1 /* Copyright (c) 2007-2018. The SimGrid Team. All rights reserved. */
3 /* This program is free software; you can redistribute it and/or modify it
4 * under the terms of the license (GNU LGPL) which comes with this package. */
6 #include "smpi_request.hpp"
8 #include "SmpiHost.hpp"
10 #include "private.hpp"
11 #include "smpi_comm.hpp"
12 #include "smpi_datatype.hpp"
13 #include "smpi_op.hpp"
14 #include "smpi_process.hpp"
15 #include "src/kernel/activity/CommImpl.hpp"
16 #include "src/mc/mc_replay.hpp"
17 #include "src/simix/ActorImpl.hpp"
18 #include "xbt/config.hpp"
23 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_request, smpi, "Logging specific to SMPI (reques)");
25 static simgrid::config::Flag<double> smpi_iprobe_sleep(
26 "smpi/iprobe", "Minimum time to inject inside a call to MPI_Iprobe", 1e-4);
27 static simgrid::config::Flag<double> smpi_test_sleep(
28 "smpi/test", "Minimum time to inject inside a call to MPI_Test", 1e-4);
30 std::vector<s_smpi_factor_t> smpi_ois_values;
32 extern void (*smpi_comm_copy_data_callback) (smx_activity_t, void*, size_t);
37 Request::Request(void* buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm, unsigned flags)
38 : buf_(buf), old_type_(datatype), src_(src), dst_(dst), tag_(tag), comm_(comm), flags_(flags)
40 void *old_buf = nullptr;
41 // FIXME Handle the case of a partial shared malloc.
42 if ((((flags & RECV) != 0) && ((flags & ACCUMULATE) != 0)) || (datatype->flags() & DT_FLAG_DERIVED)) {
43 // This part handles the problem of non-contiguous memory
48 buf_ = xbt_malloc(count*datatype->size());
49 if ((datatype->flags() & DT_FLAG_DERIVED) && ((flags & SEND) != 0)) {
50 datatype->serialize(old_buf, buf_, count);
54 // This part handles the problem of non-contiguous memory (for the unserialisation at the reception)
56 size_ = datatype->size() * count;
61 detached_sender_ = nullptr;
66 if (flags & PERSISTENT)
74 MPI_Comm Request::comm(){
94 int Request::detached(){
98 size_t Request::size(){
102 size_t Request::real_size(){
106 void Request::unref(MPI_Request* request)
108 if((*request) != MPI_REQUEST_NULL){
109 (*request)->refcount_--;
110 if((*request)->refcount_ < 0) {
111 (*request)->print_request("wrong refcount");
112 xbt_die("Whoops, wrong refcount");
114 if((*request)->refcount_==0){
115 Datatype::unref((*request)->old_type_);
116 Comm::unref((*request)->comm_);
117 (*request)->print_request("Destroying");
119 *request = MPI_REQUEST_NULL;
121 (*request)->print_request("Decrementing");
124 xbt_die("freeing an already free request");
128 int Request::match_recv(void* a, void* b, simgrid::kernel::activity::CommImpl* ignored)
130 MPI_Request ref = static_cast<MPI_Request>(a);
131 MPI_Request req = static_cast<MPI_Request>(b);
132 XBT_DEBUG("Trying to match a recv of src %d against %d, tag %d against %d",ref->src_,req->src_, ref->tag_, req->tag_);
134 xbt_assert(ref, "Cannot match recv against null reference");
135 xbt_assert(req, "Cannot match recv against null request");
136 if((ref->src_ == MPI_ANY_SOURCE || req->src_ == ref->src_)
137 && ((ref->tag_ == MPI_ANY_TAG && req->tag_ >=0) || req->tag_ == ref->tag_)){
138 //we match, we can transfer some values
139 if(ref->src_ == MPI_ANY_SOURCE)
140 ref->real_src_ = req->src_;
141 if(ref->tag_ == MPI_ANY_TAG)
142 ref->real_tag_ = req->tag_;
143 if(ref->real_size_ < req->real_size_)
145 if(req->detached_==1)
146 ref->detached_sender_=req; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
147 if(req->cancelled_==0)
148 req->cancelled_=-1;//mark as uncancellable
