1 /* Copyright (c) 2010-2023. 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/kernel/resource/HostImpl.hpp"
14 #include "src/smpi/include/smpi_actor.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. */
36 namespace simgrid::smpi {
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
61 void Comm::destroy(Comm* comm)
63 if (comm == MPI_COMM_UNINITIALIZED){
64 Comm::destroy(smpi_process()->comm_world());
67 if (comm != MPI_COMM_WORLD && not comm->deleted()) {
68 comm->cleanup_attr<Comm>();
69 comm->mark_as_deleted();
74 int Comm::dup(MPI_Comm* newcomm){
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());
81 for (auto const& [key, value] : attributes()) {
82 auto elem_it = keyvals_.find(key);
83 xbt_assert(elem_it != keyvals_.end(), "Keyval not found for Comm: %d", key);
85 smpi_key_elem& elem = elem_it->second;
86 int ret = MPI_SUCCESS;
88 void* value_out = nullptr;
89 if (elem.copy_fn.comm_copy_fn == MPI_COMM_DUP_FN) {
92 } else if (elem.copy_fn.comm_copy_fn != MPI_NULL_COPY_FN) {
93 ret = elem.copy_fn.comm_copy_fn(this, key, elem.extra_state, value, &value_out, &flag);
95 if (elem.copy_fn.comm_copy_fn_fort != MPI_NULL_COPY_FN) {
96 value_out = xbt_new(int, 1);
97 if (*(int*)*elem.copy_fn.comm_copy_fn_fort == 1) { // MPI_COMM_DUP_FN
98 memcpy(value_out, value, sizeof(int));
100 } else { // not null, nor dup
101 elem.copy_fn.comm_copy_fn_fort(this, key, elem.extra_state, value, value_out, &flag, &ret);
103 if (ret != MPI_SUCCESS)
106 if (ret != MPI_SUCCESS) {
107 Comm::destroy(*newcomm);
108 *newcomm = MPI_COMM_NULL;
113 (*newcomm)->attributes().try_emplace(key, value_out);
116 //duplicate info if present
117 if(info_!=MPI_INFO_NULL)
118 (*newcomm)->info_ = new simgrid::smpi::Info(info_);
119 //duplicate errhandler
120 if (errhandlers_ != nullptr)//MPI_COMM_WORLD, only grab our own
121 (*newcomm)->set_errhandler(errhandlers_[this->rank()]);
123 (*newcomm)->set_errhandler(errhandler_);
127 int Comm::dup_with_info(MPI_Info info, MPI_Comm* newcomm){
128 int ret = dup(newcomm);
129 if(ret != MPI_SUCCESS)
131 if((*newcomm)->info_!=MPI_INFO_NULL){
132 simgrid::smpi::Info::unref((*newcomm)->info_);
133 (*newcomm)->info_=MPI_INFO_NULL;
135 if(info != MPI_INFO_NULL){
137 (*newcomm)->info_=info;
142 MPI_Group Comm::group()
144 if (this == MPI_COMM_UNINITIALIZED)
145 return smpi_process()->comm_world()->group();
149 int Comm::size() const
151 if (this == MPI_COMM_UNINITIALIZED)
152 return smpi_process()->comm_world()->size();
153 return group_->size();
156 int Comm::rank() const
158 if (this == MPI_COMM_UNINITIALIZED)
159 return smpi_process()->comm_world()->rank();
160 return group_->rank(s4u::this_actor::get_pid());
168 void Comm::get_name(char* name, int* len) const
170 if (this == MPI_COMM_UNINITIALIZED){
171 smpi_process()->comm_world()->get_name(name, len);
174 if(this == MPI_COMM_WORLD && name_.empty()) {
175 strncpy(name, "MPI_COMM_WORLD", 15);
178 *len = snprintf(name, MPI_MAX_NAME_STRING+1, "%s", name_.c_str());
182 std::string Comm::name() const
185 std::array<char, MPI_MAX_NAME_STRING + 1> name;
186 this->get_name(name.data(), &size);
194 void Comm::set_name (const char* name)
196 if (this == MPI_COMM_UNINITIALIZED){
