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 "smpi_host.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 & MPI_REQ_RECV) != 0) && ((flags & MPI_REQ_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 & MPI_REQ_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 & MPI_REQ_PERSISTENT)
74 MPI_Comm Request::comm(){
94 int Request::detached(){
98 MPI_Datatype Request::type(){
102 size_t Request::size(){
106 size_t Request::real_size(){
110 void Request::unref(MPI_Request* request)
112 if((*request) != MPI_REQUEST_NULL){
113 (*request)->refcount_--;
114 if((*request)->refcount_ < 0) {
115 (*request)->print_request("wrong refcount");
116 xbt_die("Whoops, wrong refcount");
118 if((*request)->refcount_==0){
119 Datatype::unref((*request)->old_type_);
120 Comm::unref((*request)->comm_);
121 (*request)->print_request("Destroying");
123 *request = MPI_REQUEST_NULL;
125 (*request)->print_request("Decrementing");
128 xbt_die("freeing an already free request");
132 int Request::match_recv(void* a, void* b, simgrid::kernel::activity::CommImpl* ignored)
134 MPI_Request ref = static_cast<MPI_Request>(a);
135 MPI_Request req = static_cast<MPI_Request>(b);
136 XBT_DEBUG("Trying to match a recv of src %d against %d, tag %d against %d",ref->src_,req->src_, ref->tag_, req->tag_);
138 xbt_assert(ref, "Cannot match recv against null reference");
139 xbt_assert(req, "Cannot match recv against null request");
140 if((ref->src_ == MPI_ANY_SOURCE || req->src_ == ref->src_)
141 && ((ref->tag_ == MPI_ANY_TAG && req->tag_ >=0) || req->tag_ == ref->tag_)){
142 //we match, we can transfer some values
143 if(ref->src_ == MPI_ANY_SOURCE)
144 ref->real_src_ = req->src_;
145 if(ref->tag_ == MPI_ANY_TAG)
146 ref->real_tag_ = req->tag_;
147 if(ref->real_size_ < req->real_size_)
149 if(req->detached_==1)
150 ref->detached_sender_=req; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
151 if(req->cancelled_==0)
152 req->cancelled_=-1;//mark as uncancellable
153 XBT_DEBUG("match succeeded");
158 int Request::match_send(void* a, void* b, simgrid::kernel::activity::CommImpl* ignored)
160 MPI_Request ref = static_cast<MPI_Request>(a);
161 MPI_Request req = static_cast<MPI_Request>(b);
162 XBT_DEBUG("Trying to match a send of src %d against %d, tag %d against %d",ref->src_,req->src_, ref->tag_, req->tag_);
163 xbt_assert(ref, "Cannot match send against null reference");
164 xbt_assert(req, "Cannot match send against null request");
166 if((req->src_ == MPI_ANY_SOURCE || req->src_ == ref->src_)
167 && ((req->tag_ == MPI_ANY_TAG && ref->tag_ >=0)|| req->tag_ == ref->tag_)){
168 if(req->src_ == MPI_ANY_SOURCE)
169 req->real_src_ = ref->src_;
170 if(req->tag_ == MPI_ANY_TAG)
171 req->real_tag_ = ref->tag_;
172 if(req->real_size_ < ref->real_size_)
174 if(ref->detached_==1)
175 req->detached_sender_=ref; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
176 if(req->cancelled_==0)
177 req->cancelled_=-1;//mark as uncancellable
178 XBT_DEBUG("match succeeded");
184 void Request::print_request(const char *message)
186 XBT_VERB("%s request %p [buf = %p, size = %zu, src = %d, dst = %d, tag = %d, flags = %x]",
187 message, this, buf_, size_, src_, dst_, tag_, flags_);
191 /* factories, to hide the internal flags from the caller */
192 MPI_Request Request::send_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
195 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
196 comm->group()->actor(dst)->get_pid(), tag, comm,
197 MPI_REQ_PERSISTENT | MPI_REQ_SEND | MPI_REQ_PREPARED);
200 MPI_Request Request::ssend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
202 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
203 comm->group()->actor(dst)->get_pid(), tag, comm,
204 MPI_REQ_PERSISTENT | MPI_REQ_SSEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
207 MPI_Request Request::isend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
209 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
210 comm->group()->actor(dst)->get_pid(), tag, comm,
211 MPI_REQ_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
215 MPI_Request Request::rma_send_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, comm->group()->actor(src)->get_pid(),
221 comm->group()->actor(dst)->get_pid(), tag, comm,
222 MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
224 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src)->get_pid(),
225 comm->group()->actor(dst)->get_pid(), tag, comm,
226 MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED |
233 MPI_Request Request::recv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
235 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
236 src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(src)->get_pid(),
237 simgrid::s4u::this_actor::get_pid(), tag, comm,