149 XBT_DEBUG("match succeeded");
154 int Request::match_send(void* a, void* b, simgrid::kernel::activity::CommImpl* ignored)
156 MPI_Request ref = static_cast<MPI_Request>(a);
157 MPI_Request req = static_cast<MPI_Request>(b);
158 XBT_DEBUG("Trying to match a send of src %d against %d, tag %d against %d",ref->src_,req->src_, ref->tag_, req->tag_);
159 xbt_assert(ref, "Cannot match send against null reference");
160 xbt_assert(req, "Cannot match send against null request");
162 if((req->src_ == MPI_ANY_SOURCE || req->src_ == ref->src_)
163 && ((req->tag_ == MPI_ANY_TAG && ref->tag_ >=0)|| req->tag_ == ref->tag_)){
164 if(req->src_ == MPI_ANY_SOURCE)
165 req->real_src_ = ref->src_;
166 if(req->tag_ == MPI_ANY_TAG)
167 req->real_tag_ = ref->tag_;
168 if(req->real_size_ < ref->real_size_)
170 if(ref->detached_==1)
171 req->detached_sender_=ref; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
172 if(req->cancelled_==0)
173 req->cancelled_=-1;//mark as uncancellable
174 XBT_DEBUG("match succeeded");
180 void Request::print_request(const char *message)
182 XBT_VERB("%s request %p [buf = %p, size = %zu, src = %d, dst = %d, tag = %d, flags = %x]",
183 message, this, buf_, size_, src_, dst_, tag_, flags_);
187 /* factories, to hide the internal flags from the caller */
188 MPI_Request Request::send_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, simgrid::s4u::this_actor::get_pid(),
192 comm->group()->actor(dst)->get_pid(), tag, comm, PERSISTENT | SEND | PREPARED);
195 MPI_Request Request::ssend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
197 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
198 comm->group()->actor(dst)->get_pid(), tag, comm, PERSISTENT | SSEND | SEND | PREPARED);
201 MPI_Request Request::isend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
203 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
204 comm->group()->actor(dst)->get_pid(), tag, comm, PERSISTENT | ISEND | SEND | PREPARED);
208 MPI_Request Request::rma_send_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
211 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
214 new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src)->get_pid(),
215 comm->group()->actor(dst)->get_pid(), tag, comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED);
217 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src)->get_pid(),
218 comm->group()->actor(dst)->get_pid(), tag, comm,
219 RMA | NON_PERSISTENT | ISEND | SEND | PREPARED | ACCUMULATE);
225 MPI_Request Request::recv_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,
228 src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(src)->get_pid(),
229 simgrid::s4u::this_actor::get_pid(), tag, comm, PERSISTENT | RECV | PREPARED);
232 MPI_Request Request::rma_recv_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
235 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
237 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src)->get_pid(),
238 comm->group()->actor(dst)->get_pid(), tag, comm, RMA | NON_PERSISTENT | RECV | PREPARED);
240 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src)->get_pid(),
241 comm->group()->actor(dst)->get_pid(), tag, comm,
242 RMA | NON_PERSISTENT | RECV | PREPARED | ACCUMULATE);
248 MPI_Request Request::irecv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
250 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
251 src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(src)->get_pid(),
252 simgrid::s4u::this_actor::get_pid(), tag, comm, PERSISTENT | RECV | PREPARED);
