1 /* Copyright (c) 2010-2021. 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 "smpi_coll.hpp"
8 #include "smpi_datatype.hpp"
9 #include "smpi_request.hpp"
10 #include "smpi_win.hpp"
11 #include "smpi_info.hpp"
12 #include "src/smpi/include/smpi_actor.hpp"
13 #include "src/surf/HostImpl.hpp"
17 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_comm, smpi, "Logging specific to SMPI (comm)");
19 simgrid::smpi::Comm smpi_MPI_COMM_UNINITIALIZED;
20 MPI_Comm MPI_COMM_UNINITIALIZED=&smpi_MPI_COMM_UNINITIALIZED;
22 * Setting MPI_COMM_WORLD to MPI_COMM_UNINITIALIZED (it's a variable)
23 * is important because the implementation of MPI_Comm checks
24 * "this == MPI_COMM_UNINITIALIZED"? If yes, it uses smpi_process()->comm_world()
26 * This is basically how we only have one global variable but all processes have
27 * different communicators (the one their SMPI instance uses).
30 MPI_Comm MPI_COMM_WORLD = MPI_COMM_UNINITIALIZED;
32 /* Support for cartesian topology was added, but there are 2 other types of topology, graph et dist graph. In order to
33 * support them, we have to add a field SMPI_Topo_type, and replace the MPI_Topology field by an union. */
38 std::unordered_map<int, smpi_key_elem> Comm::keyvals_;
39 int Comm::keyval_id_=0;
41 Comm::Comm(MPI_Group group, MPI_Topology topo, bool smp, int in_id)
42 : group_(group), topo_(topo), is_smp_comm_(smp), id_(in_id)
45 //First creation of comm is done before SIMIX_run, so only do comms for others
46 if(in_id==MPI_UNDEFINED && smp==0 && this->rank()!=MPI_UNDEFINED ){
51 static int global_id_ = 0;
55 colls::bcast(&id, 1, MPI_INT, 0, this);
56 XBT_DEBUG("Communicator %p has id %d", this, id);
57 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());
71 int Comm::dup(MPI_Comm* newcomm){
72 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) {
73 // we need to switch as the called function may silently touch global variables
74 smpi_switch_data_segment(s4u::Actor::self());
76 auto* cp = new Group(this->group());
77 (*newcomm) = new Comm(cp, this->topo());
78 int ret = MPI_SUCCESS;
80 if (not attributes()->empty()) {
82 void* value_out=nullptr;
83 for (auto const& it : *attributes()) {
84 smpi_key_elem elem = keyvals_.at(it.first);
86 if( elem->copy_fn.comm_copy_fn != MPI_NULL_COPY_FN &&
87 elem->copy_fn.comm_copy_fn != MPI_COMM_DUP_FN)
88 ret = elem->copy_fn.comm_copy_fn(this, it.first, elem->extra_state, it.second, &value_out, &flag);
89 else if ( elem->copy_fn.comm_copy_fn_fort != MPI_NULL_COPY_FN &&
90 *(int*)*elem->copy_fn.comm_copy_fn_fort != 1){
91 value_out=(int*)xbt_malloc(sizeof(int));
92 elem->copy_fn.comm_copy_fn_fort(this, it.first, elem->extra_state, it.second, value_out, &flag,&ret);
94 if (ret != MPI_SUCCESS) {
95 Comm::destroy(*newcomm);
96 *newcomm = MPI_COMM_NULL;
99 if (elem->copy_fn.comm_copy_fn == MPI_COMM_DUP_FN ||
100 ((elem->copy_fn.comm_copy_fn_fort != MPI_NULL_COPY_FN) && *(int*)*elem->copy_fn.comm_copy_fn_fort == 1)){
102 (*newcomm)->attributes()->insert({it.first, it.second});
105 (*newcomm)->attributes()->insert({it.first, value_out});
110 //duplicate info if present
111 if(info_!=MPI_INFO_NULL)
112 (*newcomm)->info_ = new simgrid::smpi::Info(info_);