238 MPI_REQ_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
241 MPI_Request Request::rma_recv_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
244 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
246 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src)->get_pid(),
247 comm->group()->actor(dst)->get_pid(), tag, comm,
248 MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
250 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src)->get_pid(),
251 comm->group()->actor(dst)->get_pid(), tag, comm,
252 MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED | MPI_REQ_ACCUMULATE);
258 MPI_Request Request::irecv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
260 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
261 src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(src)->get_pid(),
262 simgrid::s4u::this_actor::get_pid(), tag, comm,
263 MPI_REQ_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
266 MPI_Request Request::isend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
268 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
269 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
270 comm->group()->actor(dst)->get_pid(), tag, comm,
271 MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND);
276 MPI_Request Request::issend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
278 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
279 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
280 comm->group()->actor(dst)->get_pid(), tag, comm,
281 MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SSEND | MPI_REQ_SEND);
287 MPI_Request Request::irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
289 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
290 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
291 src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(src)->get_pid(),
292 simgrid::s4u::this_actor::get_pid(), tag, comm, MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV);
297 void Request::recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status * status)
299 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
300 request = irecv(buf, count, datatype, src, tag, comm);
301 wait(&request,status);
305 void Request::send(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
307 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
308 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
309 comm->group()->actor(dst)->get_pid(), tag, comm, MPI_REQ_NON_PERSISTENT | MPI_REQ_SEND);
312 wait(&request, MPI_STATUS_IGNORE);
316 void Request::ssend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
318 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
319 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
320 comm->group()->actor(dst)->get_pid(), tag, comm,
321 MPI_REQ_NON_PERSISTENT | MPI_REQ_SSEND | MPI_REQ_SEND);
324 wait(&request,MPI_STATUS_IGNORE);
328 void Request::sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,int dst, int sendtag,
329 void *recvbuf, int recvcount, MPI_Datatype recvtype, int src, int recvtag,
330 MPI_Comm comm, MPI_Status * status)
332 MPI_Request requests[2];
334 int myid = simgrid::s4u::this_actor::get_pid();
335 if ((comm->group()->actor(dst)->get_pid() == myid) && (comm->group()->actor(src)->get_pid() == myid)) {
336 Datatype::copy(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype);
337 if (status != MPI_STATUS_IGNORE) {
338 status->MPI_SOURCE = src;
339 status->MPI_TAG = recvtag;
340 status->MPI_ERROR = MPI_SUCCESS;
341 status->count = sendcount * sendtype->size();
345 requests[0] = isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
346 requests[1] = irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
347 startall(2, requests);
348 waitall(2, requests, stats);
351 if(status != MPI_STATUS_IGNORE) {
352 // Copy receive status
357 void Request::start()
359 smx_mailbox_t mailbox;
361 xbt_assert(action_ == nullptr, "Cannot (re-)start unfinished communication");
362 flags_ &= ~MPI_REQ_PREPARED;
363 flags_ &= ~MPI_REQ_FINISHED;
366 if ((flags_ & MPI_REQ_RECV) != 0) {
367 this->print_request("New recv");
369 simgrid::smpi::Process* process = smpi_process_remote(simgrid::s4u::Actor::by_pid(dst_));
371 int async_small_thresh = simgrid::config::get_value<int>("smpi/async-small-thresh");
373 xbt_mutex_t mut = process->mailboxes_mutex();
374 if (async_small_thresh != 0 || (flags_ & MPI_REQ_RMA) != 0)
375 xbt_mutex_acquire(mut);
377 if (async_small_thresh == 0 && (flags_ & MPI_REQ_RMA) == 0) {
378 mailbox = process->mailbox();
379 } else if (((flags_ & MPI_REQ_RMA) != 0) || static_cast<int>(size_) < async_small_thresh) {
380 //We have to check both mailboxes (because SSEND messages are sent to the large mbox).
