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"
9 #include "src/kernel/activity/CommImpl.hpp"
10 #include "src/mc/mc_replay.h"
11 #include "SmpiHost.hpp"
13 #include "private.hpp"
14 #include "smpi_comm.hpp"
15 #include "smpi_datatype.hpp"
16 #include "smpi_op.hpp"
17 #include "smpi_process.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) : buf_(buf), old_type_(datatype), src_(src), dst_(dst), tag_(tag), comm_(comm), flags_(flags)
37 void *old_buf = nullptr;
38 // FIXME Handle the case of a partial shared malloc.
39 if ((((flags & RECV) != 0) && ((flags & ACCUMULATE) != 0)) || (datatype->flags() & DT_FLAG_DERIVED)) {
40 // This part handles the problem of non-contiguous memory
45 buf_ = xbt_malloc(count*datatype->size());
46 if ((datatype->flags() & DT_FLAG_DERIVED) && ((flags & SEND) != 0)) {
47 datatype->serialize(old_buf, buf_, count);
51 // This part handles the problem of non-contiguous memory (for the unserialisation at the reception)
53 size_ = datatype->size() * count;
58 detached_sender_ = nullptr;
63 if (flags & PERSISTENT)
70 MPI_Comm Request::comm(){
90 int Request::detached(){
94 size_t Request::size(){
98 size_t Request::real_size(){
102 void Request::unref(MPI_Request* request)
104 if((*request) != MPI_REQUEST_NULL){
105 (*request)->refcount_--;
106 if((*request)->refcount_<0) xbt_die("wrong refcount");
107 if((*request)->refcount_==0){
108 Datatype::unref((*request)->old_type_);
109 Comm::unref((*request)->comm_);
110 (*request)->print_request("Destroying");
112 *request = MPI_REQUEST_NULL;
114 (*request)->print_request("Decrementing");
117 xbt_die("freeing an already free request");
121 int Request::match_recv(void* a, void* b, simgrid::kernel::activity::CommImpl* ignored)
123 MPI_Request ref = static_cast<MPI_Request>(a);
124 MPI_Request req = static_cast<MPI_Request>(b);
125 XBT_DEBUG("Trying to match a recv of src %d against %d, tag %d against %d",ref->src_,req->src_, ref->tag_, req->tag_);
127 xbt_assert(ref, "Cannot match recv against null reference");
128 xbt_assert(req, "Cannot match recv against null request");
129 if((ref->src_ == MPI_ANY_SOURCE || req->src_ == ref->src_)
130 && ((ref->tag_ == MPI_ANY_TAG && req->tag_ >=0) || req->tag_ == ref->tag_)){
131 //we match, we can transfer some values
132 if(ref->src_ == MPI_ANY_SOURCE)
133 ref->real_src_ = req->src_;
134 if(ref->tag_ == MPI_ANY_TAG)
135 ref->real_tag_ = req->tag_;
136 if(ref->real_size_ < req->real_size_)
138 if(req->detached_==1)
139 ref->detached_sender_=req; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
140 XBT_DEBUG("match succeeded");
145 int Request::match_send(void* a, void* b, simgrid::kernel::activity::CommImpl* ignored)
147 MPI_Request ref = static_cast<MPI_Request>(a);
148 MPI_Request req = static_cast<MPI_Request>(b);
149 XBT_DEBUG("Trying to match a send of src %d against %d, tag %d against %d",ref->src_,req->src_, ref->tag_, req->tag_);
150 xbt_assert(ref, "Cannot match send against null reference");
151 xbt_assert(req, "Cannot match send against null request");
153 if((req->src_ == MPI_ANY_SOURCE || req->src_ == ref->src_)
154 && ((req->tag_ == MPI_ANY_TAG && ref->tag_ >=0)|| req->tag_ == ref->tag_)){
155 if(req->src_ == MPI_ANY_SOURCE)
156 req->real_src_ = ref->src_;
157 if(req->tag_ == MPI_ANY_TAG)
158 req->real_tag_ = ref->tag_;
159 if(req->real_size_ < ref->real_size_)
161 if(ref->detached_==1)
162 req->detached_sender_=ref; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
163 XBT_DEBUG("match succeeded");
169 void Request::print_request(const char *message)
171 XBT_VERB("%s request %p [buf = %p, size = %zu, src = %d, dst = %d, tag = %d, flags = %x]",
172 message, this, buf_, size_, src_, dst_, tag_, flags_);
176 /* factories, to hide the internal flags from the caller */
177 MPI_Request Request::send_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
180 return new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
181 comm->group()->index(dst), tag, comm, PERSISTENT | SEND | PREPARED);
184 MPI_Request Request::ssend_init(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, smpi_process()->index(),
187 comm->group()->index(dst), tag, comm, PERSISTENT | SSEND | SEND | PREPARED);
190 MPI_Request Request::isend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
192 return new Request(buf==MPI_BOTTOM ? nullptr : buf , count, datatype, smpi_process()->index(),
193 comm->group()->index(dst), tag,comm, PERSISTENT | ISEND | SEND | PREPARED);
