request, MPI_Status * status)
753 print_request("Waiting", *request);
754 if ((*request)->action != NULL) { // this is not a detached send
755 simcall_comm_wait((*request)->action, -1.0);
757 finish_wait(request, status);
758 request=MPI_REQUEST_NULL;
759 // FIXME for a detached send, finish_wait is not called:
762 int smpi_mpi_waitany(int count, MPI_Request requests[],
769 index = MPI_UNDEFINED;
771 // Wait for a request to complete
772 comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
773 map = xbt_new(int, count);
775 XBT_DEBUG("Wait for one of %d", count);
776 for(i = 0; i < count; i++) {
777 if(requests[i] != MPI_REQUEST_NULL) {
778 if (requests[i]->action != NULL) {
779 XBT_DEBUG("Waiting any %p ", requests[i]);
780 xbt_dynar_push(comms, &requests[i]->action);
784 //This is a finished detached request, let's return this one
785 size=0;//so we free the dynar but don't do the waitany call
787 finish_wait(&requests[i], status);//cleanup if refcount = 0
788 requests[i]=MPI_REQUEST_NULL;//set to null
794 i = simcall_comm_waitany(comms);
796 // not MPI_UNDEFINED, as this is a simix return code
799 finish_wait(&requests[index], status);
803 xbt_dynar_free(&comms);
806 if (index==MPI_UNDEFINED)
807 smpi_empty_status(status);
812 int smpi_mpi_waitall(int count, MPI_Request requests[],
817 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
818 int retvalue = MPI_SUCCESS;
819 //tag invalid requests in the set
820 if (status != MPI_STATUSES_IGNORE) {
821 for (c = 0; c < count; c++) {
822 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL) {
823 smpi_empty_status(&status[c]);
824 } else if (requests[c]->src == MPI_PROC_NULL) {
825 smpi_empty_status(&status[c]);
826 status[c].MPI_SOURCE = MPI_PROC_NULL;
830 for(c = 0; c < count; c++) {
831 if (MC_is_active()) {
832 smpi_mpi_wait(&requests[c], pstat);
835 index = smpi_mpi_waitany(count, requests, pstat);
836 if (index == MPI_UNDEFINED)
838 requests[index]=MPI_REQUEST_NULL;
840 if (status != MPI_STATUSES_IGNORE) {
841 status[index] = *pstat;
842 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
843 retvalue = MPI_ERR_IN_STATUS;
850 int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices,
855 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
858 for(i = 0; i < incount; i++)
860 index=smpi_mpi_waitany(incount, requests, pstat);
861 if(index!=MPI_UNDEFINED){
862 indices[count] = index;
864 if(status != MPI_STATUSES_IGNORE) {
865 status[index] = *pstat;
867 requests[index]=MPI_REQUEST_NULL;
869 return MPI_UNDEFINED;
875 int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices,
878 int i, count, count_dead;
880 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
884 for(i = 0; i < incount; i++) {
885 if((requests[i] != MPI_REQUEST_NULL)) {
886 if(smpi_mpi_test(&requests[i], pstat)) {
889 if(status != MPI_STATUSES_IGNORE) {
892 requests[i]=MPI_REQUEST_NULL;
899 if(count_dead==incount)return MPI_UNDEFINED;
903 void smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root,
906 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
907 nary_tree_bcast(buf, count, datatype, root, comm, 4);
910 void smpi_mpi_barrier(MPI_Comm comm)
912 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
913 nary_tree_barrier(comm, 4);
916 void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
917 void *recvbuf, int recvcount, MPI_Datatype recvtype,
918 int root, MPI_Comm comm)
920 int system_tag = COLL_TAG_GATHER;
921 int rank, size, src, index;
922 MPI_Aint lb = 0, recvext = 0;
923 MPI_Request *requests;
925 rank = smpi_comm_rank(comm);
926 size = smpi_comm_size(comm);
928 // Send buffer to root
929 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
931 // FIXME: check for errors
932 smpi_datatype_extent(recvtype, &lb, &recvext);
933 // Local copy from root
934 smpi_datatype_copy(sendbuf, sendcount, sendtype,
935 (char *)recvbuf + root * recvcount * recvext, recvcount, recvtype);
936 // Receive buffers from senders
937 requests = xbt_new(MPI_Request, size - 1);
939 for(src = 0; src < size; src++) {
941 requests[index] = smpi_irecv_init((char *)recvbuf + src * recvcount * recvext,
943 src, system_tag, comm);
947 // Wait for completion of irecv's.
