1 /* Copyright (c) 2007-2019. 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. */
5 #include "simgrid/s4u/Mutex.hpp"
6 #include "simgrid/s4u/ConditionVariable.hpp"
7 #include "smpi_request.hpp"
10 #include "private.hpp"
11 #include "simgrid/Exception.hpp"
12 #include "simgrid/s4u/Exec.hpp"
13 #include "smpi_comm.hpp"
14 #include "smpi_datatype.hpp"
15 #include "smpi_host.hpp"
16 #include "smpi_op.hpp"
17 #include "src/kernel/activity/CommImpl.hpp"
18 #include "src/mc/mc_replay.hpp"
19 #include "src/smpi/include/smpi_actor.hpp"
20 #include "xbt/config.hpp"
25 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_request, smpi, "Logging specific to SMPI (request)");
27 static simgrid::config::Flag<double> smpi_iprobe_sleep(
28 "smpi/iprobe", "Minimum time to inject inside a call to MPI_Iprobe", 1e-4);
29 static simgrid::config::Flag<double> smpi_test_sleep(
30 "smpi/test", "Minimum time to inject inside a call to MPI_Test", 1e-4);
32 std::vector<s_smpi_factor_t> smpi_ois_values;
34 extern void (*smpi_comm_copy_data_callback)(simgrid::kernel::activity::CommImpl*, void*, size_t);
39 Request::Request(const void* buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm, unsigned flags, MPI_Op op)
40 : buf_(const_cast<void*>(buf)), old_type_(datatype), src_(src), dst_(dst), tag_(tag), comm_(comm), flags_(flags), op_(op)
42 void *old_buf = nullptr;
43 // FIXME Handle the case of a partial shared malloc.
44 if ((((flags & MPI_REQ_RECV) != 0) && ((flags & MPI_REQ_ACCUMULATE) != 0)) || (datatype->flags() & DT_FLAG_DERIVED)) {
45 // This part handles the problem of non-contiguous memory
46 old_buf = const_cast<void*>(buf);
50 buf_ = xbt_malloc(count*datatype->size());
51 if ((datatype->flags() & DT_FLAG_DERIVED) && ((flags & MPI_REQ_SEND) != 0)) {
52 datatype->serialize(old_buf, buf_, count);
56 // This part handles the problem of non-contiguous memory (for the unserialisation at the reception)
58 size_ = datatype->size() * count;
61 if(op != MPI_REPLACE && op != MPI_OP_NULL)
65 detached_sender_ = nullptr;
70 if (flags & MPI_REQ_PERSISTENT)
75 generalized_funcs=nullptr;
76 nbc_requests_=nullptr;
84 void Request::unref(MPI_Request* request)
86 if((*request) != MPI_REQUEST_NULL){
87 (*request)->refcount_--;
88 if((*request)->refcount_ < 0) {
89 (*request)->print_request("wrong refcount");
90 xbt_die("Whoops, wrong refcount");
92 if((*request)->refcount_==0){
93 if ((*request)->flags_ & MPI_REQ_GENERALIZED){
94 ((*request)->generalized_funcs)->free_fn(((*request)->generalized_funcs)->extra_state);
95 delete (*request)->generalized_funcs;
97 Comm::unref((*request)->comm_);
98 Datatype::unref((*request)->old_type_);
100 if ((*request)->op_!=MPI_REPLACE && (*request)->op_!=MPI_OP_NULL)
101 Op::unref(&(*request)->op_);
103 (*request)->print_request("Destroying");
105 *request = MPI_REQUEST_NULL;
107 (*request)->print_request("Decrementing");
110 xbt_die("freeing an already free request");
114 int Request::match_recv(void* a, void* b, simgrid::kernel::activity::CommImpl*)
116 MPI_Request ref = static_cast<MPI_Request>(a);
117 MPI_Request req = static_cast<MPI_Request>(b);
118 XBT_DEBUG("Trying to match a recv of src %d against %d, tag %d against %d, id %d against %d",ref->src_,req->src_, ref->tag_, req->tag_,ref->comm_->id(),req->comm_->id());
120 xbt_assert(ref, "Cannot match recv against null reference");
121 xbt_assert(req, "Cannot match recv against null request");
122 if((ref->comm_->id()==MPI_UNDEFINED || req->comm_->id() == MPI_UNDEFINED || (ref->comm_->id()==req->comm_->id()))
123 && ((ref->src_ == MPI_ANY_SOURCE && (ref->comm_->group()->rank(req->src_) != MPI_UNDEFINED)) || req->src_ == ref->src_)
124 && ((ref->tag_ == MPI_ANY_TAG && req->tag_ >=0) || req->tag_ == ref->tag_)){
125 //we match, we can transfer some values
126 if(ref->src_ == MPI_ANY_SOURCE)
127 ref->real_src_ = req->src_;
128 if(ref->tag_ == MPI_ANY_TAG)
129 ref->real_tag_ = req->tag_;
130 if(ref->real_size_ < req->real_size_)
131 ref->truncated_ = true;
133 ref->detached_sender_=req; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
134 if(req->cancelled_==0)
135 req->cancelled_=-1;//mark as uncancellable
136 XBT_DEBUG("match succeeded");
141 int Request::match_send(void* a, void* b, simgrid::kernel::activity::CommImpl*)
143 MPI_Request ref = static_cast<MPI_Request>(a);
144 MPI_Request req = static_cast<MPI_Request>(b);
145 XBT_DEBUG("Trying to match a send of src %d against %d, tag %d against %d, id %d against %d",ref->src_,req->src_, ref->tag_, req->tag_,ref->comm_->id(),req->comm_->id());
146 xbt_assert(ref, "Cannot match send against null reference");
147 xbt_assert(req, "Cannot match send against null request");
149 if((ref->comm_->id()==MPI_UNDEFINED || req->comm_->id() == MPI_UNDEFINED || (ref->comm_->id()==req->comm_->id()))
150 && ((req->src_ == MPI_ANY_SOURCE && (req->comm_->group()->rank(ref->src_) != MPI_UNDEFINED)) || req->src_ == ref->src_)
151 && ((req->tag_ == MPI_ANY_TAG && ref->tag_ >=0)|| req->tag_ == ref->tag_)){
152 if(req->src_ == MPI_ANY_SOURCE)
153 req->real_src_ = ref->src_;
154 if(req->tag_ == MPI_ANY_TAG)
155 req->real_tag_ = ref->tag_;
156 if(req->real_size_ < ref->real_size_)
157 req->truncated_ = true;
159 req->detached_sender_=ref; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
