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, smx_activity_t ignored) {
121 MPI_Request ref = static_cast<MPI_Request>(a);
122 MPI_Request req = static_cast<MPI_Request>(b);
123 XBT_DEBUG("Trying to match a recv of src %d against %d, tag %d against %d",ref->src_,req->src_, ref->tag_, req->tag_);
125 xbt_assert(ref, "Cannot match recv against null reference");
126 xbt_assert(req, "Cannot match recv against null request");
127 if((ref->src_ == MPI_ANY_SOURCE || req->src_ == ref->src_)
128 && ((ref->tag_ == MPI_ANY_TAG && req->tag_ >=0) || req->tag_ == ref->tag_)){
129 //we match, we can transfer some values
130 if(ref->src_ == MPI_ANY_SOURCE)
131 ref->real_src_ = req->src_;
132 if(ref->tag_ == MPI_ANY_TAG)
133 ref->real_tag_ = req->tag_;
134 if(ref->real_size_ < req->real_size_)
136 if(req->detached_==1)
137 ref->detached_sender_=req; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
138 XBT_DEBUG("match succeeded");
143 int Request::match_send(void* a, void* b,smx_activity_t ignored) {
144 MPI_Request ref = static_cast<MPI_Request>(a);
145 MPI_Request req = static_cast<MPI_Request>(b);
146 XBT_DEBUG("Trying to match a send of src %d against %d, tag %d against %d",ref->src_,req->src_, ref->tag_, req->tag_);
147 xbt_assert(ref, "Cannot match send against null reference");
148 xbt_assert(req, "Cannot match send against null request");
150 if((req->src_ == MPI_ANY_SOURCE || req->src_ == ref->src_)
151 && ((req->tag_ == MPI_ANY_TAG && ref->tag_ >=0)|| req->tag_ == ref->tag_)){
152 if(req->src_ == MPI_ANY_SOURCE)
153 req->real_src_ = ref->src_;
154 if(req->tag_ == MPI_ANY_TAG)
155 req->real_tag_ = ref->tag_;
156 if(req->real_size_ < ref->real_size_)
158 if(ref->detached_==1)
159 req->detached_sender_=ref; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
160 XBT_DEBUG("match succeeded");
166 void Request::print_request(const char *message)
168 XBT_VERB("%s request %p [buf = %p, size = %zu, src = %d, dst = %d, tag = %d, flags = %x]",
169 message, this, buf_, size_, src_, dst_, tag_, flags_);
173 /* factories, to hide the internal flags from the caller */
174 MPI_Request Request::send_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 | SEND | PREPARED);
181 MPI_Request Request::ssend_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 | SSEND | SEND | PREPARED);
187 MPI_Request Request::isend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
189 return new Request(buf==MPI_BOTTOM ? nullptr : buf , count, datatype, smpi_process()->index(),
190 comm->group()->index(dst), tag,comm, PERSISTENT | ISEND | SEND | PREPARED);
194 MPI_Request Request::rma_send_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
197 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
199 request = new Request(buf==MPI_BOTTOM ? nullptr : buf , count, datatype, src, dst, tag,
200 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED);
202 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src, dst, tag,
203 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED | ACCUMULATE);
209 MPI_Request Request::recv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
211 return new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype,
212 src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->index(src),
213 smpi_process()->index(), tag, comm, PERSISTENT | RECV | PREPARED);
216 MPI_Request Request::rma_recv_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
219 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
221 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src, dst, tag,
222 comm, RMA | NON_PERSISTENT | RECV | PREPARED);
224 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src, dst, tag,
225 comm, RMA | NON_PERSISTENT | RECV | PREPARED | ACCUMULATE);
231 MPI_Request Request::irecv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
233 return new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
234 comm->group()->index(src), smpi_process()->index(), tag,
235 comm, PERSISTENT | RECV | PREPARED);
238 MPI_Request Request::isend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
240 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
241 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
242 comm->group()->index(dst), tag, comm, NON_PERSISTENT | ISEND | SEND);
