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 mpi_MPI_COMM_UNINITIALIZED;
20 MPI_Comm MPI_COMM_UNINITIALIZED=&mpi_MPI_COMM_UNINITIALIZED;
22 /* Support for cartesian topology was added, but there are 2 other types of topology, graph et dist graph. In order to
23 * support them, we have to add a field SMPI_Topo_type, and replace the MPI_Topology field by an union. */
28 std::unordered_map<int, smpi_key_elem> Comm::keyvals_;
29 int Comm::keyval_id_=0;
31 Comm::Comm(MPI_Group group, MPI_Topology topo, bool smp, int in_id)
32 : group_(group), topo_(topo), is_smp_comm_(smp), id_(in_id)
35 //First creation of comm is done before SIMIX_run, so only do comms for others
36 if(in_id==MPI_UNDEFINED && smp==0 && this->rank()!=MPI_UNDEFINED ){
41 static int global_id_ = 0;
45 colls::bcast(&id, 1, MPI_INT, 0, this);
46 XBT_DEBUG("Communicator %p has id %d", this, id);
47 id_=id;//only set here, as we don't want to change it in the middle of the bcast
52 void Comm::destroy(Comm* comm)
54 if (comm == MPI_COMM_UNINITIALIZED){
55 Comm::destroy(smpi_process()->comm_world());
61 int Comm::dup(MPI_Comm* newcomm){
62 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) {
63 // we need to switch as the called function may silently touch global variables
64 smpi_switch_data_segment(s4u::Actor::self());
66 auto* cp = new Group(this->group());
67 (*newcomm) = new Comm(cp, this->topo());
68 int ret = MPI_SUCCESS;
70 if (not attributes()->empty()) {
72 void* value_out=nullptr;
73 for (auto const& it : *attributes()) {
74 smpi_key_elem elem = keyvals_.at(it.first);
76 if( elem->copy_fn.comm_copy_fn != MPI_NULL_COPY_FN &&
77 elem->copy_fn.comm_copy_fn != MPI_COMM_DUP_FN)
78 ret = elem->copy_fn.comm_copy_fn(this, it.first, elem->extra_state, it.second, &value_out, &flag);
79 else if ( elem->copy_fn.comm_copy_fn_fort != MPI_NULL_COPY_FN &&
80 *(int*)*elem->copy_fn.comm_copy_fn_fort != 1){
81 value_out=(int*)xbt_malloc(sizeof(int));
82 elem->copy_fn.comm_copy_fn_fort(this, it.first, elem->extra_state, it.second, value_out, &flag,&ret);
84 if (ret != MPI_SUCCESS) {
85 Comm::destroy(*newcomm);
86 *newcomm = MPI_COMM_NULL;
89 if (elem->copy_fn.comm_copy_fn == MPI_COMM_DUP_FN ||
90 ((elem->copy_fn.comm_copy_fn_fort != MPI_NULL_COPY_FN) && *(int*)*elem->copy_fn.comm_copy_fn_fort == 1)){
92 (*newcomm)->attributes()->insert({it.first, it.second});
95 (*newcomm)->attributes()->insert({it.first, value_out});
100 //duplicate info if present
101 if(info_!=MPI_INFO_NULL)
102 (*newcomm)->info_ = new simgrid::smpi::Info(info_);
103 //duplicate errhandler
104 (*newcomm)->set_errhandler(errhandler_);
108 int Comm::dup_with_info(MPI_Info info, MPI_Comm* newcomm){
109 int ret = dup(newcomm);
110 if(ret != MPI_SUCCESS)
112 if((*newcomm)->info_!=MPI_INFO_NULL){
113 simgrid::smpi::Info::unref((*newcomm)->info_);
114 (*newcomm)->info_=MPI_INFO_NULL;
116 if(info != MPI_INFO_NULL){
118 (*newcomm)->info_=info;
123 MPI_Group Comm::group()
125 if (this == MPI_COMM_UNINITIALIZED)
126 return smpi_process()->comm_world()->group();
130 int Comm::size() const
132 if (this == MPI_COMM_UNINITIALIZED)
133 return smpi_process()->comm_world()->size();
134 return group_->size();
137 int Comm::rank() const
139 if (this == MPI_COMM_UNINITIALIZED)
140 return smpi_process()->comm_world()->rank();
141 return group_->rank(s4u::Actor::self());
149 void Comm::get_name(char* name, int* len) const
151 if (this == MPI_COMM_UNINITIALIZED){
152 smpi_process()->comm_world()->get_name(name, len);
155 if(this == MPI_COMM_WORLD && name_.empty()) {
156 strncpy(name, "MPI_COMM_WORLD", 15);
158 } else if(this == MPI_COMM_SELF && name_.empty()) {
159 strncpy(name, "MPI_COMM_SELF", 14);
