1 /* Copyright (c) 2007-2018. The SimGrid Team. All rights reserved. */
3 /* This program is free software; you can redistribute it and/or modify it
4 * under the terms of the license (GNU LGPL) which comes with this package. */
6 #include "smpi_request.hpp"
10 #include "smpi_comm.hpp"
11 #include "smpi_datatype.hpp"
12 #include "smpi_host.hpp"
13 #include "smpi_op.hpp"
14 #include "src/kernel/activity/CommImpl.hpp"
15 #include "src/mc/mc_replay.hpp"
16 #include "src/simix/ActorImpl.hpp"
17 #include "src/smpi/include/smpi_actor.hpp"
18 #include "xbt/config.hpp"
23 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_request, smpi, "Logging specific to SMPI (reques)");
25 static simgrid::config::Flag<double> smpi_iprobe_sleep(
26 "smpi/iprobe", "Minimum time to inject inside a call to MPI_Iprobe", 1e-4);
27 static simgrid::config::Flag<double> smpi_test_sleep(
28 "smpi/test", "Minimum time to inject inside a call to MPI_Test", 1e-4);
30 std::vector<s_smpi_factor_t> smpi_ois_values;
32 extern void (*smpi_comm_copy_data_callback) (smx_activity_t, void*, size_t);
37 Request::Request(void* buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm, unsigned flags)
38 : buf_(buf), old_type_(datatype), src_(src), dst_(dst), tag_(tag), comm_(comm), flags_(flags)
40 void *old_buf = nullptr;
41 // FIXME Handle the case of a partial shared malloc.
42 if ((((flags & MPI_REQ_RECV) != 0) && ((flags & MPI_REQ_ACCUMULATE) != 0)) || (datatype->flags() & DT_FLAG_DERIVED)) {
43 // This part handles the problem of non-contiguous memory
48 buf_ = xbt_malloc(count*datatype->size());
49 if ((datatype->flags() & DT_FLAG_DERIVED) && ((flags & MPI_REQ_SEND) != 0)) {
50 datatype->serialize(old_buf, buf_, count);
54 // This part handles the problem of non-contiguous memory (for the unserialisation at the reception)
56 size_ = datatype->size() * count;
61 detached_sender_ = nullptr;
66 if (flags & MPI_REQ_PERSISTENT)
74 MPI_Comm Request::comm(){
94 int Request::detached(){
98 MPI_Datatype Request::type(){
102 size_t Request::size(){
106 size_t Request::real_size(){
114 void Request::unref(MPI_Request* request)
116 if((*request) != MPI_REQUEST_NULL){
117 (*request)->refcount_--;
118 if((*request)->refcount_ < 0) {
119 (*request)->print_request("wrong refcount");
120 xbt_die("Whoops, wrong refcount");
122 if((*request)->refcount_==0){
123 Datatype::unref((*request)->old_type_);
124 Comm::unref((*request)->comm_);
125 (*request)->print_request("Destroying");
127 *request = MPI_REQUEST_NULL;
129 (*request)->print_request("Decrementing");
132 xbt_die("freeing an already free request");
136 int Request::match_recv(void* a, void* b, simgrid::kernel::activity::CommImpl* ignored)
138 MPI_Request ref = static_cast<MPI_Request>(a);
139 MPI_Request req = static_cast<MPI_Request>(b);
140 XBT_DEBUG("Trying to match a recv of src %d against %d, tag %d against %d",ref->src_,req->src_, ref->tag_, req->tag_);
142 xbt_assert(ref, "Cannot match recv against null reference");
143 xbt_assert(req, "Cannot match recv against null request");
144 if((ref->src_ == MPI_ANY_SOURCE || req->src_ == ref->src_)
145 && ((ref->tag_ == MPI_ANY_TAG && req->tag_ >=0) || req->tag_ == ref->tag_)){
146 //we match, we can transfer some values
147 if(ref->src_ == MPI_ANY_SOURCE)
148 ref->real_src_ = req->src_;
149 if(ref->tag_ == MPI_ANY_TAG)
150 ref->real_tag_ = req->tag_;
151 if(ref->real_size_ < req->real_size_)
153 if(req->detached_==1)
154 ref->detached_sender_=req; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
155 if(req->cancelled_==0)
156 req->cancelled_=-1;//mark as uncancellable
157 XBT_DEBUG("match succeeded");
162 int Request::match_send(void* a, void* b, simgrid::kernel::activity::CommImpl* ignored)
164 MPI_Request ref = static_cast<MPI_Request>(a);
165 MPI_Request req = static_cast<MPI_Request>(b);
166 XBT_DEBUG("Trying to match a send of src %d against %d, tag %d against %d",ref->src_,req->src_, ref->tag_, req->tag_);
167 xbt_assert(ref, "Cannot match send against null reference");