381 //begin with the more appropriate one : the small one.
382 mailbox = process->mailbox_small();
383 XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %p (in case of SSEND)?", mailbox);
384 smx_activity_t action = simcall_comm_iprobe(mailbox, 0, &match_recv, static_cast<void*>(this));
386 if (action == nullptr) {
387 mailbox = process->mailbox();
388 XBT_DEBUG("No, nothing in the small mailbox test the other one : %p", mailbox);
389 action = simcall_comm_iprobe(mailbox, 0, &match_recv, static_cast<void*>(this));
390 if (action == nullptr) {
391 XBT_DEBUG("Still nothing, switch back to the small mailbox : %p", mailbox);
392 mailbox = process->mailbox_small();
395 XBT_DEBUG("yes there was something for us in the large mailbox");
398 mailbox = process->mailbox_small();
399 XBT_DEBUG("Is there a corresponding send already posted the small mailbox?");
400 smx_activity_t action = simcall_comm_iprobe(mailbox, 0, &match_recv, static_cast<void*>(this));
402 if (action == nullptr) {
403 XBT_DEBUG("No, nothing in the permanent receive mailbox");
404 mailbox = process->mailbox();
406 XBT_DEBUG("yes there was something for us in the small mailbox");
410 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
412 action_ = simcall_comm_irecv(
413 process->get_actor()->get_impl(), mailbox, buf_, &real_size_, &match_recv,
414 process->replaying() ? &smpi_comm_null_copy_buffer_callback : smpi_comm_copy_data_callback, this, -1.0);
415 XBT_DEBUG("recv simcall posted");
417 if (async_small_thresh != 0 || (flags_ & MPI_REQ_RMA) != 0)
418 xbt_mutex_release(mut);
419 } else { /* the RECV flag was not set, so this is a send */
420 simgrid::smpi::Process* process = smpi_process_remote(simgrid::s4u::Actor::by_pid(dst_));
422 if (TRACE_smpi_view_internals()) {
423 TRACE_smpi_send(rank, rank, dst_, tag_, size_);
425 this->print_request("New send");
428 if ((flags_ & MPI_REQ_SSEND) == 0 &&
429 ((flags_ & MPI_REQ_RMA) != 0 ||
430 static_cast<int>(size_) < simgrid::config::get_value<int>("smpi/send-is-detached-thresh"))) {
431 void *oldbuf = nullptr;
433 XBT_DEBUG("Send request %p is detached", this);
435 if (not(old_type_->flags() & DT_FLAG_DERIVED)) {
437 if (not process->replaying() && oldbuf != nullptr && size_ != 0) {
438 if ((smpi_privatize_global_variables != SmpiPrivStrategies::None) &&
439 (static_cast<char*>(buf_) >= smpi_data_exe_start) &&
440 (static_cast<char*>(buf_) < smpi_data_exe_start + smpi_data_exe_size)) {
441 XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
442 smpi_switch_data_segment(simgrid::s4u::Actor::by_pid(src_));
444 buf = xbt_malloc(size_);
445 memcpy(buf,oldbuf,size_);
446 XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
451 //if we are giving back the control to the user without waiting for completion, we have to inject timings
452 double sleeptime = 0.0;
453 if (detached_ != 0 || ((flags_ & (MPI_REQ_ISEND | MPI_REQ_SSEND)) != 0)) { // issend should be treated as isend
454 // isend and send timings may be different
455 sleeptime = ((flags_ & MPI_REQ_ISEND) != 0)
456 ? simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::Host>()->oisend(size_)
457 : simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::Host>()->osend(size_);
461 simcall_process_sleep(sleeptime);
462 XBT_DEBUG("sending size of %zu : sleep %f ", size_, sleeptime);
465 int async_small_thresh = simgrid::config::get_value<int>("smpi/async-small-thresh");
467 xbt_mutex_t mut=process->mailboxes_mutex();
469 if (async_small_thresh != 0 || (flags_ & MPI_REQ_RMA) != 0)
470 xbt_mutex_acquire(mut);
472 if (not(async_small_thresh != 0 || (flags_ & MPI_REQ_RMA) != 0)) {
473 mailbox = process->mailbox();
474 } else if (((flags_ & MPI_REQ_RMA) != 0) || static_cast<int>(size_) < async_small_thresh) { // eager mode
475 mailbox = process->mailbox();
476 XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %p?", mailbox);
477 smx_activity_t action = simcall_comm_iprobe(mailbox, 1, &match_send, static_cast<void*>(this));
478 if (action == nullptr) {