247 MPI_Request Request::issend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
249 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
250 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
251 comm->group()->index(dst), tag,comm, NON_PERSISTENT | ISEND | SSEND | SEND);
257 MPI_Request Request::irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
259 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
260 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
261 comm->group()->index(src), smpi_process()->index(), tag, comm,
262 NON_PERSISTENT | RECV);
267 void Request::recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status * status)
269 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
270 request = irecv(buf, count, datatype, src, tag, comm);
271 wait(&request,status);
275 void Request::send(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
277 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
278 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
279 comm->group()->index(dst), tag, comm, NON_PERSISTENT | SEND);
282 wait(&request, MPI_STATUS_IGNORE);
286 void Request::ssend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
288 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
289 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
290 comm->group()->index(dst), tag, comm, NON_PERSISTENT | SSEND | SEND);
293 wait(&request,MPI_STATUS_IGNORE);
297 void Request::sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,int dst, int sendtag,
298 void *recvbuf, int recvcount, MPI_Datatype recvtype, int src, int recvtag,
299 MPI_Comm comm, MPI_Status * status)
301 MPI_Request requests[2];
303 int myid=smpi_process()->index();
304 if ((comm->group()->index(dst) == myid) && (comm->group()->index(src) == myid)){
305 Datatype::copy(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype);
308 requests[0] = isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
309 requests[1] = irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
310 startall(2, requests);
311 waitall(2, requests, stats);
314 if(status != MPI_STATUS_IGNORE) {
315 // Copy receive status
320 void Request::start()
322 smx_mailbox_t mailbox;
324 xbt_assert(action_ == nullptr, "Cannot (re-)start unfinished communication");
329 if ((flags_ & RECV) != 0) {
330 this->print_request("New recv");
332 simgrid::smpi::Process* process = smpi_process_remote(dst_);
334 int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh");
336 xbt_mutex_t mut = process->mailboxes_mutex();
337 if (async_small_thresh != 0 || (flags_ & RMA) != 0)
338 xbt_mutex_acquire(mut);
340 if (async_small_thresh == 0 && (flags_ & RMA) == 0 ) {
341 mailbox = process->mailbox();
343 else if (((flags_ & RMA) != 0) || static_cast<int>(size_) < async_small_thresh) {
344 //We have to check both mailboxes (because SSEND messages are sent to the large mbox).
345 //begin with the more appropriate one : the small one.
346 mailbox = process->mailbox_small();
347 XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %p (in case of SSEND)?", mailbox);
348 smx_activity_t action = simcall_comm_iprobe(mailbox, 0, src_,tag_, &match_recv,
349 static_cast<void*>(this));
351 if (action == nullptr) {
352 mailbox = process->mailbox();
353 XBT_DEBUG("No, nothing in the small mailbox test the other one : %p", mailbox);
354 action = simcall_comm_iprobe(mailbox, 0, src_,tag_, &match_recv, static_cast<void*>(this));
355 if (action == nullptr) {
356 XBT_DEBUG("Still nothing, switch back to the small mailbox : %p", mailbox);
357 mailbox = process->mailbox_small();
360 XBT_DEBUG("yes there was something for us in the large mailbox");
363 mailbox = process->mailbox_small();
364 XBT_DEBUG("Is there a corresponding send already posted the small mailbox?");
365 smx_activity_t action = simcall_comm_iprobe(mailbox, 0, src_,tag_, &match_recv, static_cast<void*>(this));
367 if (action == nullptr) {
368 XBT_DEBUG("No, nothing in the permanent receive mailbox");
369 mailbox = process->mailbox();
371 XBT_DEBUG("yes there was something for us in the small mailbox");
375 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
377 action_ = simcall_comm_irecv(
378 process->process(), mailbox, buf_, &real_size_, &match_recv,
379 process->replaying() ? &smpi_comm_null_copy_buffer_callback : smpi_comm_copy_data_callback, this, -1.0);