168 xbt_assert(req, "Cannot match send against null request");
170 if((req->src_ == MPI_ANY_SOURCE || req->src_ == ref->src_)
171 && ((req->tag_ == MPI_ANY_TAG && ref->tag_ >=0)|| req->tag_ == ref->tag_)){
172 if(req->src_ == MPI_ANY_SOURCE)
173 req->real_src_ = ref->src_;
174 if(req->tag_ == MPI_ANY_TAG)
175 req->real_tag_ = ref->tag_;
176 if(req->real_size_ < ref->real_size_)
178 if(ref->detached_==1)
179 req->detached_sender_=ref; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
180 if(req->cancelled_==0)
181 req->cancelled_=-1;//mark as uncancellable
182 XBT_DEBUG("match succeeded");
188 void Request::print_request(const char *message)
190 XBT_VERB("%s request %p [buf = %p, size = %zu, src = %d, dst = %d, tag = %d, flags = %x]",
191 message, this, buf_, size_, src_, dst_, tag_, flags_);
195 /* factories, to hide the internal flags from the caller */
196 MPI_Request Request::send_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
199 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
200 comm->group()->actor(dst)->get_pid(), tag, comm,
201 MPI_REQ_PERSISTENT | MPI_REQ_SEND | MPI_REQ_PREPARED);
204 MPI_Request Request::ssend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
206 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
207 comm->group()->actor(dst)->get_pid(), tag, comm,
208 MPI_REQ_PERSISTENT | MPI_REQ_SSEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
211 MPI_Request Request::isend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
213 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
214 comm->group()->actor(dst)->get_pid(), tag, comm,
215 MPI_REQ_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
219 MPI_Request Request::rma_send_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
222 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
224 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src)->get_pid(),
225 comm->group()->actor(dst)->get_pid(), tag, comm,
226 MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
228 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src)->get_pid(),
229 comm->group()->actor(dst)->get_pid(), tag, comm,
230 MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED |
237 MPI_Request Request::recv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
239 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
240 src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(src)->get_pid(),
241 simgrid::s4u::this_actor::get_pid(), tag, comm,
242 MPI_REQ_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
245 MPI_Request Request::rma_recv_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
248 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
250 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src)->get_pid(),
251 comm->group()->actor(dst)->get_pid(), tag, comm,
252 MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
254 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src)->get_pid(),
255 comm->group()->actor(dst)->get_pid(), tag, comm,
256 MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED | MPI_REQ_ACCUMULATE);
262 MPI_Request Request::irecv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
264 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
265 src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(src)->get_pid(),
266 simgrid::s4u::this_actor::get_pid(), tag, comm,
267 MPI_REQ_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
270 MPI_Request Request::isend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
272 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
273 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
274 comm->group()->actor(dst)->get_pid(), tag, comm,
275 MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND);
280 MPI_Request Request::issend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
282 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
283 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
284 comm->group()->actor(dst)->get_pid(), tag, comm,