197 MPI_Request Request::rma_send_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
200 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
202 request = new Request(buf==MPI_BOTTOM ? nullptr : buf , count, datatype, src, dst, tag,
203 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED);
205 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src, dst, tag,
206 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED | ACCUMULATE);
212 MPI_Request Request::recv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
214 return new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype,
215 src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->index(src),
216 smpi_process()->index(), tag, comm, PERSISTENT | RECV | PREPARED);
219 MPI_Request Request::rma_recv_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, src, dst, tag, comm,
225 RMA | NON_PERSISTENT | RECV | PREPARED);
227 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, src, dst, tag, comm,
228 RMA | NON_PERSISTENT | RECV | PREPARED | ACCUMULATE);
234 MPI_Request Request::irecv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
236 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
237 src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->index(src), smpi_process()->index(), tag,
238 comm, PERSISTENT | RECV | PREPARED);
241 MPI_Request Request::isend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
243 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
244 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
245 comm->group()->index(dst), tag, comm, NON_PERSISTENT | ISEND | SEND);
250 MPI_Request Request::issend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
252 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
253 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
254 comm->group()->index(dst), tag, comm, NON_PERSISTENT | ISEND | SSEND | SEND);
260 MPI_Request Request::irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
262 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
263 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
264 src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->index(src), smpi_process()->index(),
265 tag, comm, NON_PERSISTENT | RECV);
270 void Request::recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status * status)
272 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
273 request = irecv(buf, count, datatype, src, tag, comm);
274 wait(&request,status);
278 void Request::send(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
280 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
281 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
282 comm->group()->index(dst), tag, comm, NON_PERSISTENT | SEND);
285 wait(&request, MPI_STATUS_IGNORE);
289 void Request::ssend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
291 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
292 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
293 comm->group()->index(dst), tag, comm, NON_PERSISTENT | SSEND | SEND);
296 wait(&request,MPI_STATUS_IGNORE);
300 void Request::sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,int dst, int sendtag,
301 void *recvbuf, int recvcount, MPI_Datatype recvtype, int src, int recvtag,
302 MPI_Comm comm, MPI_Status * status)
304 MPI_Request requests[2];
306 int myid=smpi_process()->index();
307 if ((comm->group()->index(dst) == myid) && (comm->group()->index(src) == myid)){
308 Datatype::copy(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype);
311 requests[0] = isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
312 requests[1] = irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
313 startall(2, requests);
314 waitall(2, requests, stats);
317 if(status != MPI_STATUS_IGNORE) {
318 // Copy receive status
323 void Request::start()
325 smx_mailbox_t mailbox;
327 xbt_assert(action_ == nullptr, "Cannot (re-)start unfinished communication");
332 if ((flags_ & RECV) != 0) {
333 this->print_request("New recv");
335 simgrid::smpi::Process* process = smpi_process_remote(dst_);
337 int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh");
339 xbt_mutex_t mut = process->mailboxes_mutex();
340 if (async_small_thresh != 0 || (flags_ & RMA) != 0)
341 xbt_mutex_acquire(mut);
343 if (async_small_thresh == 0 && (flags_ & RMA) == 0 ) {
344 mailbox = process->mailbox();
346 else if (((flags_ & RMA) != 0) || static_cast<int>(size_) < async_small_thresh) {
347 //We have to check both mailboxes (because SSEND messages are sent to the large mbox).
