1 /* Copyright (c) 2010-2022. 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_comm.hpp"
7 #include "simgrid/host.h"
8 #include "smpi_coll.hpp"
9 #include "smpi_datatype.hpp"
10 #include "smpi_info.hpp"
11 #include "smpi_request.hpp"
12 #include "smpi_win.hpp"
13 #include "src/smpi/include/smpi_actor.hpp"
14 #include "src/surf/HostImpl.hpp"
18 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_comm, smpi, "Logging specific to SMPI (comm)");
20 simgrid::smpi::Comm smpi_MPI_COMM_UNINITIALIZED;
21 MPI_Comm MPI_COMM_UNINITIALIZED=&smpi_MPI_COMM_UNINITIALIZED;
23 * Setting MPI_COMM_WORLD to MPI_COMM_UNINITIALIZED (it's a variable)
24 * is important because the implementation of MPI_Comm checks
25 * "this == MPI_COMM_UNINITIALIZED"? If yes, it uses smpi_process()->comm_world()
27 * This is basically how we only have one global variable but all processes have
28 * different communicators (the one their SMPI instance uses).
31 MPI_Comm MPI_COMM_WORLD = MPI_COMM_UNINITIALIZED;
33 /* Support for cartesian topology was added, but there are 2 other types of topology, graph et dist graph. In order to
34 * support them, we have to add a field SMPI_Topo_type, and replace the MPI_Topology field by an union. */
39 std::unordered_map<int, smpi_key_elem> Comm::keyvals_;
40 int Comm::keyval_id_=0;
42 Comm::Comm(MPI_Group group, MPI_Topology topo, bool smp, int in_id)
43 : group_(group), topo_(topo), is_smp_comm_(smp), id_(in_id)
46 //First creation of comm is done before SIMIX_run, so only do comms for others
47 if(in_id==MPI_UNDEFINED && smp==0 && this->rank()!=MPI_UNDEFINED ){
52 static int global_id_ = 0;
56 colls::bcast(&id, 1, MPI_INT, 0, this);
57 XBT_DEBUG("Communicator %p has id %d", this, id);
58 id_=id;//only set here, as we don't want to change it in the middle of the bcast
62 void Comm::destroy(Comm* comm)
64 if (comm == MPI_COMM_UNINITIALIZED){
65 Comm::destroy(smpi_process()->comm_world());
68 if (comm != MPI_COMM_WORLD && not comm->deleted()) {
69 comm->cleanup_attr<Comm>();
70 comm->mark_as_deleted();
75 int Comm::dup(MPI_Comm* newcomm){
76 // we need to switch as the called function may silently touch global variables
77 smpi_switch_data_segment(s4u::Actor::self());
79 auto* cp = new Group(this->group());
80 (*newcomm) = new Comm(cp, this->topo());
82 for (auto const& it : attributes()) {
83 auto elem_it = keyvals_.find(it.first);
84 xbt_assert(elem_it != keyvals_.end(), "Keyval not found for Comm: %d", it.first);
86 smpi_key_elem& elem = elem_it->second;
87 int ret = MPI_SUCCESS;
89 void* value_out = nullptr;
90 if (elem.copy_fn.comm_copy_fn == MPI_COMM_DUP_FN) {
91 value_out = it.second;
93 } else if (elem.copy_fn.comm_copy_fn != MPI_NULL_COPY_FN) {
94 ret = elem.copy_fn.comm_copy_fn(this, it.first, elem.extra_state, it.second, &value_out, &flag);
96 if (elem.copy_fn.comm_copy_fn_fort != MPI_NULL_COPY_FN) {
97 value_out = xbt_new(int, 1);
98 if (*(int*)*elem.copy_fn.comm_copy_fn_fort == 1) { // MPI_COMM_DUP_FN
99 memcpy(value_out, it.second, sizeof(int));
101 } else { // not null, nor dup
102 elem.copy_fn.comm_copy_fn_fort(this, it.first, elem.extra_state, it.second, value_out, &flag, &ret);
104 if (ret != MPI_SUCCESS)
107 if (ret != MPI_SUCCESS) {
108 Comm::destroy(*newcomm);
109 *newcomm = MPI_COMM_NULL;
114 (*newcomm)->attributes().emplace(it.first, value_out);