479 if ((flags_ & MPI_REQ_SSEND) == 0) {
480 mailbox = process->mailbox_small();
481 XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %p", mailbox);
483 mailbox = process->mailbox_small();
484 XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %p?", mailbox);
485 action = simcall_comm_iprobe(mailbox, 1, &match_send, static_cast<void*>(this));
486 if (action == nullptr) {
487 XBT_DEBUG("No, we are first, send to large mailbox");
488 mailbox = process->mailbox();
492 XBT_DEBUG("Yes there was something for us in the large mailbox");
495 mailbox = process->mailbox();
496 XBT_DEBUG("Send request %p is in the large mailbox %p (buf: %p)",mailbox, this,buf_);
499 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
501 action_ = simcall_comm_isend(
502 simgrid::s4u::Actor::by_pid(src_)->get_impl(), mailbox, size_, -1.0, buf, real_size_, &match_send,
503 &xbt_free_f, // how to free the userdata if a detached send fails
504 not process->replaying() ? smpi_comm_copy_data_callback : &smpi_comm_null_copy_buffer_callback, this,
505 // detach if msg size < eager/rdv switch limit
507 XBT_DEBUG("send simcall posted");
509 /* FIXME: detached sends are not traceable (action_ == nullptr) */
510 if (action_ != nullptr)
511 simcall_set_category(action_, TRACE_internal_smpi_get_category());
512 if (async_small_thresh != 0 || ((flags_ & MPI_REQ_RMA) != 0))
513 xbt_mutex_release(mut);
517 void Request::startall(int count, MPI_Request * requests)
519 if(requests== nullptr)
522 for(int i = 0; i < count; i++) {
523 requests[i]->start();
527 void Request::cancel()
531 if (this->action_ != nullptr)
532 (boost::static_pointer_cast<simgrid::kernel::activity::CommImpl>(this->action_))->cancel();
535 int Request::test(MPI_Request * request, MPI_Status * status) {
536 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or testall before)
537 // to avoid deadlocks if used as a break condition, such as
538 // while (MPI_Test(request, flag, status) && flag) dostuff...
539 // because the time will not normally advance when only calls to MPI_Test are made -> deadlock
540 // multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
541 static int nsleeps = 1;
542 if(smpi_test_sleep > 0)
543 simcall_process_sleep(nsleeps*smpi_test_sleep);
545 Status::empty(status);
547 if (((*request)->flags_ & MPI_REQ_PREPARED) == 0) {
548 if ((*request)->action_ != nullptr){
550 flag = simcall_comm_test((*request)->action_);
556 finish_wait(request,status);
557 nsleeps=1;//reset the number of sleeps we will do next time
558 if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & MPI_REQ_PERSISTENT) == 0)
559 *request = MPI_REQUEST_NULL;
560 } else if (simgrid::config::get_value<bool>("smpi/grow-injected-times")) {
567 int Request::testsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
572 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
574 for (int i = 0; i < incount; i++) {
575 if (requests[i] != MPI_REQUEST_NULL) {
576 if (test(&requests[i], pstat)) {
579 if (status != MPI_STATUSES_IGNORE)
581 if ((requests[i] != MPI_REQUEST_NULL) && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
582 requests[i] = MPI_REQUEST_NULL;
588 if(count_dead==incount)
589 return MPI_UNDEFINED;
593 int Request::testany(int count, MPI_Request requests[], int *index, MPI_Status * status)
595 std::vector<simgrid::kernel::activity::ActivityImplPtr> comms;
596 comms.reserve(count);
601 *index = MPI_UNDEFINED;
603 std::vector<int> map; /** Maps all matching comms back to their location in requests **/
604 for(i = 0; i < count; i++) {
605 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action_ && not(requests[i]->flags_ & MPI_REQ_PREPARED)) {
606 comms.push_back(requests[i]->action_);
610 if (not map.empty()) {
611 //multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it
612 static int nsleeps = 1;
613 if(smpi_test_sleep > 0)
614 simcall_process_sleep(nsleeps*smpi_test_sleep);
616 i = simcall_comm_testany(comms.data(), comms.size()); // The i-th element in comms matches!