160 if(req->cancelled_==0)
161 req->cancelled_=-1;//mark as uncancellable
162 XBT_DEBUG("match succeeded");
168 void Request::print_request(const char *message)
170 XBT_VERB("%s request %p [buf = %p, size = %zu, src = %d, dst = %d, tag = %d, flags = %x]",
171 message, this, buf_, size_, src_, dst_, tag_, flags_);
175 /* factories, to hide the internal flags from the caller */
176 MPI_Request Request::send_init(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
179 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
180 comm->group()->actor(dst)->get_pid(), tag, comm,
181 MPI_REQ_PERSISTENT | MPI_REQ_SEND | MPI_REQ_PREPARED);
184 MPI_Request Request::ssend_init(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
186 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
187 comm->group()->actor(dst)->get_pid(), tag, comm,
188 MPI_REQ_PERSISTENT | MPI_REQ_SSEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
191 MPI_Request Request::isend_init(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
193 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
194 comm->group()->actor(dst)->get_pid(), tag, comm,
195 MPI_REQ_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
199 MPI_Request Request::rma_send_init(const void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
202 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
204 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src)->get_pid(),
205 comm->group()->actor(dst)->get_pid(), tag, comm,
206 MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
208 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src)->get_pid(),
209 comm->group()->actor(dst)->get_pid(), tag, comm,
210 MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED |
211 MPI_REQ_ACCUMULATE, op);
216 MPI_Request Request::recv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
218 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
219 src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(src)->get_pid(),
220 simgrid::s4u::this_actor::get_pid(), tag, comm,
221 MPI_REQ_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
224 MPI_Request Request::rma_recv_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
227 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
229 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src)->get_pid(),
230 comm->group()->actor(dst)->get_pid(), tag, comm,
231 MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
233 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src)->get_pid(),
234 comm->group()->actor(dst)->get_pid(), tag, comm,
235 MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED | MPI_REQ_ACCUMULATE, op);
240 MPI_Request Request::irecv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
242 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
243 src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(src)->get_pid(),
244 simgrid::s4u::this_actor::get_pid(), tag, comm,
245 MPI_REQ_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
248 MPI_Request Request::isend(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
250 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
251 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
252 comm->group()->actor(dst)->get_pid(), tag, comm,
253 MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND);
258 MPI_Request Request::issend(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
260 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
261 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
262 comm->group()->actor(dst)->get_pid(), tag, comm,
263 MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SSEND | MPI_REQ_SEND);
269 MPI_Request Request::irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
271 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
272 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
273 src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(src)->get_pid(),
274 simgrid::s4u::this_actor::get_pid(), tag, comm, MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV);
279 void Request::recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status * status)
281 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
282 request = irecv(buf, count, datatype, src, tag, comm);
283 wait(&request,status);
287 void Request::send(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
289 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
290 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
291 comm->group()->actor(dst)->get_pid(), tag, comm, MPI_REQ_NON_PERSISTENT | MPI_REQ_SEND);
294 wait(&request, MPI_STATUS_IGNORE);
298 void Request::ssend(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
300 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
301 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
302 comm->group()->actor(dst)->get_pid(), tag, comm,
303 MPI_REQ_NON_PERSISTENT | MPI_REQ_SSEND | MPI_REQ_SEND);
306 wait(&request,MPI_STATUS_IGNORE);
310 void Request::sendrecv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,int dst, int sendtag,
311 void *recvbuf, int recvcount, MPI_Datatype recvtype, int src, int recvtag,
312 MPI_Comm comm, MPI_Status * status)
314 MPI_Request requests[2];
316 int myid = simgrid::s4u::this_actor::get_pid();
317 if ((comm->group()->actor(dst)->get_pid() == myid) && (comm->group()->actor(src)->get_pid() == myid)) {