113 //duplicate errhandler
114 (*newcomm)->set_errhandler(errhandler_);
118 int Comm::dup_with_info(MPI_Info info, MPI_Comm* newcomm){
119 int ret = dup(newcomm);
120 if(ret != MPI_SUCCESS)
122 if((*newcomm)->info_!=MPI_INFO_NULL){
123 simgrid::smpi::Info::unref((*newcomm)->info_);
124 (*newcomm)->info_=MPI_INFO_NULL;
126 if(info != MPI_INFO_NULL){
128 (*newcomm)->info_=info;
133 MPI_Group Comm::group()
135 if (this == MPI_COMM_UNINITIALIZED)
136 return smpi_process()->comm_world()->group();
140 int Comm::size() const
142 if (this == MPI_COMM_UNINITIALIZED)
143 return smpi_process()->comm_world()->size();
144 return group_->size();
147 int Comm::rank() const
149 if (this == MPI_COMM_UNINITIALIZED)
150 return smpi_process()->comm_world()->rank();
151 return group_->rank(s4u::Actor::self());
159 void Comm::get_name(char* name, int* len) const
161 if (this == MPI_COMM_UNINITIALIZED){
162 smpi_process()->comm_world()->get_name(name, len);
165 if(this == MPI_COMM_WORLD && name_.empty()) {
166 strncpy(name, "MPI_COMM_WORLD", 15);
168 } else if(this == MPI_COMM_SELF && name_.empty()) {
169 strncpy(name, "MPI_COMM_SELF", 14);
172 *len = snprintf(name, MPI_MAX_NAME_STRING+1, "%s", name_.c_str());
176 void Comm::set_name (const char* name)
178 if (this == MPI_COMM_UNINITIALIZED){
179 smpi_process()->comm_world()->set_name(name);
182 name_.replace (0, MPI_MAX_NAME_STRING+1, name);
186 void Comm::set_leaders_comm(MPI_Comm leaders){
187 if (this == MPI_COMM_UNINITIALIZED){
188 smpi_process()->comm_world()->set_leaders_comm(leaders);
191 leaders_comm_=leaders;
194 int* Comm::get_non_uniform_map() const
196 if (this == MPI_COMM_UNINITIALIZED)
197 return smpi_process()->comm_world()->get_non_uniform_map();
198 return non_uniform_map_;
201 int* Comm::get_leaders_map() const
203 if (this == MPI_COMM_UNINITIALIZED)
204 return smpi_process()->comm_world()->get_leaders_map();
208 MPI_Comm Comm::get_leaders_comm() const
210 if (this == MPI_COMM_UNINITIALIZED)
211 return smpi_process()->comm_world()->get_leaders_comm();
212 return leaders_comm_;
215 MPI_Comm Comm::get_intra_comm() const
217 if (this == MPI_COMM_UNINITIALIZED || this==MPI_COMM_WORLD)
218 return smpi_process()->comm_intra();
219 else return intra_comm_;
222 bool Comm::is_uniform() const
224 if (this == MPI_COMM_UNINITIALIZED)
225 return smpi_process()->comm_world()->is_uniform();
226 return is_uniform_ != 0;
229 bool Comm::is_blocked() const
231 if (this == MPI_COMM_UNINITIALIZED)
232 return smpi_process()->comm_world()->is_blocked();
233 return is_blocked_ != 0;
236 bool Comm::is_smp_comm() const
238 if (this == MPI_COMM_UNINITIALIZED)
239 return smpi_process()->comm_world()->is_smp_comm();
243 MPI_Comm Comm::split(int color, int key)
245 if (this == MPI_COMM_UNINITIALIZED)
246 return smpi_process()->comm_world()->split(color, key);
247 int system_tag = -123;
249 MPI_Group group_root = nullptr;
250 MPI_Group group_out = nullptr;
251 MPI_Group group = this->group();
252 int myrank = this->rank();
253 int size = this->size();
254 /* Gather all colors and keys on rank 0 */
255 const std::array<int, 2> sendbuf = {{color, key}};
256 std::vector<int> recvbuf;