621 if (i != -1) { // -1 is not MPI_UNDEFINED but a SIMIX return code. (nothing matches)
623 finish_wait(&requests[*index],status);
626 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_NON_PERSISTENT)) {
627 requests[*index] = MPI_REQUEST_NULL;
633 //all requests are null or inactive, return true
635 Status::empty(status);
641 int Request::testall(int count, MPI_Request requests[], MPI_Status status[])
644 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
646 for(int i=0; i<count; i++){
647 if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & MPI_REQ_PREPARED)) {
648 if (test(&requests[i], pstat)!=1){
651 requests[i]=MPI_REQUEST_NULL;
654 Status::empty(pstat);
656 if(status != MPI_STATUSES_IGNORE) {
663 void Request::probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
665 //FIXME find another way to avoid busy waiting ?
666 // the issue here is that we have to wait on a nonexistent comm
668 iprobe(source, tag, comm, &flag, status);
669 XBT_DEBUG("Busy Waiting on probing : %d", flag);
673 void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
674 // to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
675 // especially when used as a break condition, such as while (MPI_Iprobe(...)) dostuff...
676 // nsleeps is a multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
677 // This can speed up the execution of certain applications by an order of magnitude, such as HPL
678 static int nsleeps = 1;
679 double speed = simgrid::s4u::Actor::self()->get_host()->getSpeed();
680 double maxrate = simgrid::config::get_value<double>("smpi/iprobe-cpu-usage");
681 MPI_Request request = new Request(nullptr, 0, MPI_CHAR,
682 source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(source)->get_pid(),
683 simgrid::s4u::this_actor::get_pid(), tag, comm, MPI_REQ_PERSISTENT | MPI_REQ_RECV);
684 if (smpi_iprobe_sleep > 0) {
685 smx_activity_t iprobe_sleep = simcall_execution_start(
686 "iprobe", /* flops to executek*/ nsleeps * smpi_iprobe_sleep * speed * maxrate, /* priority */ 1.0,
687 /* performance bound */ maxrate * speed, smpi_process()->get_actor()->get_host());
688 simcall_execution_wait(iprobe_sleep);
690 // behave like a receive, but don't do it
691 smx_mailbox_t mailbox;
693 request->print_request("New iprobe");
694 // We have to test both mailboxes as we don't know if we will receive one one or another
695 if (simgrid::config::get_value<int>("smpi/async-small-thresh") > 0) {
696 mailbox = smpi_process()->mailbox_small();
697 XBT_DEBUG("Trying to probe the perm recv mailbox");
698 request->action_ = simcall_comm_iprobe(mailbox, 0, &match_recv, static_cast<void*>(request));
701 if (request->action_ == nullptr){
702 mailbox = smpi_process()->mailbox();
703 XBT_DEBUG("trying to probe the other mailbox");
704 request->action_ = simcall_comm_iprobe(mailbox, 0, &match_recv, static_cast<void*>(request));
707 if (request->action_ != nullptr){
708 simgrid::kernel::activity::CommImplPtr sync_comm =
709 boost::static_pointer_cast<simgrid::kernel::activity::CommImpl>(request->action_);
710 MPI_Request req = static_cast<MPI_Request>(sync_comm->src_data);
712 if (status != MPI_STATUS_IGNORE && (req->flags_ & MPI_REQ_PREPARED) == 0) {
713 status->MPI_SOURCE = comm->group()->rank(req->src_);
714 status->MPI_TAG = req->tag_;
715 status->MPI_ERROR = MPI_SUCCESS;
716 status->count = req->real_size_;
718 nsleeps = 1;//reset the number of sleeps we will do next time
722 if (simgrid::config::get_value<bool>("smpi/grow-injected-times"))
728 void Request::finish_wait(MPI_Request* request, MPI_Status * status)
730 MPI_Request req = *request;
731 Status::empty(status);
733 if (req->cancelled_==1){
734 if (status!=MPI_STATUS_IGNORE)
739 if (not((req->detached_ != 0) && ((req->flags_ & MPI_REQ_SEND) != 0)) && ((req->flags_ & MPI_REQ_PREPARED) == 0)) {
740 if(status != MPI_STATUS_IGNORE) {
741 int src = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_;
742 status->MPI_SOURCE = req->comm_->group()->rank(src);
743 status->MPI_TAG = req->tag_ == MPI_ANY_TAG ? req->real_tag_ : req->tag_;
744 status->MPI_ERROR = req->truncated_ != 0 ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
745 // this handles the case were size in receive differs from size in send
746 status->count = req->real_size_;
749 req->print_request("Finishing");
750 MPI_Datatype datatype = req->old_type_;
752 // FIXME Handle the case of a partial shared malloc.
