1 /* Copyright (c) 2010-2019. 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, int smp) : group_(group), topo_(topo),is_smp_comm_(smp)
34 topoType_ = MPI_INVALID_TOPO;
35 intra_comm_ = MPI_COMM_NULL;
36 leaders_comm_ = MPI_COMM_NULL;
38 non_uniform_map_ = nullptr;
39 leaders_map_ = nullptr;
41 info_ = MPI_INFO_NULL;
44 void Comm::destroy(Comm* comm)
46 if (comm == MPI_COMM_UNINITIALIZED){
47 Comm::destroy(smpi_process()->comm_world());
50 delete comm->topo_; // there's no use count on topos
54 int Comm::dup(MPI_Comm* newcomm){
55 if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) {
56 // we need to switch as the called function may silently touch global variables
57 smpi_switch_data_segment(s4u::Actor::self());
59 MPI_Group cp = new Group(this->group());
60 (*newcomm) = new Comm(cp, this->topo());
61 int ret = MPI_SUCCESS;
63 if (not attributes()->empty()) {
65 void* value_out=nullptr;
66 for (auto const& it : *attributes()) {
67 smpi_key_elem elem = keyvals_.at(it.first);
69 if( elem->copy_fn.comm_copy_fn != MPI_NULL_COPY_FN &&
70 elem->copy_fn.comm_copy_fn != MPI_COMM_DUP_FN)
71 ret = elem->copy_fn.comm_copy_fn(this, it.first, elem->extra_state, it.second, &value_out, &flag);
72 else if ( elem->copy_fn.comm_copy_fn_fort != MPI_NULL_COPY_FN &&
73 *(int*)*elem->copy_fn.comm_copy_fn_fort != 1){
74 value_out=(int*)xbt_malloc(sizeof(int));
75 elem->copy_fn.comm_copy_fn_fort(this, it.first, elem->extra_state, it.second, value_out, &flag,&ret);
77 if (ret != MPI_SUCCESS) {
78 Comm::destroy(*newcomm);
79 *newcomm = MPI_COMM_NULL;
82 if (elem->copy_fn.comm_copy_fn == MPI_COMM_DUP_FN ||
83 ((elem->copy_fn.comm_copy_fn_fort != MPI_NULL_COPY_FN) && *(int*)*elem->copy_fn.comm_copy_fn_fort == 1)){
85 (*newcomm)->attributes()->insert({it.first, it.second});
88 (*newcomm)->attributes()->insert({it.first, value_out});
93 //duplicate info if present
94 if(info_!=MPI_INFO_NULL)
95 (*newcomm)->info_ = new simgrid::smpi::Info(info_);
99 int Comm::dup_with_info(MPI_Info info, MPI_Comm* newcomm){
100 int ret = dup(newcomm);
101 if(ret != MPI_SUCCESS)
103 if((*newcomm)->info_!=MPI_INFO_NULL){
104 simgrid::smpi::Info::unref((*newcomm)->info_);
105 (*newcomm)->info_=MPI_INFO_NULL;
107 if(info != MPI_INFO_NULL){
108 (*newcomm)->info_=info;
113 MPI_Group Comm::group()
115 if (this == MPI_COMM_UNINITIALIZED)
116 return smpi_process()->comm_world()->group();
122 if (this == MPI_COMM_UNINITIALIZED)
123 return smpi_process()->comm_world()->size();
124 return group_->size();
129 if (this == MPI_COMM_UNINITIALIZED)
130 return smpi_process()->comm_world()->rank();
131 return group_->rank(s4u::Actor::self());
134 void Comm::get_name (char* name, int* len)
136 if (this == MPI_COMM_UNINITIALIZED){
137 smpi_process()->comm_world()->get_name(name, len);
140 if(this == MPI_COMM_WORLD && name_.empty()) {
141 strncpy(name, "MPI_COMM_WORLD", 15);
143 } else if(this == MPI_COMM_SELF && name_.empty()) {
144 strncpy(name, "MPI_COMM_SELF", 14);
147 *len = snprintf(name, MPI_MAX_NAME_STRING+1, "%s", name_.c_str());
151 void Comm::set_name (const char* name)
153 if (this == MPI_COMM_UNINITIALIZED){
154 smpi_process()->comm_world()->set_name(name);
157 name_.replace (0, MPI_MAX_NAME_STRING+1, name);
161 void Comm::set_leaders_comm(MPI_Comm leaders){
162 if (this == MPI_COMM_UNINITIALIZED){
163 smpi_process()->comm_world()->set_leaders_comm(leaders);