318 Datatype::copy(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype);
319 if (status != MPI_STATUS_IGNORE) {
320 status->MPI_SOURCE = src;
321 status->MPI_TAG = recvtag;
322 status->MPI_ERROR = MPI_SUCCESS;
323 status->count = sendcount * sendtype->size();
327 requests[0] = isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
328 requests[1] = irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
329 startall(2, requests);
330 waitall(2, requests, stats);
333 if(status != MPI_STATUS_IGNORE) {
334 // Copy receive status
339 void Request::start()
341 s4u::Mailbox* mailbox;
343 xbt_assert(action_ == nullptr, "Cannot (re-)start unfinished communication");
344 flags_ &= ~MPI_REQ_PREPARED;
345 flags_ &= ~MPI_REQ_FINISHED;
348 if ((flags_ & MPI_REQ_RECV) != 0) {
349 this->print_request("New recv");
351 simgrid::smpi::ActorExt* process = smpi_process_remote(simgrid::s4u::Actor::by_pid(dst_));
353 int async_small_thresh = simgrid::config::get_value<int>("smpi/async-small-thresh");
355 simgrid::s4u::MutexPtr mut = process->mailboxes_mutex();
356 if (async_small_thresh != 0 || (flags_ & MPI_REQ_RMA) != 0)
359 if (async_small_thresh == 0 && (flags_ & MPI_REQ_RMA) == 0) {
360 mailbox = process->mailbox();
361 } else if (((flags_ & MPI_REQ_RMA) != 0) || static_cast<int>(size_) < async_small_thresh) {
362 //We have to check both mailboxes (because SSEND messages are sent to the large mbox).
363 //begin with the more appropriate one : the small one.
364 mailbox = process->mailbox_small();
365 XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %s (in case of SSEND)?",
366 mailbox->get_cname());
367 smx_activity_t action = mailbox->iprobe(0, &match_recv, static_cast<void*>(this));
369 if (action == nullptr) {
370 mailbox = process->mailbox();
371 XBT_DEBUG("No, nothing in the small mailbox test the other one : %s", mailbox->get_cname());
372 action = mailbox->iprobe(0, &match_recv, static_cast<void*>(this));
373 if (action == nullptr) {
374 XBT_DEBUG("Still nothing, switch back to the small mailbox : %s", mailbox->get_cname());
375 mailbox = process->mailbox_small();
378 XBT_DEBUG("yes there was something for us in the large mailbox");
381 mailbox = process->mailbox_small();
382 XBT_DEBUG("Is there a corresponding send already posted the small mailbox?");
383 smx_activity_t action = mailbox->iprobe(0, &match_recv, static_cast<void*>(this));
385 if (action == nullptr) {
386 XBT_DEBUG("No, nothing in the permanent receive mailbox");
387 mailbox = process->mailbox();
389 XBT_DEBUG("yes there was something for us in the small mailbox");
393 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
395 action_ = simcall_comm_irecv(
396 process->get_actor()->get_impl(), mailbox->get_impl(), buf_, &real_size_, &match_recv,
397 process->replaying() ? &smpi_comm_null_copy_buffer_callback : smpi_comm_copy_data_callback, this, -1.0);
398 XBT_DEBUG("recv simcall posted");
400 if (async_small_thresh != 0 || (flags_ & MPI_REQ_RMA) != 0)
402 } else { /* the RECV flag was not set, so this is a send */
403 simgrid::smpi::ActorExt* process = smpi_process_remote(simgrid::s4u::Actor::by_pid(dst_));
405 if (TRACE_smpi_view_internals()) {
406 TRACE_smpi_send(rank, rank, dst_, tag_, size_);
408 this->print_request("New send");
411 if ((flags_ & MPI_REQ_SSEND) == 0 &&
412 ((flags_ & MPI_REQ_RMA) != 0 ||
413 static_cast<int>(size_) < simgrid::config::get_value<int>("smpi/send-is-detached-thresh"))) {
414 void *oldbuf = nullptr;
416 XBT_DEBUG("Send request %p is detached", this);
418 if (not(old_type_->flags() & DT_FLAG_DERIVED)) {
420 if (not process->replaying() && oldbuf != nullptr && size_ != 0) {
421 if ((smpi_privatize_global_variables != SmpiPrivStrategies::NONE) &&
422 (static_cast<char*>(buf_) >= smpi_data_exe_start) &&
423 (static_cast<char*>(buf_) < smpi_data_exe_start + smpi_data_exe_size)) {
424 XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
425 smpi_switch_data_segment(simgrid::s4u::Actor::by_pid(src_));
427 buf = xbt_malloc(size_);
428 memcpy(buf,oldbuf,size_);
429 XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
434 //if we are giving back the control to the user without waiting for completion, we have to inject timings
435 double sleeptime = 0.0;
436 if (detached_ || ((flags_ & (MPI_REQ_ISEND | MPI_REQ_SSEND)) != 0)) { // issend should be treated as isend
437 // isend and send timings may be different
438 sleeptime = ((flags_ & MPI_REQ_ISEND) != 0)
439 ? simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::Host>()->oisend(size_)
440 : simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::Host>()->osend(size_);
444 simcall_process_sleep(sleeptime);
445 XBT_DEBUG("sending size of %zu : sleep %f ", size_, sleeptime);
448 int async_small_thresh = simgrid::config::get_value<int>("smpi/async-small-thresh");
450 simgrid::s4u::MutexPtr mut = process->mailboxes_mutex();
452 if (async_small_thresh != 0 || (flags_ & MPI_REQ_RMA) != 0)
455 if (not(async_small_thresh != 0 || (flags_ & MPI_REQ_RMA) != 0)) {
456 mailbox = process->mailbox();
457 } else if (((flags_ & MPI_REQ_RMA) != 0) || static_cast<int>(size_) < async_small_thresh) { // eager mode
458 mailbox = process->mailbox();
459 XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %s?", mailbox->get_cname());