348 //begin with the more appropriate one : the small one.
349 mailbox = process->mailbox_small();
350 XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %p (in case of SSEND)?", mailbox);
351 smx_activity_t action = simcall_comm_iprobe(mailbox, 0, &match_recv, static_cast<void*>(this));
353 if (action == nullptr) {
354 mailbox = process->mailbox();
355 XBT_DEBUG("No, nothing in the small mailbox test the other one : %p", mailbox);
356 action = simcall_comm_iprobe(mailbox, 0, &match_recv, static_cast<void*>(this));
357 if (action == nullptr) {
358 XBT_DEBUG("Still nothing, switch back to the small mailbox : %p", mailbox);
359 mailbox = process->mailbox_small();
362 XBT_DEBUG("yes there was something for us in the large mailbox");
365 mailbox = process->mailbox_small();
366 XBT_DEBUG("Is there a corresponding send already posted the small mailbox?");
367 smx_activity_t action = simcall_comm_iprobe(mailbox, 0, &match_recv, static_cast<void*>(this));
369 if (action == nullptr) {
370 XBT_DEBUG("No, nothing in the permanent receive mailbox");
371 mailbox = process->mailbox();
373 XBT_DEBUG("yes there was something for us in the small mailbox");
377 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
379 action_ = simcall_comm_irecv(
380 process->process(), mailbox, buf_, &real_size_, &match_recv,
381 process->replaying() ? &smpi_comm_null_copy_buffer_callback : smpi_comm_copy_data_callback, this, -1.0);
382 XBT_DEBUG("recv simcall posted");
384 if (async_small_thresh != 0 || (flags_ & RMA) != 0 )
385 xbt_mutex_release(mut);
386 } else { /* the RECV flag was not set, so this is a send */
387 simgrid::smpi::Process* process = smpi_process_remote(dst_);
389 if (TRACE_smpi_view_internals()) {
390 TRACE_smpi_send(rank, rank, dst_, tag_, size_);
392 this->print_request("New send");
395 if ((flags_ & SSEND) == 0 && ( (flags_ & RMA) != 0
396 || static_cast<int>(size_) < xbt_cfg_get_int("smpi/send-is-detached-thresh") ) ) {
397 void *oldbuf = nullptr;
399 XBT_DEBUG("Send request %p is detached", this);
401 if (not(old_type_->flags() & DT_FLAG_DERIVED)) {
403 if (not process->replaying() && oldbuf != nullptr && size_ != 0) {
404 if ((smpi_privatize_global_variables != 0) && (static_cast<char*>(buf_) >= smpi_data_exe_start) &&
405 (static_cast<char*>(buf_) < smpi_data_exe_start + smpi_data_exe_size)) {
406 XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
407 smpi_switch_data_segment(src_);
409 buf = xbt_malloc(size_);
410 memcpy(buf,oldbuf,size_);
411 XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
416 //if we are giving back the control to the user without waiting for completion, we have to inject timings
417 double sleeptime = 0.0;
418 if (detached_ != 0 || ((flags_ & (ISEND | SSEND)) != 0)) { // issend should be treated as isend
419 // isend and send timings may be different
420 sleeptime = ((flags_ & ISEND) != 0)
421 ? simgrid::s4u::Actor::self()->getHost()->extension<simgrid::smpi::SmpiHost>()->oisend(size_)
422 : simgrid::s4u::Actor::self()->getHost()->extension<simgrid::smpi::SmpiHost>()->osend(size_);
426 simcall_process_sleep(sleeptime);
427 XBT_DEBUG("sending size of %zu : sleep %f ", size_, sleeptime);
430 int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh");
432 xbt_mutex_t mut=process->mailboxes_mutex();
434 if (async_small_thresh != 0 || (flags_ & RMA) != 0)
435 xbt_mutex_acquire(mut);
437 if (not(async_small_thresh != 0 || (flags_ & RMA) != 0)) {
438 mailbox = process->mailbox();
439 } else if (((flags_ & RMA) != 0) || static_cast<int>(size_) < async_small_thresh) { // eager mode
440 mailbox = process->mailbox();
441 XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %p?", mailbox);
442 smx_activity_t action = simcall_comm_iprobe(mailbox, 1, &match_send, static_cast<void*>(this));
