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 "src/smpi/smpi_request.hpp"
9 #include "src/kernel/activity/CommImpl.hpp"
10 #include "src/mc/mc_replay.h"
11 #include "src/smpi/SmpiHost.hpp"
12 #include "src/smpi/private.h"
13 #include "src/smpi/smpi_comm.hpp"
14 #include "src/smpi/smpi_datatype.hpp"
15 #include "src/smpi/smpi_op.hpp"
16 #include "src/smpi/smpi_process.hpp"
20 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_request, smpi, "Logging specific to SMPI (reques)");
22 static simgrid::config::Flag<double> smpi_iprobe_sleep(
23 "smpi/iprobe", "Minimum time to inject inside a call to MPI_Iprobe", 1e-4);
24 static simgrid::config::Flag<double> smpi_test_sleep(
25 "smpi/test", "Minimum time to inject inside a call to MPI_Test", 1e-4);
27 std::vector<s_smpi_factor_t> smpi_ois_values;
29 extern void (*smpi_comm_copy_data_callback) (smx_activity_t, void*, size_t);
34 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)
36 void *old_buf = nullptr;
37 // FIXME Handle the case of a partial shared malloc.
38 if ((((flags & RECV) != 0) && ((flags & ACCUMULATE) != 0)) || (datatype->flags() & DT_FLAG_DERIVED)) {
39 // This part handles the problem of non-contiguous memory
44 buf_ = xbt_malloc(count*datatype->size());
45 if ((datatype->flags() & DT_FLAG_DERIVED) && ((flags & SEND) != 0)) {
46 datatype->serialize(old_buf, buf_, count);
50 // This part handles the problem of non-contiguous memory (for the unserialisation at the reception)
52 size_ = datatype->size() * count;
57 detached_sender_ = nullptr;
62 if (flags & PERSISTENT)
69 MPI_Comm Request::comm(){
89 int Request::detached(){
93 size_t Request::size(){
97 size_t Request::real_size(){
101 void Request::unref(MPI_Request* request)
103 if((*request) != MPI_REQUEST_NULL){
104 (*request)->refcount_--;
105 if((*request)->refcount_<0) xbt_die("wrong refcount");
106 if((*request)->refcount_==0){
107 Datatype::unref((*request)->old_type_);
108 Comm::unref((*request)->comm_);
109 (*request)->print_request("Destroying");
111 *request = MPI_REQUEST_NULL;
113 (*request)->print_request("Decrementing");
116 xbt_die("freeing an already free request");
120 int Request::match_recv(void* a, void* b, simgrid::kernel::activity::CommImpl* ignored)
122 MPI_Request ref = static_cast<MPI_Request>(a);
123 MPI_Request req = static_cast<MPI_Request>(b);
124 XBT_DEBUG("Trying to match a recv of src %d against %d, tag %d against %d",ref->src_,req->src_, ref->tag_, req->tag_);
126 xbt_assert(ref, "Cannot match recv against null reference");
127 xbt_assert(req, "Cannot match recv against null request");
128 if((ref->src_ == MPI_ANY_SOURCE || req->src_ == ref->src_)
129 && ((ref->tag_ == MPI_ANY_TAG && req->tag_ >=0) || req->tag_ == ref->tag_)){
130 //we match, we can transfer some values
131 if(ref->src_ == MPI_ANY_SOURCE)
132 ref->real_src_ = req->src_;
133 if(ref->tag_ == MPI_ANY_TAG)
134 ref->real_tag_ = req->tag_;
135 if(ref->real_size_ < req->real_size_)
137 if(req->detached_==1)
138 ref->detached_sender_=req; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
139 XBT_DEBUG("match succeeded");
144 int Request::match_send(void* a, void* b, simgrid::kernel::activity::CommImpl* ignored)
146 MPI_Request ref = static_cast<MPI_Request>(a);
147 MPI_Request req = static_cast<MPI_Request>(b);
148 XBT_DEBUG("Trying to match a send of src %d against %d, tag %d against %d",ref->src_,req->src_, ref->tag_, req->tag_);
149 xbt_assert(ref, "Cannot match send against null reference");
150 xbt_assert(req, "Cannot match send against null request");
152 if((req->src_ == MPI_ANY_SOURCE || req->src_ == ref->src_)
153 && ((req->tag_ == MPI_ANY_TAG && ref->tag_ >=0)|| req->tag_ == ref->tag_)){
154 if(req->src_ == MPI_ANY_SOURCE)
155 req->real_src_ = ref->src_;
156 if(req->tag_ == MPI_ANY_TAG)
157 req->real_tag_ = ref->tag_;
158 if(req->real_size_ < ref->real_size_)
160 if(ref->detached_==1)
161 req->detached_sender_=ref; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
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(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, smpi_process()->index(),
180 comm->group()->index(dst), tag, comm, PERSISTENT | SEND | PREPARED);
183 MPI_Request Request::ssend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
185 return new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
186 comm->group()->index(dst), tag, comm, PERSISTENT | SSEND | SEND | PREPARED);
189 MPI_Request Request::isend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
191 return new Request(buf==MPI_BOTTOM ? nullptr : buf , count, datatype, smpi_process()->index(),