197 smpi_process()->comm_world()->set_name(name);
200 name_.replace (0, MPI_MAX_NAME_STRING+1, name);
204 void Comm::set_leaders_comm(MPI_Comm leaders){
205 if (this == MPI_COMM_UNINITIALIZED){
206 smpi_process()->comm_world()->set_leaders_comm(leaders);
209 leaders_comm_=leaders;
212 int* Comm::get_non_uniform_map() const
214 if (this == MPI_COMM_UNINITIALIZED)
215 return smpi_process()->comm_world()->get_non_uniform_map();
216 return non_uniform_map_;
219 int* Comm::get_leaders_map() const
221 if (this == MPI_COMM_UNINITIALIZED)
222 return smpi_process()->comm_world()->get_leaders_map();
226 MPI_Comm Comm::get_leaders_comm() const
228 if (this == MPI_COMM_UNINITIALIZED)
229 return smpi_process()->comm_world()->get_leaders_comm();
230 return leaders_comm_;
233 MPI_Comm Comm::get_intra_comm() const
235 if (this == MPI_COMM_UNINITIALIZED || this==MPI_COMM_WORLD)
236 return smpi_process()->comm_intra();
237 else return intra_comm_;
240 bool Comm::is_uniform() const
242 if (this == MPI_COMM_UNINITIALIZED)
243 return smpi_process()->comm_world()->is_uniform();
247 bool Comm::is_blocked() const
249 if (this == MPI_COMM_UNINITIALIZED)
250 return smpi_process()->comm_world()->is_blocked();
254 bool Comm::is_smp_comm() const
256 if (this == MPI_COMM_UNINITIALIZED)
257 return smpi_process()->comm_world()->is_smp_comm();
261 MPI_Comm Comm::split(int color, int key)
263 if (this == MPI_COMM_UNINITIALIZED)
264 return smpi_process()->comm_world()->split(color, key);
265 int system_tag = -123;
267 MPI_Group group_root = nullptr;
268 MPI_Group group_out = nullptr;
269 const Group* group = this->group();
270 int myrank = this->rank();
271 int size = this->size();
272 /* Gather all colors and keys on rank 0 */
273 const std::array<int, 2> sendbuf = {{color, key}};
274 std::vector<int> recvbuf;
276 recvbuf.resize(2 * size);
277 gather__default(sendbuf.data(), 2, MPI_INT, recvbuf.data(), 2, MPI_INT, 0, this);
278 /* Do the actual job */
280 std::vector<MPI_Group> group_snd(size);
281 std::vector<std::pair<int, int>> rankmap;
282 rankmap.reserve(size);
283 for (int i = 0; i < size; i++) {
284 if (recvbuf[2 * i] != MPI_UNDEFINED) {
286 for (int j = i + 1; j < size; j++) {
287 if(recvbuf[2 * i] == recvbuf[2 * j]) {
288 recvbuf[2 * j] = MPI_UNDEFINED;
289 rankmap.emplace_back(recvbuf[2 * j + 1], j);
292 /* Add self in the group */
293 recvbuf[2 * i] = MPI_UNDEFINED;
294 rankmap.emplace_back(recvbuf[2 * i + 1], i);
295 std::sort(begin(rankmap), end(rankmap));
296 group_out = new Group(rankmap.size());
298 group_root = group_out; /* Save root's group */
300 for (unsigned j = 0; j < rankmap.size(); j++) {
301 aid_t actor = group->actor(rankmap[j].second);
302 group_out->set_mapping(actor, j);
304 std::vector<MPI_Request> requests(rankmap.size());
306 for (auto const& [_, rank] : rankmap) {
308 group_snd[reqs]=new Group(group_out);
309 requests[reqs] = Request::isend(&(group_snd[reqs]), 1, MPI_PTR, rank, system_tag, this);
313 if(i != 0 && group_out != MPI_COMM_WORLD->group() && group_out != MPI_GROUP_EMPTY)
314 Group::unref(group_out);
316 Request::waitall(reqs, requests.data(), MPI_STATUS_IGNORE);
319 group_out = group_root; /* exit with root's group */
321 if(color != MPI_UNDEFINED) {