443 if (action == nullptr) {
444 if ((flags_ & SSEND) == 0){
445 mailbox = process->mailbox_small();
446 XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %p", mailbox);
448 mailbox = process->mailbox_small();
449 XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %p?", mailbox);
450 action = simcall_comm_iprobe(mailbox, 1, &match_send, static_cast<void*>(this));
451 if (action == nullptr) {
452 XBT_DEBUG("No, we are first, send to large mailbox");
453 mailbox = process->mailbox();
457 XBT_DEBUG("Yes there was something for us in the large mailbox");
460 mailbox = process->mailbox();
461 XBT_DEBUG("Send request %p is in the large mailbox %p (buf: %p)",mailbox, this,buf_);
464 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
466 action_ = simcall_comm_isend(
467 SIMIX_process_from_PID(src_ + 1), mailbox, size_, -1.0, buf, real_size_, &match_send,
468 &xbt_free_f, // how to free the userdata if a detached send fails
469 not process->replaying() ? smpi_comm_copy_data_callback : &smpi_comm_null_copy_buffer_callback, this,
470 // detach if msg size < eager/rdv switch limit
472 XBT_DEBUG("send simcall posted");
474 /* FIXME: detached sends are not traceable (action_ == nullptr) */
475 if (action_ != nullptr)
476 simcall_set_category(action_, TRACE_internal_smpi_get_category());
477 if (async_small_thresh != 0 || ((flags_ & RMA)!=0))
478 xbt_mutex_release(mut);
482 void Request::startall(int count, MPI_Request * requests)
484 if(requests== nullptr)
487 for(int i = 0; i < count; i++) {
488 requests[i]->start();
492 int Request::test(MPI_Request * request, MPI_Status * status) {
493 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or testall before)
494 // to avoid deadlocks if used as a break condition, such as
495 // while (MPI_Test(request, flag, status) && flag) dostuff...
496 // because the time will not normally advance when only calls to MPI_Test are made -> deadlock
497 // multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
498 static int nsleeps = 1;
499 if(smpi_test_sleep > 0)
500 simcall_process_sleep(nsleeps*smpi_test_sleep);
502 Status::empty(status);
504 if (((*request)->flags_ & PREPARED) == 0) {
505 if ((*request)->action_ != nullptr)
506 flag = simcall_comm_test((*request)->action_);
508 finish_wait(request,status);
509 nsleeps=1;//reset the number of sleeps we will do next time
510 if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & PERSISTENT) == 0)
511 *request = MPI_REQUEST_NULL;
512 } else if (xbt_cfg_get_boolean("smpi/grow-injected-times")){
519 int Request::testsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
524 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
526 for (int i = 0; i < incount; i++) {
527 if (requests[i] != MPI_REQUEST_NULL) {
528 if (test(&requests[i], pstat)) {
531 if (status != MPI_STATUSES_IGNORE)
533 if ((requests[i] != MPI_REQUEST_NULL) && (requests[i]->flags_ & NON_PERSISTENT))
534 requests[i] = MPI_REQUEST_NULL;
540 if(count_dead==incount)
541 return MPI_UNDEFINED;
545 int Request::testany(int count, MPI_Request requests[], int *index, MPI_Status * status)
547 std::vector<simgrid::kernel::activity::ActivityImplPtr> comms;
548 comms.reserve(count);
553 *index = MPI_UNDEFINED;
555 std::vector<int> map; /** Maps all matching comms back to their location in requests **/
556 for(i = 0; i < count; i++) {
557 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action_ && not(requests[i]->flags_ & PREPARED)) {
558 comms.push_back(requests[i]->action_);
562 if (not map.empty()) {
563 //multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it
564 static int nsleeps = 1;
565 if(smpi_test_sleep > 0)
566 simcall_process_sleep(nsleeps*smpi_test_sleep);
568 i = simcall_comm_testany(comms.data(), comms.size()); // The i-th element in comms matches!