948 smpi_mpi_startall(size - 1, requests);
949 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
955 void smpi_mpi_reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts,
956 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
960 int rank = smpi_process_index();
963 /* arbitrarily choose root as rank 0 */
964 size = smpi_comm_size(comm);
966 displs = xbt_new(int, size);
967 for (i = 0; i < size; i++) {
969 count += recvcounts[i];
971 tmpbuf=(void*)xbt_malloc(count*smpi_datatype_get_extent(datatype));
972 mpi_coll_reduce_fun(sendbuf, tmpbuf, count, datatype, op, 0, comm);
973 smpi_mpi_scatterv(tmpbuf, recvcounts, displs, datatype, recvbuf,
974 recvcounts[rank], datatype, 0, comm);
979 void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
980 void *recvbuf, int *recvcounts, int *displs,
981 MPI_Datatype recvtype, int root, MPI_Comm comm)
983 int system_tag = COLL_TAG_GATHERV;
984 int rank, size, src, index;
985 MPI_Aint lb = 0, recvext = 0;
986 MPI_Request *requests;
988 rank = smpi_comm_rank(comm);
989 size = smpi_comm_size(comm);
991 // Send buffer to root
992 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
994 // FIXME: check for errors
995 smpi_datatype_extent(recvtype, &lb, &recvext);
996 // Local copy from root
997 smpi_datatype_copy(sendbuf, sendcount, sendtype,
998 (char *)recvbuf + displs[root] * recvext,
999 recvcounts[root], recvtype);
1000 // Receive buffers from senders
1001 requests = xbt_new(MPI_Request, size - 1);
1003 for(src = 0; src < size; src++) {
1006 smpi_irecv_init((char *)recvbuf + displs[src] * recvext,
1007 recvcounts[src], recvtype, src, system_tag, comm);
1011 // Wait for completion of irecv's.
1012 smpi_mpi_startall(size - 1, requests);
1013 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1018 void smpi_mpi_allgather(void *sendbuf, int sendcount,
1019 MPI_Datatype sendtype, void *recvbuf,
1020 int recvcount, MPI_Datatype recvtype,
1023 int system_tag = COLL_TAG_ALLGATHER;
1024 int rank, size, other, index;
1025 MPI_Aint lb = 0, recvext = 0;
1026 MPI_Request *requests;
1028 rank = smpi_comm_rank(comm);
1029 size = smpi_comm_size(comm);
1030 // FIXME: check for errors
1031 smpi_datatype_extent(recvtype, &lb, &recvext);
1032 // Local copy from self
1033 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1034 (char *)recvbuf + rank * recvcount * recvext, recvcount,
1036 // Send/Recv buffers to/from others;
1037 requests = xbt_new(MPI_Request, 2 * (size - 1));
1039 for(other = 0; other < size; other++) {
1042 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1045 requests[index] = smpi_irecv_init((char *)recvbuf + other * recvcount * recvext,
1046 recvcount, recvtype, other,
1051 // Wait for completion of all comms.
1052 smpi_mpi_startall(2 * (size - 1), requests);
1053 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1057 void smpi_mpi_allgatherv(void *sendbuf, int sendcount,
1058 MPI_Datatype sendtype, void *recvbuf,
1059 int *recvcounts, int *displs,
1060 MPI_Datatype recvtype, MPI_Comm comm)
1062 int system_tag = COLL_TAG_ALLGATHERV;
1063 int rank, size, other, index;
1064 MPI_Aint lb = 0, recvext = 0;
1065 MPI_Request *requests;
1067 rank = smpi_comm_rank(comm);
1068 size = smpi_comm_size(comm);
1069 // FIXME: check for errors
1070 smpi_datatype_extent(recvtype, &lb, &recvext);
1071 // Local copy from self
1072 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1073 (char *)recvbuf + displs[rank] * recvext,
1074 recvcounts[rank], recvtype);
1075 // Send buffers to others;
1076 requests = xbt_new(MPI_Request, 2 * (size - 1));
1078 for(other = 0; other < size; other++) {
1081 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1085 smpi_irecv_init((char *)recvbuf + displs[other] * recvext, recvcounts[other],
1086 recvtype, other, system_tag, comm);
1090 // Wait for completion of all comms.
