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;
43 void Comm::destroy(Comm* comm)
45 if (comm == MPI_COMM_UNINITIALIZED){
46 Comm::destroy(smpi_process()->comm_world());
49 delete comm->topo_; // there's no use count on topos
53 int Comm::dup(MPI_Comm* newcomm){
54 if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) {
55 // we need to switch as the called function may silently touch global variables
56 smpi_switch_data_segment(s4u::Actor::self());
58 MPI_Group cp = new Group(this->group());
59 (*newcomm) = new Comm(cp, this->topo());
60 int ret = MPI_SUCCESS;
62 if (not attributes()->empty()) {
64 void* value_out=nullptr;
65 for (auto const& it : *attributes()) {
66 smpi_key_elem elem = keyvals_.at(it.first);
68 if( elem->copy_fn.comm_copy_fn != MPI_NULL_COPY_FN &&
69 elem->copy_fn.comm_copy_fn != MPI_COMM_DUP_FN)
70 ret = elem->copy_fn.comm_copy_fn(this, it.first, elem->extra_state, it.second, &value_out, &flag);
71 else if ( elem->copy_fn.comm_copy_fn_fort != MPI_NULL_COPY_FN &&
72 *(int*)*elem->copy_fn.comm_copy_fn_fort != 1){
73 value_out=(int*)xbt_malloc(sizeof(int));
74 elem->copy_fn.comm_copy_fn_fort(this, it.first, elem->extra_state, it.second, value_out, &flag,&ret);
76 if (ret != MPI_SUCCESS) {
77 Comm::destroy(*newcomm);
78 *newcomm = MPI_COMM_NULL;
81 if (elem->copy_fn.comm_copy_fn == MPI_COMM_DUP_FN ||
82 ((elem->copy_fn.comm_copy_fn_fort != MPI_NULL_COPY_FN) && *(int*)*elem->copy_fn.comm_copy_fn_fort == 1)){
84 (*newcomm)->attributes()->insert({it.first, it.second});
87 (*newcomm)->attributes()->insert({it.first, value_out});
95 MPI_Group Comm::group()
97 if (this == MPI_COMM_UNINITIALIZED)
98 return smpi_process()->comm_world()->group();
104 if (this == MPI_COMM_UNINITIALIZED)
105 return smpi_process()->comm_world()->size();
106 return group_->size();
111 if (this == MPI_COMM_UNINITIALIZED)
112 return smpi_process()->comm_world()->rank();
113 return group_->rank(s4u::Actor::self());
116 void Comm::get_name (char* name, int* len)
118 if (this == MPI_COMM_UNINITIALIZED){
119 smpi_process()->comm_world()->get_name(name, len);
122 if(this == MPI_COMM_WORLD && name_.empty()) {
123 strncpy(name, "MPI_COMM_WORLD", 15);
125 } else if(this == MPI_COMM_SELF && name_.empty()) {
126 strncpy(name, "MPI_COMM_SELF", 14);
129 *len = snprintf(name, MPI_MAX_NAME_STRING+1, "%s", name_.c_str());
133 void Comm::set_name (char* name)
135 if (this == MPI_COMM_UNINITIALIZED){
136 smpi_process()->comm_world()->set_name(name);
139 name_.replace (0, MPI_MAX_NAME_STRING+1, name);
143 void Comm::set_leaders_comm(MPI_Comm leaders){
144 if (this == MPI_COMM_UNINITIALIZED){
145 smpi_process()->comm_world()->set_leaders_comm(leaders);
148 leaders_comm_=leaders;
151 int* Comm::get_non_uniform_map(){
152 if (this == MPI_COMM_UNINITIALIZED)
153 return smpi_process()->comm_world()->get_non_uniform_map();
154 return non_uniform_map_;
157 int* Comm::get_leaders_map(){
158 if (this == MPI_COMM_UNINITIALIZED)
159 return smpi_process()->comm_world()->get_leaders_map();
163 MPI_Comm Comm::get_leaders_comm(){
164 if (this == MPI_COMM_UNINITIALIZED)