258 recvbuf.resize(2 * size);
259 gather__default(sendbuf.data(), 2, MPI_INT, recvbuf.data(), 2, MPI_INT, 0, this);
260 /* Do the actual job */
262 std::vector<MPI_Group> group_snd(size);
263 std::vector<std::pair<int, int>> rankmap;
264 rankmap.reserve(size);
265 for (int i = 0; i < size; i++) {
266 if (recvbuf[2 * i] != MPI_UNDEFINED) {
268 for (int j = i + 1; j < size; j++) {
269 if(recvbuf[2 * i] == recvbuf[2 * j]) {
270 recvbuf[2 * j] = MPI_UNDEFINED;
271 rankmap.emplace_back(recvbuf[2 * j + 1], j);
274 /* Add self in the group */
275 recvbuf[2 * i] = MPI_UNDEFINED;
276 rankmap.emplace_back(recvbuf[2 * i + 1], i);
277 std::sort(begin(rankmap), end(rankmap));
278 group_out = new Group(rankmap.size());
280 group_root = group_out; /* Save root's group */
282 for (unsigned j = 0; j < rankmap.size(); j++) {
283 s4u::Actor* actor = group->actor(rankmap[j].second);
284 group_out->set_mapping(actor, j);
286 std::vector<MPI_Request> requests(rankmap.size());
288 for (auto const& rank : rankmap) {
289 if (rank.second != 0) {
290 group_snd[reqs]=new Group(group_out);
291 requests[reqs] = Request::isend(&(group_snd[reqs]), 1, MPI_PTR, rank.second, system_tag, this);
295 if(i != 0 && group_out != MPI_COMM_WORLD->group() && group_out != MPI_GROUP_EMPTY)
296 Group::unref(group_out);
298 Request::waitall(reqs, requests.data(), MPI_STATUS_IGNORE);
301 group_out = group_root; /* exit with root's group */
303 if(color != MPI_UNDEFINED) {
304 Request::recv(&group_out, 1, MPI_PTR, 0, system_tag, this, MPI_STATUS_IGNORE);
305 } /* otherwise, exit with group_out == nullptr */
307 return group_out!=nullptr ? new Comm(group_out, topo_) : MPI_COMM_NULL;
311 if (this == MPI_COMM_UNINITIALIZED){
312 smpi_process()->comm_world()->ref();
319 void Comm::cleanup_smp(){
320 if (intra_comm_ != MPI_COMM_NULL)
321 Comm::unref(intra_comm_);
322 if (leaders_comm_ != MPI_COMM_NULL)
323 Comm::unref(leaders_comm_);
324 xbt_free(non_uniform_map_);
325 delete[] leaders_map_;
328 void Comm::unref(Comm* comm){
329 if (comm == MPI_COMM_UNINITIALIZED){
330 Comm::unref(smpi_process()->comm_world());
335 if(comm->refcount_==0){
336 if(simgrid::smpi::F2C::lookup() != nullptr)
337 F2C::free_f(comm->c2f());
339 comm->cleanup_attr<Comm>();
340 if (comm->info_ != MPI_INFO_NULL)
341 simgrid::smpi::Info::unref(comm->info_);
342 if(comm->errhandlers_!=nullptr){
343 for (int i=0; i<comm->size(); i++)
344 if (comm->errhandlers_[i]!=MPI_ERRHANDLER_NULL)
345 simgrid::smpi::Errhandler::unref(comm->errhandlers_[i]);
346 delete[] comm->errhandlers_;
347 } else if (comm->errhandler_ != MPI_ERRHANDLER_NULL)
348 simgrid::smpi::Errhandler::unref(comm->errhandler_);
350 Group::unref(comm->group_);
351 if(comm->refcount_==0)
355 MPI_Comm Comm::find_intra_comm(int * leader){
356 //get the indices of all processes sharing the same simix host
357 int intra_comm_size = 0;
358 int min_index = INT_MAX; // the minimum index will be the leader
359 sg_host_self()->pimpl_->foreach_actor([this, &intra_comm_size, &min_index](auto& actor) {
360 int index = actor.get_pid();
361 if (this->group()->rank(actor.get_ciface()) != MPI_UNDEFINED) { // Is this process in the current group?