117 //duplicate info if present
118 if(info_!=MPI_INFO_NULL)
119 (*newcomm)->info_ = new simgrid::smpi::Info(info_);
120 //duplicate errhandler
121 if (errhandlers_ != nullptr)//MPI_COMM_WORLD, only grab our own
122 (*newcomm)->set_errhandler(errhandlers_[this->rank()]);
124 (*newcomm)->set_errhandler(errhandler_);
128 int Comm::dup_with_info(MPI_Info info, MPI_Comm* newcomm){
129 int ret = dup(newcomm);
130 if(ret != MPI_SUCCESS)
132 if((*newcomm)->info_!=MPI_INFO_NULL){
133 simgrid::smpi::Info::unref((*newcomm)->info_);
134 (*newcomm)->info_=MPI_INFO_NULL;
136 if(info != MPI_INFO_NULL){
138 (*newcomm)->info_=info;
143 MPI_Group Comm::group()
145 if (this == MPI_COMM_UNINITIALIZED)
146 return smpi_process()->comm_world()->group();
150 int Comm::size() const
152 if (this == MPI_COMM_UNINITIALIZED)
153 return smpi_process()->comm_world()->size();
154 return group_->size();
157 int Comm::rank() const
159 if (this == MPI_COMM_UNINITIALIZED)
160 return smpi_process()->comm_world()->rank();
161 return group_->rank(s4u::this_actor::get_pid());
169 void Comm::get_name(char* name, int* len) const
171 if (this == MPI_COMM_UNINITIALIZED){
172 smpi_process()->comm_world()->get_name(name, len);
175 if(this == MPI_COMM_WORLD && name_.empty()) {
176 strncpy(name, "MPI_COMM_WORLD", 15);
179 *len = snprintf(name, MPI_MAX_NAME_STRING+1, "%s", name_.c_str());
183 std::string Comm::name() const
186 std::array<char, MPI_MAX_NAME_STRING + 1> name;
187 this->get_name(name.data(), &size);
189 return std::string("MPI_Comm");
191 return std::string(name.data());
195 void Comm::set_name (const char* name)
197 if (this == MPI_COMM_UNINITIALIZED){
198 smpi_process()->comm_world()->set_name(name);
201 name_.replace (0, MPI_MAX_NAME_STRING+1, name);
205 void Comm::set_leaders_comm(MPI_Comm leaders){
206 if (this == MPI_COMM_UNINITIALIZED){
207 smpi_process()->comm_world()->set_leaders_comm(leaders);
210 leaders_comm_=leaders;
213 int* Comm::get_non_uniform_map() const
215 if (this == MPI_COMM_UNINITIALIZED)
216 return smpi_process()->comm_world()->get_non_uniform_map();
217 return non_uniform_map_;
220 int* Comm::get_leaders_map() const
222 if (this == MPI_COMM_UNINITIALIZED)
223 return smpi_process()->comm_world()->get_leaders_map();
227 MPI_Comm Comm::get_leaders_comm() const
229 if (this == MPI_COMM_UNINITIALIZED)
230 return smpi_process()->comm_world()->get_leaders_comm();
231 return leaders_comm_;
234 MPI_Comm Comm::get_intra_comm() const
236 if (this == MPI_COMM_UNINITIALIZED || this==MPI_COMM_WORLD)
237 return smpi_process()->comm_intra();
238 else return intra_comm_;
241 bool Comm::is_uniform() const
243 if (this == MPI_COMM_UNINITIALIZED)
244 return smpi_process()->comm_world()->is_uniform();
248 bool Comm::is_blocked() const
250 if (this == MPI_COMM_UNINITIALIZED)
251 return smpi_process()->comm_world()->is_blocked();
255 bool Comm::is_smp_comm() const
257 if (this == MPI_COMM_UNINITIALIZED)
258 return smpi_process()->comm_world()->is_smp_comm();
262 MPI_Comm Comm::split(int color, int key)
264 if (this == MPI_COMM_UNINITIALIZED)
265 return smpi_process()->comm_world()->split(color, key);
266 int system_tag = -123;
268 MPI_Group group_root = nullptr;
269 MPI_Group group_out = nullptr;
270 const Group* group = this->group();
271 int myrank = this->rank();
272 int size = this->size();
