1 /* Copyright (c) 2007-2017. 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"
8 #include "SmpiHost.hpp"
10 #include "private.hpp"
11 #include "smpi_comm.hpp"
12 #include "smpi_datatype.hpp"
13 #include "smpi_op.hpp"
14 #include "smpi_process.hpp"
15 #include "src/kernel/activity/CommImpl.hpp"
16 #include "src/mc/mc_replay.hpp"
17 #include "src/simix/ActorImpl.hpp"
21 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_request, smpi, "Logging specific to SMPI (reques)");
23 static simgrid::config::Flag<double> smpi_iprobe_sleep(
24 "smpi/iprobe", "Minimum time to inject inside a call to MPI_Iprobe", 1e-4);
25 static simgrid::config::Flag<double> smpi_test_sleep(
26 "smpi/test", "Minimum time to inject inside a call to MPI_Test", 1e-4);
28 std::vector<s_smpi_factor_t> smpi_ois_values;
30 extern void (*smpi_comm_copy_data_callback) (smx_activity_t, void*, size_t);
35 Request::Request(void* buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm, unsigned flags)
36 : buf_(buf), old_type_(datatype), src_(src), dst_(dst), tag_(tag), comm_(comm), flags_(flags)
38 void *old_buf = nullptr;
39 // FIXME Handle the case of a partial shared malloc.
40 if ((((flags & RECV) != 0) && ((flags & ACCUMULATE) != 0)) || (datatype->flags() & DT_FLAG_DERIVED)) {
41 // This part handles the problem of non-contiguous memory
46 buf_ = xbt_malloc(count*datatype->size());
47 if ((datatype->flags() & DT_FLAG_DERIVED) && ((flags & SEND) != 0)) {
48 datatype->serialize(old_buf, buf_, count);
52 // This part handles the problem of non-contiguous memory (for the unserialisation at the reception)
54 size_ = datatype->size() * count;
59 detached_sender_ = nullptr;
64 if (flags & PERSISTENT)
71 MPI_Comm Request::comm(){
91 int Request::detached(){
95 size_t Request::size(){
99 size_t Request::real_size(){
103 void Request::unref(MPI_Request* request)
105 if((*request) != MPI_REQUEST_NULL){
106 (*request)->refcount_--;
107 if((*request)->refcount_<0) xbt_die("wrong refcount");
108 if((*request)->refcount_==0){
109 Datatype::unref((*request)->old_type_);
110 Comm::unref((*request)->comm_);
111 (*request)->print_request("Destroying");
113 *request = MPI_REQUEST_NULL;
115 (*request)->print_request("Decrementing");
118 xbt_die("freeing an already free request");
122 int Request::match_recv(void* a, void* b, simgrid::kernel::activity::CommImpl* ignored)
124 MPI_Request ref = static_cast<MPI_Request>(a);
125 MPI_Request req = static_cast<MPI_Request>(b);
126 XBT_DEBUG("Trying to match a recv of src %d against %d, tag %d against %d",ref->src_,req->src_, ref->tag_, req->tag_);
128 xbt_assert(ref, "Cannot match recv against null reference");
129 xbt_assert(req, "Cannot match recv against null request");
130 if((ref->src_ == MPI_ANY_SOURCE || req->src_ == ref->src_)
131 && ((ref->tag_ == MPI_ANY_TAG && req->tag_ >=0) || req->tag_ == ref->tag_)){
132 //we match, we can transfer some values
133 if(ref->src_ == MPI_ANY_SOURCE)
134 ref->real_src_ = req->src_;
135 if(ref->tag_ == MPI_ANY_TAG)
136 ref->real_tag_ = req->tag_;
137 if(ref->real_size_ < req->real_size_)
139 if(req->detached_==1)
140 ref->detached_sender_=req; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
141 XBT_DEBUG("match succeeded");
146 int Request::match_send(void* a, void* b, simgrid::kernel::activity::CommImpl* ignored)
148 MPI_Request ref = static_cast<MPI_Request>(a);
149 MPI_Request req = static_cast<MPI_Request>(b);
150 XBT_DEBUG("Trying to match a send of src %d against %d, tag %d against %d",ref->src_,req->src_, ref->tag_, req->tag_);
151 xbt_assert(ref, "Cannot match send against null reference");
152 xbt_assert(req, "Cannot match send against null request");
154 if((req->src_ == MPI_ANY_SOURCE || req->src_ == ref->src_)
155 && ((req->tag_ == MPI_ANY_TAG && ref->tag_ >=0)|| req->tag_ == ref->tag_)){
156 if(req->src_ == MPI_ANY_SOURCE)
157 req->real_src_ = ref->src_;
158 if(req->tag_ == MPI_ANY_TAG)
159 req->real_tag_ = ref->tag_;
160 if(req->real_size_ < ref->real_size_)
162 if(ref->detached_==1)
163 req->detached_sender_=ref; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
164 XBT_DEBUG("match succeeded");
170 void Request::print_request(const char *message)
172 XBT_VERB("%s request %p [buf = %p, size = %zu, src = %d, dst = %d, tag = %d, flags = %x]",
173 message, this, buf_, size_, src_, dst_, tag_, flags_);
177 /* factories, to hide the internal flags from the caller */
178 MPI_Request Request::send_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
181 return new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
182 comm->group()->index(dst), tag, comm, PERSISTENT | SEND | PREPARED);
185 MPI_Request Request::ssend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
187 return new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
188 comm->group()->index(dst), tag, comm, PERSISTENT | SSEND | SEND | PREPARED);