162 *len = snprintf(name, MPI_MAX_NAME_STRING+1, "%s", name_.c_str());
166 void Comm::set_name (const char* name)
168 if (this == MPI_COMM_UNINITIALIZED){
169 smpi_process()->comm_world()->set_name(name);
172 name_.replace (0, MPI_MAX_NAME_STRING+1, name);
176 void Comm::set_leaders_comm(MPI_Comm leaders){
177 if (this == MPI_COMM_UNINITIALIZED){
178 smpi_process()->comm_world()->set_leaders_comm(leaders);
181 leaders_comm_=leaders;
184 int* Comm::get_non_uniform_map() const
186 if (this == MPI_COMM_UNINITIALIZED)
187 return smpi_process()->comm_world()->get_non_uniform_map();
188 return non_uniform_map_;
191 int* Comm::get_leaders_map() const
193 if (this == MPI_COMM_UNINITIALIZED)
194 return smpi_process()->comm_world()->get_leaders_map();
198 MPI_Comm Comm::get_leaders_comm() const
200 if (this == MPI_COMM_UNINITIALIZED)
201 return smpi_process()->comm_world()->get_leaders_comm();
202 return leaders_comm_;
205 MPI_Comm Comm::get_intra_comm() const
207 if (this == MPI_COMM_UNINITIALIZED || this==MPI_COMM_WORLD)
208 return smpi_process()->comm_intra();
209 else return intra_comm_;
212 bool Comm::is_uniform() const
214 if (this == MPI_COMM_UNINITIALIZED)
215 return smpi_process()->comm_world()->is_uniform();
216 return is_uniform_ != 0;
219 bool Comm::is_blocked() const
221 if (this == MPI_COMM_UNINITIALIZED)
222 return smpi_process()->comm_world()->is_blocked();
223 return is_blocked_ != 0;
226 bool Comm::is_smp_comm() const
228 if (this == MPI_COMM_UNINITIALIZED)
229 return smpi_process()->comm_world()->is_smp_comm();
233 MPI_Comm Comm::split(int color, int key)
235 if (this == MPI_COMM_UNINITIALIZED)
236 return smpi_process()->comm_world()->split(color, key);
237 int system_tag = -123;
239 MPI_Group group_root = nullptr;
240 MPI_Group group_out = nullptr;
241 MPI_Group group = this->group();
242 int myrank = this->rank();
243 int size = this->size();
244 /* Gather all colors and keys on rank 0 */
245 const std::array<int, 2> sendbuf = {{color, key}};
246 std::vector<int> recvbuf;
248 recvbuf.resize(2 * size);
249 gather__default(sendbuf.data(), 2, MPI_INT, recvbuf.data(), 2, MPI_INT, 0, this);
250 /* Do the actual job */
252 std::vector<MPI_Group> group_snd(size);
253 std::vector<std::pair<int, int>> rankmap;
254 rankmap.reserve(size);
255 for (int i = 0; i < size; i++) {
256 if (recvbuf[2 * i] != MPI_UNDEFINED) {
258 for (int j = i + 1; j < size; j++) {
259 if(recvbuf[2 * i] == recvbuf[2 * j]) {
260 recvbuf[2 * j] = MPI_UNDEFINED;
261 rankmap.emplace_back(recvbuf[2 * j + 1], j);
264 /* Add self in the group */
265 recvbuf[2 * i] = MPI_UNDEFINED;
266 rankmap.emplace_back(recvbuf[2 * i + 1], i);
267 std::sort(begin(rankmap), end(rankmap));
268 group_out = new Group(rankmap.size());
270 group_root = group_out; /* Save root's group */
272 for (unsigned j = 0; j < rankmap.size(); j++) {
273 s4u::Actor* actor = group->actor(rankmap[j].second);
274 group_out->set_mapping(actor, j);
276 std::vector<MPI_Request> requests(rankmap.size());
278 for (auto const& rank : rankmap) {
279 if (rank.second != 0) {
280 group_snd[reqs]=new Group(group_out);
281 requests[reqs] = Request::isend(&(group_snd[reqs]), 1, MPI_PTR, rank.second, system_tag, this);
285 if(i != 0 && group_out != MPI_COMM_WORLD->group() && group_out != MPI_GROUP_EMPTY)
286 Group::unref(group_out);
288 Request::waitall(reqs, requests.data(), MPI_STATUS_IGNORE);
291 group_out = group_root; /* exit with root's group */
293 if(color != MPI_UNDEFINED) {
294 Request::recv(&group_out, 1, MPI_PTR, 0, system_tag, this, MPI_STATUS_IGNORE);
295 } /* otherwise, exit with group_out == nullptr */
297 return group_out!=nullptr ? new Comm(group_out, topo_) : MPI_COMM_NULL;