273 /* Gather all colors and keys on rank 0 */
274 const std::array<int, 2> sendbuf = {{color, key}};
275 std::vector<int> recvbuf;
277 recvbuf.resize(2 * size);
278 gather__default(sendbuf.data(), 2, MPI_INT, recvbuf.data(), 2, MPI_INT, 0, this);
279 /* Do the actual job */
281 std::vector<MPI_Group> group_snd(size);
282 std::vector<std::pair<int, int>> rankmap;
283 rankmap.reserve(size);
284 for (int i = 0; i < size; i++) {
285 if (recvbuf[2 * i] != MPI_UNDEFINED) {
287 for (int j = i + 1; j < size; j++) {
288 if(recvbuf[2 * i] == recvbuf[2 * j]) {
289 recvbuf[2 * j] = MPI_UNDEFINED;
290 rankmap.emplace_back(recvbuf[2 * j + 1], j);
293 /* Add self in the group */
294 recvbuf[2 * i] = MPI_UNDEFINED;
295 rankmap.emplace_back(recvbuf[2 * i + 1], i);
296 std::sort(begin(rankmap), end(rankmap));
297 group_out = new Group(rankmap.size());
299 group_root = group_out; /* Save root's group */
301 for (unsigned j = 0; j < rankmap.size(); j++) {
302 aid_t actor = group->actor(rankmap[j].second);
303 group_out->set_mapping(actor, j);
305 std::vector<MPI_Request> requests(rankmap.size());
307 for (auto const& rank : rankmap) {
308 if (rank.second != 0) {
309 group_snd[reqs]=new Group(group_out);
310 requests[reqs] = Request::isend(&(group_snd[reqs]), 1, MPI_PTR, rank.second, system_tag, this);
314 if(i != 0 && group_out != MPI_COMM_WORLD->group() && group_out != MPI_GROUP_EMPTY)
315 Group::unref(group_out);
317 Request::waitall(reqs, requests.data(), MPI_STATUS_IGNORE);
320 group_out = group_root; /* exit with root's group */
322 if(color != MPI_UNDEFINED) {
323 Request::recv(&group_out, 1, MPI_PTR, 0, system_tag, this, MPI_STATUS_IGNORE);
324 } /* otherwise, exit with group_out == nullptr */
326 return group_out!=nullptr ? new Comm(group_out, topo_) : MPI_COMM_NULL;
330 if (this == MPI_COMM_UNINITIALIZED){
331 smpi_process()->comm_world()->ref();
338 void Comm::cleanup_smp(){
339 if (intra_comm_ != MPI_COMM_NULL)
340 Comm::unref(intra_comm_);
341 if (leaders_comm_ != MPI_COMM_NULL)
342 Comm::unref(leaders_comm_);
343 xbt_free(non_uniform_map_);
344 delete[] leaders_map_;
347 void Comm::unref(Comm* comm){
348 if (comm == MPI_COMM_UNINITIALIZED){
349 Comm::unref(smpi_process()->comm_world());
354 if(comm->refcount_==0){
355 if(simgrid::smpi::F2C::lookup() != nullptr)
356 F2C::free_f(comm->f2c_id());
358 comm->cleanup_attr<Comm>();
359 if (comm->info_ != MPI_INFO_NULL)
360 simgrid::smpi::Info::unref(comm->info_);
361 if(comm->errhandlers_!=nullptr){
362 for (int i=0; i<comm->size(); i++)
363 if (comm->errhandlers_[i]!=MPI_ERRHANDLER_NULL)
364 simgrid::smpi::Errhandler::unref(comm->errhandlers_[i]);
365 delete[] comm->errhandlers_;
366 } else if (comm->errhandler_ != MPI_ERRHANDLER_NULL)
367 simgrid::smpi::Errhandler::unref(comm->errhandler_);
369 Group::unref(comm->group_);
370 if(comm->refcount_==0)
374 MPI_Comm Comm::find_intra_comm(int * leader){
375 //get the indices of all processes sharing the same simix host
376 int intra_comm_size = 0;
377 aid_t min_index = std::numeric_limits<aid_t>::max(); // the minimum index will be the leader
378 sg_host_self()->get_impl()->foreach_actor([this, &intra_comm_size, &min_index](auto& actor) {
379 aid_t index = actor.get_pid();
380 if (this->group()->rank(index) != MPI_UNDEFINED) { // Is this process in the current group?
