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. */
8 #include "src/mc/mc_replay.h"
9 #include "src/smpi/SmpiHost.hpp"
10 #include "src/kernel/activity/SynchroComm.hpp"
14 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_request, smpi, "Logging specific to SMPI (reques)");
16 static simgrid::config::Flag<double> smpi_iprobe_sleep(
17 "smpi/iprobe", "Minimum time to inject inside a call to MPI_Iprobe", 1e-4);
18 static simgrid::config::Flag<double> smpi_test_sleep(
19 "smpi/test", "Minimum time to inject inside a call to MPI_Test", 1e-4);
21 std::vector<s_smpi_factor_t> smpi_ois_values;
23 extern void (*smpi_comm_copy_data_callback) (smx_activity_t, void*, size_t);
28 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)
30 void *old_buf = nullptr;
31 // FIXME Handle the case of a partial shared malloc.
32 if ((((flags & RECV) != 0) && ((flags & ACCUMULATE) != 0)) || (datatype->flags() & DT_FLAG_DERIVED)) {
33 // This part handles the problem of non-contiguous memory
38 buf_ = xbt_malloc(count*datatype->size());
39 if ((datatype->flags() & DT_FLAG_DERIVED) && ((flags & SEND) != 0)) {
40 datatype->serialize(old_buf, buf_, count);
44 // This part handles the problem of non-contiguous memory (for the unserialisation at the reception)
46 size_ = datatype->size() * count;
51 detached_sender_ = nullptr;
56 if (flags & PERSISTENT)
63 MPI_Comm Request::comm(){
83 int Request::detached(){
87 size_t Request::size(){
91 size_t Request::real_size(){
95 void Request::unref(MPI_Request* request)
97 if((*request) != MPI_REQUEST_NULL){
98 (*request)->refcount_--;
99 if((*request)->refcount_<0) xbt_die("wrong refcount");
100 if((*request)->refcount_==0){
101 Datatype::unref((*request)->old_type_);
102 Comm::unref((*request)->comm_);
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, smx_activity_t ignored) {
115 MPI_Request ref = static_cast<MPI_Request>(a);
116 MPI_Request req = static_cast<MPI_Request>(b);
117 XBT_DEBUG("Trying to match a recv of src %d against %d, tag %d against %d",ref->src_,req->src_, ref->tag_, req->tag_);
119 xbt_assert(ref, "Cannot match recv against null reference");
120 xbt_assert(req, "Cannot match recv against null request");
121 if((ref->src_ == MPI_ANY_SOURCE || req->src_ == ref->src_)
122 && ((ref->tag_ == MPI_ANY_TAG && req->tag_ >=0) || req->tag_ == ref->tag_)){
123 //we match, we can transfer some values
124 if(ref->src_ == MPI_ANY_SOURCE)
125 ref->real_src_ = req->src_;
126 if(ref->tag_ == MPI_ANY_TAG)
127 ref->real_tag_ = req->tag_;
128 if(ref->real_size_ < req->real_size_)
130 if(req->detached_==1)
131 ref->detached_sender_=req; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
132 XBT_DEBUG("match succeeded");
137 int Request::match_send(void* a, void* b,smx_activity_t ignored) {
138 MPI_Request ref = static_cast<MPI_Request>(a);
139 MPI_Request req = static_cast<MPI_Request>(b);
140 XBT_DEBUG("Trying to match a send of src %d against %d, tag %d against %d",ref->src_,req->src_, ref->tag_, req->tag_);
141 xbt_assert(ref, "Cannot match send against null reference");
142 xbt_assert(req, "Cannot match send against null request");
144 if((req->src_ == MPI_ANY_SOURCE || req->src_ == ref->src_)
145 && ((req->tag_ == MPI_ANY_TAG && ref->tag_ >=0)|| req->tag_ == ref->tag_)){
146 if(req->src_ == MPI_ANY_SOURCE)
147 req->real_src_ = ref->src_;
148 if(req->tag_ == MPI_ANY_TAG)
149 req->real_tag_ = ref->tag_;
150 if(req->real_size_ < ref->real_size_)
152 if(ref->detached_==1)
153 req->detached_sender_=ref; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
154 XBT_DEBUG("match succeeded");
160 void Request::print_request(const char *message)
162 XBT_VERB("%s request %p [buf = %p, size = %zu, src = %d, dst = %d, tag = %d, flags = %x]",
163 message, this, buf_, size_, src_, dst_, tag_, flags_);
167 /* factories, to hide the internal flags from the caller */
168 MPI_Request Request::send_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
171 return new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
172 comm->group()->index(dst), tag, comm, PERSISTENT | SEND | PREPARED);
175 MPI_Request Request::ssend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
177 return new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
178 comm->group()->index(dst), tag, comm, PERSISTENT | SSEND | SEND | PREPARED);
181 MPI_Request Request::isend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
183 return new Request(buf==MPI_BOTTOM ? nullptr : buf , count, datatype, smpi_process()->index(),