1091 smpi_mpi_startall(2 * (size - 1), requests);
1092 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1096 void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1097 void *recvbuf, int recvcount, MPI_Datatype recvtype,
1098 int root, MPI_Comm comm)
1100 int system_tag = COLL_TAG_SCATTER;
1101 int rank, size, dst, index;
1102 MPI_Aint lb = 0, sendext = 0;
1103 MPI_Request *requests;
1105 rank = smpi_comm_rank(comm);
1106 size = smpi_comm_size(comm);
1108 // Recv buffer from root
1109 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1112 // FIXME: check for errors
1113 smpi_datatype_extent(sendtype, &lb, &sendext);
1114 // Local copy from root
1115 if(recvbuf!=MPI_IN_PLACE){
1116 smpi_datatype_copy((char *)sendbuf + root * sendcount * sendext,
1117 sendcount, sendtype, recvbuf, recvcount, recvtype);
1119 // Send buffers to receivers
1120 requests = xbt_new(MPI_Request, size - 1);
1122 for(dst = 0; dst < size; dst++) {
1124 requests[index] = smpi_isend_init((char *)sendbuf + dst * sendcount * sendext,
1125 sendcount, sendtype, dst,
1130 // Wait for completion of isend's.
1131 smpi_mpi_startall(size - 1, requests);
1132 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1137 void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs,
1138 MPI_Datatype sendtype, void *recvbuf, int recvcount,
1139 MPI_Datatype recvtype, int root, MPI_Comm comm)
1141 int system_tag = COLL_TAG_SCATTERV;
1142 int rank, size, dst, index;
1143 MPI_Aint lb = 0, sendext = 0;
1144 MPI_Request *requests;
1146 rank = smpi_comm_rank(comm);
1147 size = smpi_comm_size(comm);
1149 // Recv buffer from root
1150 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1153 // FIXME: check for errors
1154 smpi_datatype_extent(sendtype, &lb, &sendext);
1155 // Local copy from root
1156 if(recvbuf!=MPI_IN_PLACE){
1157 smpi_datatype_copy((char *)sendbuf + displs[root] * sendext, sendcounts[root],
1158 sendtype, recvbuf, recvcount, recvtype);
1160 // Send buffers to receivers
1161 requests = xbt_new(MPI_Request, size - 1);
1163 for(dst = 0; dst < size; dst++) {
1166 smpi_isend_init((char *)sendbuf + displs[dst] * sendext, sendcounts[dst],
1167 sendtype, dst, system_tag, comm);
1171 // Wait for completion of isend's.
1172 smpi_mpi_startall(size - 1, requests);
1173 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1178 void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count,
1179 MPI_Datatype datatype, MPI_Op op, int root,
1182 int system_tag = COLL_TAG_REDUCE;
1183 int rank, size, src, index;
1184 MPI_Aint lb = 0, dataext = 0;
1185 MPI_Request *requests;
1189 char* sendtmpbuf = (char*) sendbuf;
1190 if( sendbuf == MPI_IN_PLACE ) {
1191 sendtmpbuf = (char *)xbt_malloc(count*smpi_datatype_get_extent(datatype));
1192 smpi_datatype_copy(recvbuf, count, datatype,sendtmpbuf, count, datatype);
1195 rank = smpi_comm_rank(comm);
1196 size = smpi_comm_size(comm);
1197 //non commutative case, use a working algo from openmpi
1198 if(!smpi_op_is_commute(op)){
1199 smpi_coll_tuned_reduce_ompi_basic_linear(sendtmpbuf, recvbuf, count,
1200 datatype, op, root, comm);
1205 // Send buffer to root
1206 smpi_mpi_send(sendtmpbuf, count, datatype, root, system_tag, comm);
1208 // FIXME: check for errors
1209 smpi_datatype_extent(datatype, &lb, &dataext);
1210 // Local copy from root
1211 if (sendtmpbuf && recvbuf)
1212 smpi_datatype_copy(sendtmpbuf, count, datatype, recvbuf, count, datatype);
1213 // Receive buffers from senders
1214 //TODO: make a MPI_barrier here ?