380 XBT_DEBUG("recv simcall posted");
382 if (async_small_thresh != 0 || (flags_ & RMA) != 0 )
383 xbt_mutex_release(mut);
384 } else { /* the RECV flag was not set, so this is a send */
385 simgrid::smpi::Process* process = smpi_process_remote(dst_);
387 if (TRACE_smpi_view_internals()) {
388 TRACE_smpi_send(rank, rank, dst_, tag_, size_);
390 this->print_request("New send");
393 if ((flags_ & SSEND) == 0 && ( (flags_ & RMA) != 0
394 || static_cast<int>(size_) < xbt_cfg_get_int("smpi/send-is-detached-thresh") ) ) {
395 void *oldbuf = nullptr;
397 XBT_DEBUG("Send request %p is detached", this);
399 if (not(old_type_->flags() & DT_FLAG_DERIVED)) {
401 if (not process->replaying() && oldbuf != nullptr && size_ != 0) {
402 if((smpi_privatize_global_variables != 0)
403 && (static_cast<char*>(buf_) >= smpi_start_data_exe)
404 && (static_cast<char*>(buf_) < smpi_start_data_exe + smpi_size_data_exe )){
405 XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
406 smpi_switch_data_segment(src_);
408 buf = xbt_malloc(size_);
409 memcpy(buf,oldbuf,size_);
410 XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
415 //if we are giving back the control to the user without waiting for completion, we have to inject timings
416 double sleeptime = 0.0;
417 if (detached_ != 0 || ((flags_ & (ISEND | SSEND)) != 0)) { // issend should be treated as isend
418 // isend and send timings may be different
419 sleeptime = ((flags_ & ISEND) != 0)
420 ? simgrid::s4u::Actor::self()->host()->extension<simgrid::smpi::SmpiHost>()->oisend(size_)
421 : simgrid::s4u::Actor::self()->host()->extension<simgrid::smpi::SmpiHost>()->osend(size_);
425 simcall_process_sleep(sleeptime);
426 XBT_DEBUG("sending size of %zu : sleep %f ", size_, sleeptime);
429 int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh");
431 xbt_mutex_t mut=process->mailboxes_mutex();
433 if (async_small_thresh != 0 || (flags_ & RMA) != 0)
434 xbt_mutex_acquire(mut);
436 if (not(async_small_thresh != 0 || (flags_ & RMA) != 0)) {
437 mailbox = process->mailbox();
438 } else if (((flags_ & RMA) != 0) || static_cast<int>(size_) < async_small_thresh) { // eager mode
439 mailbox = process->mailbox();
440 XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %p?", mailbox);
441 smx_activity_t action = simcall_comm_iprobe(mailbox, 1,dst_, tag_, &match_send,
442 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,dst_, tag_, &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()->host()->speed();
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, request->src_, request->tag_, &match_recv,
644 static_cast<void*>(request));
647 if (request->action_ == nullptr){
648 mailbox = smpi_process()->mailbox();
649 XBT_DEBUG("trying to probe the other mailbox");
650 request->action_ = simcall_comm_iprobe(mailbox, 0, request->src_,request->tag_, &match_recv,
651 static_cast<void*>(request));
654 if (request->action_ != nullptr){
655 simgrid::kernel::activity::CommImplPtr sync_comm =
656 boost::static_pointer_cast<simgrid::kernel::activity::CommImpl>(request->action_);
657 MPI_Request req = static_cast<MPI_Request>(sync_comm->src_data);
659 if(status != MPI_STATUS_IGNORE && (req->flags_ & PREPARED) == 0) {
660 status->MPI_SOURCE = comm->group()->rank(req->src_);
661 status->MPI_TAG = req->tag_;
662 status->MPI_ERROR = MPI_SUCCESS;
663 status->count = req->real_size_;
665 nsleeps = 1;//reset the number of sleeps we will do next time
669 if (xbt_cfg_get_boolean("smpi/grow-injected-times"))
675 void Request::finish_wait(MPI_Request* request, MPI_Status * status)
677 MPI_Request req = *request;
678 Status::empty(status);
680 if (not((req->detached_ != 0) && ((req->flags_ & SEND) != 0)) && ((req->flags_ & PREPARED) == 0)) {
681 if(status != MPI_STATUS_IGNORE) {
682 int src = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_;
683 status->MPI_SOURCE = req->comm_->group()->rank(src);
684 status->MPI_TAG = req->tag_ == MPI_ANY_TAG ? req->real_tag_ : req->tag_;
685 status->MPI_ERROR = req->truncated_ != 0 ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
686 // this handles the case were size in receive differs from size in send
687 status->count = req->real_size_;
690 req->print_request("Finishing");
691 MPI_Datatype datatype = req->old_type_;
693 // FIXME Handle the case of a partial shared malloc.