322 Request::recv(&group_out, 1, MPI_PTR, 0, system_tag, this, MPI_STATUS_IGNORE);
323 } /* otherwise, exit with group_out == nullptr */
325 return group_out!=nullptr ? new Comm(group_out, topo_) : MPI_COMM_NULL;
329 if (this == MPI_COMM_UNINITIALIZED){
330 smpi_process()->comm_world()->ref();
337 void Comm::cleanup_smp(){
338 if (intra_comm_ != MPI_COMM_NULL)
339 Comm::unref(intra_comm_);
340 if (leaders_comm_ != MPI_COMM_NULL)
341 Comm::unref(leaders_comm_);
342 xbt_free(non_uniform_map_);
343 delete[] leaders_map_;
346 void Comm::unref(Comm* comm){
347 if (comm == MPI_COMM_UNINITIALIZED){
348 Comm::unref(smpi_process()->comm_world());
353 if(comm->refcount_==0){
354 if(simgrid::smpi::F2C::lookup() != nullptr)
355 F2C::free_f(comm->f2c_id());
357 comm->cleanup_attr<Comm>();
358 if (comm->info_ != MPI_INFO_NULL)
359 simgrid::smpi::Info::unref(comm->info_);
360 if(comm->errhandlers_!=nullptr){
361 for (int i=0; i<comm->size(); i++)
362 if (comm->errhandlers_[i]!=MPI_ERRHANDLER_NULL)
363 simgrid::smpi::Errhandler::unref(comm->errhandlers_[i]);
364 delete[] comm->errhandlers_;
365 } else if (comm->errhandler_ != MPI_ERRHANDLER_NULL)
366 simgrid::smpi::Errhandler::unref(comm->errhandler_);
368 Group::unref(comm->group_);
369 if(comm->refcount_==0)
373 MPI_Comm Comm::find_intra_comm(int * leader){
374 //get the indices of all processes sharing the same simix host
375 int intra_comm_size = 0;
376 aid_t min_index = std::numeric_limits<aid_t>::max(); // the minimum index will be the leader
377 sg_host_self()->get_impl()->foreach_actor([this, &intra_comm_size, &min_index](auto& actor) {
378 aid_t index = actor.get_pid();
379 if (this->group()->rank(index) != MPI_UNDEFINED) { // Is this process in the current group?
381 if (index < min_index)
385 XBT_DEBUG("number of processes deployed on my node : %d", intra_comm_size);
386 auto* group_intra = new Group(intra_comm_size);
388 sg_host_self()->get_impl()->foreach_actor([this, group_intra, &i](auto& actor) {
389 if (this->group()->rank(actor.get_pid()) != MPI_UNDEFINED) {
390 group_intra->set_mapping(actor.get_pid(), i);
395 return new Comm(group_intra, nullptr, true);
398 void Comm::init_smp(){
401 if (this == MPI_COMM_UNINITIALIZED)
402 smpi_process()->comm_world()->init_smp();
404 int comm_size = this->size();
406 // If we are in replay - perform an ugly hack
407 // tell SimGrid we are not in replay for a while, because we need the buffers to be copied for the following calls
408 bool replaying = false; //cache data to set it back again after
409 if(smpi_process()->replaying()){
411 smpi_process()->set_replaying(false);
414 // we need to switch as the called function may silently touch global variables
415 smpi_switch_data_segment(s4u::Actor::self());
417 // identify neighbors in comm
418 MPI_Comm comm_intra = find_intra_comm(&leader);
420 auto* leaders_map = new int[comm_size];
421 auto* leader_list = new int[comm_size];
422 std::fill_n(leaders_map, comm_size, 0);
423 std::fill_n(leader_list, comm_size, -1);
425 allgather__ring(&leader, 1, MPI_INT , leaders_map, 1, MPI_INT, this);
427 if(leaders_map_==nullptr){
428 leaders_map_= leaders_map;
430 delete[] leaders_map;
432 int leader_group_size = 0;