1215 requests = xbt_new(MPI_Request, size - 1);
1216 tmpbufs = xbt_new(void *, size - 1);
1218 for(src = 0; src < size; src++) {
1220 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1222 tmpbufs[index] = xbt_malloc(count * dataext);
1224 smpi_irecv_init(tmpbufs[index], count, datatype, src,
1229 // Wait for completion of irecv's.
1230 smpi_mpi_startall(size - 1, requests);
1231 for(src = 0; src < size - 1; src++) {
1232 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1233 XBT_DEBUG("finished waiting any request with index %d", index);
1234 if(index == MPI_UNDEFINED) {
1237 if(op) /* op can be MPI_OP_NULL that does nothing */
1238 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1240 for(index = 0; index < size - 1; index++) {
1241 xbt_free(tmpbufs[index]);
1246 if( sendbuf == MPI_IN_PLACE ) {
1247 xbt_free(sendtmpbuf);
1252 void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count,
1253 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1255 smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, 0, comm);
1256 smpi_mpi_bcast(recvbuf, count, datatype, 0, comm);
1259 void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count,
1260 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1262 int system_tag = -888;
1263 int rank, size, other, index;
1264 MPI_Aint lb = 0, dataext = 0;
1265 MPI_Request *requests;
1268 rank = smpi_comm_rank(comm);
1269 size = smpi_comm_size(comm);
1271 // FIXME: check for errors
1272 smpi_datatype_extent(datatype, &lb, &dataext);
1274 // Local copy from self
1275 smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
1277 // Send/Recv buffers to/from others;
1278 requests = xbt_new(MPI_Request, size - 1);
1279 tmpbufs = xbt_new(void *, rank);
1281 for(other = 0; other < rank; other++) {
1282 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1284 tmpbufs[index] = xbt_malloc(count * dataext);
1286 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1290 for(other = rank + 1; other < size; other++) {
1292 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1295 // Wait for completion of all comms.
1296 smpi_mpi_startall(size - 1, requests);
1298 if(smpi_op_is_commute(op)){
1299 for(other = 0; other < size - 1; other++) {
1300 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1301 if(index == MPI_UNDEFINED) {
1305 // #Request is below rank: it's a irecv
1306 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1310 //non commutative case, wait in order
1311 for(other = 0; other < size - 1; other++) {
1312 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1314 smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1318 for(index = 0; index < rank; index++) {
1319 xbt_free(tmpbufs[index]);
1325 void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count,
1326 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1328 int system_tag = -888;
1329 int rank, size, other, index;
1330 MPI_Aint lb = 0, dataext = 0;
1331 MPI_Request *requests;
1333 int recvbuf_is_empty=1;
1334 rank = smpi_comm_rank(comm);
1335 size = smpi_comm_size(comm);
1337 // FIXME: check for errors
1338 smpi_datatype_extent(datatype, &lb, &dataext);
1340 // Send/Recv buffers to/from others;
1341 requests = xbt_new(MPI_Request, size - 1);
1342 tmpbufs = xbt_new(void *, rank);
1344 for(other = 0; other < rank; other++) {
1345 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1347 tmpbufs[index] = xbt_malloc(count * dataext);
1349 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1353 for(other = rank + 1; other < size; other++) {
1355 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1358 // Wait for completion of all comms.
1359 smpi_mpi_startall(size - 1, requests);
1360 if(smpi_op_is_commute(op)){
1361 for(other = 0; other < size - 1; other++) {
1362 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1363 if(index == MPI_UNDEFINED) {
1367 if(recvbuf_is_empty){
1368 smpi_datatype_copy(tmpbufs[index], count, datatype, recvbuf, count, datatype);
1371 // #Request is below rank: it's a irecv
1372 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1376 //non commutative case, wait in order
1377 for(other = 0; other < size - 1; other++) {
1378 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1380 if(recvbuf_is_empty){
1381 smpi_datatype_copy(tmpbufs[other], count, datatype, recvbuf, count, datatype);
1383 }else smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1387 for(index = 0; index < rank; index++) {
1388 xbt_free(tmpbufs[index]);