363 if (index < min_index)
367 XBT_DEBUG("number of processes deployed on my node : %d", intra_comm_size);
368 auto* group_intra = new Group(intra_comm_size);
370 sg_host_self()->pimpl_->foreach_actor([this, group_intra, &i](auto& actor) {
371 if (this->group()->rank(actor.get_ciface()) != MPI_UNDEFINED) {
372 group_intra->set_mapping(actor.get_ciface(), i);
377 return new Comm(group_intra, nullptr, true);
380 void Comm::init_smp(){
383 if (this == MPI_COMM_UNINITIALIZED)
384 smpi_process()->comm_world()->init_smp();
386 int comm_size = this->size();
388 // If we are in replay - perform an ugly hack
389 // tell SimGrid we are not in replay for a while, because we need the buffers to be copied for the following calls
390 bool replaying = false; //cache data to set it back again after
391 if(smpi_process()->replaying()){
393 smpi_process()->set_replaying(false);
396 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) {
397 // we need to switch as the called function may silently touch global variables
398 smpi_switch_data_segment(s4u::Actor::self());
400 // identify neighbors in comm
401 MPI_Comm comm_intra = find_intra_comm(&leader);
403 auto* leaders_map = new int[comm_size];
404 auto* leader_list = new int[comm_size];
405 std::fill_n(leaders_map, comm_size, 0);
406 std::fill_n(leader_list, comm_size, -1);
408 allgather__ring(&leader, 1, MPI_INT , leaders_map, 1, MPI_INT, this);
410 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) {
411 // we need to switch as the called function may silently touch global variables
412 smpi_switch_data_segment(s4u::Actor::self());
415 if(leaders_map_==nullptr){
416 leaders_map_= leaders_map;
418 delete[] leaders_map;
420 int leader_group_size = 0;
421 for(i=0; i<comm_size; i++){
422 int already_done = 0;
423 for (int j = 0; j < leader_group_size; j++) {
424 if (leaders_map_[i] == leader_list[j]) {
428 if (already_done == 0) {
429 leader_list[leader_group_size] = leaders_map_[i];
433 xbt_assert(leader_group_size > 0);
434 std::sort(leader_list, leader_list + leader_group_size);
436 auto* leaders_group = new Group(leader_group_size);
438 MPI_Comm leader_comm = MPI_COMM_NULL;
439 if(MPI_COMM_WORLD!=MPI_COMM_UNINITIALIZED && this!=MPI_COMM_WORLD){
440 //create leader_communicator
441 for (i=0; i< leader_group_size;i++)
442 leaders_group->set_mapping(s4u::Actor::by_pid(leader_list[i]).get(), i);
443 leader_comm = new Comm(leaders_group, nullptr, true);
444 this->set_leaders_comm(leader_comm);
445 this->set_intra_comm(comm_intra);
447 // create intracommunicator
449 for (i=0; i< leader_group_size;i++)
450 leaders_group->set_mapping(s4u::Actor::by_pid(leader_list[i]).get(), i);
452 if(this->get_leaders_comm()==MPI_COMM_NULL){
453 leader_comm = new Comm(leaders_group, nullptr, true);
454 this->set_leaders_comm(leader_comm);
456 leader_comm=this->get_leaders_comm();
457 Group::unref(leaders_group);
459 smpi_process()->set_comm_intra(comm_intra);
462 // Are the nodes uniform ? = same number of process/node
463 int my_local_size=comm_intra->size();
464 if(comm_intra->rank()==0) {
466 int* non_uniform_map = xbt_new0(int,leader_group_size);
467 allgather__ring(&my_local_size, 1, MPI_INT,
468 non_uniform_map, 1, MPI_INT, leader_comm);
469 for(i=0; i < leader_group_size; i++) {
470 if(non_uniform_map[0] != non_uniform_map[i]) {