753 if (((req->flags_ & MPI_REQ_ACCUMULATE) != 0) ||
754 (datatype->flags() & DT_FLAG_DERIVED)) { // && (not smpi_is_shared(req->old_buf_))){
756 if (not smpi_process()->replaying() && smpi_privatize_global_variables != SmpiPrivStrategies::None &&
757 static_cast<char*>(req->old_buf_) >= smpi_data_exe_start &&
758 static_cast<char*>(req->old_buf_) < smpi_data_exe_start + smpi_data_exe_size) {
759 XBT_VERB("Privatization : We are unserializing to a zone in global memory Switch data segment ");
760 smpi_switch_data_segment(simgrid::s4u::Actor::self());
763 if(datatype->flags() & DT_FLAG_DERIVED){
764 // This part handles the problem of non-contignous memory the unserialization at the reception
765 if ((req->flags_ & MPI_REQ_RECV) && datatype->size() != 0)
766 datatype->unserialize(req->buf_, req->old_buf_, req->real_size_/datatype->size() , req->op_);
768 } else if (req->flags_ & MPI_REQ_RECV) { // apply op on contiguous buffer for accumulate
769 if (datatype->size() != 0) {
770 int n = req->real_size_ / datatype->size();
771 req->op_->apply(req->buf_, req->old_buf_, &n, datatype);
778 if (TRACE_smpi_view_internals() && ((req->flags_ & MPI_REQ_RECV) != 0)) {
779 int rank = simgrid::s4u::this_actor::get_pid();
780 int src_traced = (req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_);
781 TRACE_smpi_recv(src_traced, rank,req->tag_);
783 if(req->detached_sender_ != nullptr){
784 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
786 simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::Host>()->orecv(req->real_size());
788 simcall_process_sleep(sleeptime);
789 XBT_DEBUG("receiving size of %zu : sleep %f ", req->real_size_, sleeptime);
791 unref(&(req->detached_sender_));
793 if (req->flags_ & MPI_REQ_PERSISTENT)
794 req->action_ = nullptr;
795 req->flags_ |= MPI_REQ_FINISHED;
799 void Request::wait(MPI_Request * request, MPI_Status * status)
801 (*request)->print_request("Waiting");
802 if ((*request)->flags_ & MPI_REQ_PREPARED) {
803 Status::empty(status);
807 if ((*request)->action_ != nullptr){
809 // this is not a detached send
810 simcall_comm_wait((*request)->action_, -1.0);
812 XBT_VERB("Request cancelled");
817 finish_wait(request,status);
818 if (*request != MPI_REQUEST_NULL && (((*request)->flags_ & MPI_REQ_NON_PERSISTENT) != 0))
819 *request = MPI_REQUEST_NULL;
822 int Request::waitany(int count, MPI_Request requests[], MPI_Status * status)
824 s_xbt_dynar_t comms; // Keep it on stack to save some extra mallocs
825 int index = MPI_UNDEFINED;
829 // Wait for a request to complete
830 xbt_dynar_init(&comms, sizeof(smx_activity_t), [](void*ptr){
831 intrusive_ptr_release(*(simgrid::kernel::activity::ActivityImpl**)ptr);
833 int *map = xbt_new(int, count);
834 XBT_DEBUG("Wait for one of %d", count);
835 for(int i = 0; i < count; i++) {
836 if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & MPI_REQ_PREPARED) &&
837 not(requests[i]->flags_ & MPI_REQ_FINISHED)) {
838 if (requests[i]->action_ != nullptr) {
839 XBT_DEBUG("Waiting any %p ", requests[i]);
840 intrusive_ptr_add_ref(requests[i]->action_.get());
841 xbt_dynar_push_as(&comms, simgrid::kernel::activity::ActivityImpl*, requests[i]->action_.get());
845 // This is a finished detached request, let's return this one
846 size = 0; // so we free the dynar but don't do the waitany call
848 finish_wait(&requests[i], status); // cleanup if refcount = 0
849 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
850 requests[i] = MPI_REQUEST_NULL; // set to null
856 XBT_DEBUG("Enter waitany for %lu comms", xbt_dynar_length(&comms));
859 // this is not a detached send
860 i = simcall_comm_waitany(&comms, -1);
862 XBT_INFO("request %d cancelled ",i);
866 // not MPI_UNDEFINED, as this is a simix return code
869 //in case of an accumulate, we have to wait the end of all requests to apply the operation, ordered correctly.