255 MPI_Request Request::isend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
257 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
258 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
259 comm->group()->actor(dst)->get_pid(), tag, comm, NON_PERSISTENT | ISEND | SEND);
264 MPI_Request Request::issend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
266 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
267 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
268 comm->group()->actor(dst)->get_pid(), tag, comm, NON_PERSISTENT | ISEND | SSEND | SEND);
274 MPI_Request Request::irecv(void *buf, int count, MPI_Datatype datatype, int src, 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,
278 src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(src)->get_pid(),
279 simgrid::s4u::this_actor::get_pid(), tag, comm, NON_PERSISTENT | RECV);
284 void Request::recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status * status)
286 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
287 request = irecv(buf, count, datatype, src, tag, comm);
288 wait(&request,status);
292 void Request::send(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
294 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
295 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
296 comm->group()->actor(dst)->get_pid(), tag, comm, NON_PERSISTENT | SEND);
299 wait(&request, MPI_STATUS_IGNORE);
303 void Request::ssend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
305 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
306 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
307 comm->group()->actor(dst)->get_pid(), tag, comm, NON_PERSISTENT | SSEND | SEND);
310 wait(&request,MPI_STATUS_IGNORE);
314 void Request::sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,int dst, int sendtag,
315 void *recvbuf, int recvcount, MPI_Datatype recvtype, int src, int recvtag,
316 MPI_Comm comm, MPI_Status * status)
318 MPI_Request requests[2];
320 int myid = simgrid::s4u::this_actor::get_pid();
321 if ((comm->group()->actor(dst)->get_pid() == myid) && (comm->group()->actor(src)->get_pid() == myid)) {
322 Datatype::copy(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype);
323 if (status != MPI_STATUS_IGNORE) {
324 status->MPI_SOURCE = src;
325 status->MPI_TAG = recvtag;
326 status->MPI_ERROR = MPI_SUCCESS;
327 status->count = sendcount * sendtype->size();
331 requests[0] = isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
332 requests[1] = irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
333 startall(2, requests);
334 waitall(2, requests, stats);
337 if(status != MPI_STATUS_IGNORE) {
338 // Copy receive status
343 void Request::start()
345 smx_mailbox_t mailbox;
347 xbt_assert(action_ == nullptr, "Cannot (re-)start unfinished communication");
352 if ((flags_ & RECV) != 0) {
353 this->print_request("New recv");
355 simgrid::smpi::Process* process = smpi_process_remote(simgrid::s4u::Actor::by_pid(dst_));
357 int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh");
359 xbt_mutex_t mut = process->mailboxes_mutex();
360 if (async_small_thresh != 0 || (flags_ & RMA) != 0)
361 xbt_mutex_acquire(mut);
363 if (async_small_thresh == 0 && (flags_ & RMA) == 0 ) {
364 mailbox = process->mailbox();
366 else if (((flags_ & RMA) != 0) || static_cast<int>(size_) < async_small_thresh) {
367 //We have to check both mailboxes (because SSEND messages are sent to the large mbox).
368 //begin with the more appropriate one : the small one.