460 smx_activity_t action = mailbox->iprobe(1, &match_send, static_cast<void*>(this));
461 if (action == nullptr) {
462 if ((flags_ & MPI_REQ_SSEND) == 0) {
463 mailbox = process->mailbox_small();
464 XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %s",
465 mailbox->get_cname());
467 mailbox = process->mailbox_small();
468 XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %s?",
469 mailbox->get_cname());
470 action = mailbox->iprobe(1, &match_send, static_cast<void*>(this));
471 if (action == nullptr) {
472 XBT_DEBUG("No, we are first, send to large mailbox");
473 mailbox = process->mailbox();
477 XBT_DEBUG("Yes there was something for us in the large mailbox");
480 mailbox = process->mailbox();
481 XBT_DEBUG("Send request %p is in the large mailbox %s (buf: %p)", this, mailbox->get_cname(), buf_);
484 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
486 action_ = simcall_comm_isend(
487 simgrid::s4u::Actor::by_pid(src_)->get_impl(), mailbox->get_impl(), size_, -1.0, buf, real_size_, &match_send,
488 &xbt_free_f, // how to free the userdata if a detached send fails
489 not process->replaying() ? smpi_comm_copy_data_callback : &smpi_comm_null_copy_buffer_callback, this,
490 // detach if msg size < eager/rdv switch limit
492 XBT_DEBUG("send simcall posted");
494 /* FIXME: detached sends are not traceable (action_ == nullptr) */
495 if (action_ != nullptr) {
496 boost::static_pointer_cast<kernel::activity::CommImpl>(action_)->set_tracing_category(
497 smpi_process()->get_tracing_category());
500 if (async_small_thresh != 0 || ((flags_ & MPI_REQ_RMA) != 0))
505 void Request::startall(int count, MPI_Request * requests)
507 if(requests== nullptr)
510 for(int i = 0; i < count; i++) {
511 requests[i]->start();
515 void Request::cancel()
519 if (this->action_ != nullptr)
520 (boost::static_pointer_cast<simgrid::kernel::activity::CommImpl>(this->action_))->cancel();
523 int Request::test(MPI_Request * request, MPI_Status * status, int* flag) {
524 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or testall before)
525 // to avoid deadlocks if used as a break condition, such as
526 // while (MPI_Test(request, flag, status) && flag) dostuff...
527 // because the time will not normally advance when only calls to MPI_Test are made -> deadlock
528 // multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
529 static int nsleeps = 1;
530 int ret = MPI_SUCCESS;
532 // Are we testing a request meant for non blocking collectives ?
533 // If so, test all the subrequests.
534 if ((*request)->nbc_requests_size_>0){
535 ret = testall((*request)->nbc_requests_size_, (*request)->nbc_requests_, flag, MPI_STATUSES_IGNORE);
537 delete[] (*request)->nbc_requests_;
538 (*request)->nbc_requests_size_=0;
544 if(smpi_test_sleep > 0)
545 simcall_process_sleep(nsleeps*smpi_test_sleep);
547 Status::empty(status);
549 if (((*request)->flags_ & MPI_REQ_PREPARED) == 0) {
550 if ((*request)->action_ != nullptr){
552 *flag = simcall_comm_test((*request)->action_);
553 } catch (const Exception&) {
558 if (*request != MPI_REQUEST_NULL &&
559 ((*request)->flags_ & MPI_REQ_GENERALIZED)
560 && !((*request)->flags_ & MPI_REQ_COMPLETE))
563 finish_wait(request,status);
564 if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & MPI_REQ_GENERALIZED)){
565 MPI_Status* mystatus;
566 if(status==MPI_STATUS_IGNORE){
567 mystatus=new MPI_Status();
568 Status::empty(mystatus);
572 ret = ((*request)->generalized_funcs)->query_fn(((*request)->generalized_funcs)->extra_state, mystatus);
573 if(status==MPI_STATUS_IGNORE)
576 nsleeps=1;//reset the number of sleeps we will do next time
577 if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & MPI_REQ_PERSISTENT) == 0)
578 *request = MPI_REQUEST_NULL;
579 } else if (simgrid::config::get_value<bool>("smpi/grow-injected-times")) {
586 int Request::testsome(int incount, MPI_Request requests[], int *count, int *indices, MPI_Status status[])
588 int ret = MPI_SUCCESS;
593 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
596 for (int i = 0; i < incount; i++) {
597 if (requests[i] != MPI_REQUEST_NULL) {
598 ret = test(&requests[i], pstat, &flag);
604 if (status != MPI_STATUSES_IGNORE)
606 if ((requests[i] != MPI_REQUEST_NULL) && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
607 requests[i] = MPI_REQUEST_NULL;
613 if(count_dead==incount)*count=MPI_UNDEFINED;
615 return MPI_ERR_IN_STATUS;
620 int Request::testany(int count, MPI_Request requests[], int *index, int* flag, MPI_Status * status)
622 std::vector<simgrid::kernel::activity::CommImpl*> comms;
623 comms.reserve(count);
627 int ret = MPI_SUCCESS;
628 *index = MPI_UNDEFINED;
630 std::vector<int> map; /** Maps all matching comms back to their location in requests **/
631 for(i = 0; i < count; i++) {
632 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action_ && not(requests[i]->flags_ & MPI_REQ_PREPARED)) {
633 comms.push_back(static_cast<simgrid::kernel::activity::CommImpl*>(requests[i]->action_.get()));
637 if (not map.empty()) {
638 //multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it
639 static int nsleeps = 1;
640 if(smpi_test_sleep > 0)
641 simcall_process_sleep(nsleeps*smpi_test_sleep);
643 i = simcall_comm_testany(comms.data(), comms.size()); // The i-th element in comms matches!