192 comm->group()->index(dst), tag,comm, PERSISTENT | ISEND | SEND | PREPARED);
196 MPI_Request Request::rma_send_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
199 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
201 request = new Request(buf==MPI_BOTTOM ? nullptr : buf , count, datatype, src, dst, tag,
202 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED);
204 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src, dst, tag,
205 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED | ACCUMULATE);
211 MPI_Request Request::recv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
213 return new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype,
214 src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->index(src),
215 smpi_process()->index(), tag, comm, PERSISTENT | RECV | PREPARED);
218 MPI_Request Request::rma_recv_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
221 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
223 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src, dst, tag,
224 comm, RMA | NON_PERSISTENT | RECV | PREPARED);
226 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src, dst, tag,
227 comm, RMA | NON_PERSISTENT | RECV | PREPARED | ACCUMULATE);
233 MPI_Request Request::irecv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
235 return new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
236 comm->group()->index(src), smpi_process()->index(), tag,
237 comm, PERSISTENT | RECV | PREPARED);
240 MPI_Request Request::isend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
242 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
243 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
244 comm->group()->index(dst), tag, comm, NON_PERSISTENT | ISEND | SEND);
249 MPI_Request Request::issend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
251 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
252 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
253 comm->group()->index(dst), tag,comm, NON_PERSISTENT | ISEND | SSEND | SEND);
259 MPI_Request Request::irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
261 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
262 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
263 comm->group()->index(src), smpi_process()->index(), tag, comm,
264 NON_PERSISTENT | RECV);
269 void Request::recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status * status)
271 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
272 request = irecv(buf, count, datatype, src, tag, comm);
273 wait(&request,status);
277 void Request::send(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
279 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
280 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
281 comm->group()->index(dst), tag, comm, NON_PERSISTENT | SEND);
284 wait(&request, MPI_STATUS_IGNORE);
288 void Request::ssend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
290 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
291 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
292 comm->group()->index(dst), tag, comm, NON_PERSISTENT | SSEND | SEND);
295 wait(&request,MPI_STATUS_IGNORE);
299 void Request::sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,int dst, int sendtag,
300 void *recvbuf, int recvcount, MPI_Datatype recvtype, int src, int recvtag,
301 MPI_Comm comm, MPI_Status * status)
303 MPI_Request requests[2];
305 int myid=smpi_process()->index();
306 if ((comm->group()->index(dst) == myid) && (comm->group()->index(src) == myid)){
307 Datatype::copy(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype);
310 requests[0] = isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
311 requests[1] = irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
312 startall(2, requests);
313 waitall(2, requests, stats);
316 if(status != MPI_STATUS_IGNORE) {
317 // Copy receive status
322 void Request::start()
324 smx_mailbox_t mailbox;
326 xbt_assert(action_ == nullptr, "Cannot (re-)start unfinished communication");
331 if ((flags_ & RECV) != 0) {
332 this->print_request("New recv");
334 simgrid::smpi::Process* process = smpi_process_remote(dst_);
336 int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh");
338 xbt_mutex_t mut = process->mailboxes_mutex();
339 if (async_small_thresh != 0 || (flags_ & RMA) != 0)
340 xbt_mutex_acquire(mut);
342 if (async_small_thresh == 0 && (flags_ & RMA) == 0 ) {
343 mailbox = process->mailbox();
345 else if (((flags_ & RMA) != 0) || static_cast<int>(size_) < async_small_thresh) {
346 //We have to check both mailboxes (because SSEND messages are sent to the large mbox).