191 MPI_Request Request::isend_init(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, smpi_process()->index(),
194 comm->group()->index(dst), tag,comm, PERSISTENT | ISEND | SEND | PREPARED);
198 MPI_Request Request::rma_send_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
201 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
203 request = new Request(buf==MPI_BOTTOM ? nullptr : buf , count, datatype, src, dst, tag,
204 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED);
206 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src, dst, tag,
207 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED | ACCUMULATE);
213 MPI_Request Request::recv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
215 return new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype,
216 src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->index(src),
217 smpi_process()->index(), tag, comm, PERSISTENT | RECV | PREPARED);
220 MPI_Request Request::rma_recv_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
223 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
225 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, src, dst, tag, comm,
226 RMA | NON_PERSISTENT | RECV | PREPARED);
228 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, src, dst, tag, comm,
229 RMA | NON_PERSISTENT | RECV | PREPARED | ACCUMULATE);
235 MPI_Request Request::irecv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
237 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
238 src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->index(src), smpi_process()->index(), tag,
239 comm, PERSISTENT | RECV | PREPARED);
242 MPI_Request Request::isend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
244 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
245 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
246 comm->group()->index(dst), tag, comm, NON_PERSISTENT | ISEND | SEND);
251 MPI_Request Request::issend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
253 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
254 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
255 comm->group()->index(dst), tag, comm, NON_PERSISTENT | ISEND | SSEND | SEND);
261 MPI_Request Request::irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
263 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
264 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
265 src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->index(src), smpi_process()->index(),
266 tag, comm, NON_PERSISTENT | RECV);
271 void Request::recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status * status)
273 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
274 request = irecv(buf, count, datatype, src, tag, comm);
275 wait(&request,status);
279 void Request::send(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
281 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
282 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
283 comm->group()->index(dst), tag, comm, NON_PERSISTENT | SEND);
286 wait(&request, MPI_STATUS_IGNORE);
290 void Request::ssend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
292 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
293 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
294 comm->group()->index(dst), tag, comm, NON_PERSISTENT | SSEND | SEND);
297 wait(&request,MPI_STATUS_IGNORE);
301 void Request::sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,int dst, int sendtag,
302 void *recvbuf, int recvcount, MPI_Datatype recvtype, int src, int recvtag,
303 MPI_Comm comm, MPI_Status * status)
305 MPI_Request requests[2];
307 int myid=smpi_process()->index();
308 if ((comm->group()->index(dst) == myid) && (comm->group()->index(src) == myid)){
309 Datatype::copy(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype);
312 requests[0] = isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
313 requests[1] = irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
314 startall(2, requests);
315 waitall(2, requests, stats);
318 if(status != MPI_STATUS_IGNORE) {
319 // Copy receive status
324 void Request::start()
326 smx_mailbox_t mailbox;
328 xbt_assert(action_ == nullptr, "Cannot (re-)start unfinished communication");
333 if ((flags_ & RECV) != 0) {
334 this->print_request("New recv");
336 simgrid::smpi::Process* process = smpi_process_remote(dst_);
338 int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh");
340 xbt_mutex_t mut = process->mailboxes_mutex();
341 if (async_small_thresh != 0 || (flags_ & RMA) != 0)
342 xbt_mutex_acquire(mut);
344 if (async_small_thresh == 0 && (flags_ & RMA) == 0 ) {
345 mailbox = process->mailbox();
347 else if (((flags_ & RMA) != 0) || static_cast<int>(size_) < async_small_thresh) {
348 //We have to check both mailboxes (because SSEND messages are sent to the large mbox).