644 } catch (const Exception&) {
645 XBT_DEBUG("Exception in testany");
649 if (i != -1) { // -1 is not MPI_UNDEFINED but a SIMIX return code. (nothing matches)
651 if (requests[*index] != MPI_REQUEST_NULL &&
652 (requests[*index]->flags_ & MPI_REQ_GENERALIZED)
653 && !(requests[*index]->flags_ & MPI_REQ_COMPLETE)) {
656 finish_wait(&requests[*index],status);
657 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_GENERALIZED)){
658 MPI_Status* mystatus;
659 if(status==MPI_STATUS_IGNORE){
660 mystatus=new MPI_Status();
661 Status::empty(mystatus);
665 ret=(requests[*index]->generalized_funcs)->query_fn((requests[*index]->generalized_funcs)->extra_state, mystatus);
666 if(status==MPI_STATUS_IGNORE)
670 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_NON_PERSISTENT))
671 requests[*index] = MPI_REQUEST_NULL;
672 XBT_DEBUG("Testany - returning with index %d", *index);
680 XBT_DEBUG("Testany on inactive handles, returning flag=1 but empty status");
681 //all requests are null or inactive, return true
683 *index = MPI_UNDEFINED;
684 Status::empty(status);
690 int Request::testall(int count, MPI_Request requests[], int* outflag, MPI_Status status[])
693 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
698 for(int i=0; i<count; i++){
699 if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & MPI_REQ_PREPARED)) {
700 ret = test(&requests[i], pstat, &flag);
703 requests[i]=MPI_REQUEST_NULL;
707 if (ret != MPI_SUCCESS)
710 Status::empty(pstat);
712 if(status != MPI_STATUSES_IGNORE) {
717 return MPI_ERR_IN_STATUS;
722 void Request::probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
724 //FIXME find another way to avoid busy waiting ?
725 // the issue here is that we have to wait on a nonexistent comm
727 iprobe(source, tag, comm, &flag, status);
728 XBT_DEBUG("Busy Waiting on probing : %d", flag);
732 void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
733 // to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
734 // especially when used as a break condition, such as while (MPI_Iprobe(...)) dostuff...
735 // nsleeps is a multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
736 // This can speed up the execution of certain applications by an order of magnitude, such as HPL
737 static int nsleeps = 1;
738 double speed = s4u::this_actor::get_host()->get_speed();
739 double maxrate = simgrid::config::get_value<double>("smpi/iprobe-cpu-usage");
740 MPI_Request request = new Request(nullptr, 0, MPI_CHAR,
741 source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(source)->get_pid(),
742 simgrid::s4u::this_actor::get_pid(), tag, comm, MPI_REQ_PERSISTENT | MPI_REQ_RECV);
743 if (smpi_iprobe_sleep > 0) {
744 /** Compute the number of flops we will sleep **/
745 s4u::this_actor::exec_init(/*nsleeps: See comment above */ nsleeps *
746 /*(seconds * flop/s -> total flops)*/ smpi_iprobe_sleep * speed * maxrate)
748 /* Not the entire CPU can be used when iprobing: This is important for
749 * the energy consumption caused by polling with iprobes.