166 leaders_comm_=leaders;
169 int* Comm::get_non_uniform_map(){
170 if (this == MPI_COMM_UNINITIALIZED)
171 return smpi_process()->comm_world()->get_non_uniform_map();
172 return non_uniform_map_;
175 int* Comm::get_leaders_map(){
176 if (this == MPI_COMM_UNINITIALIZED)
177 return smpi_process()->comm_world()->get_leaders_map();
181 MPI_Comm Comm::get_leaders_comm(){
182 if (this == MPI_COMM_UNINITIALIZED)
183 return smpi_process()->comm_world()->get_leaders_comm();
184 return leaders_comm_;
187 MPI_Comm Comm::get_intra_comm(){
188 if (this == MPI_COMM_UNINITIALIZED || this==MPI_COMM_WORLD)
189 return smpi_process()->comm_intra();
190 else return intra_comm_;
193 int Comm::is_uniform(){
194 if (this == MPI_COMM_UNINITIALIZED)
195 return smpi_process()->comm_world()->is_uniform();
199 int Comm::is_blocked(){
200 if (this == MPI_COMM_UNINITIALIZED)
201 return smpi_process()->comm_world()->is_blocked();
205 int Comm::is_smp_comm(){
206 if (this == MPI_COMM_UNINITIALIZED)
207 return smpi_process()->comm_world()->is_smp_comm();
211 MPI_Comm Comm::split(int color, int key)
213 if (this == MPI_COMM_UNINITIALIZED)
214 return smpi_process()->comm_world()->split(color, key);
215 int system_tag = -123;
218 MPI_Group group_root = nullptr;
219 MPI_Group group_out = nullptr;
220 MPI_Group group = this->group();
221 int myrank = this->rank();
222 int size = this->size();
223 /* Gather all colors and keys on rank 0 */
224 int* sendbuf = xbt_new(int, 2);
228 recvbuf = xbt_new(int, 2 * size);
232 Coll_gather_default::gather(sendbuf, 2, MPI_INT, recvbuf, 2, MPI_INT, 0, this);
234 /* Do the actual job */
236 MPI_Group* group_snd = xbt_new(MPI_Group, size);
237 std::vector<std::pair<int, int>> rankmap;
238 rankmap.reserve(size);
239 for (int i = 0; i < size; i++) {
240 if (recvbuf[2 * i] != MPI_UNDEFINED) {
242 for (int j = i + 1; j < size; j++) {
243 if(recvbuf[2 * i] == recvbuf[2 * j]) {
244 recvbuf[2 * j] = MPI_UNDEFINED;
245 rankmap.push_back({recvbuf[2 * j + 1], j});
248 /* Add self in the group */
249 recvbuf[2 * i] = MPI_UNDEFINED;
250 rankmap.push_back({recvbuf[2 * i + 1], i});
251 std::sort(begin(rankmap), end(rankmap));
252 group_out = new Group(rankmap.size());
254 group_root = group_out; /* Save root's group */
256 for (unsigned j = 0; j < rankmap.size(); j++) {
257 s4u::ActorPtr actor = group->actor(rankmap[j].second);
258 group_out->set_mapping(actor, j);
260 MPI_Request* requests = xbt_new(MPI_Request, rankmap.size());
262 for (auto const& rank : rankmap) {
263 if (rank.second != 0) {
264 group_snd[reqs]=new Group(group_out);
265 requests[reqs] = Request::isend(&(group_snd[reqs]), 1, MPI_PTR, rank.second, system_tag, this);
269 if(i != 0 && group_out != MPI_COMM_WORLD->group() && group_out != MPI_GROUP_EMPTY)
270 Group::unref(group_out);
272 Request::waitall(reqs, requests, MPI_STATUS_IGNORE);
278 group_out = group_root; /* exit with root's group */
280 if(color != MPI_UNDEFINED) {
281 Request::recv(&group_out, 1, MPI_PTR, 0, system_tag, this, MPI_STATUS_IGNORE);
282 } /* otherwise, exit with group_out == nullptr */
284 return group_out!=nullptr ? new Comm(group_out, nullptr) : MPI_COMM_NULL;
288 if (this == MPI_COMM_UNINITIALIZED){
289 smpi_process()->comm_world()->ref();
296 void Comm::cleanup_smp(){
297 if (intra_comm_ != MPI_COMM_NULL)
298 Comm::unref(intra_comm_);