569 if (i != -1) { // -1 is not MPI_UNDEFINED but a SIMIX return code. (nothing matches)
571 finish_wait(&requests[*index],status);
574 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & NON_PERSISTENT)) {
575 requests[*index] = MPI_REQUEST_NULL;
581 //all requests are null or inactive, return true
583 Status::empty(status);
589 int Request::testall(int count, MPI_Request requests[], MPI_Status status[])
592 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
594 for(int i=0; i<count; i++){
595 if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & PREPARED)) {
596 if (test(&requests[i], pstat)!=1){
599 requests[i]=MPI_REQUEST_NULL;
602 Status::empty(pstat);
604 if(status != MPI_STATUSES_IGNORE) {
611 void Request::probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
613 //FIXME find another way to avoid busy waiting ?
614 // the issue here is that we have to wait on a nonexistent comm
616 iprobe(source, tag, comm, &flag, status);
617 XBT_DEBUG("Busy Waiting on probing : %d", flag);
621 void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
622 // to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
623 // especially when used as a break condition, such as while (MPI_Iprobe(...)) dostuff...
624 // nsleeps is a multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
625 // This can speed up the execution of certain applications by an order of magnitude, such as HPL
626 static int nsleeps = 1;
627 double speed = simgrid::s4u::Actor::self()->getHost()->getSpeed();
628 double maxrate = xbt_cfg_get_double("smpi/iprobe-cpu-usage");
629 MPI_Request request = new Request(nullptr, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
630 comm->group()->index(source), comm->rank(), tag, comm, PERSISTENT | RECV);
631 if (smpi_iprobe_sleep > 0) {
632 smx_activity_t iprobe_sleep = simcall_execution_start("iprobe", /* flops to executek*/nsleeps*smpi_iprobe_sleep*speed*maxrate, /* priority */1.0, /* performance bound */maxrate*speed);
633 simcall_execution_wait(iprobe_sleep);
635 // behave like a receive, but don't do it
636 smx_mailbox_t mailbox;
638 request->print_request("New iprobe");
639 // We have to test both mailboxes as we don't know if we will receive one one or another
640 if (xbt_cfg_get_int("smpi/async-small-thresh") > 0){
641 mailbox = smpi_process()->mailbox_small();
642 XBT_DEBUG("Trying to probe the perm recv mailbox");
643 request->action_ = simcall_comm_iprobe(mailbox, 0, &match_recv, static_cast<void*>(request));
646 if (request->action_ == nullptr){
647 mailbox = smpi_process()->mailbox();
648 XBT_DEBUG("trying to probe the other mailbox");
649 request->action_ = simcall_comm_iprobe(mailbox, 0, &match_recv, static_cast<void*>(request));
652 if (request->action_ != nullptr){
653 simgrid::kernel::activity::CommImplPtr sync_comm =
654 boost::static_pointer_cast<simgrid::kernel::activity::CommImpl>(request->action_);
655 MPI_Request req = static_cast<MPI_Request>(sync_comm->src_data);
657 if(status != MPI_STATUS_IGNORE && (req->flags_ & PREPARED) == 0) {
658 status->MPI_SOURCE = comm->group()->rank(req->src_);
659 status->MPI_TAG = req->tag_;
660 status->MPI_ERROR = MPI_SUCCESS;
661 status->count = req->real_size_;
663 nsleeps = 1;//reset the number of sleeps we will do next time
667 if (xbt_cfg_get_boolean("smpi/grow-injected-times"))
673 void Request::finish_wait(MPI_Request* request, MPI_Status * status)
675 MPI_Request req = *request;
676 Status::empty(status);
678 if (not((req->detached_ != 0) && ((req->flags_ & SEND) != 0)) && ((req->flags_ & PREPARED) == 0)) {
679 if(status != MPI_STATUS_IGNORE) {
680 int src = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_;
681 status->MPI_SOURCE = req->comm_->group()->rank(src);
682 status->MPI_TAG = req->tag_ == MPI_ANY_TAG ? req->real_tag_ : req->tag_;
683 status->MPI_ERROR = req->truncated_ != 0 ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
684 // this handles the case were size in receive differs from size in send
685 status->count = req->real_size_;
688 req->print_request("Finishing");
689 MPI_Datatype datatype = req->old_type_;
691 // FIXME Handle the case of a partial shared malloc.