750 * Note also that the number of flops that was
751 * computed above contains a maxrate factor and is hence reduced (maxrate < 1)
753 ->set_bound(maxrate*speed)
757 // behave like a receive, but don't do it
758 s4u::Mailbox* mailbox;
760 request->print_request("New iprobe");
761 // We have to test both mailboxes as we don't know if we will receive one one or another
762 if (simgrid::config::get_value<int>("smpi/async-small-thresh") > 0) {
763 mailbox = smpi_process()->mailbox_small();
764 XBT_DEBUG("Trying to probe the perm recv mailbox");
765 request->action_ = mailbox->iprobe(0, &match_recv, static_cast<void*>(request));
768 if (request->action_ == nullptr){
769 mailbox = smpi_process()->mailbox();
770 XBT_DEBUG("trying to probe the other mailbox");
771 request->action_ = mailbox->iprobe(0, &match_recv, static_cast<void*>(request));
774 if (request->action_ != nullptr){
775 kernel::activity::CommImplPtr sync_comm = boost::static_pointer_cast<kernel::activity::CommImpl>(request->action_);
776 MPI_Request req = static_cast<MPI_Request>(sync_comm->src_data_);
778 if (status != MPI_STATUS_IGNORE && (req->flags_ & MPI_REQ_PREPARED) == 0) {
779 status->MPI_SOURCE = comm->group()->rank(req->src_);
780 status->MPI_TAG = req->tag_;
781 status->MPI_ERROR = MPI_SUCCESS;
782 status->count = req->real_size_;
784 nsleeps = 1;//reset the number of sleeps we will do next time
788 if (simgrid::config::get_value<bool>("smpi/grow-injected-times"))
792 xbt_assert(request == MPI_REQUEST_NULL);
795 void Request::finish_wait(MPI_Request* request, MPI_Status * status)
797 MPI_Request req = *request;
798 Status::empty(status);
800 if (req->cancelled_==1){
801 if (status!=MPI_STATUS_IGNORE)
803 if(req->detached_sender_ != nullptr)
804 unref(&(req->detached_sender_));
809 if (not(req->detached_ && ((req->flags_ & MPI_REQ_SEND) != 0)) && ((req->flags_ & MPI_REQ_PREPARED) == 0) &&
810 ((req->flags_ & MPI_REQ_GENERALIZED) == 0)) {
811 if(status != MPI_STATUS_IGNORE) {
812 int src = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_;
813 status->MPI_SOURCE = req->comm_->group()->rank(src);
814 status->MPI_TAG = req->tag_ == MPI_ANY_TAG ? req->real_tag_ : req->tag_;
815 status->MPI_ERROR = req->truncated_ ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
816 // this handles the case were size in receive differs from size in send
817 status->count = req->real_size_;
820 req->print_request("Finishing");
821 MPI_Datatype datatype = req->old_type_;
823 // FIXME Handle the case of a partial shared malloc.
824 if (((req->flags_ & MPI_REQ_ACCUMULATE) != 0) ||
825 (datatype->flags() & DT_FLAG_DERIVED)) { // && (not smpi_is_shared(req->old_buf_))){
827 if (not smpi_process()->replaying() && smpi_privatize_global_variables != SmpiPrivStrategies::NONE &&
828 static_cast<char*>(req->old_buf_) >= smpi_data_exe_start &&
829 static_cast<char*>(req->old_buf_) < smpi_data_exe_start + smpi_data_exe_size) {
830 XBT_VERB("Privatization : We are unserializing to a zone in global memory Switch data segment ");
831 smpi_switch_data_segment(simgrid::s4u::Actor::self());
834 if(datatype->flags() & DT_FLAG_DERIVED){
835 // This part handles the problem of non-contignous memory the unserialization at the reception
836 if ((req->flags_ & MPI_REQ_RECV) && datatype->size() != 0)
837 datatype->unserialize(req->buf_, req->old_buf_, req->real_size_/datatype->size() , req->op_);
839 } else if (req->flags_ & MPI_REQ_RECV) { // apply op on contiguous buffer for accumulate
840 if (datatype->size() != 0) {
841 int n = req->real_size_ / datatype->size();
842 req->op_->apply(req->buf_, req->old_buf_, &n, datatype);
849 if (TRACE_smpi_view_internals() && ((req->flags_ & MPI_REQ_RECV) != 0)) {
850 int rank = simgrid::s4u::this_actor::get_pid();
851 int src_traced = (req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_);
852 TRACE_smpi_recv(src_traced, rank,req->tag_);
854 if(req->detached_sender_ != nullptr){
855 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
857 simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::Host>()->orecv(req->real_size());
859 simcall_process_sleep(sleeptime);
860 XBT_DEBUG("receiving size of %zu : sleep %f ", req->real_size_, sleeptime);
862 unref(&(req->detached_sender_));
864 if (req->flags_ & MPI_REQ_PERSISTENT)
865 req->action_ = nullptr;
866 req->flags_ |= MPI_REQ_FINISHED;
870 int Request::wait(MPI_Request * request, MPI_Status * status)
873 // Are we waiting on a request meant for non blocking collectives ?
874 // If so, wait for all the subrequests.