299 if (leaders_comm_ != MPI_COMM_NULL)
300 Comm::unref(leaders_comm_);
301 xbt_free(non_uniform_map_);
302 delete[] leaders_map_;
305 void Comm::unref(Comm* comm){
306 if (comm == MPI_COMM_UNINITIALIZED){
307 Comm::unref(smpi_process()->comm_world());
311 Group::unref(comm->group_);
313 if(comm->refcount_==0){
315 comm->cleanup_attr<Comm>();
320 MPI_Comm Comm::find_intra_comm(int * leader){
321 //get the indices of all processes sharing the same simix host
322 auto& process_list = sg_host_self()->pimpl_->process_list_;
323 int intra_comm_size = 0;
324 int min_index = INT_MAX; // the minimum index will be the leader
325 for (auto& actor : process_list) {
326 int index = actor.get_pid();
327 if (this->group()->rank(actor.iface()) != MPI_UNDEFINED) { // Is this process in the current group?
329 if (index < min_index)
333 XBT_DEBUG("number of processes deployed on my node : %d", intra_comm_size);
334 MPI_Group group_intra = new Group(intra_comm_size);
336 for (auto& actor : process_list) {
337 if (this->group()->rank(actor.iface()) != MPI_UNDEFINED) {
338 group_intra->set_mapping(actor.iface(), i);
343 return new Comm(group_intra, nullptr, 1);
346 void Comm::init_smp(){
349 if (this == MPI_COMM_UNINITIALIZED)
350 smpi_process()->comm_world()->init_smp();
352 int comm_size = this->size();
354 // If we are in replay - perform an ugly hack
355 // tell SimGrid we are not in replay for a while, because we need the buffers to be copied for the following calls
356 bool replaying = false; //cache data to set it back again after
357 if(smpi_process()->replaying()){
359 smpi_process()->set_replaying(false);
362 if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) {
363 // we need to switch as the called function may silently touch global variables
364 smpi_switch_data_segment(s4u::Actor::self());
366 //identify neighbours in comm
367 MPI_Comm comm_intra = find_intra_comm(&leader);
370 int* leaders_map = new int[comm_size];
371 int* leader_list = new int[comm_size];
372 std::fill_n(leaders_map, comm_size, 0);
373 std::fill_n(leader_list, comm_size, -1);
375 Coll_allgather_ring::allgather(&leader, 1, MPI_INT , leaders_map, 1, MPI_INT, this);
377 if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) {
378 // we need to switch as the called function may silently touch global variables
379 smpi_switch_data_segment(s4u::Actor::self());
382 if(leaders_map_==nullptr){
383 leaders_map_= leaders_map;
385 delete[] leaders_map;
387 int leader_group_size = 0;
388 for(i=0; i<comm_size; i++){
389 int already_done = 0;
390 for (int j = 0; j < leader_group_size; j++) {
391 if (leaders_map_[i] == leader_list[j]) {
395 if (already_done == 0) {
396 leader_list[leader_group_size] = leaders_map_[i];
400 std::sort(leader_list, leader_list + leader_group_size);
402 MPI_Group leaders_group = new Group(leader_group_size);
404 MPI_Comm leader_comm = MPI_COMM_NULL;
405 if(MPI_COMM_WORLD!=MPI_COMM_UNINITIALIZED && this!=MPI_COMM_WORLD){
406 //create leader_communicator
407 for (i=0; i< leader_group_size;i++)
408 leaders_group->set_mapping(s4u::Actor::by_pid(leader_list[i]), i);
409 leader_comm = new Comm(leaders_group, nullptr,1);
410 this->set_leaders_comm(leader_comm);
411 this->set_intra_comm(comm_intra);
413 // create intracommunicator
415 for (i=0; i< leader_group_size;i++)
416 leaders_group->set_mapping(s4u::Actor::by_pid(leader_list[i]), i);
