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());
68 if(comm != MPI_COMM_WORLD)
69 comm->mark_as_deleted();
73 int Comm::dup(MPI_Comm* newcomm){
74 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) {
75 // we need to switch as the called function may silently touch global variables
76 smpi_switch_data_segment(s4u::Actor::self());
78 auto* cp = new Group(this->group());
79 (*newcomm) = new Comm(cp, this->topo());
80 int ret = MPI_SUCCESS;
82 for (auto const& it : attributes()) {
83 smpi_key_elem elem = keyvals_.at(it.first);
84 if (elem != nullptr) {
86 void* value_out = nullptr;
87 if (elem->copy_fn.comm_copy_fn != MPI_NULL_COPY_FN && 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 && *(int*)*elem->copy_fn.comm_copy_fn_fort != 1) {
90 value_out = (int*)xbt_malloc(sizeof(int));
91 elem->copy_fn.comm_copy_fn_fort(this, it.first, elem->extra_state, it.second, value_out, &flag, &ret);
93 if (ret != MPI_SUCCESS) {
94 Comm::destroy(*newcomm);
95 *newcomm = MPI_COMM_NULL;
98 if (elem->copy_fn.comm_copy_fn == MPI_COMM_DUP_FN ||
99 ((elem->copy_fn.comm_copy_fn_fort != MPI_NULL_COPY_FN) && *(int*)*elem->copy_fn.comm_copy_fn_fort == 1)) {
101 (*newcomm)->attributes().insert({it.first, it.second});
104 (*newcomm)->attributes().insert({it.first, value_out});
108 //duplicate info if present
109 if(info_!=MPI_INFO_NULL)
110 (*newcomm)->info_ = new simgrid::smpi::Info(info_);
111 //duplicate errhandler
112 if (errhandlers_ != nullptr)//MPI_COMM_WORLD, only grab our own
113 (*newcomm)->set_errhandler(errhandlers_[this->rank()]);
115 (*newcomm)->set_errhandler(errhandler_);
119 int Comm::dup_with_info(MPI_Info info, MPI_Comm* newcomm){
120 int ret = dup(newcomm);
121 if(ret != MPI_SUCCESS)
123 if((*newcomm)->info_!=MPI_INFO_NULL){
124 simgrid::smpi::Info::unref((*newcomm)->info_);
125 (*newcomm)->info_=MPI_INFO_NULL;
127 if(info != MPI_INFO_NULL){
129 (*newcomm)->info_=info;
134 MPI_Group Comm::group()
136 if (this == MPI_COMM_UNINITIALIZED)
137 return smpi_process()->comm_world()->group();
141 int Comm::size() const
143 if (this == MPI_COMM_UNINITIALIZED)
144 return smpi_process()->comm_world()->size();
145 return group_->size();
148 int Comm::rank() const
150 if (this == MPI_COMM_UNINITIALIZED)
151 return smpi_process()->comm_world()->rank();
152 return group_->rank(s4u::this_actor::get_pid());
160 void Comm::get_name(char* name, int* len) const
162 if (this == MPI_COMM_UNINITIALIZED){
163 smpi_process()->comm_world()->get_name(name, len);
166 if(this == MPI_COMM_WORLD && name_.empty()) {
167 strncpy(name, "MPI_COMM_WORLD", 15);
170 *len = snprintf(name, MPI_MAX_NAME_STRING+1, "%s", name_.c_str());
174 std::string Comm::name() const
177 char name[MPI_MAX_NAME_STRING+1];
178 this->get_name(name, &size);
180 return std::string("MPI_Comm");
182 return std::string(name);
186 void Comm::set_name (const char* name)
188 if (this == MPI_COMM_UNINITIALIZED){
189 smpi_process()->comm_world()->set_name(name);
192 name_.replace (0, MPI_MAX_NAME_STRING+1, name);
196 void Comm::set_leaders_comm(MPI_Comm leaders){
197 if (this == MPI_COMM_UNINITIALIZED){
198 smpi_process()->comm_world()->set_leaders_comm(leaders);
201 leaders_comm_=leaders;
204 int* Comm::get_non_uniform_map() const
206 if (this == MPI_COMM_UNINITIALIZED)
207 return smpi_process()->comm_world()->get_non_uniform_map();
208 return non_uniform_map_;
211 int* Comm::get_leaders_map() const
213 if (this == MPI_COMM_UNINITIALIZED)
214 return smpi_process()->comm_world()->get_leaders_map();
218 MPI_Comm Comm::get_leaders_comm() const
220 if (this == MPI_COMM_UNINITIALIZED)
221 return smpi_process()->comm_world()->get_leaders_comm();
222 return leaders_comm_;
225 MPI_Comm Comm::get_intra_comm() const
227 if (this == MPI_COMM_UNINITIALIZED || this==MPI_COMM_WORLD)
228 return smpi_process()->comm_intra();
229 else return intra_comm_;
232 bool Comm::is_uniform() const
234 if (this == MPI_COMM_UNINITIALIZED)
235 return smpi_process()->comm_world()->is_uniform();