184 comm->group()->index(dst), tag,comm, PERSISTENT | ISEND | SEND | PREPARED);
188 MPI_Request Request::rma_send_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
191 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
193 request = new Request(buf==MPI_BOTTOM ? nullptr : buf , count, datatype, src, dst, tag,
194 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED);
196 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src, dst, tag,
197 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED | ACCUMULATE);
203 MPI_Request Request::recv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
205 return new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype,
206 src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->index(src),
207 smpi_process()->index(), tag, comm, PERSISTENT | RECV | PREPARED);
210 MPI_Request Request::rma_recv_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
213 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
215 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src, dst, tag,
216 comm, RMA | NON_PERSISTENT | RECV | PREPARED);
218 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src, dst, tag,
219 comm, RMA | NON_PERSISTENT | RECV | PREPARED | ACCUMULATE);
225 MPI_Request Request::irecv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
227 return new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
228 comm->group()->index(src), smpi_process()->index(), tag,
229 comm, PERSISTENT | RECV | PREPARED);
232 MPI_Request Request::isend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
234 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
235 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
236 comm->group()->index(dst), tag, comm, NON_PERSISTENT | ISEND | SEND);
241 MPI_Request Request::issend(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 | SSEND | SEND);
251 MPI_Request Request::irecv(void *buf, int count, MPI_Datatype datatype, int src, 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, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
255 comm->group()->index(src), smpi_process()->index(), tag, comm,
256 NON_PERSISTENT | RECV);
261 void Request::recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status * status)
263 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
264 request = irecv(buf, count, datatype, src, tag, comm);
265 wait(&request,status);
269 void Request::send(void *buf, int count, MPI_Datatype datatype, int dst, 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, smpi_process()->index(),
273 comm->group()->index(dst), tag, comm, NON_PERSISTENT | SEND);
276 wait(&request, MPI_STATUS_IGNORE);
280 void Request::ssend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
282 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
283 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
284 comm->group()->index(dst), tag, comm, NON_PERSISTENT | SSEND | SEND);
287 wait(&request,MPI_STATUS_IGNORE);
291 void Request::sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,int dst, int sendtag,
292 void *recvbuf, int recvcount, MPI_Datatype recvtype, int src, int recvtag,
293 MPI_Comm comm, MPI_Status * status)
295 MPI_Request requests[2];
297 int myid=smpi_process()->index();
298 if ((comm->group()->index(dst) == myid) && (comm->group()->index(src) == myid)){
299 Datatype::copy(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype);
302 requests[0] = isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
303 requests[1] = irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
304 startall(2, requests);
305 waitall(2, requests, stats);
308 if(status != MPI_STATUS_IGNORE) {
309 // Copy receive status
314 void Request::start()
316 smx_mailbox_t mailbox;
318 xbt_assert(action_ == nullptr, "Cannot (re-)start unfinished communication");
323 if ((flags_ & RECV) != 0) {
324 this->print_request("New recv");
326 simgrid::smpi::Process* process = smpi_process_remote(dst_);
328 int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh");
330 xbt_mutex_t mut = process->mailboxes_mutex();
331 if (async_small_thresh != 0 || (flags_ & RMA) != 0)
332 xbt_mutex_acquire(mut);
334 if (async_small_thresh == 0 && (flags_ & RMA) == 0 ) {
335 mailbox = process->mailbox();
337 else if (((flags_ & RMA) != 0) || static_cast<int>(size_) < async_small_thresh) {
338 //We have to check both mailboxes (because SSEND messages are sent to the large mbox).
