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. */
7 #include "src/mc/mc_replay.h"
8 #include "src/smpi/SmpiHost.hpp"
9 #include "src/kernel/activity/SynchroComm.hpp"
10 #include "src/smpi/private.h"
11 #include "src/smpi/smpi_comm.hpp"
12 #include "src/smpi/smpi_datatype.hpp"
13 #include "src/smpi/smpi_op.hpp"
14 #include "src/smpi/smpi_process.hpp"
15 #include "src/smpi/smpi_request.hpp"
19 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_request, smpi, "Logging specific to SMPI (reques)");
21 static simgrid::config::Flag<double> smpi_iprobe_sleep(
22 "smpi/iprobe", "Minimum time to inject inside a call to MPI_Iprobe", 1e-4);
23 static simgrid::config::Flag<double> smpi_test_sleep(
24 "smpi/test", "Minimum time to inject inside a call to MPI_Test", 1e-4);
26 std::vector<s_smpi_factor_t> smpi_ois_values;
28 extern void (*smpi_comm_copy_data_callback) (smx_activity_t, void*, size_t);
33 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)
35 void *old_buf = nullptr;
36 // FIXME Handle the case of a partial shared malloc.
37 if ((((flags & RECV) != 0) && ((flags & ACCUMULATE) != 0)) || (datatype->flags() & DT_FLAG_DERIVED)) {
38 // This part handles the problem of non-contiguous memory
43 buf_ = xbt_malloc(count*datatype->size());
44 if ((datatype->flags() & DT_FLAG_DERIVED) && ((flags & SEND) != 0)) {
45 datatype->serialize(old_buf, buf_, count);
49 // This part handles the problem of non-contiguous memory (for the unserialisation at the reception)
51 size_ = datatype->size() * count;
56 detached_sender_ = nullptr;
61 if (flags & PERSISTENT)
68 MPI_Comm Request::comm(){
88 int Request::detached(){
92 size_t Request::size(){
96 size_t Request::real_size(){
100 void Request::unref(MPI_Request* request)
102 if((*request) != MPI_REQUEST_NULL){
103 (*request)->refcount_--;
104 if((*request)->refcount_<0) xbt_die("wrong refcount");
105 if((*request)->refcount_==0){
106 Datatype::unref((*request)->old_type_);
107 Comm::unref((*request)->comm_);
108 (*request)->print_request("Destroying");
110 *request = MPI_REQUEST_NULL;
112 (*request)->print_request("Decrementing");
115 xbt_die("freeing an already free request");
119 int Request::match_recv(void* a, void* b, smx_activity_t ignored) {
120 MPI_Request ref = static_cast<MPI_Request>(a);
121 MPI_Request req = static_cast<MPI_Request>(b);
122 XBT_DEBUG("Trying to match a recv of src %d against %d, tag %d against %d",ref->src_,req->src_, ref->tag_, req->tag_);
124 xbt_assert(ref, "Cannot match recv against null reference");
125 xbt_assert(req, "Cannot match recv against null request");
126 if((ref->src_ == MPI_ANY_SOURCE || req->src_ == ref->src_)
127 && ((ref->tag_ == MPI_ANY_TAG && req->tag_ >=0) || req->tag_ == ref->tag_)){
128 //we match, we can transfer some values
129 if(ref->src_ == MPI_ANY_SOURCE)
130 ref->real_src_ = req->src_;
131 if(ref->tag_ == MPI_ANY_TAG)
132 ref->real_tag_ = req->tag_;
133 if(ref->real_size_ < req->real_size_)
135 if(req->detached_==1)
136 ref->detached_sender_=req; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
137 XBT_DEBUG("match succeeded");
142 int Request::match_send(void* a, void* b,smx_activity_t ignored) {
143 MPI_Request ref = static_cast<MPI_Request>(a);
144 MPI_Request req = static_cast<MPI_Request>(b);
145 XBT_DEBUG("Trying to match a send of src %d against %d, tag %d against %d",ref->src_,req->src_, ref->tag_, req->tag_);
146 xbt_assert(ref, "Cannot match send against null reference");
147 xbt_assert(req, "Cannot match send against null request");
149 if((req->src_ == MPI_ANY_SOURCE || req->src_ == ref->src_)
150 && ((req->tag_ == MPI_ANY_TAG && ref->tag_ >=0)|| req->tag_ == ref->tag_)){
151 if(req->src_ == MPI_ANY_SOURCE)
152 req->real_src_ = ref->src_;
153 if(req->tag_ == MPI_ANY_TAG)
154 req->real_tag_ = ref->tag_;
155 if(req->real_size_ < ref->real_size_)
157 if(ref->detached_==1)
158 req->detached_sender_=ref; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
159 XBT_DEBUG("match succeeded");
165 void Request::print_request(const char *message)
167 XBT_VERB("%s request %p [buf = %p, size = %zu, src = %d, dst = %d, tag = %d, flags = %x]",
168 message, this, buf_, size_, src_, dst_, tag_, flags_);
172 /* factories, to hide the internal flags from the caller */
173 MPI_Request Request::send_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
176 return new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
177 comm->group()->index(dst), tag, comm, PERSISTENT | SEND | PREPARED);
180 MPI_Request Request::ssend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
182 return new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
183 comm->group()->index(dst), tag, comm, PERSISTENT | SSEND | SEND | PREPARED);
186 MPI_Request Request::isend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