236 return is_uniform_ != 0;
239 bool Comm::is_blocked() const
241 if (this == MPI_COMM_UNINITIALIZED)
242 return smpi_process()->comm_world()->is_blocked();
243 return is_blocked_ != 0;
246 bool Comm::is_smp_comm() const
248 if (this == MPI_COMM_UNINITIALIZED)
249 return smpi_process()->comm_world()->is_smp_comm();
253 MPI_Comm Comm::split(int color, int key)
255 if (this == MPI_COMM_UNINITIALIZED)
256 return smpi_process()->comm_world()->split(color, key);
257 int system_tag = -123;
259 MPI_Group group_root = nullptr;
260 MPI_Group group_out = nullptr;
261 const Group* group = this->group();
262 int myrank = this->rank();
263 int size = this->size();
264 /* Gather all colors and keys on rank 0 */
265 const std::array<int, 2> sendbuf = {{color, key}};
266 std::vector<int> recvbuf;
268 recvbuf.resize(2 * size);
269 gather__default(sendbuf.data(), 2, MPI_INT, recvbuf.data(), 2, MPI_INT, 0, this);
270 /* Do the actual job */
272 std::vector<MPI_Group> group_snd(size);
273 std::vector<std::pair<int, int>> rankmap;
274 rankmap.reserve(size);
275 for (int i = 0; i < size; i++) {
276 if (recvbuf[2 * i] != MPI_UNDEFINED) {
278 for (int j = i + 1; j < size; j++) {
279 if(recvbuf[2 * i] == recvbuf[2 * j]) {
280 recvbuf[2 * j] = MPI_UNDEFINED;
281 rankmap.emplace_back(recvbuf[2 * j + 1], j);
284 /* Add self in the group */
285 recvbuf[2 * i] = MPI_UNDEFINED;
286 rankmap.emplace_back(recvbuf[2 * i + 1], i);
287 std::sort(begin(rankmap), end(rankmap));
288 group_out = new Group(rankmap.size());
290 group_root = group_out; /* Save root's group */
292 for (unsigned j = 0; j < rankmap.size(); j++) {
293 aid_t actor = group->actor(rankmap[j].second);
294 group_out->set_mapping(actor, j);
296 std::vector<MPI_Request> requests(rankmap.size());
298 for (auto const& rank : rankmap) {
299 if (rank.second != 0) {
300 group_snd[reqs]=new Group(group_out);
301 requests[reqs] = Request::isend(&(group_snd[reqs]), 1, MPI_PTR, rank.second, system_tag, this);
305 if(i != 0 && group_out != MPI_COMM_WORLD->group() && group_out != MPI_GROUP_EMPTY)
306 Group::unref(group_out);
308 Request::waitall(reqs, requests.data(), MPI_STATUS_IGNORE);
311 group_out = group_root; /* exit with root's group */
313 if(color != MPI_UNDEFINED) {
314 Request::recv(&group_out, 1, MPI_PTR, 0, system_tag, this, MPI_STATUS_IGNORE);
315 } /* otherwise, exit with group_out == nullptr */
317 return group_out!=nullptr ? new Comm(group_out, topo_) : MPI_COMM_NULL;
321 if (this == MPI_COMM_UNINITIALIZED){
322 smpi_process()->comm_world()->ref();
329 void Comm::cleanup_smp(){
330 if (intra_comm_ != MPI_COMM_NULL)
331 Comm::unref(intra_comm_);
332 if (leaders_comm_ != MPI_COMM_NULL)
333 Comm::unref(leaders_comm_);
334 xbt_free(non_uniform_map_);
335 delete[] leaders_map_;
338 void Comm::unref(Comm* comm){
339 if (comm == MPI_COMM_UNINITIALIZED){
340 Comm::unref(smpi_process()->comm_world());
345 if(comm->refcount_==0){
346 if(simgrid::smpi::F2C::lookup() != nullptr)
347 F2C::free_f(comm->f2c_id());
349 comm->cleanup_attr<Comm>();
350 if (comm->info_ != MPI_INFO_NULL)
351 simgrid::smpi::Info::unref(comm->info_);
352 if(comm->errhandlers_!=nullptr){
353 for (int i=0; i<comm->size(); i++)
354 if (comm->errhandlers_[i]!=MPI_ERRHANDLER_NULL)
355 simgrid::smpi::Errhandler::unref(comm->errhandlers_[i]);
356 delete[] comm->errhandlers_;
357 } else if (comm->errhandler_ != MPI_ERRHANDLER_NULL)
358 simgrid::smpi::Errhandler::unref(comm->errhandler_);
360 Group::unref(comm->group_);
361 if(comm->refcount_==0)
365 MPI_Comm Comm::find_intra_comm(int * leader){
366 //get the indices of all processes sharing the same simix host
367 int intra_comm_size = 0;
368 aid_t min_index = std::numeric_limits<aid_t>::max(); // the minimum index will be the leader
369 sg_host_self()->get_impl()->foreach_actor([this, &intra_comm_size, &min_index](auto& actor) {
370 aid_t index = actor.get_pid();
371 if (this->group()->rank(index) != MPI_UNDEFINED) { // Is this process in the current group?