369 mailbox = process->mailbox_small();
370 XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %p (in case of SSEND)?", mailbox);
371 smx_activity_t action = simcall_comm_iprobe(mailbox, 0, &match_recv, static_cast<void*>(this));
373 if (action == nullptr) {
374 mailbox = process->mailbox();
375 XBT_DEBUG("No, nothing in the small mailbox test the other one : %p", mailbox);
376 action = simcall_comm_iprobe(mailbox, 0, &match_recv, static_cast<void*>(this));
377 if (action == nullptr) {
378 XBT_DEBUG("Still nothing, switch back to the small mailbox : %p", mailbox);
379 mailbox = process->mailbox_small();
382 XBT_DEBUG("yes there was something for us in the large mailbox");
385 mailbox = process->mailbox_small();
386 XBT_DEBUG("Is there a corresponding send already posted the small mailbox?");
387 smx_activity_t action = simcall_comm_iprobe(mailbox, 0, &match_recv, static_cast<void*>(this));
389 if (action == nullptr) {
390 XBT_DEBUG("No, nothing in the permanent receive mailbox");
391 mailbox = process->mailbox();
393 XBT_DEBUG("yes there was something for us in the small mailbox");
397 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
399 action_ = simcall_comm_irecv(
400 process->process()->get_impl(), mailbox, buf_, &real_size_, &match_recv,
401 process->replaying() ? &smpi_comm_null_copy_buffer_callback : smpi_comm_copy_data_callback, this, -1.0);
402 XBT_DEBUG("recv simcall posted");
404 if (async_small_thresh != 0 || (flags_ & RMA) != 0 )
405 xbt_mutex_release(mut);
406 } else { /* the RECV flag was not set, so this is a send */
407 simgrid::smpi::Process* process = smpi_process_remote(simgrid::s4u::Actor::by_pid(dst_));
409 if (TRACE_smpi_view_internals()) {
410 TRACE_smpi_send(rank, rank, dst_, tag_, size_);
412 this->print_request("New send");
415 if ((flags_ & SSEND) == 0 && ( (flags_ & RMA) != 0
416 || static_cast<int>(size_) < xbt_cfg_get_int("smpi/send-is-detached-thresh") ) ) {
417 void *oldbuf = nullptr;
419 XBT_DEBUG("Send request %p is detached", this);
421 if (not(old_type_->flags() & DT_FLAG_DERIVED)) {
423 if (not process->replaying() && oldbuf != nullptr && size_ != 0) {
424 if ((smpi_privatize_global_variables != SmpiPrivStrategies::None) &&
425 (static_cast<char*>(buf_) >= smpi_data_exe_start) &&
426 (static_cast<char*>(buf_) < smpi_data_exe_start + smpi_data_exe_size)) {
427 XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
428 smpi_switch_data_segment(simgrid::s4u::Actor::by_pid(src_));
430 buf = xbt_malloc(size_);
431 memcpy(buf,oldbuf,size_);
432 XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
437 //if we are giving back the control to the user without waiting for completion, we have to inject timings
438 double sleeptime = 0.0;
439 if (detached_ != 0 || ((flags_ & (ISEND | SSEND)) != 0)) { // issend should be treated as isend
440 // isend and send timings may be different
441 sleeptime = ((flags_ & ISEND) != 0)
442 ? simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::SmpiHost>()->oisend(size_)
443 : simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::SmpiHost>()->osend(size_);
447 simcall_process_sleep(sleeptime);
448 XBT_DEBUG("sending size of %zu : sleep %f ", size_, sleeptime);
451 int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh");
453 xbt_mutex_t mut=process->mailboxes_mutex();
455 if (async_small_thresh != 0 || (flags_ & RMA) != 0)
456 xbt_mutex_acquire(mut);
458 if (not(async_small_thresh != 0 || (flags_ & RMA) != 0)) {
459 mailbox = process->mailbox();
460 } else if (((flags_ & RMA) != 0) || static_cast<int>(size_) < async_small_thresh) { // eager mode
461 mailbox = process->mailbox();
462 XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %p?", mailbox);
463 smx_activity_t action = simcall_comm_iprobe(mailbox, 1, &match_send, static_cast<void*>(this));
464 if (action == nullptr) {
465 if ((flags_ & SSEND) == 0){
466 mailbox = process->mailbox_small();
467 XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %p", mailbox);
469 mailbox = process->mailbox_small();
470 XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %p?", mailbox);
471 action = simcall_comm_iprobe(mailbox, 1, &match_send, static_cast<void*>(this));
472 if (action == nullptr) {
473 XBT_DEBUG("No, we are first, send to large mailbox");
474 mailbox = process->mailbox();