165 return smpi_process()->comm_world()->get_leaders_comm();
166 return leaders_comm_;
169 MPI_Comm Comm::get_intra_comm(){
170 if (this == MPI_COMM_UNINITIALIZED || this==MPI_COMM_WORLD)
171 return smpi_process()->comm_intra();
172 else return intra_comm_;
175 int Comm::is_uniform(){
176 if (this == MPI_COMM_UNINITIALIZED)
177 return smpi_process()->comm_world()->is_uniform();
181 int Comm::is_blocked(){
182 if (this == MPI_COMM_UNINITIALIZED)
183 return smpi_process()->comm_world()->is_blocked();
187 int Comm::is_smp_comm(){
188 if (this == MPI_COMM_UNINITIALIZED)
189 return smpi_process()->comm_world()->is_smp_comm();
193 MPI_Comm Comm::split(int color, int key)
195 if (this == MPI_COMM_UNINITIALIZED)
196 return smpi_process()->comm_world()->split(color, key);
197 int system_tag = 123;
200 MPI_Group group_root = nullptr;
201 MPI_Group group_out = nullptr;
202 MPI_Group group = this->group();
203 int myrank = this->rank();
204 int size = this->size();
205 /* Gather all colors and keys on rank 0 */
206 int* sendbuf = xbt_new(int, 2);
210 recvbuf = xbt_new(int, 2 * size);
214 Coll_gather_default::gather(sendbuf, 2, MPI_INT, recvbuf, 2, MPI_INT, 0, this);
216 /* Do the actual job */
218 MPI_Group* group_snd = xbt_new(MPI_Group, size);
219 std::vector<std::pair<int, int>> rankmap;
220 rankmap.reserve(size);
221 for (int i = 0; i < size; i++) {
222 if (recvbuf[2 * i] != MPI_UNDEFINED) {
224 for (int j = i + 1; j < size; j++) {
225 if(recvbuf[2 * i] == recvbuf[2 * j]) {
226 recvbuf[2 * j] = MPI_UNDEFINED;
227 rankmap.push_back({recvbuf[2 * j + 1], j});
230 /* Add self in the group */
231 recvbuf[2 * i] = MPI_UNDEFINED;
232 rankmap.push_back({recvbuf[2 * i + 1], i});
233 std::sort(begin(rankmap), end(rankmap));
234 group_out = new Group(rankmap.size());
236 group_root = group_out; /* Save root's group */
238 for (unsigned j = 0; j < rankmap.size(); j++) {
239 s4u::ActorPtr actor = group->actor(rankmap[j].second);
240 group_out->set_mapping(actor, j);
242 MPI_Request* requests = xbt_new(MPI_Request, rankmap.size());
244 for (auto const& rank : rankmap) {
245 if (rank.second != 0) {
246 group_snd[reqs]=new Group(group_out);
247 requests[reqs] = Request::isend(&(group_snd[reqs]), 1, MPI_PTR, rank.second, system_tag, this);
251 if(i != 0 && group_out != MPI_COMM_WORLD->group() && group_out != MPI_GROUP_EMPTY)
252 Group::unref(group_out);
254 Request::waitall(reqs, requests, MPI_STATUS_IGNORE);
260 group_out = group_root; /* exit with root's group */
262 if(color != MPI_UNDEFINED) {
263 Request::recv(&group_out, 1, MPI_PTR, 0, system_tag, this, MPI_STATUS_IGNORE);
264 } /* otherwise, exit with group_out == nullptr */
266 return group_out!=nullptr ? new Comm(group_out, nullptr) : MPI_COMM_NULL;
270 if (this == MPI_COMM_UNINITIALIZED){
271 smpi_process()->comm_world()->ref();
278 void Comm::cleanup_smp(){
279 if (intra_comm_ != MPI_COMM_NULL)
280 Comm::unref(intra_comm_);
281 if (leaders_comm_ != MPI_COMM_NULL)
282 Comm::unref(leaders_comm_);
283 xbt_free(non_uniform_map_);
284 delete[] leaders_map_;
287 void Comm::unref(Comm* comm){
288 if (comm == MPI_COMM_UNINITIALIZED){