692 if (((req->flags_ & ACCUMULATE) != 0) ||
693 (datatype->flags() & DT_FLAG_DERIVED)) { // && (not smpi_is_shared(req->old_buf_))){
695 if (not smpi_process()->replaying()) {
696 if (smpi_privatize_global_variables != 0 && (static_cast<char*>(req->old_buf_) >= smpi_data_exe_start) &&
697 ((char*)req->old_buf_ < smpi_data_exe_start + smpi_data_exe_size)) {
698 XBT_VERB("Privatization : We are unserializing to a zone in global memory Switch data segment ");
699 smpi_switch_data_segment(smpi_process()->index());
703 if(datatype->flags() & DT_FLAG_DERIVED){
704 // This part handles the problem of non-contignous memory the unserialization at the reception
705 if((req->flags_ & RECV) && datatype->size()!=0)
706 datatype->unserialize(req->buf_, req->old_buf_, req->real_size_/datatype->size() , req->op_);
708 }else if(req->flags_ & RECV){//apply op on contiguous buffer for accumulate
709 if(datatype->size()!=0){
710 int n =req->real_size_/datatype->size();
711 req->op_->apply(req->buf_, req->old_buf_, &n, datatype);
718 if (TRACE_smpi_view_internals() && ((req->flags_ & RECV) != 0)){
719 int rank = smpi_process()->index();
720 int src_traced = (req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_);
721 TRACE_smpi_recv(src_traced, rank,req->tag_);
723 if(req->detached_sender_ != nullptr){
724 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
726 simgrid::s4u::Actor::self()->getHost()->extension<simgrid::smpi::SmpiHost>()->orecv(req->real_size());
728 simcall_process_sleep(sleeptime);
729 XBT_DEBUG("receiving size of %zu : sleep %f ", req->real_size_, sleeptime);
731 unref(&(req->detached_sender_));
733 if(req->flags_ & PERSISTENT)
734 req->action_ = nullptr;
735 req->flags_ |= FINISHED;
739 void Request::wait(MPI_Request * request, MPI_Status * status)
741 (*request)->print_request("Waiting");
742 if ((*request)->flags_ & PREPARED) {
743 Status::empty(status);
747 if ((*request)->action_ != nullptr)
748 // this is not a detached send
749 simcall_comm_wait((*request)->action_, -1.0);
751 finish_wait(request,status);
752 if (*request != MPI_REQUEST_NULL && (((*request)->flags_ & NON_PERSISTENT)!=0))
753 *request = MPI_REQUEST_NULL;
756 int Request::waitany(int count, MPI_Request requests[], MPI_Status * status)
758 s_xbt_dynar_t comms; // Keep it on stack to save some extra mallocs
759 int index = MPI_UNDEFINED;
763 // Wait for a request to complete
764 xbt_dynar_init(&comms, sizeof(smx_activity_t), [](void*ptr){
765 intrusive_ptr_release(*(simgrid::kernel::activity::ActivityImpl**)ptr);
767 int *map = xbt_new(int, count);
768 XBT_DEBUG("Wait for one of %d", count);
769 for(int i = 0; i < count; i++) {
770 if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & PREPARED) &&
771 not(requests[i]->flags_ & FINISHED)) {
772 if (requests[i]->action_ != nullptr) {
773 XBT_DEBUG("Waiting any %p ", requests[i]);
774 intrusive_ptr_add_ref(requests[i]->action_.get());
775 xbt_dynar_push_as(&comms, simgrid::kernel::activity::ActivityImpl*, requests[i]->action_.get());
779 // This is a finished detached request, let's return this one
780 size = 0; // so we free the dynar but don't do the waitany call
782 finish_wait(&requests[i], status); // cleanup if refcount = 0
783 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & NON_PERSISTENT))
784 requests[i] = MPI_REQUEST_NULL; // set to null
790 XBT_DEBUG("Enter waitany for %lu comms", xbt_dynar_length(&comms));
791 int i = simcall_comm_waitany(&comms, -1);
793 // not MPI_UNDEFINED, as this is a simix return code
796 //in case of an accumulate, we have to wait the end of all requests to apply the operation, ordered correctly.