694 if (((req->flags_ & ACCUMULATE) != 0) ||
695 (datatype->flags() & DT_FLAG_DERIVED)) { // && (not smpi_is_shared(req->old_buf_))){
697 if (not smpi_process()->replaying()) {
698 if( smpi_privatize_global_variables != 0 && (static_cast<char*>(req->old_buf_) >= smpi_start_data_exe)
699 && ((char*)req->old_buf_ < smpi_start_data_exe + smpi_size_data_exe )){
700 XBT_VERB("Privatization : We are unserializing to a zone in global memory Switch data segment ");
701 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(rank, 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
727 double sleeptime = simgrid::s4u::Actor::self()->host()->extension<simgrid::smpi::SmpiHost>()->orecv(req->real_size());
729 simcall_process_sleep(sleeptime);
730 XBT_DEBUG("receiving size of %zu : sleep %f ", req->real_size_, sleeptime);
732 unref(&(req->detached_sender_));
734 if(req->flags_ & PERSISTENT)
735 req->action_ = nullptr;
736 req->flags_ |= FINISHED;
740 void Request::wait(MPI_Request * request, MPI_Status * status)
742 (*request)->print_request("Waiting");
743 if ((*request)->flags_ & PREPARED) {
744 Status::empty(status);
748 if ((*request)->action_ != nullptr)
749 // this is not a detached send
750 simcall_comm_wait((*request)->action_, -1.0);
752 finish_wait(request,status);
753 if (*request != MPI_REQUEST_NULL && (((*request)->flags_ & NON_PERSISTENT)!=0))
754 *request = MPI_REQUEST_NULL;
757 int Request::waitany(int count, MPI_Request requests[], MPI_Status * status)
759 s_xbt_dynar_t comms; // Keep it on stack to save some extra mallocs
760 int index = MPI_UNDEFINED;
764 // Wait for a request to complete
765 xbt_dynar_init(&comms, sizeof(smx_activity_t), [](void*ptr){
766 intrusive_ptr_release(*(simgrid::kernel::activity::ActivityImpl**)ptr);
768 int *map = xbt_new(int, count);
769 XBT_DEBUG("Wait for one of %d", count);
770 for(int i = 0; i < count; i++) {
771 if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & PREPARED) &&
772 not(requests[i]->flags_ & FINISHED)) {
773 if (requests[i]->action_ != nullptr) {
774 XBT_DEBUG("Waiting any %p ", requests[i]);
775 intrusive_ptr_add_ref(requests[i]->action_.get());
776 xbt_dynar_push_as(&comms, simgrid::kernel::activity::ActivityImpl*, requests[i]->action_.get());
780 // This is a finished detached request, let's return this one
781 size = 0; // so we free the dynar but don't do the waitany call
783 finish_wait(&requests[i], status); // cleanup if refcount = 0
784 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & NON_PERSISTENT))
785 requests[i] = MPI_REQUEST_NULL; // set to null
791 XBT_DEBUG("Enter waitany for %lu comms", xbt_dynar_length(&comms));
792 int i = simcall_comm_waitany(&comms, -1);
794 // not MPI_UNDEFINED, as this is a simix return code
797 //in case of an accumulate, we have to wait the end of all requests to apply the operation, ordered correctly.