285 MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SSEND | MPI_REQ_SEND);
291 MPI_Request Request::irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
293 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
294 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
295 src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(src)->get_pid(),
296 simgrid::s4u::this_actor::get_pid(), tag, comm, MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV);
301 void Request::recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status * status)
303 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
304 request = irecv(buf, count, datatype, src, tag, comm);
305 wait(&request,status);
309 void Request::send(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
311 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
312 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
313 comm->group()->actor(dst)->get_pid(), tag, comm, MPI_REQ_NON_PERSISTENT | MPI_REQ_SEND);
316 wait(&request, MPI_STATUS_IGNORE);
320 void Request::ssend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
322 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
323 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
324 comm->group()->actor(dst)->get_pid(), tag, comm,
325 MPI_REQ_NON_PERSISTENT | MPI_REQ_SSEND | MPI_REQ_SEND);
328 wait(&request,MPI_STATUS_IGNORE);
332 void Request::sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,int dst, int sendtag,
333 void *recvbuf, int recvcount, MPI_Datatype recvtype, int src, int recvtag,
334 MPI_Comm comm, MPI_Status * status)
336 MPI_Request requests[2];
338 int myid = simgrid::s4u::this_actor::get_pid();
339 if ((comm->group()->actor(dst)->get_pid() == myid) && (comm->group()->actor(src)->get_pid() == myid)) {
340 Datatype::copy(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype);
341 if (status != MPI_STATUS_IGNORE) {
342 status->MPI_SOURCE = src;
343 status->MPI_TAG = recvtag;
344 status->MPI_ERROR = MPI_SUCCESS;
345 status->count = sendcount * sendtype->size();
349 requests[0] = isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
350 requests[1] = irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
351 startall(2, requests);
352 waitall(2, requests, stats);
355 if(status != MPI_STATUS_IGNORE) {
356 // Copy receive status
361 void Request::start()
363 smx_mailbox_t mailbox;
365 xbt_assert(action_ == nullptr, "Cannot (re-)start unfinished communication");
366 flags_ &= ~MPI_REQ_PREPARED;
367 flags_ &= ~MPI_REQ_FINISHED;
370 if ((flags_ & MPI_REQ_RECV) != 0) {
371 this->print_request("New recv");
373 simgrid::smpi::ActorExt* process = smpi_process_remote(simgrid::s4u::Actor::by_pid(dst_));
375 int async_small_thresh = simgrid::config::get_value<int>("smpi/async-small-thresh");
377 xbt_mutex_t mut = process->mailboxes_mutex();
378 if (async_small_thresh != 0 || (flags_ & MPI_REQ_RMA) != 0)
379 xbt_mutex_acquire(mut);
381 if (async_small_thresh == 0 && (flags_ & MPI_REQ_RMA) == 0) {
382 mailbox = process->mailbox();
383 } else if (((flags_ & MPI_REQ_RMA) != 0) || static_cast<int>(size_) < async_small_thresh) {
384 //We have to check both mailboxes (because SSEND messages are sent to the large mbox).
385 //begin with the more appropriate one : the small one.
386 mailbox = process->mailbox_small();
387 XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %p (in case of SSEND)?", mailbox);
388 smx_activity_t action = simcall_comm_iprobe(mailbox, 0, &match_recv, static_cast<void*>(this));
390 if (action == nullptr) {
391 mailbox = process->mailbox();
392 XBT_DEBUG("No, nothing in the small mailbox test the other one : %p", mailbox);
393 action = simcall_comm_iprobe(mailbox, 0, &match_recv, static_cast<void*>(this));
394 if (action == nullptr) {
395 XBT_DEBUG("Still nothing, switch back to the small mailbox : %p", mailbox);
396 mailbox = process->mailbox_small();
399 XBT_DEBUG("yes there was something for us in the large mailbox");
402 mailbox = process->mailbox_small();
403 XBT_DEBUG("Is there a corresponding send already posted the small mailbox?");