339 //begin with the more appropriate one : the small one.
340 mailbox = process->mailbox_small();
341 XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %p (in case of SSEND)?", mailbox);
342 smx_activity_t action = simcall_comm_iprobe(mailbox, 0, src_,tag_, &match_recv,
343 static_cast<void*>(this));
345 if (action == nullptr) {
346 mailbox = process->mailbox();
347 XBT_DEBUG("No, nothing in the small mailbox test the other one : %p", mailbox);
348 action = simcall_comm_iprobe(mailbox, 0, src_,tag_, &match_recv, static_cast<void*>(this));
349 if (action == nullptr) {
350 XBT_DEBUG("Still nothing, switch back to the small mailbox : %p", mailbox);
351 mailbox = process->mailbox_small();
354 XBT_DEBUG("yes there was something for us in the large mailbox");
357 mailbox = process->mailbox_small();
358 XBT_DEBUG("Is there a corresponding send already posted the small mailbox?");
359 smx_activity_t action = simcall_comm_iprobe(mailbox, 0, src_,tag_, &match_recv, static_cast<void*>(this));
361 if (action == nullptr) {
362 XBT_DEBUG("No, nothing in the permanent receive mailbox");
363 mailbox = process->mailbox();
365 XBT_DEBUG("yes there was something for us in the small mailbox");
369 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
371 action_ = simcall_comm_irecv(process->process(), mailbox, buf_, &real_size_, &match_recv,
372 ! process->replaying()? smpi_comm_copy_data_callback
373 : &smpi_comm_null_copy_buffer_callback, this, -1.0);
374 XBT_DEBUG("recv simcall posted");
376 if (async_small_thresh != 0 || (flags_ & RMA) != 0 )
377 xbt_mutex_release(mut);
378 } else { /* the RECV flag was not set, so this is a send */
379 simgrid::smpi::Process* process = smpi_process_remote(dst_);
381 if (TRACE_smpi_view_internals()) {
382 TRACE_smpi_send(rank, rank, dst_, tag_, size_);
384 this->print_request("New send");
387 if ((flags_ & SSEND) == 0 && ( (flags_ & RMA) != 0
388 || static_cast<int>(size_) < xbt_cfg_get_int("smpi/send-is-detached-thresh") ) ) {
389 void *oldbuf = nullptr;
391 XBT_DEBUG("Send request %p is detached", this);
393 if(!(old_type_->flags() & DT_FLAG_DERIVED)){
395 if (!process->replaying() && oldbuf != nullptr && size_!=0){
396 if((smpi_privatize_global_variables != 0)
397 && (static_cast<char*>(buf_) >= smpi_start_data_exe)
398 && (static_cast<char*>(buf_) < smpi_start_data_exe + smpi_size_data_exe )){
399 XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
400 smpi_switch_data_segment(src_);
402 buf = xbt_malloc(size_);
403 memcpy(buf,oldbuf,size_);
404 XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
409 //if we are giving back the control to the user without waiting for completion, we have to inject timings
410 double sleeptime = 0.0;
411 if (detached_ != 0 || ((flags_ & (ISEND | SSEND)) != 0)) { // issend should be treated as isend
412 // isend and send timings may be different
413 sleeptime = ((flags_ & ISEND) != 0)
414 ? simgrid::s4u::Actor::self()->host()->extension<simgrid::smpi::SmpiHost>()->oisend(size_)
415 : simgrid::s4u::Actor::self()->host()->extension<simgrid::smpi::SmpiHost>()->osend(size_);
419 simcall_process_sleep(sleeptime);
420 XBT_DEBUG("sending size of %zu : sleep %f ", size_, sleeptime);
423 int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh");
425 xbt_mutex_t mut=process->mailboxes_mutex();
427 if (async_small_thresh != 0 || (flags_ & RMA) != 0)
428 xbt_mutex_acquire(mut);
430 if (!(async_small_thresh != 0 || (flags_ & RMA) !=0)) {
431 mailbox = process->mailbox();
432 } else if (((flags_ & RMA) != 0) || static_cast<int>(size_) < async_small_thresh) { // eager mode
433 mailbox = process->mailbox();
434 XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %p?", mailbox);
435 smx_activity_t action = simcall_comm_iprobe(mailbox, 1,dst_, tag_, &match_send,