349 //begin with the more appropriate one : the small one.
350 mailbox = process->mailbox_small();
351 XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %p (in case of SSEND)?", mailbox);
352 smx_activity_t action = simcall_comm_iprobe(mailbox, 0, &match_recv, static_cast<void*>(this));
354 if (action == nullptr) {
355 mailbox = process->mailbox();
356 XBT_DEBUG("No, nothing in the small mailbox test the other one : %p", mailbox);
357 action = simcall_comm_iprobe(mailbox, 0, &match_recv, static_cast<void*>(this));
358 if (action == nullptr) {
359 XBT_DEBUG("Still nothing, switch back to the small mailbox : %p", mailbox);
360 mailbox = process->mailbox_small();
363 XBT_DEBUG("yes there was something for us in the large mailbox");
366 mailbox = process->mailbox_small();
367 XBT_DEBUG("Is there a corresponding send already posted the small mailbox?");
368 smx_activity_t action = simcall_comm_iprobe(mailbox, 0, &match_recv, static_cast<void*>(this));
370 if (action == nullptr) {
371 XBT_DEBUG("No, nothing in the permanent receive mailbox");
372 mailbox = process->mailbox();
374 XBT_DEBUG("yes there was something for us in the small mailbox");
378 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
380 action_ = simcall_comm_irecv(
381 process->process(), mailbox, buf_, &real_size_, &match_recv,
382 process->replaying() ? &smpi_comm_null_copy_buffer_callback : smpi_comm_copy_data_callback, this, -1.0);
383 XBT_DEBUG("recv simcall posted");
385 if (async_small_thresh != 0 || (flags_ & RMA) != 0 )
386 xbt_mutex_release(mut);
387 } else { /* the RECV flag was not set, so this is a send */
388 simgrid::smpi::Process* process = smpi_process_remote(dst_);
390 if (TRACE_smpi_view_internals()) {
391 TRACE_smpi_send(rank, rank, dst_, tag_, size_);
393 this->print_request("New send");
396 if ((flags_ & SSEND) == 0 && ( (flags_ & RMA) != 0
397 || static_cast<int>(size_) < xbt_cfg_get_int("smpi/send-is-detached-thresh") ) ) {
398 void *oldbuf = nullptr;
400 XBT_DEBUG("Send request %p is detached", this);
402 if (not(old_type_->flags() & DT_FLAG_DERIVED)) {
404 if (not process->replaying() && oldbuf != nullptr && size_ != 0) {
405 if ((smpi_privatize_global_variables != 0) && (static_cast<char*>(buf_) >= smpi_data_exe_start) &&
406 (static_cast<char*>(buf_) < smpi_data_exe_start + smpi_data_exe_size)) {
407 XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
408 smpi_switch_data_segment(src_);
410 buf = xbt_malloc(size_);
411 memcpy(buf,oldbuf,size_);
412 XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
417 //if we are giving back the control to the user without waiting for completion, we have to inject timings
418 double sleeptime = 0.0;
419 if (detached_ != 0 || ((flags_ & (ISEND | SSEND)) != 0)) { // issend should be treated as isend
420 // isend and send timings may be different
421 sleeptime = ((flags_ & ISEND) != 0)
422 ? simgrid::s4u::Actor::self()->getHost()->extension<simgrid::smpi::SmpiHost>()->oisend(size_)
423 : simgrid::s4u::Actor::self()->getHost()->extension<simgrid::smpi::SmpiHost>()->osend(size_);
427 simcall_process_sleep(sleeptime);
428 XBT_DEBUG("sending size of %zu : sleep %f ", size_, sleeptime);
431 int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh");
433 xbt_mutex_t mut=process->mailboxes_mutex();
435 if (async_small_thresh != 0 || (flags_ & RMA) != 0)
436 xbt_mutex_acquire(mut);
438 if (not(async_small_thresh != 0 || (flags_ & RMA) != 0)) {
439 mailbox = process->mailbox();
440 } else if (((flags_ & RMA) != 0) || static_cast<int>(size_) < async_small_thresh) { // eager mode
441 mailbox = process->mailbox();
442 XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %p?", mailbox);
443 smx_activity_t action = simcall_comm_iprobe(mailbox, 1, &match_send, static_cast<void*>(this));