188 return new Request(buf==MPI_BOTTOM ? nullptr : buf , count, datatype, smpi_process()->index(),
189 comm->group()->index(dst), tag,comm, PERSISTENT | ISEND | SEND | PREPARED);
193 MPI_Request Request::rma_send_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
196 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
198 request = new Request(buf==MPI_BOTTOM ? nullptr : buf , count, datatype, src, dst, tag,
199 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED);
201 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src, dst, tag,
202 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED | ACCUMULATE);
208 MPI_Request Request::recv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
210 return new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype,
211 src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->index(src),
212 smpi_process()->index(), tag, comm, PERSISTENT | RECV | PREPARED);
215 MPI_Request Request::rma_recv_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
218 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
220 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src, dst, tag,
221 comm, RMA | NON_PERSISTENT | RECV | PREPARED);
223 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src, dst, tag,
224 comm, RMA | NON_PERSISTENT | RECV | PREPARED | ACCUMULATE);
230 MPI_Request Request::irecv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
232 return new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
233 comm->group()->index(src), smpi_process()->index(), tag,
234 comm, PERSISTENT | RECV | PREPARED);
237 MPI_Request Request::isend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
239 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
240 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
241 comm->group()->index(dst), tag, comm, NON_PERSISTENT | ISEND | SEND);
246 MPI_Request Request::issend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
248 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
249 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
250 comm->group()->index(dst), tag,comm, NON_PERSISTENT | ISEND | SSEND | SEND);
256 MPI_Request Request::irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
258 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
259 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
260 comm->group()->index(src), smpi_process()->index(), tag, comm,
261 NON_PERSISTENT | RECV);
266 void Request::recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status * status)
268 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
269 request = irecv(buf, count, datatype, src, tag, comm);
270 wait(&request,status);
274 void Request::send(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
276 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
277 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
278 comm->group()->index(dst), tag, comm, NON_PERSISTENT | SEND);
281 wait(&request, MPI_STATUS_IGNORE);
285 void Request::ssend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
287 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
288 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
289 comm->group()->index(dst), tag, comm, NON_PERSISTENT | SSEND | SEND);
292 wait(&request,MPI_STATUS_IGNORE);
296 void Request::sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,int dst, int sendtag,
297 void *recvbuf, int recvcount, MPI_Datatype recvtype, int src, int recvtag,
298 MPI_Comm comm, MPI_Status * status)
300 MPI_Request requests[2];
302 int myid=smpi_process()->index();
303 if ((comm->group()->index(dst) == myid) && (comm->group()->index(src) == myid)){
304 Datatype::copy(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype);
307 requests[0] = isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
308 requests[1] = irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
309 startall(2, requests);
310 waitall(2, requests, stats);
313 if(status != MPI_STATUS_IGNORE) {
314 // Copy receive status
319 void Request::start()
321 smx_mailbox_t mailbox;
323 xbt_assert(action_ == nullptr, "Cannot (re-)start unfinished communication");
328 if ((flags_ & RECV) != 0) {
329 this->print_request("New recv");
331 simgrid::smpi::Process* process = smpi_process_remote(dst_);
333 int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh");
335 xbt_mutex_t mut = process->mailboxes_mutex();
336 if (async_small_thresh != 0 || (flags_ & RMA) != 0)
337 xbt_mutex_acquire(mut);
339 if (async_small_thresh == 0 && (flags_ & RMA) == 0 ) {
340 mailbox = process->mailbox();
342 else if (((flags_ & RMA) != 0) || static_cast<int>(size_) < async_small_thresh) {
343 //We have to check both mailboxes (because SSEND messages are sent to the large mbox).