478 XBT_DEBUG("Yes there was something for us in the large mailbox");
481 mailbox = process->mailbox();
482 XBT_DEBUG("Send request %p is in the large mailbox %p (buf: %p)",mailbox, this,buf_);
485 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
487 action_ = simcall_comm_isend(
488 simgrid::s4u::Actor::by_pid(src_)->get_impl(), mailbox, size_, -1.0, buf, real_size_, &match_send,
489 &xbt_free_f, // how to free the userdata if a detached send fails
490 not process->replaying() ? smpi_comm_copy_data_callback : &smpi_comm_null_copy_buffer_callback, this,
491 // detach if msg size < eager/rdv switch limit
493 XBT_DEBUG("send simcall posted");
495 /* FIXME: detached sends are not traceable (action_ == nullptr) */
496 if (action_ != nullptr)
497 simcall_set_category(action_, TRACE_internal_smpi_get_category());
498 if (async_small_thresh != 0 || ((flags_ & RMA)!=0))
499 xbt_mutex_release(mut);
503 void Request::startall(int count, MPI_Request * requests)
505 if(requests== nullptr)
508 for(int i = 0; i < count; i++) {
509 requests[i]->start();
513 void Request::cancel()
517 if (this->action_ != nullptr)
518 (boost::static_pointer_cast<simgrid::kernel::activity::CommImpl>(this->action_))->cancel();
521 int Request::test(MPI_Request * request, MPI_Status * status) {
522 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or testall before)
523 // to avoid deadlocks if used as a break condition, such as
524 // while (MPI_Test(request, flag, status) && flag) dostuff...
525 // because the time will not normally advance when only calls to MPI_Test are made -> deadlock
526 // multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
527 static int nsleeps = 1;
528 if(smpi_test_sleep > 0)
529 simcall_process_sleep(nsleeps*smpi_test_sleep);
531 Status::empty(status);
533 if (((*request)->flags_ & PREPARED) == 0) {
534 if ((*request)->action_ != nullptr){
536 flag = simcall_comm_test((*request)->action_);
542 finish_wait(request,status);
543 nsleeps=1;//reset the number of sleeps we will do next time
544 if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & PERSISTENT) == 0)
545 *request = MPI_REQUEST_NULL;
546 } else if (xbt_cfg_get_boolean("smpi/grow-injected-times")){
553 int Request::testsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
558 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
560 for (int i = 0; i < incount; i++) {
561 if (requests[i] != MPI_REQUEST_NULL) {
562 if (test(&requests[i], pstat)) {
565 if (status != MPI_STATUSES_IGNORE)
567 if ((requests[i] != MPI_REQUEST_NULL) && (requests[i]->flags_ & NON_PERSISTENT))
568 requests[i] = MPI_REQUEST_NULL;
574 if(count_dead==incount)
575 return MPI_UNDEFINED;
579 int Request::testany(int count, MPI_Request requests[], int *index, MPI_Status * status)
581 std::vector<simgrid::kernel::activity::ActivityImplPtr> comms;
582 comms.reserve(count);
587 *index = MPI_UNDEFINED;
589 std::vector<int> map; /** Maps all matching comms back to their location in requests **/
590 for(i = 0; i < count; i++) {
591 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action_ && not(requests[i]->flags_ & PREPARED)) {
592 comms.push_back(requests[i]->action_);
596 if (not map.empty()) {
597 //multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it
598 static int nsleeps = 1;
599 if(smpi_test_sleep > 0)
600 simcall_process_sleep(nsleeps*smpi_test_sleep);
602 i = simcall_comm_testany(comms.data(), comms.size()); // The i-th element in comms matches!
607 if (i != -1) { // -1 is not MPI_UNDEFINED but a SIMIX return code. (nothing matches)
609 finish_wait(&requests[*index],status);
612 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & NON_PERSISTENT)) {
613 requests[*index] = MPI_REQUEST_NULL;
619 //all requests are null or inactive, return true
621 Status::empty(status);
627 int Request::testall(int count, MPI_Request requests[], MPI_Status status[])
630 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
632 for(int i=0; i<count; i++){
633 if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & PREPARED)) {
634 if (test(&requests[i], pstat)!=1){
637 requests[i]=MPI_REQUEST_NULL;
640 Status::empty(pstat);
642 if(status != MPI_STATUSES_IGNORE) {
649 void Request::probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
651 //FIXME find another way to avoid busy waiting ?