444 if (action == nullptr) {
445 if ((flags_ & SSEND) == 0){
446 mailbox = process->mailbox_small();
447 XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %p", mailbox);
449 mailbox = process->mailbox_small();
450 XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %p?", mailbox);
451 action = simcall_comm_iprobe(mailbox, 1, &match_send, static_cast<void*>(this));
452 if (action == nullptr) {
453 XBT_DEBUG("No, we are first, send to large mailbox");
454 mailbox = process->mailbox();
458 XBT_DEBUG("Yes there was something for us in the large mailbox");
461 mailbox = process->mailbox();
462 XBT_DEBUG("Send request %p is in the large mailbox %p (buf: %p)",mailbox, this,buf_);
465 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
467 action_ = simcall_comm_isend(
468 SIMIX_process_from_PID(src_ + 1), mailbox, size_, -1.0, buf, real_size_, &match_send,
469 &xbt_free_f, // how to free the userdata if a detached send fails
470 not process->replaying() ? smpi_comm_copy_data_callback : &smpi_comm_null_copy_buffer_callback, this,
471 // detach if msg size < eager/rdv switch limit
473 XBT_DEBUG("send simcall posted");
475 /* FIXME: detached sends are not traceable (action_ == nullptr) */
476 if (action_ != nullptr)
477 simcall_set_category(action_, TRACE_internal_smpi_get_category());
478 if (async_small_thresh != 0 || ((flags_ & RMA)!=0))
479 xbt_mutex_release(mut);
483 void Request::startall(int count, MPI_Request * requests)
485 if(requests== nullptr)
488 for(int i = 0; i < count; i++) {
489 requests[i]->start();
493 int Request::test(MPI_Request * request, MPI_Status * status) {
494 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or testall before)
495 // to avoid deadlocks if used as a break condition, such as
496 // while (MPI_Test(request, flag, status) && flag) dostuff...
497 // because the time will not normally advance when only calls to MPI_Test are made -> deadlock
498 // multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
499 static int nsleeps = 1;
500 if(smpi_test_sleep > 0)
501 simcall_process_sleep(nsleeps*smpi_test_sleep);
503 Status::empty(status);
505 if (((*request)->flags_ & PREPARED) == 0) {
506 if ((*request)->action_ != nullptr)
507 flag = simcall_comm_test((*request)->action_);
509 finish_wait(request,status);
510 nsleeps=1;//reset the number of sleeps we will do next time
511 if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & PERSISTENT) == 0)
512 *request = MPI_REQUEST_NULL;
513 } else if (xbt_cfg_get_boolean("smpi/grow-injected-times")){
520 int Request::testsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
525 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
527 for (int i = 0; i < incount; i++) {
528 if (requests[i] != MPI_REQUEST_NULL) {
529 if (test(&requests[i], pstat)) {
532 if (status != MPI_STATUSES_IGNORE)
534 if ((requests[i] != MPI_REQUEST_NULL) && (requests[i]->flags_ & NON_PERSISTENT))
535 requests[i] = MPI_REQUEST_NULL;
541 if(count_dead==incount)
542 return MPI_UNDEFINED;
546 int Request::testany(int count, MPI_Request requests[], int *index, MPI_Status * status)
548 std::vector<simgrid::kernel::activity::ActivityImplPtr> comms;
549 comms.reserve(count);
554 *index = MPI_UNDEFINED;
556 std::vector<int> map; /** Maps all matching comms back to their location in requests **/
557 for(i = 0; i < count; i++) {
558 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action_ && not(requests[i]->flags_ & PREPARED)) {
559 comms.push_back(requests[i]->action_);
563 if (not map.empty()) {
564 //multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it
565 static int nsleeps = 1;
566 if(smpi_test_sleep > 0)
567 simcall_process_sleep(nsleeps*smpi_test_sleep);
569 i = simcall_comm_testany(comms.data(), comms.size()); // The i-th element in comms matches!