344 //begin with the more appropriate one : the small one.
345 mailbox = process->mailbox_small();
346 XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %p (in case of SSEND)?", mailbox);
347 smx_activity_t action = simcall_comm_iprobe(mailbox, 0, src_,tag_, &match_recv,
348 static_cast<void*>(this));
350 if (action == nullptr) {
351 mailbox = process->mailbox();
352 XBT_DEBUG("No, nothing in the small mailbox test the other one : %p", mailbox);
353 action = simcall_comm_iprobe(mailbox, 0, src_,tag_, &match_recv, static_cast<void*>(this));
354 if (action == nullptr) {
355 XBT_DEBUG("Still nothing, switch back to the small mailbox : %p", mailbox);
356 mailbox = process->mailbox_small();
359 XBT_DEBUG("yes there was something for us in the large mailbox");
362 mailbox = process->mailbox_small();
363 XBT_DEBUG("Is there a corresponding send already posted the small mailbox?");
364 smx_activity_t action = simcall_comm_iprobe(mailbox, 0, src_,tag_, &match_recv, static_cast<void*>(this));
366 if (action == nullptr) {
367 XBT_DEBUG("No, nothing in the permanent receive mailbox");
368 mailbox = process->mailbox();
370 XBT_DEBUG("yes there was something for us in the small mailbox");
374 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
376 action_ = simcall_comm_irecv(process->process(), mailbox, buf_, &real_size_, &match_recv,
377 ! process->replaying()? smpi_comm_copy_data_callback
378 : &smpi_comm_null_copy_buffer_callback, this, -1.0);
379 XBT_DEBUG("recv simcall posted");
381 if (async_small_thresh != 0 || (flags_ & RMA) != 0 )
382 xbt_mutex_release(mut);
383 } else { /* the RECV flag was not set, so this is a send */
384 simgrid::smpi::Process* process = smpi_process_remote(dst_);
386 if (TRACE_smpi_view_internals()) {
387 TRACE_smpi_send(rank, rank, dst_, tag_, size_);
389 this->print_request("New send");
392 if ((flags_ & SSEND) == 0 && ( (flags_ & RMA) != 0
393 || static_cast<int>(size_) < xbt_cfg_get_int("smpi/send-is-detached-thresh") ) ) {
394 void *oldbuf = nullptr;
396 XBT_DEBUG("Send request %p is detached", this);
398 if(!(old_type_->flags() & DT_FLAG_DERIVED)){
400 if (!process->replaying() && oldbuf != nullptr && size_!=0){
401 if((smpi_privatize_global_variables != 0)
402 && (static_cast<char*>(buf_) >= smpi_start_data_exe)
403 && (static_cast<char*>(buf_) < smpi_start_data_exe + smpi_size_data_exe )){
404 XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
405 smpi_switch_data_segment(src_);
407 buf = xbt_malloc(size_);
408 memcpy(buf,oldbuf,size_);
409 XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
414 //if we are giving back the control to the user without waiting for completion, we have to inject timings
415 double sleeptime = 0.0;
416 if (detached_ != 0 || ((flags_ & (ISEND | SSEND)) != 0)) { // issend should be treated as isend
417 // isend and send timings may be different
418 sleeptime = ((flags_ & ISEND) != 0)
419 ? simgrid::s4u::Actor::self()->host()->extension<simgrid::smpi::SmpiHost>()->oisend(size_)
420 : simgrid::s4u::Actor::self()->host()->extension<simgrid::smpi::SmpiHost>()->osend(size_);
424 simcall_process_sleep(sleeptime);
425 XBT_DEBUG("sending size of %zu : sleep %f ", size_, sleeptime);
428 int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh");