418 if(this->get_leaders_comm()==MPI_COMM_NULL){
419 leader_comm = new Comm(leaders_group, nullptr,1);
420 this->set_leaders_comm(leader_comm);
422 leader_comm=this->get_leaders_comm();
423 Group::unref(leaders_group);
425 smpi_process()->set_comm_intra(comm_intra);
428 // Are the nodes uniform ? = same number of process/node
429 int my_local_size=comm_intra->size();
430 if(comm_intra->rank()==0) {
432 int* non_uniform_map = xbt_new0(int,leader_group_size);
433 Coll_allgather_ring::allgather(&my_local_size, 1, MPI_INT,
434 non_uniform_map, 1, MPI_INT, leader_comm);
435 for(i=0; i < leader_group_size; i++) {
436 if(non_uniform_map[0] != non_uniform_map[i]) {
441 if(is_uniform==0 && this->is_uniform()!=0){
442 non_uniform_map_ = non_uniform_map;
444 xbt_free(non_uniform_map);
446 is_uniform_=is_uniform;
448 Coll_bcast_scatter_LR_allgather::bcast(&(is_uniform_),1, MPI_INT, 0, comm_intra );
450 if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) {
451 // we need to switch as the called function may silently touch global variables
452 smpi_switch_data_segment(s4u::Actor::self());
454 // Are the ranks blocked ? = allocated contiguously on the SMP nodes
456 int prev=this->group()->rank(comm_intra->group()->actor(0));
457 for (i = 1; i < my_local_size; i++) {
458 int that = this->group()->rank(comm_intra->group()->actor(i));
459 if (that != prev + 1) {
467 Coll_allreduce_default::allreduce(&is_blocked, &(global_blocked), 1, MPI_INT, MPI_LAND, this);
469 if(MPI_COMM_WORLD==MPI_COMM_UNINITIALIZED || this==MPI_COMM_WORLD){
471 is_blocked_ = global_blocked;
474 is_blocked_=global_blocked;
476 delete[] leader_list;
479 smpi_process()->set_replaying(true);
482 MPI_Comm Comm::f2c(int id) {
484 return MPI_COMM_SELF;
486 return MPI_COMM_WORLD;
487 } else if(F2C::f2c_lookup() != nullptr && id >= 0) {
489 const auto& lookup = F2C::f2c_lookup();
490 auto comm = lookup->find(get_key_id(key, id));
491 return comm == lookup->end() ? MPI_COMM_NULL : static_cast<MPI_Comm>(comm->second);
493 return MPI_COMM_NULL;
497 void Comm::free_f(int id) {
499 F2C::f2c_lookup()->erase(id == 0 ? get_key(key, id) : get_key_id(key, id));
503 if(F2C::f2c_lookup()==nullptr){
504 F2C::set_f2c_lookup(new std::unordered_map<std::string, F2C*>);
507 (*(F2C::f2c_lookup()))[this == MPI_COMM_WORLD ? get_key(key, F2C::f2c_id()) : get_key_id(key, F2C::f2c_id())] = this;
509 return F2C::f2c_id()-1;
512 void Comm::add_rma_win(MPI_Win win){
513 rma_wins_.push_back(win);
516 void Comm::remove_rma_win(MPI_Win win){
517 rma_wins_.remove(win);
520 void Comm::finish_rma_calls(){
521 for (auto const& it : rma_wins_) {
522 if(it->rank()==this->rank()){//is it ours (for MPI_COMM_WORLD)?
523 int finished = it->finish_comms();
524 XBT_DEBUG("Barrier for rank %d - Finished %d RMA calls",this->rank(), finished);
529 MPI_Info Comm::info(){
530 if(info_== MPI_INFO_NULL)
536 void Comm::set_info(MPI_Info info){
537 if(info_!= MPI_INFO_NULL)
542 MPI_Comm Comm::split_type(int type, int /*key*/, MPI_Info)
544 //MPI_UNDEFINED can be given to some nodes... but we need them to still perform the smp part which is collective
545 if(type != MPI_COMM_TYPE_SHARED && type != MPI_UNDEFINED){
546 return MPI_COMM_NULL;
549 MPI_Comm res= this->find_intra_comm(&leader);
550 if(type != MPI_UNDEFINED)
554 return MPI_COMM_NULL;
559 } // namespace simgrid