404 smx_activity_t action = simcall_comm_iprobe(mailbox, 0, &match_recv, static_cast<void*>(this));
406 if (action == nullptr) {
407 XBT_DEBUG("No, nothing in the permanent receive mailbox");
408 mailbox = process->mailbox();
410 XBT_DEBUG("yes there was something for us in the small mailbox");
414 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
416 action_ = simcall_comm_irecv(
417 process->get_actor()->get_impl(), mailbox, buf_, &real_size_, &match_recv,
418 process->replaying() ? &smpi_comm_null_copy_buffer_callback : smpi_comm_copy_data_callback, this, -1.0);
419 XBT_DEBUG("recv simcall posted");
421 if (async_small_thresh != 0 || (flags_ & MPI_REQ_RMA) != 0)
422 xbt_mutex_release(mut);
423 } else { /* the RECV flag was not set, so this is a send */
424 simgrid::smpi::ActorExt* process = smpi_process_remote(simgrid::s4u::Actor::by_pid(dst_));
426 if (TRACE_smpi_view_internals()) {
427 TRACE_smpi_send(rank, rank, dst_, tag_, size_);
429 this->print_request("New send");
432 if ((flags_ & MPI_REQ_SSEND) == 0 &&
433 ((flags_ & MPI_REQ_RMA) != 0 ||
434 static_cast<int>(size_) < simgrid::config::get_value<int>("smpi/send-is-detached-thresh"))) {
435 void *oldbuf = nullptr;
437 XBT_DEBUG("Send request %p is detached", this);
439 if (not(old_type_->flags() & DT_FLAG_DERIVED)) {
441 if (not process->replaying() && oldbuf != nullptr && size_ != 0) {
442 if ((smpi_privatize_global_variables != SmpiPrivStrategies::NONE) &&
443 (static_cast<char*>(buf_) >= smpi_data_exe_start) &&
444 (static_cast<char*>(buf_) < smpi_data_exe_start + smpi_data_exe_size)) {
445 XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
446 smpi_switch_data_segment(simgrid::s4u::Actor::by_pid(src_));
448 buf = xbt_malloc(size_);
449 memcpy(buf,oldbuf,size_);
450 XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
455 //if we are giving back the control to the user without waiting for completion, we have to inject timings
456 double sleeptime = 0.0;
457 if (detached_ != 0 || ((flags_ & (MPI_REQ_ISEND | MPI_REQ_SSEND)) != 0)) { // issend should be treated as isend
458 // isend and send timings may be different
459 sleeptime = ((flags_ & MPI_REQ_ISEND) != 0)
460 ? simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::Host>()->oisend(size_)
461 : simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::Host>()->osend(size_);
465 simcall_process_sleep(sleeptime);
466 XBT_DEBUG("sending size of %zu : sleep %f ", size_, sleeptime);
469 int async_small_thresh = simgrid::config::get_value<int>("smpi/async-small-thresh");
471 xbt_mutex_t mut=process->mailboxes_mutex();
473 if (async_small_thresh != 0 || (flags_ & MPI_REQ_RMA) != 0)
474 xbt_mutex_acquire(mut);
476 if (not(async_small_thresh != 0 || (flags_ & MPI_REQ_RMA) != 0)) {
477 mailbox = process->mailbox();
478 } else if (((flags_ & MPI_REQ_RMA) != 0) || static_cast<int>(size_) < async_small_thresh) { // eager mode
479 mailbox = process->mailbox();
480 XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %p?", mailbox);
481 smx_activity_t action = simcall_comm_iprobe(mailbox, 1, &match_send, static_cast<void*>(this));
482 if (action == nullptr) {
483 if ((flags_ & MPI_REQ_SSEND) == 0) {
484 mailbox = process->mailbox_small();
485 XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %p", mailbox);
487 mailbox = process->mailbox_small();
488 XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %p?", mailbox);
489 action = simcall_comm_iprobe(mailbox, 1, &match_send, static_cast<void*>(this));
490 if (action == nullptr) {
491 XBT_DEBUG("No, we are first, send to large mailbox");
492 mailbox = process->mailbox();
496 XBT_DEBUG("Yes there was something for us in the large mailbox");
499 mailbox = process->mailbox();
500 XBT_DEBUG("Send request %p is in the large mailbox %p (buf: %p)",mailbox, this,buf_);
503 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
505 action_ = simcall_comm_isend(
506 simgrid::s4u::Actor::by_pid(src_)->get_impl(), mailbox, size_, -1.0, buf, real_size_, &match_send,