301 if (this == MPI_COMM_UNINITIALIZED){
302 smpi_process()->comm_world()->ref();
309 void Comm::cleanup_smp(){
310 if (intra_comm_ != MPI_COMM_NULL)
311 Comm::unref(intra_comm_);
312 if (leaders_comm_ != MPI_COMM_NULL)
313 Comm::unref(leaders_comm_);
314 xbt_free(non_uniform_map_);
315 delete[] leaders_map_;
318 void Comm::unref(Comm* comm){
319 if (comm == MPI_COMM_UNINITIALIZED){
320 Comm::unref(smpi_process()->comm_world());
324 Group::unref(comm->group_);
326 if(comm->refcount_==0){
327 if(simgrid::smpi::F2C::lookup() != nullptr)
328 F2C::free_f(comm->c2f());
330 comm->cleanup_attr<Comm>();
331 if (comm->info_ != MPI_INFO_NULL)
332 simgrid::smpi::Info::unref(comm->info_);
333 if (comm->errhandler_ != MPI_ERRHANDLER_NULL)
334 simgrid::smpi::Errhandler::unref(comm->errhandler_);
339 MPI_Comm Comm::find_intra_comm(int * leader){
340 //get the indices of all processes sharing the same simix host
341 auto& actor_list = sg_host_self()->pimpl_->actor_list_;
342 int intra_comm_size = 0;
343 int min_index = INT_MAX; // the minimum index will be the leader
344 for (auto& actor : actor_list) {
345 int index = actor.get_pid();
346 if (this->group()->rank(actor.get_ciface()) != MPI_UNDEFINED) { // Is this process in the current group?
348 if (index < min_index)
352 XBT_DEBUG("number of processes deployed on my node : %d", intra_comm_size);
353 auto* group_intra = new Group(intra_comm_size);
355 for (auto& actor : actor_list) {
356 if (this->group()->rank(actor.get_ciface()) != MPI_UNDEFINED) {
357 group_intra->set_mapping(actor.get_ciface(), i);
362 return new Comm(group_intra, nullptr, true);
365 void Comm::init_smp(){
368 if (this == MPI_COMM_UNINITIALIZED)
369 smpi_process()->comm_world()->init_smp();
371 int comm_size = this->size();
373 // If we are in replay - perform an ugly hack
374 // tell SimGrid we are not in replay for a while, because we need the buffers to be copied for the following calls
375 bool replaying = false; //cache data to set it back again after
376 if(smpi_process()->replaying()){
378 smpi_process()->set_replaying(false);
381 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) {
382 // we need to switch as the called function may silently touch global variables
383 smpi_switch_data_segment(s4u::Actor::self());
385 // identify neighbors in comm
386 MPI_Comm comm_intra = find_intra_comm(&leader);
388 auto* leaders_map = new int[comm_size];
389 auto* leader_list = new int[comm_size];
390 std::fill_n(leaders_map, comm_size, 0);
391 std::fill_n(leader_list, comm_size, -1);
393 allgather__ring(&leader, 1, MPI_INT , leaders_map, 1, MPI_INT, this);
395 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) {
396 // we need to switch as the called function may silently touch global variables
397 smpi_switch_data_segment(s4u::Actor::self());
400 if(leaders_map_==nullptr){
401 leaders_map_= leaders_map;
403 delete[] leaders_map;
405 int leader_group_size = 0;
406 for(i=0; i<comm_size; i++){
407 int already_done = 0;
408 for (int j = 0; j < leader_group_size; j++) {
409 if (leaders_map_[i] == leader_list[j]) {
413 if (already_done == 0) {
414 leader_list[leader_group_size] = leaders_map_[i];
418 xbt_assert(leader_group_size > 0);
419 std::sort(leader_list, leader_list + leader_group_size);
421 auto* leaders_group = new Group(leader_group_size);
423 MPI_Comm leader_comm = MPI_COMM_NULL;
424 if(MPI_COMM_WORLD!=MPI_COMM_UNINITIALIZED && this!=MPI_COMM_WORLD){
425 //create leader_communicator
426 for (i=0; i< leader_group_size;i++)
427 leaders_group->set_mapping(s4u::Actor::by_pid(leader_list[i]).get(), i);
428 leader_comm = new Comm(leaders_group, nullptr, true);