373 if (index < min_index)
377 XBT_DEBUG("number of processes deployed on my node : %d", intra_comm_size);
378 auto* group_intra = new Group(intra_comm_size);
380 sg_host_self()->get_impl()->foreach_actor([this, group_intra, &i](auto& actor) {
381 if (this->group()->rank(actor.get_pid()) != MPI_UNDEFINED) {
382 group_intra->set_mapping(actor.get_pid(), i);
387 return new Comm(group_intra, nullptr, true);
390 void Comm::init_smp(){
393 if (this == MPI_COMM_UNINITIALIZED)
394 smpi_process()->comm_world()->init_smp();
396 int comm_size = this->size();
398 // If we are in replay - perform an ugly hack
399 // tell SimGrid we are not in replay for a while, because we need the buffers to be copied for the following calls
400 bool replaying = false; //cache data to set it back again after
401 if(smpi_process()->replaying()){
403 smpi_process()->set_replaying(false);
406 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) {
407 // we need to switch as the called function may silently touch global variables
408 smpi_switch_data_segment(s4u::Actor::self());
410 // identify neighbors in comm
411 MPI_Comm comm_intra = find_intra_comm(&leader);
413 auto* leaders_map = new int[comm_size];
414 auto* leader_list = new int[comm_size];
415 std::fill_n(leaders_map, comm_size, 0);
416 std::fill_n(leader_list, comm_size, -1);
418 allgather__ring(&leader, 1, MPI_INT , leaders_map, 1, MPI_INT, this);
420 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) {
421 // we need to switch as the called function may silently touch global variables
422 smpi_switch_data_segment(s4u::Actor::self());
425 if(leaders_map_==nullptr){
426 leaders_map_= leaders_map;
428 delete[] leaders_map;
430 int leader_group_size = 0;
431 for(i=0; i<comm_size; i++){
432 int already_done = 0;
433 for (int j = 0; j < leader_group_size; j++) {
434 if (leaders_map_[i] == leader_list[j]) {
438 if (already_done == 0) {
439 leader_list[leader_group_size] = leaders_map_[i];
443 xbt_assert(leader_group_size > 0);
444 std::sort(leader_list, leader_list + leader_group_size);
446 auto* leaders_group = new Group(leader_group_size);
448 MPI_Comm leader_comm = MPI_COMM_NULL;
449 if(MPI_COMM_WORLD!=MPI_COMM_UNINITIALIZED && this!=MPI_COMM_WORLD){
450 //create leader_communicator
451 for (i=0; i< leader_group_size;i++)
452 leaders_group->set_mapping(leader_list[i], i);
453 leader_comm = new Comm(leaders_group, nullptr, true);
454 this->set_leaders_comm(leader_comm);
455 this->set_intra_comm(comm_intra);
457 // create intracommunicator
459 for (i=0; i< leader_group_size;i++)
460 leaders_group->set_mapping(leader_list[i], i);
462 if(this->get_leaders_comm()==MPI_COMM_NULL){
463 leader_comm = new Comm(leaders_group, nullptr, true);
464 this->set_leaders_comm(leader_comm);
466 leader_comm=this->get_leaders_comm();
467 Group::unref(leaders_group);
469 smpi_process()->set_comm_intra(comm_intra);
472 // Are the nodes uniform ? = same number of process/node
473 int my_local_size=comm_intra->size();
474 if(comm_intra->rank()==0) {
476 int* non_uniform_map = xbt_new0(int,leader_group_size);
477 allgather__ring(&my_local_size, 1, MPI_INT,
478 non_uniform_map, 1, MPI_INT, leader_comm);
479 for(i=0; i < leader_group_size; i++) {
480 if(non_uniform_map[0] != non_uniform_map[i]) {
485 if (is_uniform == 0 && this->is_uniform()) {