475 if (is_uniform == 0 && this->is_uniform()) {
476 non_uniform_map_ = non_uniform_map;
478 xbt_free(non_uniform_map);
480 is_uniform_=is_uniform;
482 bcast__scatter_LR_allgather(&(is_uniform_),1, MPI_INT, 0, comm_intra );
484 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) {
485 // we need to switch as the called function may silently touch global variables
486 smpi_switch_data_segment(s4u::Actor::self());
488 // Are the ranks blocked ? = allocated contiguously on the SMP nodes
490 int prev=this->group()->rank(comm_intra->group()->actor(0));
491 for (i = 1; i < my_local_size; i++) {
492 int that = this->group()->rank(comm_intra->group()->actor(i));
493 if (that != prev + 1) {
501 allreduce__default(&is_blocked, &(global_blocked), 1, MPI_INT, MPI_LAND, this);
503 if(MPI_COMM_WORLD==MPI_COMM_UNINITIALIZED || this==MPI_COMM_WORLD){
505 is_blocked_ = global_blocked;
508 is_blocked_=global_blocked;
510 delete[] leader_list;
513 smpi_process()->set_replaying(true);
516 MPI_Comm Comm::f2c(int id) {
518 return MPI_COMM_SELF;
520 return MPI_COMM_WORLD;
521 } else if (F2C::lookup() != nullptr && id >= 0) {
522 const auto& lookup = F2C::lookup();
523 auto comm = lookup->find(id);
524 return comm == lookup->end() ? MPI_COMM_NULL : static_cast<MPI_Comm>(comm->second);
526 return MPI_COMM_NULL;
530 void Comm::free_f(int id) {
531 F2C::lookup()->erase(id);
534 void Comm::add_rma_win(MPI_Win win){
535 rma_wins_.push_back(win);
538 void Comm::remove_rma_win(MPI_Win win){
539 rma_wins_.remove(win);
542 void Comm::finish_rma_calls() const
544 for (auto const& it : rma_wins_) {
545 if(it->rank()==this->rank()){//is it ours (for MPI_COMM_WORLD)?
546 int finished = it->finish_comms();
547 XBT_DEBUG("Barrier for rank %d - Finished %d RMA calls",this->rank(), finished);
552 MPI_Info Comm::info()
554 if (info_ == MPI_INFO_NULL)
560 void Comm::set_info(MPI_Info info)
562 if (info_ != MPI_INFO_NULL)
563 simgrid::smpi::Info::unref(info);
565 if (info_ != MPI_INFO_NULL)
569 MPI_Errhandler Comm::errhandler()
571 if (this != MPI_COMM_WORLD){
572 if (errhandler_ != MPI_ERRHANDLER_NULL)
576 if(errhandlers_==nullptr)
577 return MPI_ERRORS_ARE_FATAL;
579 if(errhandlers_[this->rank()] != MPI_ERRHANDLER_NULL)
580 errhandlers_[this->rank()]->ref();
581 return errhandlers_[this->rank()];
586 void Comm::set_errhandler(MPI_Errhandler errhandler)
588 if(this != MPI_COMM_WORLD){
589 if (errhandler_ != MPI_ERRHANDLER_NULL)
590 simgrid::smpi::Errhandler::unref(errhandler_);
591 errhandler_ = errhandler;
593 if(errhandlers_==nullptr)
594 errhandlers_= new MPI_Errhandler[this->size()]{MPI_ERRHANDLER_NULL};
595 if(errhandlers_[this->rank()] != MPI_ERRHANDLER_NULL)
596 simgrid::smpi::Errhandler::unref(errhandlers_[this->rank()]);
597 errhandlers_[this->rank()]=errhandler;
599 if (errhandler != MPI_ERRHANDLER_NULL)
603 MPI_Comm Comm::split_type(int type, int /*key*/, const Info*)
605 //MPI_UNDEFINED can be given to some nodes... but we need them to still perform the smp part which is collective
606 if(type != MPI_COMM_TYPE_SHARED && type != MPI_UNDEFINED){
607 return MPI_COMM_NULL;
610 MPI_Comm res= this->find_intra_comm(&leader);
611 if(type != MPI_UNDEFINED)
614 xbt_assert(res->refcount_ == 1); // ensure the next call to Comm::destroy really frees the comm
616 return MPI_COMM_NULL;
621 } // namespace simgrid