429 this->set_leaders_comm(leader_comm);
430 this->set_intra_comm(comm_intra);
432 // create intracommunicator
434 for (i=0; i< leader_group_size;i++)
435 leaders_group->set_mapping(s4u::Actor::by_pid(leader_list[i]).get(), i);
437 if(this->get_leaders_comm()==MPI_COMM_NULL){
438 leader_comm = new Comm(leaders_group, nullptr, true);
439 this->set_leaders_comm(leader_comm);
441 leader_comm=this->get_leaders_comm();
442 Group::unref(leaders_group);
444 smpi_process()->set_comm_intra(comm_intra);
447 // Are the nodes uniform ? = same number of process/node
448 int my_local_size=comm_intra->size();
449 if(comm_intra->rank()==0) {
451 int* non_uniform_map = xbt_new0(int,leader_group_size);
452 allgather__ring(&my_local_size, 1, MPI_INT,
453 non_uniform_map, 1, MPI_INT, leader_comm);
454 for(i=0; i < leader_group_size; i++) {
455 if(non_uniform_map[0] != non_uniform_map[i]) {
460 if (is_uniform == 0 && this->is_uniform()) {
461 non_uniform_map_ = non_uniform_map;
463 xbt_free(non_uniform_map);
465 is_uniform_=is_uniform;
467 bcast__scatter_LR_allgather(&(is_uniform_),1, MPI_INT, 0, comm_intra );
469 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) {
470 // we need to switch as the called function may silently touch global variables
471 smpi_switch_data_segment(s4u::Actor::self());
473 // Are the ranks blocked ? = allocated contiguously on the SMP nodes
475 int prev=this->group()->rank(comm_intra->group()->actor(0));
476 for (i = 1; i < my_local_size; i++) {
477 int that = this->group()->rank(comm_intra->group()->actor(i));
478 if (that != prev + 1) {
486 allreduce__default(&is_blocked, &(global_blocked), 1, MPI_INT, MPI_LAND, this);
488 if(MPI_COMM_WORLD==MPI_COMM_UNINITIALIZED || this==MPI_COMM_WORLD){
490 is_blocked_ = global_blocked;
493 is_blocked_=global_blocked;
495 delete[] leader_list;
498 smpi_process()->set_replaying(true);
501 MPI_Comm Comm::f2c(int id) {
503 return MPI_COMM_SELF;
505 return MPI_COMM_WORLD;
506 } else if (F2C::lookup() != nullptr && id >= 0) {
507 const auto& lookup = F2C::lookup();
508 auto comm = lookup->find(id);
509 return comm == lookup->end() ? MPI_COMM_NULL : static_cast<MPI_Comm>(comm->second);
511 return MPI_COMM_NULL;
515 void Comm::free_f(int id) {
516 F2C::lookup()->erase(id);
519 void Comm::add_rma_win(MPI_Win win){
520 rma_wins_.push_back(win);
523 void Comm::remove_rma_win(MPI_Win win){
524 rma_wins_.remove(win);
527 void Comm::finish_rma_calls() const
529 for (auto const& it : rma_wins_) {
530 if(it->rank()==this->rank()){//is it ours (for MPI_COMM_WORLD)?
531 int finished = it->finish_comms();
532 XBT_DEBUG("Barrier for rank %d - Finished %d RMA calls",this->rank(), finished);
537 MPI_Info Comm::info()
539 if (info_ == MPI_INFO_NULL)
545 void Comm::set_info(MPI_Info info)
547 if (info_ != MPI_INFO_NULL)
548 simgrid::smpi::Info::unref(info);
550 if (info_ != MPI_INFO_NULL)
554 MPI_Errhandler Comm::errhandler()
556 if (errhandler_ != MPI_ERRHANDLER_NULL)
561 void Comm::set_errhandler(MPI_Errhandler errhandler)
563 if (errhandler_ != MPI_ERRHANDLER_NULL)
564 simgrid::smpi::Errhandler::unref(errhandler_);
565 errhandler_ = errhandler;
566 if (errhandler_ != MPI_ERRHANDLER_NULL)
570 MPI_Comm Comm::split_type(int type, int /*key*/, const Info*)
572 //MPI_UNDEFINED can be given to some nodes... but we need them to still perform the smp part which is collective
573 if(type != MPI_COMM_TYPE_SHARED && type != MPI_UNDEFINED){
574 return MPI_COMM_NULL;
577 MPI_Comm res= this->find_intra_comm(&leader);
578 if(type != MPI_UNDEFINED)
581 xbt_assert(res->refcount_ == 1); // ensure the next call to Comm::destroy really frees the comm
583 return MPI_COMM_NULL;
588 } // namespace simgrid