347 //begin with the more appropriate one : the small one.
348 mailbox = process->mailbox_small();
349 XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %p (in case of SSEND)?", mailbox);
350 smx_activity_t action = simcall_comm_iprobe(mailbox, 0, src_,tag_, &match_recv,
351 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, src_,tag_, &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, src_,tag_, &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)
405 && (static_cast<char*>(buf_) >= smpi_start_data_exe)
406 && (static_cast<char*>(buf_) < smpi_start_data_exe + smpi_size_data_exe )){
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()->host()->extension<simgrid::smpi::SmpiHost>()->oisend(size_)
423 : simgrid::s4u::Actor::self()->host()->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,dst_, tag_, &match_send,
444 static_cast<void*>(this));
445 if (action == nullptr) {
446 if ((flags_ & SSEND) == 0){
447 mailbox = process->mailbox_small();
448 XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %p", mailbox);
450 mailbox = process->mailbox_small();
451 XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %p?", mailbox);
452 action = simcall_comm_iprobe(mailbox, 1,dst_, tag_, &match_send, static_cast<void*>(this));
453 if (action == nullptr) {
454 XBT_DEBUG("No, we are first, send to large mailbox");
455 mailbox = process->mailbox();
459 XBT_DEBUG("Yes there was something for us in the large mailbox");
462 mailbox = process->mailbox();
463 XBT_DEBUG("Send request %p is in the large mailbox %p (buf: %p)",mailbox, this,buf_);
466 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
468 action_ = simcall_comm_isend(
469 SIMIX_process_from_PID(src_ + 1), mailbox, size_, -1.0, buf, real_size_, &match_send,
470 &xbt_free_f, // how to free the userdata if a detached send fails
471 not process->replaying() ? smpi_comm_copy_data_callback : &smpi_comm_null_copy_buffer_callback, this,
472 // detach if msg size < eager/rdv switch limit
474 XBT_DEBUG("send simcall posted");
476 /* FIXME: detached sends are not traceable (action_ == nullptr) */
477 if (action_ != nullptr)
478 simcall_set_category(action_, TRACE_internal_smpi_get_category());
479 if (async_small_thresh != 0 || ((flags_ & RMA)!=0))
480 xbt_mutex_release(mut);
484 void Request::startall(int count, MPI_Request * requests)
486 if(requests== nullptr)
489 for(int i = 0; i < count; i++) {
490 requests[i]->start();
494 int Request::test(MPI_Request * request, MPI_Status * status) {
495 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or testall before)
496 // to avoid deadlocks if used as a break condition, such as
497 // while (MPI_Test(request, flag, status) && flag) dostuff...
498 // because the time will not normally advance when only calls to MPI_Test are made -> deadlock
499 // multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
500 static int nsleeps = 1;
501 if(smpi_test_sleep > 0)
502 simcall_process_sleep(nsleeps*smpi_test_sleep);
504 Status::empty(status);
506 if (((*request)->flags_ & PREPARED) == 0) {
507 if ((*request)->action_ != nullptr)
508 flag = simcall_comm_test((*request)->action_);
510 finish_wait(request,status);
511 nsleeps=1;//reset the number of sleeps we will do next time
512 if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & PERSISTENT)==0)
513 *request = MPI_REQUEST_NULL;
514 } else if (xbt_cfg_get_boolean("smpi/grow-injected-times")){
521 int Request::testsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
526 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
528 for (int i = 0; i < incount; i++) {
529 if (requests[i] != MPI_REQUEST_NULL) {
530 if (test(&requests[i], pstat)) {
533 if (status != MPI_STATUSES_IGNORE)
535 if ((requests[i] != MPI_REQUEST_NULL) && (requests[i]->flags_ & NON_PERSISTENT))
536 requests[i] = MPI_REQUEST_NULL;
542 if(count_dead==incount)
543 return MPI_UNDEFINED;
547 int Request::testany(int count, MPI_Request requests[], int *index, MPI_Status * status)
549 std::vector<simgrid::kernel::activity::ActivityImplPtr> comms;
550 comms.reserve(count);
555 *index = MPI_UNDEFINED;
557 std::vector<int> map; /** Maps all matching comms back to their location in requests **/
558 for(i = 0; i < count; i++) {
559 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action_ && not(requests[i]->flags_ & PREPARED)) {
560 comms.push_back(requests[i]->action_);
564 if (not map.empty()) {
565 //multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it
566 static int nsleeps = 1;
567 if(smpi_test_sleep > 0)
568 simcall_process_sleep(nsleeps*smpi_test_sleep);
570 i = simcall_comm_testany(comms.data(), comms.size()); // The i-th element in comms matches!