507 &xbt_free_f, // how to free the userdata if a detached send fails
508 not process->replaying() ? smpi_comm_copy_data_callback : &smpi_comm_null_copy_buffer_callback, this,
509 // detach if msg size < eager/rdv switch limit
511 XBT_DEBUG("send simcall posted");
513 /* FIXME: detached sends are not traceable (action_ == nullptr) */
514 if (action_ != nullptr)
515 simcall_set_category(action_, TRACE_internal_smpi_get_category());
516 if (async_small_thresh != 0 || ((flags_ & MPI_REQ_RMA) != 0))
517 xbt_mutex_release(mut);
521 void Request::startall(int count, MPI_Request * requests)
523 if(requests== nullptr)
526 for(int i = 0; i < count; i++) {
527 requests[i]->start();
531 void Request::cancel()
535 if (this->action_ != nullptr)
536 (boost::static_pointer_cast<simgrid::kernel::activity::CommImpl>(this->action_))->cancel();
539 int Request::test(MPI_Request * request, MPI_Status * status) {
540 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or testall before)
541 // to avoid deadlocks if used as a break condition, such as
542 // while (MPI_Test(request, flag, status) && flag) dostuff...
543 // because the time will not normally advance when only calls to MPI_Test are made -> deadlock
544 // multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
545 static int nsleeps = 1;
546 if(smpi_test_sleep > 0)
547 simcall_process_sleep(nsleeps*smpi_test_sleep);
549 Status::empty(status);
551 if (((*request)->flags_ & MPI_REQ_PREPARED) == 0) {
552 if ((*request)->action_ != nullptr){
554 flag = simcall_comm_test((*request)->action_);
560 finish_wait(request,status);
561 nsleeps=1;//reset the number of sleeps we will do next time
562 if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & MPI_REQ_PERSISTENT) == 0)
563 *request = MPI_REQUEST_NULL;
564 } else if (simgrid::config::get_value<bool>("smpi/grow-injected-times")) {
571 int Request::testsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
576 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
578 for (int i = 0; i < incount; i++) {
579 if (requests[i] != MPI_REQUEST_NULL) {
580 if (test(&requests[i], pstat)) {
583 if (status != MPI_STATUSES_IGNORE)
585 if ((requests[i] != MPI_REQUEST_NULL) && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
586 requests[i] = MPI_REQUEST_NULL;
592 if(count_dead==incount)
593 return MPI_UNDEFINED;
597 int Request::testany(int count, MPI_Request requests[], int *index, MPI_Status * status)
599 std::vector<simgrid::kernel::activity::ActivityImplPtr> comms;
600 comms.reserve(count);
605 *index = MPI_UNDEFINED;
607 std::vector<int> map; /** Maps all matching comms back to their location in requests **/
608 for(i = 0; i < count; i++) {
609 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action_ && not(requests[i]->flags_ & MPI_REQ_PREPARED)) {
610 comms.push_back(requests[i]->action_);
614 if (not map.empty()) {
615 //multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it
616 static int nsleeps = 1;
617 if(smpi_test_sleep > 0)
618 simcall_process_sleep(nsleeps*smpi_test_sleep);
620 i = simcall_comm_testany(comms.data(), comms.size()); // The i-th element in comms matches!
625 if (i != -1) { // -1 is not MPI_UNDEFINED but a SIMIX return code. (nothing matches)
627 finish_wait(&requests[*index],status);
630 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_NON_PERSISTENT)) {
631 requests[*index] = MPI_REQUEST_NULL;
637 //all requests are null or inactive, return true
639 Status::empty(status);
645 int Request::testall(int count, MPI_Request requests[], MPI_Status status[])
648 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
650 for(int i=0; i<count; i++){
651 if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & MPI_REQ_PREPARED)) {
652 if (test(&requests[i], pstat)!=1){
655 requests[i]=MPI_REQUEST_NULL;
658 Status::empty(pstat);
660 if(status != MPI_STATUSES_IGNORE) {
667 void Request::probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
669 //FIXME find another way to avoid busy waiting ?