436 static_cast<void*>(this));
437 if (action == nullptr) {
438 if ((flags_ & SSEND) == 0){
439 mailbox = process->mailbox_small();
440 XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %p", mailbox);
442 mailbox = process->mailbox_small();
443 XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %p?", mailbox);
444 action = simcall_comm_iprobe(mailbox, 1,dst_, tag_, &match_send, static_cast<void*>(this));
445 if (action == nullptr) {
446 XBT_DEBUG("No, we are first, send to large mailbox");
447 mailbox = process->mailbox();
451 XBT_DEBUG("Yes there was something for us in the large mailbox");
454 mailbox = process->mailbox();
455 XBT_DEBUG("Send request %p is in the large mailbox %p (buf: %p)",mailbox, this,buf_);
458 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
460 action_ = simcall_comm_isend(SIMIX_process_from_PID(src_+1), mailbox, size_, -1.0,
461 buf, real_size_, &match_send,
462 &xbt_free_f, // how to free the userdata if a detached send fails
463 !process->replaying() ? smpi_comm_copy_data_callback
464 : &smpi_comm_null_copy_buffer_callback, this,
465 // detach if msg size < eager/rdv switch limit
467 XBT_DEBUG("send simcall posted");
469 /* FIXME: detached sends are not traceable (action_ == nullptr) */
470 if (action_ != nullptr)
471 simcall_set_category(action_, TRACE_internal_smpi_get_category());
472 if (async_small_thresh != 0 || ((flags_ & RMA)!=0))
473 xbt_mutex_release(mut);
477 void Request::startall(int count, MPI_Request * requests)
479 if(requests== nullptr)
482 for(int i = 0; i < count; i++) {
483 requests[i]->start();
487 int Request::test(MPI_Request * request, MPI_Status * status) {
488 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or testall before)
489 // to avoid deadlocks if used as a break condition, such as
490 // while (MPI_Test(request, flag, status) && flag) dostuff...
491 // because the time will not normally advance when only calls to MPI_Test are made -> deadlock
492 // multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
493 static int nsleeps = 1;
494 if(smpi_test_sleep > 0)
495 simcall_process_sleep(nsleeps*smpi_test_sleep);
497 Status::empty(status);
499 if (((*request)->flags_ & PREPARED) == 0) {
500 if ((*request)->action_ != nullptr)
501 flag = simcall_comm_test((*request)->action_);
503 finish_wait(request,status);
504 nsleeps=1;//reset the number of sleeps we will do next time
505 if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & PERSISTENT)==0)
506 *request = MPI_REQUEST_NULL;
507 } else if (xbt_cfg_get_boolean("smpi/grow-injected-times")){
514 int Request::testsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
519 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
521 for (int i = 0; i < incount; i++) {
522 if (requests[i] != MPI_REQUEST_NULL) {
523 if (test(&requests[i], pstat)) {
526 if (status != MPI_STATUSES_IGNORE)
528 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->flags_ & NON_PERSISTENT)
529 requests[i] = MPI_REQUEST_NULL;
535 if(count_dead==incount)
536 return MPI_UNDEFINED;
540 int Request::testany(int count, MPI_Request requests[], int *index, MPI_Status * status)
542 std::vector<simgrid::kernel::activity::ActivityImpl*> comms;
543 comms.reserve(count);
548 *index = MPI_UNDEFINED;
550 std::vector<int> map; /** Maps all matching comms back to their location in requests **/
551 for(i = 0; i < count; i++) {
552 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action_ && !(requests[i]->flags_ & PREPARED)) {
553 comms.push_back(requests[i]->action_);
558 //multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it
559 static int nsleeps = 1;
560 if(smpi_test_sleep > 0)
561 simcall_process_sleep(nsleeps*smpi_test_sleep);
563 i = simcall_comm_testany(comms.data(), comms.size()); // The i-th element in comms matches!