382 if (index < min_index)
386 XBT_DEBUG("number of processes deployed on my node : %d", intra_comm_size);
387 auto* group_intra = new Group(intra_comm_size);
389 sg_host_self()->get_impl()->foreach_actor([this, group_intra, &i](auto& actor) {
390 if (this->group()->rank(actor.get_pid()) != MPI_UNDEFINED) {
391 group_intra->set_mapping(actor.get_pid(), i);
396 return new Comm(group_intra, nullptr, true);
399 void Comm::init_smp(){
402 if (this == MPI_COMM_UNINITIALIZED)
403 smpi_process()->comm_world()->init_smp();
405 int comm_size = this->size();
407 // If we are in replay - perform an ugly hack
408 // tell SimGrid we are not in replay for a while, because we need the buffers to be copied for the following calls
409 bool replaying = false; //cache data to set it back again after
410 if(smpi_process()->replaying()){
412 smpi_process()->set_replaying(false);
415 // we need to switch as the called function may silently touch global variables
416 smpi_switch_data_segment(s4u::Actor::self());
418 // identify neighbors in comm
419 MPI_Comm comm_intra = find_intra_comm(&leader);
421 auto* leaders_map = new int[comm_size];
422 auto* leader_list = new int[comm_size];
423 std::fill_n(leaders_map, comm_size, 0);
424 std::fill_n(leader_list, comm_size, -1);
426 allgather__ring(&leader, 1, MPI_INT , leaders_map, 1, MPI_INT, this);
428 if(leaders_map_==nullptr){
429 leaders_map_= leaders_map;
431 delete[] leaders_map;
433 int leader_group_size = 0;
434 for(i=0; i<comm_size; i++){
435 int already_done = 0;
436 for (int j = 0; j < leader_group_size; j++) {
437 if (leaders_map_[i] == leader_list[j]) {
441 if (already_done == 0) {
442 leader_list[leader_group_size] = leaders_map_[i];
446 xbt_assert(leader_group_size > 0);
447 std::sort(leader_list, leader_list + leader_group_size);
449 auto* leaders_group = new Group(leader_group_size);
451 MPI_Comm leader_comm = MPI_COMM_NULL;
452 if(MPI_COMM_WORLD!=MPI_COMM_UNINITIALIZED && this!=MPI_COMM_WORLD){
453 //create leader_communicator
454 for (i=0; i< leader_group_size;i++)
455 leaders_group->set_mapping(leader_list[i], i);
456 leader_comm = new Comm(leaders_group, nullptr, true);
457 this->set_leaders_comm(leader_comm);
458 this->set_intra_comm(comm_intra);
460 // create intracommunicator
462 for (i=0; i< leader_group_size;i++)
463 leaders_group->set_mapping(leader_list[i], i);
465 if(this->get_leaders_comm()==MPI_COMM_NULL){
466 leader_comm = new Comm(leaders_group, nullptr, true);
467 this->set_leaders_comm(leader_comm);
469 leader_comm=this->get_leaders_comm();
470 Group::unref(leaders_group);
472 smpi_process()->set_comm_intra(comm_intra);
475 // Are the nodes uniform ? = same number of process/node
476 int my_local_size=comm_intra->size();
478 if(comm_intra->rank()==0) {
480 auto* non_uniform_map = xbt_new0(int, leader_group_size);
481 allgather__ring(&my_local_size, 1, MPI_INT,
482 non_uniform_map, 1, MPI_INT, leader_comm);
483 for(i=0; i < leader_group_size; i++) {
484 if(non_uniform_map[0] != non_uniform_map[i]) {
489 if (is_uniform == 0 && this->is_uniform()) {
490 non_uniform_map_ = non_uniform_map;
492 xbt_free(non_uniform_map);
495 bcast__scatter_LR_allgather(&is_uniform, 1, MPI_INT, 0, comm_intra);
496 is_uniform_ = (is_uniform != 0);
498 // we need to switch as the called function may silently touch global variables
499 smpi_switch_data_segment(s4u::Actor::self());
501 // Are the ranks blocked ? = allocated contiguously on the SMP nodes
503 int prev = this->group()->rank(comm_intra->group()->actor(0));
504 for (i = 1; i < my_local_size; i++) {
505 int that = this->group()->rank(comm_intra->group()->actor(i));
506 if (that != prev + 1) {
514 allreduce__default(&is_blocked, &global_blocked, 1, MPI_INT, MPI_LAND, this);
516 if ((MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED && this != MPI_COMM_WORLD) || this->rank() == 0)
517 is_blocked_ = (global_blocked != 0);
518 delete[] leader_list;
521 smpi_process()->set_replaying(true);
524 MPI_Comm Comm::f2c(int id) {
526 return MPI_COMM_SELF;
528 return MPI_COMM_WORLD;
529 } else if (F2C::lookup() != nullptr && id >= 0) {
530 const auto& lookup = F2C::lookup();
531 auto comm = lookup->find(id);
532 return comm == lookup->end() ? MPI_COMM_NULL : static_cast<MPI_Comm>(comm->second);
534 return MPI_COMM_NULL;
538 void Comm::free_f(int id) {
539 F2C::lookup()->erase(id);
542 void Comm::add_rma_win(MPI_Win win){
543 rma_wins_.push_back(win);
546 void Comm::remove_rma_win(MPI_Win win){
547 rma_wins_.remove(win);
550 void Comm::finish_rma_calls() const
552 const int myrank = rank();
553 for (auto const& it : rma_wins_) {
554 if (it->rank() == myrank) { // is it ours (for MPI_COMM_WORLD)?