670 // the issue here is that we have to wait on a nonexistent comm
672 iprobe(source, tag, comm, &flag, status);
673 XBT_DEBUG("Busy Waiting on probing : %d", flag);
677 void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
678 // to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
679 // especially when used as a break condition, such as while (MPI_Iprobe(...)) dostuff...
680 // nsleeps is a multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
681 // This can speed up the execution of certain applications by an order of magnitude, such as HPL
682 static int nsleeps = 1;
683 double speed = simgrid::s4u::Actor::self()->get_host()->get_speed();
684 double maxrate = simgrid::config::get_value<double>("smpi/iprobe-cpu-usage");
685 MPI_Request request = new Request(nullptr, 0, MPI_CHAR,
686 source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(source)->get_pid(),
687 simgrid::s4u::this_actor::get_pid(), tag, comm, MPI_REQ_PERSISTENT | MPI_REQ_RECV);
688 if (smpi_iprobe_sleep > 0) {
689 smx_activity_t iprobe_sleep = simcall_execution_start(
690 "iprobe", /* flops to executek*/ nsleeps * smpi_iprobe_sleep * speed * maxrate, /* priority */ 1.0,
691 /* performance bound */ maxrate * speed, smpi_process()->get_actor()->get_host());
692 simcall_execution_wait(iprobe_sleep);
694 // behave like a receive, but don't do it
695 smx_mailbox_t mailbox;
697 request->print_request("New iprobe");
698 // We have to test both mailboxes as we don't know if we will receive one one or another
699 if (simgrid::config::get_value<int>("smpi/async-small-thresh") > 0) {
700 mailbox = smpi_process()->mailbox_small();
701 XBT_DEBUG("Trying to probe the perm recv mailbox");
702 request->action_ = simcall_comm_iprobe(mailbox, 0, &match_recv, static_cast<void*>(request));
705 if (request->action_ == nullptr){
706 mailbox = smpi_process()->mailbox();
707 XBT_DEBUG("trying to probe the other mailbox");
708 request->action_ = simcall_comm_iprobe(mailbox, 0, &match_recv, static_cast<void*>(request));
711 if (request->action_ != nullptr){
712 simgrid::kernel::activity::CommImplPtr sync_comm =
713 boost::static_pointer_cast<simgrid::kernel::activity::CommImpl>(request->action_);
714 MPI_Request req = static_cast<MPI_Request>(sync_comm->src_data);
716 if (status != MPI_STATUS_IGNORE && (req->flags_ & MPI_REQ_PREPARED) == 0) {
717 status->MPI_SOURCE = comm->group()->rank(req->src_);
718 status->MPI_TAG = req->tag_;
719 status->MPI_ERROR = MPI_SUCCESS;
720 status->count = req->real_size_;
722 nsleeps = 1;//reset the number of sleeps we will do next time
726 if (simgrid::config::get_value<bool>("smpi/grow-injected-times"))
732 void Request::finish_wait(MPI_Request* request, MPI_Status * status)
734 MPI_Request req = *request;
735 Status::empty(status);
737 if (req->cancelled_==1){
738 if (status!=MPI_STATUS_IGNORE)
740 if(req->detached_sender_ != nullptr)
741 unref(&(req->detached_sender_));
746 if (not((req->detached_ != 0) && ((req->flags_ & MPI_REQ_SEND) != 0)) && ((req->flags_ & MPI_REQ_PREPARED) == 0)) {
747 if(status != MPI_STATUS_IGNORE) {
748 int src = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_;
749 status->MPI_SOURCE = req->comm_->group()->rank(src);
750 status->MPI_TAG = req->tag_ == MPI_ANY_TAG ? req->real_tag_ : req->tag_;
751 status->MPI_ERROR = req->truncated_ != 0 ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
752 // this handles the case were size in receive differs from size in send
753 status->count = req->real_size_;
756 req->print_request("Finishing");
757 MPI_Datatype datatype = req->old_type_;
759 // FIXME Handle the case of a partial shared malloc.