433 for(i=0; i<comm_size; i++){
434 int already_done = 0;
435 for (int j = 0; j < leader_group_size; j++) {
436 if (leaders_map_[i] == leader_list[j]) {
440 if (already_done == 0) {
441 leader_list[leader_group_size] = leaders_map_[i];
445 xbt_assert(leader_group_size > 0);
446 std::sort(leader_list, leader_list + leader_group_size);
448 auto* leaders_group = new Group(leader_group_size);
450 MPI_Comm leader_comm = MPI_COMM_NULL;
451 if(MPI_COMM_WORLD!=MPI_COMM_UNINITIALIZED && this!=MPI_COMM_WORLD){
452 //create leader_communicator
453 for (i=0; i< leader_group_size;i++)
454 leaders_group->set_mapping(leader_list[i], i);
455 leader_comm = new Comm(leaders_group, nullptr, true);
456 this->set_leaders_comm(leader_comm);
457 this->set_intra_comm(comm_intra);
459 // create intracommunicator
461 for (i=0; i< leader_group_size;i++)
462 leaders_group->set_mapping(leader_list[i], i);
464 if(this->get_leaders_comm()==MPI_COMM_NULL){
465 leader_comm = new Comm(leaders_group, nullptr, true);
466 this->set_leaders_comm(leader_comm);
468 leader_comm=this->get_leaders_comm();
469 Group::unref(leaders_group);
471 smpi_process()->set_comm_intra(comm_intra);
474 // Are the nodes uniform ? = same number of process/node
475 int my_local_size=comm_intra->size();
477 if(comm_intra->rank()==0) {
479 auto* non_uniform_map = xbt_new0(int, leader_group_size);
480 allgather__ring(&my_local_size, 1, MPI_INT,
481 non_uniform_map, 1, MPI_INT, leader_comm);
482 for(i=0; i < leader_group_size; i++) {
483 if(non_uniform_map[0] != non_uniform_map[i]) {
488 if (is_uniform == 0 && this->is_uniform()) {
489 non_uniform_map_ = non_uniform_map;
491 xbt_free(non_uniform_map);
494 bcast__scatter_LR_allgather(&is_uniform, 1, MPI_INT, 0, comm_intra);
495 is_uniform_ = (is_uniform != 0);
497 // we need to switch as the called function may silently touch global variables
498 smpi_switch_data_segment(s4u::Actor::self());
500 // Are the ranks blocked ? = allocated contiguously on the SMP nodes
502 int prev = this->group()->rank(comm_intra->group()->actor(0));
503 for (i = 1; i < my_local_size; i++) {
504 int that = this->group()->rank(comm_intra->group()->actor(i));
505 if (that != prev + 1) {
513 allreduce__default(&is_blocked, &global_blocked, 1, MPI_INT, MPI_LAND, this);
515 if ((MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED && this != MPI_COMM_WORLD) || this->rank() == 0)
516 is_blocked_ = (global_blocked != 0);
517 delete[] leader_list;
520 smpi_process()->set_replaying(true);
523 MPI_Comm Comm::f2c(int id) {
525 return MPI_COMM_SELF;
527 return MPI_COMM_WORLD;
528 } else if (F2C::lookup() != nullptr && id >= 0) {
529 const auto& lookup = F2C::lookup();
530 auto comm = lookup->find(id);
531 return comm == lookup->end() ? MPI_COMM_NULL : static_cast<MPI_Comm>(comm->second);
533 return MPI_COMM_NULL;
537 void Comm::free_f(int id) {
538 F2C::lookup()->erase(id);
541 void Comm::add_rma_win(MPI_Win win){
542 rma_wins_.push_back(win);
545 void Comm::remove_rma_win(MPI_Win win){
546 rma_wins_.remove(win);
549 void Comm::finish_rma_calls() const
551 const int myrank = rank();
552 for (auto const& it : rma_wins_) {
553 if (it->rank() == myrank) { // is it ours (for MPI_COMM_WORLD)?