571 if (i != -1) { // -1 is not MPI_UNDEFINED but a SIMIX return code. (nothing matches)
573 finish_wait(&requests[*index],status);
576 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & NON_PERSISTENT)) {
577 requests[*index] = MPI_REQUEST_NULL;
583 //all requests are null or inactive, return true
585 Status::empty(status);
591 int Request::testall(int count, MPI_Request requests[], MPI_Status status[])
594 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
596 for(int i=0; i<count; i++){
597 if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & PREPARED)) {
598 if (test(&requests[i], pstat)!=1){
601 requests[i]=MPI_REQUEST_NULL;
604 Status::empty(pstat);
606 if(status != MPI_STATUSES_IGNORE) {
613 void Request::probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
615 //FIXME find another way to avoid busy waiting ?
616 // the issue here is that we have to wait on a nonexistent comm
618 iprobe(source, tag, comm, &flag, status);
619 XBT_DEBUG("Busy Waiting on probing : %d", flag);
623 void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
624 // to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
625 // especially when used as a break condition, such as while (MPI_Iprobe(...)) dostuff...
626 // nsleeps is a multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
627 // This can speed up the execution of certain applications by an order of magnitude, such as HPL
628 static int nsleeps = 1;
629 double speed = simgrid::s4u::Actor::self()->host()->speed();
630 double maxrate = xbt_cfg_get_double("smpi/iprobe-cpu-usage");
631 MPI_Request request = new Request(nullptr, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
632 comm->group()->index(source), comm->rank(), tag, comm, PERSISTENT | RECV);
633 if (smpi_iprobe_sleep > 0) {
634 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);
635 simcall_execution_wait(iprobe_sleep);
637 // behave like a receive, but don't do it
638 smx_mailbox_t mailbox;
640 request->print_request("New iprobe");
641 // We have to test both mailboxes as we don't know if we will receive one one or another
642 if (xbt_cfg_get_int("smpi/async-small-thresh") > 0){
643 mailbox = smpi_process()->mailbox_small();
644 XBT_DEBUG("Trying to probe the perm recv mailbox");
645 request->action_ = simcall_comm_iprobe(mailbox, 0, request->src_, request->tag_, &match_recv,
646 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, request->src_,request->tag_, &match_recv,
653 static_cast<void*>(request));
656 if (request->action_ != nullptr){
657 simgrid::kernel::activity::CommImplPtr sync_comm =
658 boost::static_pointer_cast<simgrid::kernel::activity::CommImpl>(request->action_);
659 MPI_Request req = static_cast<MPI_Request>(sync_comm->src_data);
661 if(status != MPI_STATUS_IGNORE && (req->flags_ & PREPARED) == 0) {
662 status->MPI_SOURCE = comm->group()->rank(req->src_);
663 status->MPI_TAG = req->tag_;
664 status->MPI_ERROR = MPI_SUCCESS;
665 status->count = req->real_size_;
667 nsleeps = 1;//reset the number of sleeps we will do next time
671 if (xbt_cfg_get_boolean("smpi/grow-injected-times"))
677 void Request::finish_wait(MPI_Request* request, MPI_Status * status)
679 MPI_Request req = *request;
680 Status::empty(status);
682 if (not((req->detached_ != 0) && ((req->flags_ & SEND) != 0)) && ((req->flags_ & PREPARED) == 0)) {
683 if(status != MPI_STATUS_IGNORE) {
684 int src = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_;
685 status->MPI_SOURCE = req->comm_->group()->rank(src);
686 status->MPI_TAG = req->tag_ == MPI_ANY_TAG ? req->real_tag_ : req->tag_;
687 status->MPI_ERROR = req->truncated_ != 0 ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
688 // this handles the case were size in receive differs from size in send
689 status->count = req->real_size_;
692 req->print_request("Finishing");
693 MPI_Datatype datatype = req->old_type_;
695 // FIXME Handle the case of a partial shared malloc.