430 xbt_mutex_t mut=process->mailboxes_mutex();
432 if (async_small_thresh != 0 || (flags_ & RMA) != 0)
433 xbt_mutex_acquire(mut);
435 if (!(async_small_thresh != 0 || (flags_ & RMA) !=0)) {
436 mailbox = process->mailbox();
437 } else if (((flags_ & RMA) != 0) || static_cast<int>(size_) < async_small_thresh) { // eager mode
438 mailbox = process->mailbox();
439 XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %p?", mailbox);
440 smx_activity_t action = simcall_comm_iprobe(mailbox, 1,dst_, tag_, &match_send,
441 static_cast<void*>(this));
442 if (action == nullptr) {
443 if ((flags_ & SSEND) == 0){
444 mailbox = process->mailbox_small();
445 XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %p", mailbox);
447 mailbox = process->mailbox_small();
448 XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %p?", mailbox);
449 action = simcall_comm_iprobe(mailbox, 1,dst_, tag_, &match_send, static_cast<void*>(this));
450 if (action == nullptr) {
451 XBT_DEBUG("No, we are first, send to large mailbox");
452 mailbox = process->mailbox();
456 XBT_DEBUG("Yes there was something for us in the large mailbox");
459 mailbox = process->mailbox();
460 XBT_DEBUG("Send request %p is in the large mailbox %p (buf: %p)",mailbox, this,buf_);
463 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
465 action_ = simcall_comm_isend(SIMIX_process_from_PID(src_+1), mailbox, size_, -1.0,
466 buf, real_size_, &match_send,
467 &xbt_free_f, // how to free the userdata if a detached send fails
468 !process->replaying() ? smpi_comm_copy_data_callback
469 : &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::ActivityImpl*> 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_ && !(requests[i]->flags_ & PREPARED)) {
558 comms.push_back(requests[i]->action_);
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 && !(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::Comm *sync_comm = static_cast<simgrid::kernel::activity::Comm*>(request->action_);
656 MPI_Request req = static_cast<MPI_Request>(sync_comm->src_data);
658 if(status != MPI_STATUS_IGNORE && (req->flags_ & PREPARED) == 0) {
659 status->MPI_SOURCE = comm->group()->rank(req->src_);
660 status->MPI_TAG = req->tag_;
661 status->MPI_ERROR = MPI_SUCCESS;
662 status->count = req->real_size_;
664 nsleeps = 1;//reset the number of sleeps we will do next time
668 if (xbt_cfg_get_boolean("smpi/grow-injected-times"))
674 void Request::finish_wait(MPI_Request* request, MPI_Status * status)
676 MPI_Request req = *request;
677 Status::empty(status);
679 if(!((req->detached_ != 0) && ((req->flags_ & SEND) != 0)) && ((req->flags_ & PREPARED) == 0)){
680 if(status != MPI_STATUS_IGNORE) {
681 int src = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_;
682 status->MPI_SOURCE = req->comm_->group()->rank(src);
683 status->MPI_TAG = req->tag_ == MPI_ANY_TAG ? req->real_tag_ : req->tag_;
684 status->MPI_ERROR = req->truncated_ != 0 ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
685 // this handles the case were size in receive differs from size in send
686 status->count = req->real_size_;
689 req->print_request("Finishing");
690 MPI_Datatype datatype = req->old_type_;
692 // FIXME Handle the case of a partial shared malloc.