875 if ((*request)->nbc_requests_size_>0){
876 ret = waitall((*request)->nbc_requests_size_, (*request)->nbc_requests_, MPI_STATUSES_IGNORE);
877 for (int i = 0; i < (*request)->nbc_requests_size_; i++) {
878 if((*request)->buf_!=nullptr && (*request)->nbc_requests_[i]!=MPI_REQUEST_NULL){//reduce case
879 void * buf=(*request)->nbc_requests_[i]->buf_;
880 if((*request)->old_type_->flags() & DT_FLAG_DERIVED)
881 buf=(*request)->nbc_requests_[i]->old_buf_;
882 if((*request)->nbc_requests_[i]->flags_ & MPI_REQ_RECV ){
883 if((*request)->op_!=MPI_OP_NULL){
884 int count=(*request)->size_/ (*request)->old_type_->size();
885 (*request)->op_->apply(buf, (*request)->buf_, &count, (*request)->old_type_);
887 smpi_free_tmp_buffer(static_cast<unsigned char*>(buf));
890 if((*request)->nbc_requests_[i]!=MPI_REQUEST_NULL)
891 Request::unref(&((*request)->nbc_requests_[i]));
893 delete[] (*request)->nbc_requests_;
894 (*request)->nbc_requests_size_=0;
896 (*request)=MPI_REQUEST_NULL;
900 (*request)->print_request("Waiting");
901 if ((*request)->flags_ & MPI_REQ_PREPARED) {
902 Status::empty(status);
906 if ((*request)->action_ != nullptr){
908 // this is not a detached send
909 simcall_comm_wait((*request)->action_, -1.0);
910 } catch (const Exception&) {
911 XBT_VERB("Request cancelled");
915 if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & MPI_REQ_GENERALIZED)){
916 MPI_Status* mystatus;
917 if(!((*request)->flags_ & MPI_REQ_COMPLETE)){
918 ((*request)->generalized_funcs)->mutex->lock();
919 ((*request)->generalized_funcs)->cond->wait(((*request)->generalized_funcs)->mutex);
920 ((*request)->generalized_funcs)->mutex->unlock();
922 if(status==MPI_STATUS_IGNORE){
923 mystatus=new MPI_Status();
924 Status::empty(mystatus);
928 ret = ((*request)->generalized_funcs)->query_fn(((*request)->generalized_funcs)->extra_state, mystatus);
929 if(status==MPI_STATUS_IGNORE)
933 finish_wait(request,status);
934 if (*request != MPI_REQUEST_NULL && (((*request)->flags_ & MPI_REQ_NON_PERSISTENT) != 0))
935 *request = MPI_REQUEST_NULL;
939 int Request::waitany(int count, MPI_Request requests[], MPI_Status * status)
941 std::vector<simgrid::kernel::activity::CommImpl*> comms;
942 comms.reserve(count);
943 int index = MPI_UNDEFINED;
946 // Wait for a request to complete
947 std::vector<int> map;
948 XBT_DEBUG("Wait for one of %d", count);
949 for(int i = 0; i < count; i++) {
950 if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & MPI_REQ_PREPARED) &&
951 not(requests[i]->flags_ & MPI_REQ_FINISHED)) {
952 if (requests[i]->action_ != nullptr) {
953 XBT_DEBUG("Waiting any %p ", requests[i]);
954 comms.push_back(static_cast<simgrid::kernel::activity::CommImpl*>(requests[i]->action_.get()));
957 // This is a finished detached request, let's return this one
958 comms.clear(); // so we free don't do the waitany call
960 finish_wait(&requests[i], status); // cleanup if refcount = 0
961 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
962 requests[i] = MPI_REQUEST_NULL; // set to null
967 if (not comms.empty()) {
968 XBT_DEBUG("Enter waitany for %zu comms", comms.size());
971 // this is not a detached send
972 i = simcall_comm_waitany(comms.data(), comms.size(), -1);
973 } catch (const Exception&) {
974 XBT_INFO("request %d cancelled ", i);
978 // not MPI_UNDEFINED, as this is a simix return code
981 //in case of an accumulate, we have to wait the end of all requests to apply the operation, ordered correctly.