564 if (i != -1) { // -1 is not MPI_UNDEFINED but a SIMIX return code. (nothing matches)
566 finish_wait(&requests[*index],status);
569 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & NON_PERSISTENT)) {
570 requests[*index] = MPI_REQUEST_NULL;
576 //all requests are null or inactive, return true
578 Status::empty(status);
584 int Request::testall(int count, MPI_Request requests[], MPI_Status status[])
587 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
589 for(int i=0; i<count; i++){
590 if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags_ & PREPARED)) {
591 if (test(&requests[i], pstat)!=1){
594 requests[i]=MPI_REQUEST_NULL;
597 Status::empty(pstat);
599 if(status != MPI_STATUSES_IGNORE) {
606 void Request::probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
608 //FIXME find another way to avoid busy waiting ?
609 // the issue here is that we have to wait on a nonexistent comm
611 iprobe(source, tag, comm, &flag, status);
612 XBT_DEBUG("Busy Waiting on probing : %d", flag);
616 void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
617 // to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
618 // especially when used as a break condition, such as while (MPI_Iprobe(...)) dostuff...
619 // nsleeps is a multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
620 // This can speed up the execution of certain applications by an order of magnitude, such as HPL
621 static int nsleeps = 1;
622 double speed = simgrid::s4u::Actor::self()->host()->speed();
623 double maxrate = xbt_cfg_get_double("smpi/iprobe-cpu-usage");
624 MPI_Request request = new Request(nullptr, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
625 comm->group()->index(source), comm->rank(), tag, comm, PERSISTENT | RECV);
626 if (smpi_iprobe_sleep > 0) {
627 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);
628 simcall_execution_wait(iprobe_sleep);
630 // behave like a receive, but don't do it
631 smx_mailbox_t mailbox;
633 request->print_request("New iprobe");
634 // We have to test both mailboxes as we don't know if we will receive one one or another
635 if (xbt_cfg_get_int("smpi/async-small-thresh") > 0){
636 mailbox = smpi_process()->mailbox_small();
637 XBT_DEBUG("Trying to probe the perm recv mailbox");
638 request->action_ = simcall_comm_iprobe(mailbox, 0, request->src_, request->tag_, &match_recv,
639 static_cast<void*>(request));
642 if (request->action_ == nullptr){
643 mailbox = smpi_process()->mailbox();
644 XBT_DEBUG("trying to probe the other 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 simgrid::kernel::activity::Comm *sync_comm = static_cast<simgrid::kernel::activity::Comm*>(request->action_);
651 MPI_Request req = static_cast<MPI_Request>(sync_comm->src_data);
653 if(status != MPI_STATUS_IGNORE && (req->flags_ & PREPARED) == 0) {
654 status->MPI_SOURCE = comm->group()->rank(req->src_);
655 status->MPI_TAG = req->tag_;
656 status->MPI_ERROR = MPI_SUCCESS;
657 status->count = req->real_size_;
659 nsleeps = 1;//reset the number of sleeps we will do next time
663 if (xbt_cfg_get_boolean("smpi/grow-injected-times"))
669 void Request::finish_wait(MPI_Request* request, MPI_Status * status)
671 MPI_Request req = *request;
672 Status::empty(status);
674 if(!((req->detached_ != 0) && ((req->flags_ & SEND) != 0)) && ((req->flags_ & PREPARED) == 0)){
675 if(status != MPI_STATUS_IGNORE) {
676 int src = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_;
677 status->MPI_SOURCE = req->comm_->group()->rank(src);
678 status->MPI_TAG = req->tag_ == MPI_ANY_TAG ? req->real_tag_ : req->tag_;
679 status->MPI_ERROR = req->truncated_ != 0 ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
680 // this handles the case were size in receive differs from size in send
681 status->count = req->real_size_;
684 req->print_request("Finishing");
685 MPI_Datatype datatype = req->old_type_;
687 // FIXME Handle the case of a partial shared malloc.