760 if (((req->flags_ & MPI_REQ_ACCUMULATE) != 0) ||
761 (datatype->flags() & DT_FLAG_DERIVED)) { // && (not smpi_is_shared(req->old_buf_))){
763 if (not smpi_process()->replaying() && smpi_privatize_global_variables != SmpiPrivStrategies::NONE &&
764 static_cast<char*>(req->old_buf_) >= smpi_data_exe_start &&
765 static_cast<char*>(req->old_buf_) < smpi_data_exe_start + smpi_data_exe_size) {
766 XBT_VERB("Privatization : We are unserializing to a zone in global memory Switch data segment ");
767 smpi_switch_data_segment(simgrid::s4u::Actor::self());
770 if(datatype->flags() & DT_FLAG_DERIVED){
771 // This part handles the problem of non-contignous memory the unserialization at the reception
772 if ((req->flags_ & MPI_REQ_RECV) && datatype->size() != 0)
773 datatype->unserialize(req->buf_, req->old_buf_, req->real_size_/datatype->size() , req->op_);
775 } else if (req->flags_ & MPI_REQ_RECV) { // apply op on contiguous buffer for accumulate
776 if (datatype->size() != 0) {
777 int n = req->real_size_ / datatype->size();
778 req->op_->apply(req->buf_, req->old_buf_, &n, datatype);
785 if (TRACE_smpi_view_internals() && ((req->flags_ & MPI_REQ_RECV) != 0)) {
786 int rank = simgrid::s4u::this_actor::get_pid();
787 int src_traced = (req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_);
788 TRACE_smpi_recv(src_traced, rank,req->tag_);
790 if(req->detached_sender_ != nullptr){
791 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
793 simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::Host>()->orecv(req->real_size());
795 simcall_process_sleep(sleeptime);
796 XBT_DEBUG("receiving size of %zu : sleep %f ", req->real_size_, sleeptime);
798 unref(&(req->detached_sender_));
800 if (req->flags_ & MPI_REQ_PERSISTENT)
801 req->action_ = nullptr;
802 req->flags_ |= MPI_REQ_FINISHED;
806 void Request::wait(MPI_Request * request, MPI_Status * status)
808 (*request)->print_request("Waiting");
809 if ((*request)->flags_ & MPI_REQ_PREPARED) {
810 Status::empty(status);
814 if ((*request)->action_ != nullptr){
816 // this is not a detached send
817 simcall_comm_wait((*request)->action_, -1.0);
819 XBT_VERB("Request cancelled");
824 finish_wait(request,status);
825 if (*request != MPI_REQUEST_NULL && (((*request)->flags_ & MPI_REQ_NON_PERSISTENT) != 0))
826 *request = MPI_REQUEST_NULL;
829 int Request::waitany(int count, MPI_Request requests[], MPI_Status * status)
831 s_xbt_dynar_t comms; // Keep it on stack to save some extra mallocs
832 int index = MPI_UNDEFINED;
836 // Wait for a request to complete
837 xbt_dynar_init(&comms, sizeof(smx_activity_t), [](void*ptr){
838 intrusive_ptr_release(*(simgrid::kernel::activity::ActivityImpl**)ptr);
840 int *map = xbt_new(int, count);
841 XBT_DEBUG("Wait for one of %d", count);
842 for(int i = 0; i < count; i++) {
843 if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & MPI_REQ_PREPARED) &&
844 not(requests[i]->flags_ & MPI_REQ_FINISHED)) {
845 if (requests[i]->action_ != nullptr) {
846 XBT_DEBUG("Waiting any %p ", requests[i]);
847 intrusive_ptr_add_ref(requests[i]->action_.get());
848 xbt_dynar_push_as(&comms, simgrid::kernel::activity::ActivityImpl*, requests[i]->action_.get());
852 // This is a finished detached request, let's return this one
853 size = 0; // so we free the dynar but don't do the waitany call
855 finish_wait(&requests[i], status); // cleanup if refcount = 0
856 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
857 requests[i] = MPI_REQUEST_NULL; // set to null
863 XBT_DEBUG("Enter waitany for %lu comms", xbt_dynar_length(&comms));
866 // this is not a detached send
867 i = simcall_comm_waitany(&comms, -1);
869 XBT_INFO("request %d cancelled ",i);
873 // not MPI_UNDEFINED, as this is a simix return code
876 //in case of an accumulate, we have to wait the end of all requests to apply the operation, ordered correctly.