982 if ((requests[index] == MPI_REQUEST_NULL) ||
983 (not((requests[index]->flags_ & MPI_REQ_ACCUMULATE) && (requests[index]->flags_ & MPI_REQ_RECV)))) {
984 finish_wait(&requests[index],status);
985 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_NON_PERSISTENT))
986 requests[index] = MPI_REQUEST_NULL;
992 if (index==MPI_UNDEFINED)
993 Status::empty(status);
998 static int sort_accumulates(MPI_Request a, MPI_Request b)
1000 return (a->tag() > b->tag());
1003 int Request::waitall(int count, MPI_Request requests[], MPI_Status status[])
1005 std::vector<MPI_Request> accumulates;
1008 MPI_Status *pstat = (status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat);
1009 int retvalue = MPI_SUCCESS;
1010 //tag invalid requests in the set
1011 if (status != MPI_STATUSES_IGNORE) {
1012 for (int c = 0; c < count; c++) {
1013 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst_ == MPI_PROC_NULL ||
1014 (requests[c]->flags_ & MPI_REQ_PREPARED)) {
1015 Status::empty(&status[c]);
1016 } else if (requests[c]->src_ == MPI_PROC_NULL) {
1017 Status::empty(&status[c]);
1018 status[c].MPI_SOURCE = MPI_PROC_NULL;
1022 for (int c = 0; c < count; c++) {
1023 if (MC_is_active() || MC_record_replay_is_active()) {
1024 wait(&requests[c],pstat);
1027 index = waitany(count, (MPI_Request*)requests, pstat);
1029 if (index == MPI_UNDEFINED)
1032 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_RECV) &&
1033 (requests[index]->flags_ & MPI_REQ_ACCUMULATE))
1034 accumulates.push_back(requests[index]);
1035 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_NON_PERSISTENT))
1036 requests[index] = MPI_REQUEST_NULL;
1038 if (status != MPI_STATUSES_IGNORE) {
1039 status[index] = *pstat;
1040 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
1041 retvalue = MPI_ERR_IN_STATUS;
1045 if (not accumulates.empty()) {
1046 std::sort(accumulates.begin(), accumulates.end(), sort_accumulates);
1047 for (auto& req : accumulates) {
1048 finish_wait(&req, status);
1055 int Request::waitsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
1061 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
1063 index = waitany(incount, (MPI_Request*)requests, pstat);
1064 if(index==MPI_UNDEFINED) return MPI_UNDEFINED;
1065 if(status != MPI_STATUSES_IGNORE) {
1066 status[count] = *pstat;
1068 indices[count] = index;
1070 for (int i = 0; i < incount; i++) {
1071 if (requests[i] != MPI_REQUEST_NULL) {
1072 test(&requests[i], pstat,&flag);
1075 if(status != MPI_STATUSES_IGNORE) {
1076 status[count] = *pstat;
1078 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
1079 requests[i]=MPI_REQUEST_NULL;
1087 MPI_Request Request::f2c(int id) {
1089 if(id==MPI_FORTRAN_REQUEST_NULL)
1090 return static_cast<MPI_Request>(MPI_REQUEST_NULL);
1091 return static_cast<MPI_Request>(F2C::f2c_lookup()->at(get_key_id(key, id)));
1094 int Request::add_f()
1096 if (F2C::f2c_lookup() == nullptr) {
1097 F2C::set_f2c_lookup(new std::unordered_map<std::string, F2C*>);
1100 (*(F2C::f2c_lookup()))[get_key_id(key, F2C::f2c_id())] = this;
1101 F2C::f2c_id_increment();
1102 return F2C::f2c_id()-1;
1105 void Request::free_f(int id)
1107 if (id != MPI_FORTRAN_REQUEST_NULL) {
1109 F2C::f2c_lookup()->erase(get_key_id(key, id));
1114 int Request::get_status(MPI_Request req, int* flag, MPI_Status * status){
1117 if(req != MPI_REQUEST_NULL && req->action_ != nullptr) {
1118 req->iprobe(req->src_, req->tag_, req->comm_, flag, status);
1122 if (req != MPI_REQUEST_NULL &&
1123 (req->flags_ & MPI_REQ_GENERALIZED)
1124 && !(req->flags_ & MPI_REQ_COMPLETE)) {
1130 if(req != MPI_REQUEST_NULL &&
1131 status != MPI_STATUS_IGNORE) {
1132 int src = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_;
1133 status->MPI_SOURCE = req->comm_->group()->rank(src);
1134 status->MPI_TAG = req->tag_ == MPI_ANY_TAG ? req->real_tag_ : req->tag_;
1135 status->MPI_ERROR = req->truncated_ ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
1136 status->count = req->real_size_;
1141 int Request::grequest_start( MPI_Grequest_query_function *query_fn, MPI_Grequest_free_function *free_fn, MPI_Grequest_cancel_function *cancel_fn, void *extra_state, MPI_Request *request){
1143 *request = new Request();
1144 (*request)->flags_ |= MPI_REQ_GENERALIZED;
1145 (*request)->flags_ |= MPI_REQ_PERSISTENT;
1146 (*request)->refcount_ = 1;
1147 ((*request)->generalized_funcs) = new s_smpi_mpi_generalized_request_funcs_t;
1148 ((*request)->generalized_funcs)->query_fn=query_fn;
1149 ((*request)->generalized_funcs)->free_fn=free_fn;
1150 ((*request)->generalized_funcs)->cancel_fn=cancel_fn;
1151 ((*request)->generalized_funcs)->extra_state=extra_state;
1152 ((*request)->generalized_funcs)->cond = simgrid::s4u::ConditionVariable::create();
1153 ((*request)->generalized_funcs)->mutex = simgrid::s4u::Mutex::create();
1157 int Request::grequest_complete( MPI_Request request){
1158 if ((!(request->flags_ & MPI_REQ_GENERALIZED)) || request->generalized_funcs->mutex==NULL)
1159 return MPI_ERR_REQUEST;
1160 request->generalized_funcs->mutex->lock();
1161 request->flags_ |= MPI_REQ_COMPLETE; // in case wait would be called after complete
1162 request->generalized_funcs->cond->notify_one();
1163 request->generalized_funcs->mutex->unlock();
1167 void Request::set_nbc_requests(MPI_Request* reqs, int size){
1168 nbc_requests_size_ = size;
1170 nbc_requests_ = reqs;
1173 nbc_requests_ = nullptr;
1177 int Request::get_nbc_requests_size(){
1178 return nbc_requests_size_;
1181 MPI_Request* Request::get_nbc_requests(){
1182 return nbc_requests_;