652 // the issue here is that we have to wait on a nonexistent comm
654 iprobe(source, tag, comm, &flag, status);
655 XBT_DEBUG("Busy Waiting on probing : %d", flag);
659 void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
660 // to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
661 // especially when used as a break condition, such as while (MPI_Iprobe(...)) dostuff...
662 // nsleeps is a multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
663 // This can speed up the execution of certain applications by an order of magnitude, such as HPL
664 static int nsleeps = 1;
665 double speed = simgrid::s4u::Actor::self()->get_host()->getSpeed();
666 double maxrate = xbt_cfg_get_double("smpi/iprobe-cpu-usage");
667 MPI_Request request = new Request(nullptr, 0, MPI_CHAR,
668 source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(source)->get_pid(),
669 simgrid::s4u::this_actor::get_pid(), tag, comm, PERSISTENT | RECV);
670 if (smpi_iprobe_sleep > 0) {
671 smx_activity_t iprobe_sleep = simcall_execution_start(
672 "iprobe", /* flops to executek*/ nsleeps * smpi_iprobe_sleep * speed * maxrate, /* priority */ 1.0,
673 /* performance bound */ maxrate * speed, smpi_process()->process()->get_impl()->host);
674 simcall_execution_wait(iprobe_sleep);
676 // behave like a receive, but don't do it
677 smx_mailbox_t mailbox;
679 request->print_request("New iprobe");
680 // We have to test both mailboxes as we don't know if we will receive one one or another
681 if (xbt_cfg_get_int("smpi/async-small-thresh") > 0){
682 mailbox = smpi_process()->mailbox_small();
683 XBT_DEBUG("Trying to probe the perm recv mailbox");
684 request->action_ = simcall_comm_iprobe(mailbox, 0, &match_recv, static_cast<void*>(request));
687 if (request->action_ == nullptr){
688 mailbox = smpi_process()->mailbox();
689 XBT_DEBUG("trying to probe the other mailbox");
690 request->action_ = simcall_comm_iprobe(mailbox, 0, &match_recv, static_cast<void*>(request));
693 if (request->action_ != nullptr){
694 simgrid::kernel::activity::CommImplPtr sync_comm =
695 boost::static_pointer_cast<simgrid::kernel::activity::CommImpl>(request->action_);
696 MPI_Request req = static_cast<MPI_Request>(sync_comm->src_data);
698 if(status != MPI_STATUS_IGNORE && (req->flags_ & PREPARED) == 0) {
699 status->MPI_SOURCE = comm->group()->rank(req->src_);
700 status->MPI_TAG = req->tag_;
701 status->MPI_ERROR = MPI_SUCCESS;
702 status->count = req->real_size_;
704 nsleeps = 1;//reset the number of sleeps we will do next time
708 if (xbt_cfg_get_boolean("smpi/grow-injected-times"))
714 void Request::finish_wait(MPI_Request* request, MPI_Status * status)
716 MPI_Request req = *request;
717 Status::empty(status);
719 if (req->cancelled_==1){
720 if (status!=MPI_STATUS_IGNORE)
725 if (not((req->detached_ != 0) && ((req->flags_ & SEND) != 0)) && ((req->flags_ & PREPARED) == 0)) {
726 if(status != MPI_STATUS_IGNORE) {
727 int src = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_;
728 status->MPI_SOURCE = req->comm_->group()->rank(src);
729 status->MPI_TAG = req->tag_ == MPI_ANY_TAG ? req->real_tag_ : req->tag_;
730 status->MPI_ERROR = req->truncated_ != 0 ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
731 // this handles the case were size in receive differs from size in send
732 status->count = req->real_size_;
735 req->print_request("Finishing");
736 MPI_Datatype datatype = req->old_type_;
738 // FIXME Handle the case of a partial shared malloc.