688 if (((req->flags_ & ACCUMULATE) != 0) ||
689 (datatype->flags() & DT_FLAG_DERIVED)) { // && (!smpi_is_shared(req->old_buf_))){
691 if (!smpi_process()->replaying()){
692 if( smpi_privatize_global_variables != 0 && (static_cast<char*>(req->old_buf_) >= smpi_start_data_exe)
693 && ((char*)req->old_buf_ < smpi_start_data_exe + smpi_size_data_exe )){
694 XBT_VERB("Privatization : We are unserializing to a zone in global memory Switch data segment ");
695 smpi_switch_data_segment(smpi_process()->index());
699 if(datatype->flags() & DT_FLAG_DERIVED){
700 // This part handles the problem of non-contignous memory the unserialization at the reception
701 if((req->flags_ & RECV) && datatype->size()!=0)
702 datatype->unserialize(req->buf_, req->old_buf_, req->real_size_/datatype->size() , req->op_);
704 }else if(req->flags_ & RECV){//apply op on contiguous buffer for accumulate
705 if(datatype->size()!=0){
706 int n =req->real_size_/datatype->size();
707 req->op_->apply(req->buf_, req->old_buf_, &n, datatype);
714 if (TRACE_smpi_view_internals() && ((req->flags_ & RECV) != 0)){
715 int rank = smpi_process()->index();
716 int src_traced = (req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_);
717 TRACE_smpi_recv(rank, src_traced, rank,req->tag_);
719 if(req->detached_sender_ != nullptr){
720 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
721 double sleeptime = simgrid::s4u::Actor::self()->host()->extension<simgrid::smpi::SmpiHost>()->orecv(req->real_size());
723 simcall_process_sleep(sleeptime);
724 XBT_DEBUG("receiving size of %zu : sleep %f ", req->real_size_, sleeptime);
726 unref(&(req->detached_sender_));
728 if(req->flags_ & PERSISTENT)
729 req->action_ = nullptr;
730 req->flags_ |= FINISHED;
734 void Request::wait(MPI_Request * request, MPI_Status * status)
736 (*request)->print_request("Waiting");
737 if ((*request)->flags_ & PREPARED) {
738 Status::empty(status);
742 if ((*request)->action_ != nullptr)
743 // this is not a detached send
744 simcall_comm_wait((*request)->action_, -1.0);
746 finish_wait(request,status);
747 if (*request != MPI_REQUEST_NULL && (((*request)->flags_ & NON_PERSISTENT)!=0))
748 *request = MPI_REQUEST_NULL;
751 int Request::waitany(int count, MPI_Request requests[], MPI_Status * status)
753 s_xbt_dynar_t comms; // Keep it on stack to save some extra mallocs
756 int index = MPI_UNDEFINED;
760 // Wait for a request to complete
761 xbt_dynar_init(&comms, sizeof(smx_activity_t), nullptr);
762 map = xbt_new(int, count);
763 XBT_DEBUG("Wait for one of %d", count);
764 for(i = 0; i < count; i++) {
765 if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags_ & PREPARED) && !(requests[i]->flags_ & FINISHED)) {
766 if (requests[i]->action_ != nullptr) {
767 XBT_DEBUG("Waiting any %p ", requests[i]);
768 xbt_dynar_push(&comms, &requests[i]->action_);
772 // This is a finished detached request, let's return this one
773 size = 0; // so we free the dynar but don't do the waitany call
775 finish_wait(&requests[i], status); // cleanup if refcount = 0
776 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & NON_PERSISTENT))
777 requests[i] = MPI_REQUEST_NULL; // set to null
783 i = simcall_comm_waitany(&comms, -1);
785 // not MPI_UNDEFINED, as this is a simix return code
788 //in case of an accumulate, we have to wait the end of all requests to apply the operation, ordered correctly.