570 if (i != -1) { // -1 is not MPI_UNDEFINED but a SIMIX return code. (nothing matches)
572 finish_wait(&requests[*index],status);
575 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & NON_PERSISTENT)) {
576 requests[*index] = MPI_REQUEST_NULL;
582 //all requests are null or inactive, return true
584 Status::empty(status);
590 int Request::testall(int count, MPI_Request requests[], MPI_Status status[])
593 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
595 for(int i=0; i<count; i++){
596 if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & PREPARED)) {
597 if (test(&requests[i], pstat)!=1){
600 requests[i]=MPI_REQUEST_NULL;
603 Status::empty(pstat);
605 if(status != MPI_STATUSES_IGNORE) {
612 void Request::probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
614 //FIXME find another way to avoid busy waiting ?
615 // the issue here is that we have to wait on a nonexistent comm
617 iprobe(source, tag, comm, &flag, status);
618 XBT_DEBUG("Busy Waiting on probing : %d", flag);
622 void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
623 // to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
624 // especially when used as a break condition, such as while (MPI_Iprobe(...)) dostuff...
625 // nsleeps is a multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
626 // This can speed up the execution of certain applications by an order of magnitude, such as HPL
627 static int nsleeps = 1;
628 double speed = simgrid::s4u::Actor::self()->getHost()->getSpeed();
629 double maxrate = xbt_cfg_get_double("smpi/iprobe-cpu-usage");
630 MPI_Request request = new Request(nullptr, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
631 comm->group()->index(source), comm->rank(), tag, comm, PERSISTENT | RECV);
632 if (smpi_iprobe_sleep > 0) {
633 smx_activity_t iprobe_sleep = simcall_execution_start(
634 "iprobe", /* flops to executek*/ nsleeps * smpi_iprobe_sleep * speed * maxrate, /* priority */ 1.0,
635 /* performance bound */ maxrate * speed, smpi_process()->process()->host);
636 simcall_execution_wait(iprobe_sleep);
638 // behave like a receive, but don't do it
639 smx_mailbox_t mailbox;
641 request->print_request("New iprobe");
642 // We have to test both mailboxes as we don't know if we will receive one one or another
643 if (xbt_cfg_get_int("smpi/async-small-thresh") > 0){
644 mailbox = smpi_process()->mailbox_small();
645 XBT_DEBUG("Trying to probe the perm recv mailbox");
646 request->action_ = simcall_comm_iprobe(mailbox, 0, &match_recv, static_cast<void*>(request));
649 if (request->action_ == nullptr){
650 mailbox = smpi_process()->mailbox();
651 XBT_DEBUG("trying to probe the other mailbox");
652 request->action_ = simcall_comm_iprobe(mailbox, 0, &match_recv, static_cast<void*>(request));
655 if (request->action_ != nullptr){
656 simgrid::kernel::activity::CommImplPtr sync_comm =
657 boost::static_pointer_cast<simgrid::kernel::activity::CommImpl>(request->action_);
658 MPI_Request req = static_cast<MPI_Request>(sync_comm->src_data);
660 if(status != MPI_STATUS_IGNORE && (req->flags_ & PREPARED) == 0) {
661 status->MPI_SOURCE = comm->group()->rank(req->src_);
662 status->MPI_TAG = req->tag_;
663 status->MPI_ERROR = MPI_SUCCESS;
664 status->count = req->real_size_;
666 nsleeps = 1;//reset the number of sleeps we will do next time
670 if (xbt_cfg_get_boolean("smpi/grow-injected-times"))
676 void Request::finish_wait(MPI_Request* request, MPI_Status * status)
678 MPI_Request req = *request;
679 Status::empty(status);
681 if (not((req->detached_ != 0) && ((req->flags_ & SEND) != 0)) && ((req->flags_ & PREPARED) == 0)) {
682 if(status != MPI_STATUS_IGNORE) {
683 int src = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_;
684 status->MPI_SOURCE = req->comm_->group()->rank(src);
685 status->MPI_TAG = req->tag_ == MPI_ANY_TAG ? req->real_tag_ : req->tag_;
686 status->MPI_ERROR = req->truncated_ != 0 ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
687 // this handles the case were size in receive differs from size in send
688 status->count = req->real_size_;
691 req->print_request("Finishing");
692 MPI_Datatype datatype = req->old_type_;
694 // FIXME Handle the case of a partial shared malloc.