289 Comm::unref(smpi_process()->comm_world());
293 Group::unref(comm->group_);
295 if(comm->refcount_==0){
297 comm->cleanup_attr<Comm>();
302 void Comm::init_smp(){
305 if (this == MPI_COMM_UNINITIALIZED)
306 smpi_process()->comm_world()->init_smp();
308 int comm_size = this->size();
310 // If we are in replay - perform an ugly hack
311 // tell SimGrid we are not in replay for a while, because we need the buffers to be copied for the following calls
312 bool replaying = false; //cache data to set it back again after
313 if(smpi_process()->replaying()){
315 smpi_process()->set_replaying(false);
318 if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) {
319 // we need to switch as the called function may silently touch global variables
320 smpi_switch_data_segment(s4u::Actor::self());
322 //identify neighbours in comm
323 //get the indices of all processes sharing the same simix host
324 auto& process_list = sg_host_self()->pimpl_->process_list_;
325 int intra_comm_size = 0;
326 int min_index = INT_MAX; // the minimum index will be the leader
327 for (auto& actor : process_list) {
328 int index = actor.get_pid();
329 if (this->group()->rank(actor.iface()) != MPI_UNDEFINED) { // Is this process in the current group?
331 if (index < min_index)
335 XBT_DEBUG("number of processes deployed on my node : %d", intra_comm_size);
336 MPI_Group group_intra = new Group(intra_comm_size);
338 for (auto& actor : process_list) {
339 if (this->group()->rank(actor.iface()) != MPI_UNDEFINED) {
340 group_intra->set_mapping(actor.iface(), i);
345 MPI_Comm comm_intra = new Comm(group_intra, nullptr, 1);
348 int* leaders_map = new int[comm_size];
349 int* leader_list = new int[comm_size];
350 std::fill_n(leaders_map, comm_size, 0);
351 std::fill_n(leader_list, comm_size, -1);
353 Coll_allgather_ring::allgather(&leader, 1, MPI_INT , leaders_map, 1, MPI_INT, this);
355 if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) {
356 // we need to switch as the called function may silently touch global variables
357 smpi_switch_data_segment(s4u::Actor::self());
360 if(leaders_map_==nullptr){
361 leaders_map_= leaders_map;
363 delete[] leaders_map;
365 int leader_group_size = 0;
366 for(i=0; i<comm_size; i++){
367 int already_done = 0;
368 for (int j = 0; j < leader_group_size; j++) {
369 if (leaders_map_[i] == leader_list[j]) {
373 if (already_done == 0) {
374 leader_list[leader_group_size] = leaders_map_[i];
378 std::sort(leader_list, leader_list + leader_group_size);
380 MPI_Group leaders_group = new Group(leader_group_size);
382 MPI_Comm leader_comm = MPI_COMM_NULL;
383 if(MPI_COMM_WORLD!=MPI_COMM_UNINITIALIZED && this!=MPI_COMM_WORLD){
384 //create leader_communicator
385 for (i=0; i< leader_group_size;i++)
386 leaders_group->set_mapping(s4u::Actor::by_pid(leader_list[i]), i);
387 leader_comm = new Comm(leaders_group, nullptr,1);
388 this->set_leaders_comm(leader_comm);
389 this->set_intra_comm(comm_intra);
391 // create intracommunicator
393 for (i=0; i< leader_group_size;i++)
394 leaders_group->set_mapping(s4u::Actor::by_pid(leader_list[i]), i);
396 if(this->get_leaders_comm()==MPI_COMM_NULL){