696 if (((req->flags_ & ACCUMULATE) != 0) ||
697 (datatype->flags() & DT_FLAG_DERIVED)) { // && (not smpi_is_shared(req->old_buf_))){
699 if (not smpi_process()->replaying()) {
700 if( smpi_privatize_global_variables != 0 && (static_cast<char*>(req->old_buf_) >= smpi_start_data_exe)
701 && ((char*)req->old_buf_ < smpi_start_data_exe + smpi_size_data_exe )){
702 XBT_VERB("Privatization : We are unserializing to a zone in global memory Switch data segment ");
703 smpi_switch_data_segment(smpi_process()->index());
707 if(datatype->flags() & DT_FLAG_DERIVED){
708 // This part handles the problem of non-contignous memory the unserialization at the reception
709 if((req->flags_ & RECV) && datatype->size()!=0)
710 datatype->unserialize(req->buf_, req->old_buf_, req->real_size_/datatype->size() , req->op_);
712 }else if(req->flags_ & RECV){//apply op on contiguous buffer for accumulate
713 if(datatype->size()!=0){
714 int n =req->real_size_/datatype->size();
715 req->op_->apply(req->buf_, req->old_buf_, &n, datatype);
722 if (TRACE_smpi_view_internals() && ((req->flags_ & RECV) != 0)){
723 int rank = smpi_process()->index();
724 int src_traced = (req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_);
725 TRACE_smpi_recv(rank, src_traced, rank,req->tag_);
727 if(req->detached_sender_ != nullptr){
728 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
729 double sleeptime = simgrid::s4u::Actor::self()->host()->extension<simgrid::smpi::SmpiHost>()->orecv(req->real_size());
731 simcall_process_sleep(sleeptime);
732 XBT_DEBUG("receiving size of %zu : sleep %f ", req->real_size_, sleeptime);
734 unref(&(req->detached_sender_));
736 if(req->flags_ & PERSISTENT)
737 req->action_ = nullptr;
738 req->flags_ |= FINISHED;
742 void Request::wait(MPI_Request * request, MPI_Status * status)
744 (*request)->print_request("Waiting");
745 if ((*request)->flags_ & PREPARED) {
746 Status::empty(status);
750 if ((*request)->action_ != nullptr)
751 // this is not a detached send
752 simcall_comm_wait((*request)->action_, -1.0);
754 finish_wait(request,status);
755 if (*request != MPI_REQUEST_NULL && (((*request)->flags_ & NON_PERSISTENT)!=0))
756 *request = MPI_REQUEST_NULL;
759 int Request::waitany(int count, MPI_Request requests[], MPI_Status * status)
761 s_xbt_dynar_t comms; // Keep it on stack to save some extra mallocs
762 int index = MPI_UNDEFINED;
766 // Wait for a request to complete
767 xbt_dynar_init(&comms, sizeof(smx_activity_t), [](void*ptr){
768 intrusive_ptr_release(*(simgrid::kernel::activity::ActivityImpl**)ptr);
770 int *map = xbt_new(int, count);
771 XBT_DEBUG("Wait for one of %d", count);
772 for(int i = 0; i < count; i++) {
773 if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & PREPARED) &&
774 not(requests[i]->flags_ & FINISHED)) {
775 if (requests[i]->action_ != nullptr) {
776 XBT_DEBUG("Waiting any %p ", requests[i]);
777 intrusive_ptr_add_ref(requests[i]->action_.get());
778 xbt_dynar_push_as(&comms, simgrid::kernel::activity::ActivityImpl*, requests[i]->action_.get());
782 // This is a finished detached request, let's return this one
783 size = 0; // so we free the dynar but don't do the waitany call
785 finish_wait(&requests[i], status); // cleanup if refcount = 0
786 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & NON_PERSISTENT))
787 requests[i] = MPI_REQUEST_NULL; // set to null
793 XBT_DEBUG("Enter waitany for %lu comms", xbt_dynar_length(&comms));
794 int i = simcall_comm_waitany(&comms, -1);
796 // not MPI_UNDEFINED, as this is a simix return code
799 //in case of an accumulate, we have to wait the end of all requests to apply the operation, ordered correctly.