695 if (((req->flags_ & ACCUMULATE) != 0) ||
696 (datatype->flags() & DT_FLAG_DERIVED)) { // && (not smpi_is_shared(req->old_buf_))){
698 if (not smpi_process()->replaying() && smpi_privatize_global_variables != 0 &&
699 static_cast<char*>(req->old_buf_) >= smpi_data_exe_start &&
700 static_cast<char*>(req->old_buf_) < smpi_data_exe_start + smpi_data_exe_size) {
701 XBT_VERB("Privatization : We are unserializing to a zone in global memory Switch data segment ");
702 smpi_switch_data_segment(smpi_process()->index());
705 if(datatype->flags() & DT_FLAG_DERIVED){
706 // This part handles the problem of non-contignous memory the unserialization at the reception
707 if((req->flags_ & RECV) && datatype->size()!=0)
708 datatype->unserialize(req->buf_, req->old_buf_, req->real_size_/datatype->size() , req->op_);
710 }else if(req->flags_ & RECV){//apply op on contiguous buffer for accumulate
711 if(datatype->size()!=0){
712 int n =req->real_size_/datatype->size();
713 req->op_->apply(req->buf_, req->old_buf_, &n, datatype);
720 if (TRACE_smpi_view_internals() && ((req->flags_ & RECV) != 0)){
721 int rank = smpi_process()->index();
722 int src_traced = (req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_);
723 TRACE_smpi_recv(src_traced, rank,req->tag_);
725 if(req->detached_sender_ != nullptr){
726 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
728 simgrid::s4u::Actor::self()->getHost()->extension<simgrid::smpi::SmpiHost>()->orecv(req->real_size());
730 simcall_process_sleep(sleeptime);
731 XBT_DEBUG("receiving size of %zu : sleep %f ", req->real_size_, sleeptime);
733 unref(&(req->detached_sender_));
735 if(req->flags_ & PERSISTENT)
736 req->action_ = nullptr;
737 req->flags_ |= FINISHED;
741 void Request::wait(MPI_Request * request, MPI_Status * status)
743 (*request)->print_request("Waiting");
744 if ((*request)->flags_ & PREPARED) {
745 Status::empty(status);
749 if ((*request)->action_ != nullptr)
750 // this is not a detached send
751 simcall_comm_wait((*request)->action_, -1.0);
753 finish_wait(request,status);
754 if (*request != MPI_REQUEST_NULL && (((*request)->flags_ & NON_PERSISTENT)!=0))
755 *request = MPI_REQUEST_NULL;
758 int Request::waitany(int count, MPI_Request requests[], MPI_Status * status)
760 s_xbt_dynar_t comms; // Keep it on stack to save some extra mallocs
761 int index = MPI_UNDEFINED;
765 // Wait for a request to complete
766 xbt_dynar_init(&comms, sizeof(smx_activity_t), [](void*ptr){
767 intrusive_ptr_release(*(simgrid::kernel::activity::ActivityImpl**)ptr);
769 int *map = xbt_new(int, count);
770 XBT_DEBUG("Wait for one of %d", count);
771 for(int i = 0; i < count; i++) {
772 if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & PREPARED) &&
773 not(requests[i]->flags_ & FINISHED)) {
774 if (requests[i]->action_ != nullptr) {
775 XBT_DEBUG("Waiting any %p ", requests[i]);
776 intrusive_ptr_add_ref(requests[i]->action_.get());
777 xbt_dynar_push_as(&comms, simgrid::kernel::activity::ActivityImpl*, requests[i]->action_.get());
781 // This is a finished detached request, let's return this one
782 size = 0; // so we free the dynar but don't do the waitany call
784 finish_wait(&requests[i], status); // cleanup if refcount = 0
785 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & NON_PERSISTENT))
786 requests[i] = MPI_REQUEST_NULL; // set to null
792 XBT_DEBUG("Enter waitany for %lu comms", xbt_dynar_length(&comms));
793 int i = simcall_comm_waitany(&comms, -1);
795 // not MPI_UNDEFINED, as this is a simix return code
798 //in case of an accumulate, we have to wait the end of all requests to apply the operation, ordered correctly.