739 if (((req->flags_ & ACCUMULATE) != 0) ||
740 (datatype->flags() & DT_FLAG_DERIVED)) { // && (not smpi_is_shared(req->old_buf_))){
742 if (not smpi_process()->replaying() && smpi_privatize_global_variables != SmpiPrivStrategies::None &&
743 static_cast<char*>(req->old_buf_) >= smpi_data_exe_start &&
744 static_cast<char*>(req->old_buf_) < smpi_data_exe_start + smpi_data_exe_size) {
745 XBT_VERB("Privatization : We are unserializing to a zone in global memory Switch data segment ");
746 smpi_switch_data_segment(simgrid::s4u::Actor::self());
749 if(datatype->flags() & DT_FLAG_DERIVED){
750 // This part handles the problem of non-contignous memory the unserialization at the reception
751 if((req->flags_ & RECV) && datatype->size()!=0)
752 datatype->unserialize(req->buf_, req->old_buf_, req->real_size_/datatype->size() , req->op_);
754 }else if(req->flags_ & RECV){//apply op on contiguous buffer for accumulate
755 if(datatype->size()!=0){
756 int n =req->real_size_/datatype->size();
757 req->op_->apply(req->buf_, req->old_buf_, &n, datatype);
764 if (TRACE_smpi_view_internals() && ((req->flags_ & RECV) != 0)){
765 int rank = simgrid::s4u::this_actor::get_pid();
766 int src_traced = (req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_);
767 TRACE_smpi_recv(src_traced, rank,req->tag_);
769 if(req->detached_sender_ != nullptr){
770 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
772 simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::SmpiHost>()->orecv(req->real_size());
774 simcall_process_sleep(sleeptime);
775 XBT_DEBUG("receiving size of %zu : sleep %f ", req->real_size_, sleeptime);
777 unref(&(req->detached_sender_));
779 if(req->flags_ & PERSISTENT)
780 req->action_ = nullptr;
781 req->flags_ |= FINISHED;
785 void Request::wait(MPI_Request * request, MPI_Status * status)
787 (*request)->print_request("Waiting");
788 if ((*request)->flags_ & PREPARED) {
789 Status::empty(status);
793 if ((*request)->action_ != nullptr){
795 // this is not a detached send
796 simcall_comm_wait((*request)->action_, -1.0);
798 XBT_VERB("Request cancelled");
803 finish_wait(request,status);
804 if (*request != MPI_REQUEST_NULL && (((*request)->flags_ & NON_PERSISTENT)!=0))
805 *request = MPI_REQUEST_NULL;
808 int Request::waitany(int count, MPI_Request requests[], MPI_Status * status)
810 s_xbt_dynar_t comms; // Keep it on stack to save some extra mallocs
811 int index = MPI_UNDEFINED;
815 // Wait for a request to complete
816 xbt_dynar_init(&comms, sizeof(smx_activity_t), [](void*ptr){
817 intrusive_ptr_release(*(simgrid::kernel::activity::ActivityImpl**)ptr);
819 int *map = xbt_new(int, count);
820 XBT_DEBUG("Wait for one of %d", count);
821 for(int i = 0; i < count; i++) {
822 if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & PREPARED) &&
823 not(requests[i]->flags_ & FINISHED)) {
824 if (requests[i]->action_ != nullptr) {
825 XBT_DEBUG("Waiting any %p ", requests[i]);
826 intrusive_ptr_add_ref(requests[i]->action_.get());
827 xbt_dynar_push_as(&comms, simgrid::kernel::activity::ActivityImpl*, requests[i]->action_.get());
831 // This is a finished detached request, let's return this one
832 size = 0; // so we free the dynar but don't do the waitany call
834 finish_wait(&requests[i], status); // cleanup if refcount = 0
835 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & NON_PERSISTENT))
836 requests[i] = MPI_REQUEST_NULL; // set to null
842 XBT_DEBUG("Enter waitany for %lu comms", xbt_dynar_length(&comms));
845 // this is not a detached send
846 i = simcall_comm_waitany(&comms, -1);
848 XBT_INFO("request %d cancelled ",i);
852 // not MPI_UNDEFINED, as this is a simix return code
855 //in case of an accumulate, we have to wait the end of all requests to apply the operation, ordered correctly.