798 if ((requests[index] == MPI_REQUEST_NULL) ||
799 (not((requests[index]->flags_ & ACCUMULATE) && (requests[index]->flags_ & RECV)))) {
800 finish_wait(&requests[index],status);
801 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & NON_PERSISTENT))
802 requests[index] = MPI_REQUEST_NULL;
807 xbt_dynar_free_data(&comms);
811 if (index==MPI_UNDEFINED)
812 Status::empty(status);
817 static int sort_accumulates(MPI_Request a, MPI_Request b)
819 return (a->tag() > b->tag());
822 int Request::waitall(int count, MPI_Request requests[], MPI_Status status[])
824 std::vector<MPI_Request> accumulates;
827 MPI_Status *pstat = (status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat);
828 int retvalue = MPI_SUCCESS;
829 //tag invalid requests in the set
830 if (status != MPI_STATUSES_IGNORE) {
831 for (int c = 0; c < count; c++) {
832 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst_ == MPI_PROC_NULL || (requests[c]->flags_ & PREPARED)) {
833 Status::empty(&status[c]);
834 } else if (requests[c]->src_ == MPI_PROC_NULL) {
835 Status::empty(&status[c]);
836 status[c].MPI_SOURCE = MPI_PROC_NULL;
840 for (int c = 0; c < count; c++) {
841 if (MC_is_active() || MC_record_replay_is_active()) {
842 wait(&requests[c],pstat);
845 index = waitany(count, (MPI_Request*)requests, pstat);
846 if (index == MPI_UNDEFINED)
849 if (requests[index] != MPI_REQUEST_NULL
850 && (requests[index]->flags_ & RECV)
851 && (requests[index]->flags_ & ACCUMULATE))
852 accumulates.push_back(requests[index]);
853 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & NON_PERSISTENT))
854 requests[index] = MPI_REQUEST_NULL;
856 if (status != MPI_STATUSES_IGNORE) {
857 status[index] = *pstat;
858 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
859 retvalue = MPI_ERR_IN_STATUS;
863 if (not accumulates.empty()) {
864 std::sort(accumulates.begin(), accumulates.end(), sort_accumulates);
865 for (auto req : accumulates) {
866 finish_wait(&req, status);
873 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(i = 0; i < incount; i++)
883 index=waitany(incount, requests, pstat);
884 if(index!=MPI_UNDEFINED){
885 indices[count] = index;
887 if(status != MPI_STATUSES_IGNORE) {
888 status[index] = *pstat;
890 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & NON_PERSISTENT))
891 requests[index]=MPI_REQUEST_NULL;
893 return MPI_UNDEFINED;
899 MPI_Request Request::f2c(int id) {
901 if(id==MPI_FORTRAN_REQUEST_NULL)
902 return static_cast<MPI_Request>(MPI_REQUEST_NULL);
903 return static_cast<MPI_Request>(xbt_dict_get(F2C::f2c_lookup(), get_key_id(key, id)));
906 int Request::add_f() {
907 if(F2C::f2c_lookup()==nullptr){
908 F2C::set_f2c_lookup(xbt_dict_new_homogeneous(nullptr));
911 xbt_dict_set(F2C::f2c_lookup(), get_key_id(key, F2C::f2c_id()), this, nullptr);
912 F2C::f2c_id_increment();
913 return F2C::f2c_id()-1;
916 void Request::free_f(int id) {
917 if (id != MPI_FORTRAN_REQUEST_NULL) {
919 xbt_dict_remove(F2C::f2c_lookup(), get_key_id(key, id));