789 if ((requests[index] == MPI_REQUEST_NULL)
790 || (!((requests[index]->flags_ & ACCUMULATE) && (requests[index]->flags_ & RECV)))){
791 finish_wait(&requests[index],status);
792 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & NON_PERSISTENT))
793 requests[index] = MPI_REQUEST_NULL;
798 xbt_dynar_free_data(&comms);
802 if (index==MPI_UNDEFINED)
803 Status::empty(status);
808 static int sort_accumulates(MPI_Request a, MPI_Request b)
810 return (a->tag() > b->tag());
813 int Request::waitall(int count, MPI_Request requests[], MPI_Status status[])
815 std::vector<MPI_Request> accumulates;
818 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
819 int retvalue = MPI_SUCCESS;
820 //tag invalid requests in the set
821 if (status != MPI_STATUSES_IGNORE) {
822 for (int c = 0; c < count; c++) {
823 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst_ == MPI_PROC_NULL || (requests[c]->flags_ & PREPARED)) {
824 Status::empty(&status[c]);
825 } else if (requests[c]->src_ == MPI_PROC_NULL) {
826 Status::empty(&status[c]);
827 status[c].MPI_SOURCE = MPI_PROC_NULL;
831 for (int c = 0; c < count; c++) {
832 if (MC_is_active() || MC_record_replay_is_active()) {
833 wait(&requests[c],pstat);
836 index = waitany(count, requests, pstat);
837 if (index == MPI_UNDEFINED)
840 if (requests[index] != MPI_REQUEST_NULL
841 && (requests[index]->flags_ & RECV)
842 && (requests[index]->flags_ & ACCUMULATE))
843 accumulates.push_back(requests[index]);
844 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & NON_PERSISTENT))
845 requests[index] = MPI_REQUEST_NULL;
847 if (status != MPI_STATUSES_IGNORE) {
848 status[index] = *pstat;
849 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
850 retvalue = MPI_ERR_IN_STATUS;
854 if (!accumulates.empty()) {
855 std::sort(accumulates.begin(), accumulates.end(), sort_accumulates);
856 for (auto req : accumulates) {
857 finish_wait(&req, status);
864 int Request::waitsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
870 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
872 for(i = 0; i < incount; i++)
874 index=waitany(incount, requests, pstat);
875 if(index!=MPI_UNDEFINED){
876 indices[count] = index;
878 if(status != MPI_STATUSES_IGNORE) {
879 status[index] = *pstat;
881 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & NON_PERSISTENT))
882 requests[index]=MPI_REQUEST_NULL;
884 return MPI_UNDEFINED;
890 MPI_Request Request::f2c(int id) {
892 if(id==MPI_FORTRAN_REQUEST_NULL)
893 return static_cast<MPI_Request>(MPI_REQUEST_NULL);
894 return static_cast<MPI_Request>(xbt_dict_get(F2C::f2c_lookup(), get_key_id(key, id)));
897 int Request::add_f() {
898 if(F2C::f2c_lookup()==nullptr){
899 F2C::set_f2c_lookup(xbt_dict_new_homogeneous(nullptr));
902 xbt_dict_set(F2C::f2c_lookup(), get_key_id(key, F2C::f2c_id()), this, nullptr);
903 F2C::f2c_id_increment();
904 return F2C::f2c_id()-1;
907 void Request::free_f(int id) {
909 if(id!=MPI_FORTRAN_REQUEST_NULL)
910 xbt_dict_remove(F2C::f2c_lookup(), get_key_id(key, id));