397 leader_comm = new Comm(leaders_group, nullptr,1);
398 this->set_leaders_comm(leader_comm);
400 leader_comm=this->get_leaders_comm();
401 Group::unref(leaders_group);
403 smpi_process()->set_comm_intra(comm_intra);
406 // Are the nodes uniform ? = same number of process/node
407 int my_local_size=comm_intra->size();
408 if(comm_intra->rank()==0) {
410 int* non_uniform_map = xbt_new0(int,leader_group_size);
411 Coll_allgather_ring::allgather(&my_local_size, 1, MPI_INT,
412 non_uniform_map, 1, MPI_INT, leader_comm);
413 for(i=0; i < leader_group_size; i++) {
414 if(non_uniform_map[0] != non_uniform_map[i]) {
419 if(is_uniform==0 && this->is_uniform()!=0){
420 non_uniform_map_ = non_uniform_map;
422 xbt_free(non_uniform_map);
424 is_uniform_=is_uniform;
426 Coll_bcast_scatter_LR_allgather::bcast(&(is_uniform_),1, MPI_INT, 0, comm_intra );
428 if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) {
429 // we need to switch as the called function may silently touch global variables
430 smpi_switch_data_segment(s4u::Actor::self());
432 // Are the ranks blocked ? = allocated contiguously on the SMP nodes
434 int prev=this->group()->rank(comm_intra->group()->actor(0));
435 for (i = 1; i < my_local_size; i++) {
436 int that = this->group()->rank(comm_intra->group()->actor(i));
437 if (that != prev + 1) {
445 Coll_allreduce_default::allreduce(&is_blocked, &(global_blocked), 1, MPI_INT, MPI_LAND, this);
447 if(MPI_COMM_WORLD==MPI_COMM_UNINITIALIZED || this==MPI_COMM_WORLD){
449 is_blocked_ = global_blocked;
452 is_blocked_=global_blocked;
454 delete[] leader_list;
457 smpi_process()->set_replaying(true);
460 MPI_Comm Comm::f2c(int id) {
462 return MPI_COMM_SELF;
464 return MPI_COMM_WORLD;
465 } else if(F2C::f2c_lookup() != nullptr && id >= 0) {
467 const auto& lookup = F2C::f2c_lookup();
468 auto comm = lookup->find(get_key_id(key, id));
469 return comm == lookup->end() ? MPI_COMM_NULL : static_cast<MPI_Comm>(comm->second);
471 return MPI_COMM_NULL;
475 void Comm::free_f(int id) {
477 F2C::f2c_lookup()->erase(id == 0 ? get_key(key, id) : get_key_id(key, id));
481 if(F2C::f2c_lookup()==nullptr){
482 F2C::set_f2c_lookup(new std::unordered_map<std::string, F2C*>);
485 (*(F2C::f2c_lookup()))[this == MPI_COMM_WORLD ? get_key(key, F2C::f2c_id()) : get_key_id(key, F2C::f2c_id())] = this;
487 return F2C::f2c_id()-1;
490 void Comm::add_rma_win(MPI_Win win){
491 rma_wins_.push_back(win);
494 void Comm::remove_rma_win(MPI_Win win){
495 rma_wins_.remove(win);
498 void Comm::finish_rma_calls(){
499 for (auto const& it : rma_wins_) {
500 if(it->rank()==this->rank()){//is it ours (for MPI_COMM_WORLD)?
501 int finished = it->finish_comms();
502 XBT_DEBUG("Barrier for rank %d - Finished %d RMA calls",this->rank(), finished);
507 MPI_Info Comm::info(){
508 if(info_== MPI_INFO_NULL)
514 void Comm::set_info(MPI_Info info){
515 if(info_!= MPI_INFO_NULL)
520 MPI_Comm Comm::split_type(int type, int key, MPI_Info info)
522 if(type != MPI_COMM_TYPE_SHARED){
523 return MPI_COMM_NULL;
527 this->get_intra_comm()->ref();
528 return this->get_intra_comm();
532 } // namespace simgrid