856 if ((requests[index] == MPI_REQUEST_NULL) ||
857 (not((requests[index]->flags_ & ACCUMULATE) && (requests[index]->flags_ & RECV)))) {
858 finish_wait(&requests[index],status);
859 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & NON_PERSISTENT))
860 requests[index] = MPI_REQUEST_NULL;
865 xbt_dynar_free_data(&comms);
869 if (index==MPI_UNDEFINED)
870 Status::empty(status);
875 static int sort_accumulates(MPI_Request a, MPI_Request b)
877 return (a->tag() > b->tag());
880 int Request::waitall(int count, MPI_Request requests[], MPI_Status status[])
882 std::vector<MPI_Request> accumulates;
885 MPI_Status *pstat = (status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat);
886 int retvalue = MPI_SUCCESS;
887 //tag invalid requests in the set
888 if (status != MPI_STATUSES_IGNORE) {
889 for (int c = 0; c < count; c++) {
890 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst_ == MPI_PROC_NULL || (requests[c]->flags_ & PREPARED)) {
891 Status::empty(&status[c]);
892 } else if (requests[c]->src_ == MPI_PROC_NULL) {
893 Status::empty(&status[c]);
894 status[c].MPI_SOURCE = MPI_PROC_NULL;
898 for (int c = 0; c < count; c++) {
899 if (MC_is_active() || MC_record_replay_is_active()) {
900 wait(&requests[c],pstat);
903 index = waitany(count, (MPI_Request*)requests, pstat);
905 if (index == MPI_UNDEFINED)
908 if (requests[index] != MPI_REQUEST_NULL
909 && (requests[index]->flags_ & RECV)
910 && (requests[index]->flags_ & ACCUMULATE))
911 accumulates.push_back(requests[index]);
912 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & NON_PERSISTENT))
913 requests[index] = MPI_REQUEST_NULL;
915 if (status != MPI_STATUSES_IGNORE) {
916 status[index] = *pstat;
917 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
918 retvalue = MPI_ERR_IN_STATUS;
922 if (not accumulates.empty()) {
923 std::sort(accumulates.begin(), accumulates.end(), sort_accumulates);
924 for (auto& req : accumulates) {
925 finish_wait(&req, status);
932 int Request::waitsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
936 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
938 for (int i = 0; i < incount; i++) {
939 int index = waitany(incount, requests, pstat);
940 if(index!=MPI_UNDEFINED){
941 indices[count] = index;
943 if(status != MPI_STATUSES_IGNORE) {
944 status[index] = *pstat;
946 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & NON_PERSISTENT))
947 requests[index] = MPI_REQUEST_NULL;
949 return MPI_UNDEFINED;
955 MPI_Request Request::f2c(int id) {
957 if(id==MPI_FORTRAN_REQUEST_NULL)
958 return static_cast<MPI_Request>(MPI_REQUEST_NULL);
959 return static_cast<MPI_Request>(F2C::f2c_lookup()->at(get_key_id(key, id)));
964 if (F2C::f2c_lookup() == nullptr) {
965 F2C::set_f2c_lookup(new std::unordered_map<std::string, F2C*>);
968 (*(F2C::f2c_lookup()))[get_key_id(key, F2C::f2c_id())] = this;
969 F2C::f2c_id_increment();
970 return F2C::f2c_id()-1;
973 void Request::free_f(int id)
975 if (id != MPI_FORTRAN_REQUEST_NULL) {
977 F2C::f2c_lookup()->erase(get_key_id(key, id));