877 if ((requests[index] == MPI_REQUEST_NULL) ||
878 (not((requests[index]->flags_ & MPI_REQ_ACCUMULATE) && (requests[index]->flags_ & MPI_REQ_RECV)))) {
879 finish_wait(&requests[index],status);
880 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_NON_PERSISTENT))
881 requests[index] = MPI_REQUEST_NULL;
886 xbt_dynar_free_data(&comms);
890 if (index==MPI_UNDEFINED)
891 Status::empty(status);
896 static int sort_accumulates(MPI_Request a, MPI_Request b)
898 return (a->tag() > b->tag());
901 int Request::waitall(int count, MPI_Request requests[], MPI_Status status[])
903 std::vector<MPI_Request> accumulates;
906 MPI_Status *pstat = (status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat);
907 int retvalue = MPI_SUCCESS;
908 //tag invalid requests in the set
909 if (status != MPI_STATUSES_IGNORE) {
910 for (int c = 0; c < count; c++) {
911 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst_ == MPI_PROC_NULL ||
912 (requests[c]->flags_ & MPI_REQ_PREPARED)) {
913 Status::empty(&status[c]);
914 } else if (requests[c]->src_ == MPI_PROC_NULL) {
915 Status::empty(&status[c]);
916 status[c].MPI_SOURCE = MPI_PROC_NULL;
920 for (int c = 0; c < count; c++) {
921 if (MC_is_active() || MC_record_replay_is_active()) {
922 wait(&requests[c],pstat);
925 index = waitany(count, (MPI_Request*)requests, pstat);
927 if (index == MPI_UNDEFINED)
930 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_RECV) &&
931 (requests[index]->flags_ & MPI_REQ_ACCUMULATE))
932 accumulates.push_back(requests[index]);
933 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_NON_PERSISTENT))
934 requests[index] = MPI_REQUEST_NULL;
936 if (status != MPI_STATUSES_IGNORE) {
937 status[index] = *pstat;
938 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
939 retvalue = MPI_ERR_IN_STATUS;
943 if (not accumulates.empty()) {
944 std::sort(accumulates.begin(), accumulates.end(), sort_accumulates);
945 for (auto& req : accumulates) {
946 finish_wait(&req, status);
953 int Request::waitsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
957 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
959 for (int i = 0; i < incount; i++) {
960 int index = waitany(incount, requests, pstat);
961 if(index!=MPI_UNDEFINED){
962 indices[count] = index;
964 if(status != MPI_STATUSES_IGNORE) {
965 status[index] = *pstat;
967 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_NON_PERSISTENT))
968 requests[index] = MPI_REQUEST_NULL;
970 return MPI_UNDEFINED;
976 MPI_Request Request::f2c(int id) {
978 if(id==MPI_FORTRAN_REQUEST_NULL)
979 return static_cast<MPI_Request>(MPI_REQUEST_NULL);
980 return static_cast<MPI_Request>(F2C::f2c_lookup()->at(get_key_id(key, id)));
985 if (F2C::f2c_lookup() == nullptr) {
986 F2C::set_f2c_lookup(new std::unordered_map<std::string, F2C*>);
989 (*(F2C::f2c_lookup()))[get_key_id(key, F2C::f2c_id())] = this;
990 F2C::f2c_id_increment();
991 return F2C::f2c_id()-1;
994 void Request::free_f(int id)
996 if (id != MPI_FORTRAN_REQUEST_NULL) {
998 F2C::f2c_lookup()->erase(get_key_id(key, id));