486 non_uniform_map_ = non_uniform_map;
488 xbt_free(non_uniform_map);
490 is_uniform_=is_uniform;
492 bcast__scatter_LR_allgather(&is_uniform_, 1, MPI_INT, 0, comm_intra);
494 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) {
495 // we need to switch as the called function may silently touch global variables
496 smpi_switch_data_segment(s4u::Actor::self());
498 // Are the ranks blocked ? = allocated contiguously on the SMP nodes
500 int prev = this->group()->rank(comm_intra->group()->actor(0));
501 for (i = 1; i < my_local_size; i++) {
502 int that = this->group()->rank(comm_intra->group()->actor(i));
503 if (that != prev + 1) {
511 allreduce__default(&is_blocked, &global_blocked, 1, MPI_INT, MPI_LAND, this);
513 if(MPI_COMM_WORLD==MPI_COMM_UNINITIALIZED || this==MPI_COMM_WORLD){
515 is_blocked_ = global_blocked;
518 is_blocked_=global_blocked;
520 delete[] leader_list;
523 smpi_process()->set_replaying(true);
526 MPI_Comm Comm::f2c(int id) {
528 return MPI_COMM_SELF;
530 return MPI_COMM_WORLD;
531 } else if (F2C::lookup() != nullptr && id >= 0) {
532 const auto& lookup = F2C::lookup();
533 auto comm = lookup->find(id);
534 return comm == lookup->end() ? MPI_COMM_NULL : static_cast<MPI_Comm>(comm->second);
536 return MPI_COMM_NULL;
540 void Comm::free_f(int id) {
541 F2C::lookup()->erase(id);
544 void Comm::add_rma_win(MPI_Win win){
545 rma_wins_.push_back(win);
548 void Comm::remove_rma_win(MPI_Win win){
549 rma_wins_.remove(win);
552 void Comm::finish_rma_calls() const
554 for (auto const& it : rma_wins_) {
555 if(it->rank()==this->rank()){//is it ours (for MPI_COMM_WORLD)?
556 int finished = it->finish_comms();
557 XBT_DEBUG("Barrier for rank %d - Finished %d RMA calls",this->rank(), finished);
562 MPI_Info Comm::info()
564 if (info_ == MPI_INFO_NULL)
570 void Comm::set_info(MPI_Info info)
572 if (info_ != MPI_INFO_NULL)
573 simgrid::smpi::Info::unref(info);
575 if (info_ != MPI_INFO_NULL)
579 MPI_Errhandler Comm::errhandler()
581 if (this != MPI_COMM_WORLD){
582 if (errhandler_ != MPI_ERRHANDLER_NULL)
586 if(errhandlers_==nullptr)
587 return MPI_ERRORS_ARE_FATAL;
589 if(errhandlers_[this->rank()] != MPI_ERRHANDLER_NULL)
590 errhandlers_[this->rank()]->ref();
591 return errhandlers_[this->rank()];
596 void Comm::set_errhandler(MPI_Errhandler errhandler)
598 if(this != MPI_COMM_WORLD){
599 if (errhandler_ != MPI_ERRHANDLER_NULL)
600 simgrid::smpi::Errhandler::unref(errhandler_);
601 errhandler_ = errhandler;
603 if(errhandlers_==nullptr)
604 errhandlers_= new MPI_Errhandler[this->size()]{MPI_ERRHANDLER_NULL};
605 if(errhandlers_[this->rank()] != MPI_ERRHANDLER_NULL)
606 simgrid::smpi::Errhandler::unref(errhandlers_[this->rank()]);
607 errhandlers_[this->rank()]=errhandler;
609 if (errhandler != MPI_ERRHANDLER_NULL)
613 MPI_Comm Comm::split_type(int type, int /*key*/, const Info*)
615 //MPI_UNDEFINED can be given to some nodes... but we need them to still perform the smp part which is collective
616 if(type != MPI_COMM_TYPE_SHARED && type != MPI_UNDEFINED){
617 return MPI_COMM_NULL;
620 MPI_Comm res= this->find_intra_comm(&leader);
621 if(type != MPI_UNDEFINED)
624 xbt_assert(res->refcount_ == 1); // ensure the next call to Comm::destroy really frees the comm
626 return MPI_COMM_NULL;
631 } // namespace simgrid