799 if ((requests[index] == MPI_REQUEST_NULL) ||
800 (not((requests[index]->flags_ & ACCUMULATE) && (requests[index]->flags_ & RECV)))) {
801 finish_wait(&requests[index],status);
802 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & NON_PERSISTENT))
803 requests[index] = MPI_REQUEST_NULL;
808 xbt_dynar_free_data(&comms);
812 if (index==MPI_UNDEFINED)
813 Status::empty(status);
818 static int sort_accumulates(MPI_Request a, MPI_Request b)
820 return (a->tag() > b->tag());
823 int Request::waitall(int count, MPI_Request requests[], MPI_Status status[])
825 std::vector<MPI_Request> accumulates;
828 MPI_Status *pstat = (status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat);
829 int retvalue = MPI_SUCCESS;
830 //tag invalid requests in the set
831 if (status != MPI_STATUSES_IGNORE) {
832 for (int c = 0; c < count; c++) {
833 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst_ == MPI_PROC_NULL || (requests[c]->flags_ & PREPARED)) {
834 Status::empty(&status[c]);
835 } else if (requests[c]->src_ == MPI_PROC_NULL) {
836 Status::empty(&status[c]);
837 status[c].MPI_SOURCE = MPI_PROC_NULL;
841 for (int c = 0; c < count; c++) {
842 if (MC_is_active() || MC_record_replay_is_active()) {
843 wait(&requests[c],pstat);
846 index = waitany(count, (MPI_Request*)requests, pstat);
847 if (index == MPI_UNDEFINED)
850 if (requests[index] != MPI_REQUEST_NULL
851 && (requests[index]->flags_ & RECV)
852 && (requests[index]->flags_ & ACCUMULATE))
853 accumulates.push_back(requests[index]);
854 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & NON_PERSISTENT))
855 requests[index] = MPI_REQUEST_NULL;
857 if (status != MPI_STATUSES_IGNORE) {
858 status[index] = *pstat;
859 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
860 retvalue = MPI_ERR_IN_STATUS;
864 if (not accumulates.empty()) {
865 std::sort(accumulates.begin(), accumulates.end(), sort_accumulates);
866 for (auto& req : accumulates) {
867 finish_wait(&req, status);
874 int Request::waitsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
878 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
880 for (int i = 0; i < incount; i++) {
881 int index = waitany(incount, requests, pstat);
882 if(index!=MPI_UNDEFINED){
883 indices[count] = index;
885 if(status != MPI_STATUSES_IGNORE) {
886 status[index] = *pstat;
888 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & NON_PERSISTENT))
889 requests[index] = MPI_REQUEST_NULL;
891 return MPI_UNDEFINED;
897 MPI_Request Request::f2c(int id) {
899 if(id==MPI_FORTRAN_REQUEST_NULL)
900 return static_cast<MPI_Request>(MPI_REQUEST_NULL);
901 return static_cast<MPI_Request>(F2C::f2c_lookup()->at(get_key_id(key, id)));
906 if (F2C::f2c_lookup() == nullptr) {
907 F2C::set_f2c_lookup(new std::unordered_map<std::string, F2C*>);
910 (*(F2C::f2c_lookup()))[get_key_id(key, F2C::f2c_id())] = this;
911 F2C::f2c_id_increment();
912 return F2C::f2c_id()-1;
915 void Request::free_f(int id)
917 if (id != MPI_FORTRAN_REQUEST_NULL) {
919 F2C::f2c_lookup()->erase(get_key_id(key, id));