800 if ((requests[index] == MPI_REQUEST_NULL) ||
801 (not((requests[index]->flags_ & ACCUMULATE) && (requests[index]->flags_ & RECV)))) {
802 finish_wait(&requests[index],status);
803 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & NON_PERSISTENT))
804 requests[index] = MPI_REQUEST_NULL;
809 xbt_dynar_free_data(&comms);
813 if (index==MPI_UNDEFINED)
814 Status::empty(status);
819 static int sort_accumulates(MPI_Request a, MPI_Request b)
821 return (a->tag() > b->tag());
824 int Request::waitall(int count, MPI_Request requests[], MPI_Status status[])
826 std::vector<MPI_Request> accumulates;
829 MPI_Status *pstat = (status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat);
830 int retvalue = MPI_SUCCESS;
831 //tag invalid requests in the set
832 if (status != MPI_STATUSES_IGNORE) {
833 for (int c = 0; c < count; c++) {
834 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst_ == MPI_PROC_NULL || (requests[c]->flags_ & PREPARED)) {
835 Status::empty(&status[c]);
836 } else if (requests[c]->src_ == MPI_PROC_NULL) {
837 Status::empty(&status[c]);
838 status[c].MPI_SOURCE = MPI_PROC_NULL;
842 for (int c = 0; c < count; c++) {
843 if (MC_is_active() || MC_record_replay_is_active()) {
844 wait(&requests[c],pstat);
847 index = waitany(count, (MPI_Request*)requests, pstat);
848 if (index == MPI_UNDEFINED)
851 if (requests[index] != MPI_REQUEST_NULL
852 && (requests[index]->flags_ & RECV)
853 && (requests[index]->flags_ & ACCUMULATE))
854 accumulates.push_back(requests[index]);
855 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & NON_PERSISTENT))
856 requests[index] = MPI_REQUEST_NULL;
858 if (status != MPI_STATUSES_IGNORE) {
859 status[index] = *pstat;
860 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
861 retvalue = MPI_ERR_IN_STATUS;
865 if (not accumulates.empty()) {
866 std::sort(accumulates.begin(), accumulates.end(), sort_accumulates);
867 for (auto req : accumulates) {
868 finish_wait(&req, status);
875 int Request::waitsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
879 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
881 for (int i = 0; i < incount; i++) {
882 int index = waitany(incount, requests, pstat);
883 if(index!=MPI_UNDEFINED){
884 indices[count] = index;
886 if(status != MPI_STATUSES_IGNORE) {
887 status[index] = *pstat;
889 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & NON_PERSISTENT))
890 requests[index]=MPI_REQUEST_NULL;
892 return MPI_UNDEFINED;
898 MPI_Request Request::f2c(int id) {
900 if(id==MPI_FORTRAN_REQUEST_NULL)
901 return static_cast<MPI_Request>(MPI_REQUEST_NULL);
902 return static_cast<MPI_Request>(xbt_dict_get(F2C::f2c_lookup(), get_key_id(key, id)));
905 int Request::add_f() {
906 if(F2C::f2c_lookup()==nullptr){
907 F2C::set_f2c_lookup(xbt_dict_new_homogeneous(nullptr));
910 xbt_dict_set(F2C::f2c_lookup(), get_key_id(key, F2C::f2c_id()), this, nullptr);
911 F2C::f2c_id_increment();
912 return F2C::f2c_id()-1;
915 void Request::free_f(int id) {
916 if (id != MPI_FORTRAN_REQUEST_NULL) {
918 xbt_dict_remove(F2C::f2c_lookup(), get_key_id(key, id));