797 if ((requests[index] == MPI_REQUEST_NULL) ||
798 (not((requests[index]->flags_ & ACCUMULATE) && (requests[index]->flags_ & RECV)))) {
799 finish_wait(&requests[index],status);
800 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & NON_PERSISTENT))
801 requests[index] = MPI_REQUEST_NULL;
806 xbt_dynar_free_data(&comms);
810 if (index==MPI_UNDEFINED)
811 Status::empty(status);
816 static int sort_accumulates(MPI_Request a, MPI_Request b)
818 return (a->tag() > b->tag());
821 int Request::waitall(int count, MPI_Request requests[], MPI_Status status[])
823 std::vector<MPI_Request> accumulates;
826 MPI_Status *pstat = (status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat);
827 int retvalue = MPI_SUCCESS;
828 //tag invalid requests in the set
829 if (status != MPI_STATUSES_IGNORE) {
830 for (int c = 0; c < count; c++) {
831 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst_ == MPI_PROC_NULL || (requests[c]->flags_ & PREPARED)) {
832 Status::empty(&status[c]);
833 } else if (requests[c]->src_ == MPI_PROC_NULL) {
834 Status::empty(&status[c]);
835 status[c].MPI_SOURCE = MPI_PROC_NULL;
839 for (int c = 0; c < count; c++) {
840 if (MC_is_active() || MC_record_replay_is_active()) {
841 wait(&requests[c],pstat);
844 index = waitany(count, (MPI_Request*)requests, pstat);
845 if (index == MPI_UNDEFINED)
848 if (requests[index] != MPI_REQUEST_NULL
849 && (requests[index]->flags_ & RECV)
850 && (requests[index]->flags_ & ACCUMULATE))
851 accumulates.push_back(requests[index]);
852 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & NON_PERSISTENT))
853 requests[index] = MPI_REQUEST_NULL;
855 if (status != MPI_STATUSES_IGNORE) {
856 status[index] = *pstat;
857 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
858 retvalue = MPI_ERR_IN_STATUS;
862 if (not accumulates.empty()) {
863 std::sort(accumulates.begin(), accumulates.end(), sort_accumulates);
864 for (auto& req : accumulates) {
865 finish_wait(&req, status);
872 int Request::waitsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
876 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
878 for (int i = 0; i < incount; i++) {
879 int index = waitany(incount, requests, pstat);
880 if(index!=MPI_UNDEFINED){
881 indices[count] = index;
883 if(status != MPI_STATUSES_IGNORE) {
884 status[index] = *pstat;
886 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & NON_PERSISTENT))
887 requests[index] = MPI_REQUEST_NULL;
889 return MPI_UNDEFINED;
895 MPI_Request Request::f2c(int id) {
897 if(id==MPI_FORTRAN_REQUEST_NULL)
898 return static_cast<MPI_Request>(MPI_REQUEST_NULL);
899 return static_cast<MPI_Request>(F2C::f2c_lookup()->at(get_key_id(key, id)));
904 if (F2C::f2c_lookup() == nullptr) {
905 F2C::set_f2c_lookup(new std::unordered_map<std::string, F2C*>);
908 (*(F2C::f2c_lookup()))[get_key_id(key, F2C::f2c_id())] = this;
909 F2C::f2c_id_increment();
910 return F2C::f2c_id()-1;
913 void Request::free_f(int id)
915 if (id != MPI_FORTRAN_REQUEST_NULL) {
917 F2C::f2c_lookup()->erase(get_key_id(key, id));