693 if (((req->flags_ & ACCUMULATE) != 0) ||
694 (datatype->flags() & DT_FLAG_DERIVED)) { // && (!smpi_is_shared(req->old_buf_))){
696 if (!smpi_process()->replaying()){
697 if( smpi_privatize_global_variables != 0 && (static_cast<char*>(req->old_buf_) >= smpi_start_data_exe)
698 && ((char*)req->old_buf_ < smpi_start_data_exe + smpi_size_data_exe )){
699 XBT_VERB("Privatization : We are unserializing to a zone in global memory Switch data segment ");
700 smpi_switch_data_segment(smpi_process()->index());
704 if(datatype->flags() & DT_FLAG_DERIVED){
705 // This part handles the problem of non-contignous memory the unserialization at the reception
706 if((req->flags_ & RECV) && datatype->size()!=0)
707 datatype->unserialize(req->buf_, req->old_buf_, req->real_size_/datatype->size() , req->op_);
709 }else if(req->flags_ & RECV){//apply op on contiguous buffer for accumulate
710 if(datatype->size()!=0){
711 int n =req->real_size_/datatype->size();
712 req->op_->apply(req->buf_, req->old_buf_, &n, datatype);
719 if (TRACE_smpi_view_internals() && ((req->flags_ & RECV) != 0)){
720 int rank = smpi_process()->index();
721 int src_traced = (req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_);
722 TRACE_smpi_recv(rank, src_traced, rank,req->tag_);
724 if(req->detached_sender_ != nullptr){
725 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
726 double sleeptime = simgrid::s4u::Actor::self()->host()->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
761 int index = MPI_UNDEFINED;
765 // Wait for a request to complete
766 xbt_dynar_init(&comms, sizeof(smx_activity_t), nullptr);
767 map = xbt_new(int, count);
768 XBT_DEBUG("Wait for one of %d", count);
769 for(i = 0; i < count; i++) {
770 if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags_ & PREPARED) && !(requests[i]->flags_ & FINISHED)) {
771 if (requests[i]->action_ != nullptr) {
772 XBT_DEBUG("Waiting any %p ", requests[i]);
773 xbt_dynar_push(&comms, &requests[i]->action_);
777 // This is a finished detached request, let's return this one
778 size = 0; // so we free the dynar but don't do the waitany call
780 finish_wait(&requests[i], status); // cleanup if refcount = 0
781 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & NON_PERSISTENT))
782 requests[i] = MPI_REQUEST_NULL; // set to null
788 i = simcall_comm_waitany(&comms, -1);
790 // not MPI_UNDEFINED, as this is a simix return code
793 //in case of an accumulate, we have to wait the end of all requests to apply the operation, ordered correctly.
794 if ((requests[index] == MPI_REQUEST_NULL)
795 || (!((requests[index]->flags_ & ACCUMULATE) && (requests[index]->flags_ & RECV)))){
796 finish_wait(&requests[index],status);
797 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & NON_PERSISTENT))
798 requests[index] = MPI_REQUEST_NULL;
803 xbt_dynar_free_data(&comms);
807 if (index==MPI_UNDEFINED)
808 Status::empty(status);
813 static int sort_accumulates(MPI_Request a, MPI_Request b)
815 return (a->tag() > b->tag());
818 int Request::waitall(int count, MPI_Request requests[], MPI_Status status[])
820 std::vector<MPI_Request> accumulates;
823 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
824 int retvalue = MPI_SUCCESS;
825 //tag invalid requests in the set
826 if (status != MPI_STATUSES_IGNORE) {
827 for (int c = 0; c < count; c++) {
828 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst_ == MPI_PROC_NULL || (requests[c]->flags_ & PREPARED)) {
829 Status::empty(&status[c]);
830 } else if (requests[c]->src_ == MPI_PROC_NULL) {
831 Status::empty(&status[c]);
832 status[c].MPI_SOURCE = MPI_PROC_NULL;
836 for (int c = 0; c < count; c++) {
837 if (MC_is_active() || MC_record_replay_is_active()) {
838 wait(&requests[c],pstat);
841 index = waitany(count, requests, pstat);
842 if (index == MPI_UNDEFINED)
845 if (requests[index] != MPI_REQUEST_NULL
846 && (requests[index]->flags_ & RECV)
847 && (requests[index]->flags_ & ACCUMULATE))
848 accumulates.push_back(requests[index]);
849 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & NON_PERSISTENT))
850 requests[index] = MPI_REQUEST_NULL;
852 if (status != MPI_STATUSES_IGNORE) {
853 status[index] = *pstat;
854 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
855 retvalue = MPI_ERR_IN_STATUS;
859 if (!accumulates.empty()) {
860 std::sort(accumulates.begin(), accumulates.end(), sort_accumulates);
861 for (auto req : accumulates) {
862 finish_wait(&req, status);
869 int Request::waitsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
875 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
877 for(i = 0; i < incount; i++)
879 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>(xbt_dict_get(F2C::f2c_lookup(), get_key_id(key, id)));
902 int Request::add_f() {
903 if(F2C::f2c_lookup()==nullptr){
904 F2C::set_f2c_lookup(xbt_dict_new_homogeneous(nullptr));
907 xbt_dict_set(F2C::f2c_lookup(), get_key_id(key, F2C::f2c_id()), this, nullptr);
908 F2C::f2c_id_increment();
909 return F2C::f2c_id()-1;
912 void Request::free_f(int id) {
914 if(id!=MPI_FORTRAN_REQUEST_NULL)
915 xbt_dict_remove(F2C::f2c_lookup(), get_key_id(key, id));