555 int finished = it->finish_comms();
556 XBT_DEBUG("Barrier for rank %d - Finished %d RMA calls", myrank, finished);
561 MPI_Info Comm::info()
566 void Comm::set_info(MPI_Info info)
568 if (info_ != MPI_INFO_NULL)
569 simgrid::smpi::Info::unref(info);
571 if (info_ != MPI_INFO_NULL)
575 MPI_Errhandler Comm::errhandler()
577 if (this != MPI_COMM_WORLD){
578 if (errhandler_ != MPI_ERRHANDLER_NULL)
582 if(errhandlers_==nullptr){
583 if (_smpi_cfg_default_errhandler_is_error)
584 return MPI_ERRORS_ARE_FATAL;
586 return MPI_ERRORS_RETURN;
588 if(errhandlers_[this->rank()] != MPI_ERRHANDLER_NULL)
589 errhandlers_[this->rank()]->ref();
590 return errhandlers_[this->rank()];
595 void Comm::set_errhandler(MPI_Errhandler errhandler)
597 if(this != MPI_COMM_WORLD){
598 if (errhandler_ != MPI_ERRHANDLER_NULL)
599 simgrid::smpi::Errhandler::unref(errhandler_);
600 errhandler_ = errhandler;
602 if(errhandlers_==nullptr)
603 errhandlers_= new MPI_Errhandler[this->size()]{MPI_ERRHANDLER_NULL};
604 if(errhandlers_[this->rank()] != MPI_ERRHANDLER_NULL)
605 simgrid::smpi::Errhandler::unref(errhandlers_[this->rank()]);
606 errhandlers_[this->rank()]=errhandler;
608 if (errhandler != MPI_ERRHANDLER_NULL)
612 MPI_Comm Comm::split_type(int type, int /*key*/, const Info*)
614 //MPI_UNDEFINED can be given to some nodes... but we need them to still perform the smp part which is collective
615 if(type != MPI_COMM_TYPE_SHARED && type != MPI_UNDEFINED){
616 return MPI_COMM_NULL;
619 MPI_Comm res= this->find_intra_comm(&leader);
620 if(type != MPI_UNDEFINED)
623 xbt_assert(res->refcount_ == 1); // ensure the next call to Comm::destroy really frees the comm
625 return MPI_COMM_NULL;
629 static inline std::string hash_message(int src, int dst, int tag){
630 return std::to_string(tag) + '_' + std::to_string(src) + '_' + std::to_string(dst);
633 unsigned int Comm::get_sent_messages_count(int src, int dst, int tag)
635 return sent_messages_[hash_message(src, dst, tag)];
638 void Comm::increment_sent_messages_count(int src, int dst, int tag)
640 sent_messages_[hash_message(src, dst, tag)]++;
643 unsigned int Comm::get_received_messages_count(int src, int dst, int tag)
645 return recv_messages_[hash_message(src, dst, tag)];
648 void Comm::increment_received_messages_count(int src, int dst, int tag)
650 recv_messages_[hash_message(src, dst, tag)]++;
653 unsigned int Comm::get_collectives_count()
655 if (this==MPI_COMM_UNINITIALIZED){
656 return smpi_process()->comm_world()->get_collectives_count();
657 }else if(this == MPI_COMM_WORLD || this == smpi_process()->comm_world()){
658 if (collectives_counts_.empty())
659 collectives_counts_.resize(this->size());
660 return collectives_counts_[this->rank()];
662 return collectives_count_;
666 void Comm::increment_collectives_count()
668 if (this==MPI_COMM_UNINITIALIZED){
669 smpi_process()->comm_world()->increment_collectives_count();
670 }else if (this == MPI_COMM_WORLD || this == smpi_process()->comm_world()){
671 if (collectives_counts_.empty())
672 collectives_counts_.resize(this->size());
673 collectives_counts_[this->rank()]++;
675 collectives_count_++;
680 } // namespace simgrid