554 int finished = it->finish_comms();
555 XBT_DEBUG("Barrier for rank %d - Finished %d RMA calls", myrank, finished);
560 MPI_Info Comm::info()
565 void Comm::set_info(MPI_Info info)
567 if (info_ != MPI_INFO_NULL)
568 simgrid::smpi::Info::unref(info);
570 if (info_ != MPI_INFO_NULL)
574 MPI_Errhandler Comm::errhandler()
576 if (this != MPI_COMM_WORLD){
577 if (errhandler_ != MPI_ERRHANDLER_NULL)
581 if(errhandlers_==nullptr){
582 if (_smpi_cfg_default_errhandler_is_error)
583 return MPI_ERRORS_ARE_FATAL;
585 return MPI_ERRORS_RETURN;
587 if(errhandlers_[this->rank()] != MPI_ERRHANDLER_NULL)
588 errhandlers_[this->rank()]->ref();
589 return errhandlers_[this->rank()];
594 void Comm::set_errhandler(MPI_Errhandler errhandler)
596 if(this != MPI_COMM_WORLD){
597 if (errhandler_ != MPI_ERRHANDLER_NULL)
598 simgrid::smpi::Errhandler::unref(errhandler_);
599 errhandler_ = errhandler;
601 if(errhandlers_==nullptr)
602 errhandlers_= new MPI_Errhandler[this->size()]{MPI_ERRHANDLER_NULL};
603 if(errhandlers_[this->rank()] != MPI_ERRHANDLER_NULL)
604 simgrid::smpi::Errhandler::unref(errhandlers_[this->rank()]);
605 errhandlers_[this->rank()]=errhandler;
607 if (errhandler != MPI_ERRHANDLER_NULL)
611 MPI_Comm Comm::split_type(int type, int /*key*/, const Info*)
613 //MPI_UNDEFINED can be given to some nodes... but we need them to still perform the smp part which is collective
614 if(type != MPI_COMM_TYPE_SHARED && type != MPI_UNDEFINED){
615 return MPI_COMM_NULL;
618 MPI_Comm res= this->find_intra_comm(&leader);
619 if(type != MPI_UNDEFINED)
622 xbt_assert(res->refcount_ == 1); // ensure the next call to Comm::destroy really frees the comm
624 return MPI_COMM_NULL;
628 static inline std::string hash_message(int src, int dst, int tag){
629 return std::to_string(tag) + '_' + std::to_string(src) + '_' + std::to_string(dst);
632 unsigned int Comm::get_sent_messages_count(int src, int dst, int tag)
634 return sent_messages_[hash_message(src, dst, tag)];
637 void Comm::increment_sent_messages_count(int src, int dst, int tag)
639 sent_messages_[hash_message(src, dst, tag)]++;
642 unsigned int Comm::get_received_messages_count(int src, int dst, int tag)
644 return recv_messages_[hash_message(src, dst, tag)];
647 void Comm::increment_received_messages_count(int src, int dst, int tag)
649 recv_messages_[hash_message(src, dst, tag)]++;
652 unsigned int Comm::get_collectives_count()
654 if (this==MPI_COMM_UNINITIALIZED){
655 return smpi_process()->comm_world()->get_collectives_count();
656 }else if(this == MPI_COMM_WORLD || this == smpi_process()->comm_world()){
657 if (collectives_counts_.empty())
658 collectives_counts_.resize(this->size());
659 return collectives_counts_[this->rank()];
661 return collectives_count_;
665 void Comm::increment_collectives_count()
667 if (this==MPI_COMM_UNINITIALIZED){
668 smpi_process()->comm_world()->increment_collectives_count();
669 }else if (this == MPI_COMM_WORLD || this == smpi_process()->comm_world()){
670 if (collectives_counts_.empty())
671 collectives_counts_.resize(this->size());
672 collectives_counts_[this->rank()]++;
674 collectives_count_++;
678 } // namespace simgrid::smpi