870 if ((requests[index] == MPI_REQUEST_NULL) ||
871 (not((requests[index]->flags_ & MPI_REQ_ACCUMULATE) && (requests[index]->flags_ & MPI_REQ_RECV)))) {
872 finish_wait(&requests[index],status);
873 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
874 requests[index] = MPI_REQUEST_NULL;
879 xbt_dynar_free_data(&comms);
883 if (index==MPI_UNDEFINED)
884 Status::empty(status);
889 static int sort_accumulates(MPI_Request a, MPI_Request b)
891 return (a->tag() > b->tag());
894 int Request::waitall(int count, MPI_Request requests[], MPI_Status status[])
896 std::vector<MPI_Request> accumulates;
899 MPI_Status *pstat = (status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat);
900 int retvalue = MPI_SUCCESS;
901 //tag invalid requests in the set
902 if (status != MPI_STATUSES_IGNORE) {
903 for (int c = 0; c < count; c++) {
904 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst_ == MPI_PROC_NULL ||
905 (requests[c]->flags_ & MPI_REQ_PREPARED)) {
906 Status::empty(&status[c]);
907 } else if (requests[c]->src_ == MPI_PROC_NULL) {
908 Status::empty(&status[c]);
909 status[c].MPI_SOURCE = MPI_PROC_NULL;
913 for (int c = 0; c < count; c++) {
914 if (MC_is_active() || MC_record_replay_is_active()) {
915 wait(&requests[c],pstat);
918 index = waitany(count, (MPI_Request*)requests, pstat);
920 if (index == MPI_UNDEFINED)
923 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_RECV) &&
924 (requests[index]->flags_ & MPI_REQ_ACCUMULATE))
925 accumulates.push_back(requests[index]);
926 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_NON_PERSISTENT))
927 requests[index] = MPI_REQUEST_NULL;
929 if (status != MPI_STATUSES_IGNORE) {
930 status[index] = *pstat;
931 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
932 retvalue = MPI_ERR_IN_STATUS;
936 if (not accumulates.empty()) {
937 std::sort(accumulates.begin(), accumulates.end(), sort_accumulates);
938 for (auto& req : accumulates) {
939 finish_wait(&req, status);
946 int Request::waitsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
950 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
952 for (int i = 0; i < incount; i++) {
953 int index = waitany(incount, requests, pstat);
954 if(index!=MPI_UNDEFINED){
955 indices[count] = index;
957 if(status != MPI_STATUSES_IGNORE) {
958 status[index] = *pstat;
960 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_NON_PERSISTENT))
961 requests[index] = MPI_REQUEST_NULL;
963 return MPI_UNDEFINED;
969 MPI_Request Request::f2c(int id) {
971 if(id==MPI_FORTRAN_REQUEST_NULL)
972 return static_cast<MPI_Request>(MPI_REQUEST_NULL);
973 return static_cast<MPI_Request>(F2C::f2c_lookup()->at(get_key_id(key, id)));
978 if (F2C::f2c_lookup() == nullptr) {
979 F2C::set_f2c_lookup(new std::unordered_map<std::string, F2C*>);
982 (*(F2C::f2c_lookup()))[get_key_id(key, F2C::f2c_id())] = this;
983 F2C::f2c_id_increment();
984 return F2C::f2c_id()-1;
987 void Request::free_f(int id)
989 if (id != MPI_FORTRAN_REQUEST_NULL) {
991 F2C::f2c_lookup()->erase(get_key_id(key, id));