1 /* Copyright (c) 2010-2017. 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"
8 #include "simgrid/s4u/Host.hpp"
9 #include "smpi_coll.hpp"
10 #include "smpi_datatype.hpp"
11 #include "smpi_process.hpp"
12 #include "smpi_request.hpp"
13 #include "smpi_status.hpp"
14 #include "smpi_win.hpp"
15 #include "src/simix/smx_private.hpp"
20 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_comm, smpi, "Logging specific to SMPI (comm)");
22 simgrid::smpi::Comm mpi_MPI_COMM_UNINITIALIZED;
23 MPI_Comm MPI_COMM_UNINITIALIZED=&mpi_MPI_COMM_UNINITIALIZED;
25 /* Support for cartesian topology was added, but there are 2 other types of topology, graph et dist graph. In order to
26 * support them, we have to add a field SMPI_Topo_type, and replace the MPI_Topology field by an union. */
31 std::unordered_map<int, smpi_key_elem> Comm::keyvals_;
32 int Comm::keyval_id_=0;
34 Comm::Comm(MPI_Group group, MPI_Topology topo) : group_(group), topo_(topo)
37 topoType_ = MPI_INVALID_TOPO;
38 intra_comm_ = MPI_COMM_NULL;
39 leaders_comm_ = MPI_COMM_NULL;
41 non_uniform_map_ = nullptr;
42 leaders_map_ = nullptr;
46 void Comm::destroy(Comm* comm)
48 if (comm == MPI_COMM_UNINITIALIZED){
49 Comm::destroy(smpi_process()->comm_world());
52 delete comm->topo_; // there's no use count on topos
56 int Comm::dup(MPI_Comm* newcomm){
57 if(smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP){ //we need to switch as the called function may silently touch global variables
58 smpi_switch_data_segment(smpi_process()->index());
60 MPI_Group cp = new Group(this->group());
61 (*newcomm) = new Comm(cp, this->topo());
62 int ret = MPI_SUCCESS;
64 if (not attributes()->empty()) {
67 for (auto const& it : *attributes()) {
68 smpi_key_elem elem = keyvals_.at(it.first);
69 if (elem != nullptr && elem->copy_fn.comm_copy_fn != MPI_NULL_COPY_FN) {
70 ret = elem->copy_fn.comm_copy_fn(this, it.first, nullptr, it.second, &value_out, &flag);
71 if (ret != MPI_SUCCESS) {
72 Comm::destroy(*newcomm);
73 *newcomm = MPI_COMM_NULL;
78 (*newcomm)->attributes()->insert({it.first, value_out});
86 MPI_Group Comm::group()
88 if (this == MPI_COMM_UNINITIALIZED)
89 return smpi_process()->comm_world()->group();
93 MPI_Topology Comm::topo() {
99 if (this == MPI_COMM_UNINITIALIZED)
100 return smpi_process()->comm_world()->size();
101 return group_->size();
106 if (this == MPI_COMM_UNINITIALIZED)
107 return smpi_process()->comm_world()->rank();
108 return group_->rank(smpi_process()->index());
111 void Comm::get_name (char* name, int* len)
113 if (this == MPI_COMM_UNINITIALIZED){
114 smpi_process()->comm_world()->get_name(name, len);
117 if(this == MPI_COMM_WORLD) {
118 strncpy(name, "WORLD",5);
121 *len = snprintf(name, MPI_MAX_NAME_STRING, "%p", this);
125 void Comm::set_leaders_comm(MPI_Comm leaders){
126 if (this == MPI_COMM_UNINITIALIZED){
127 smpi_process()->comm_world()->set_leaders_comm(leaders);
130 leaders_comm_=leaders;
133 void Comm::set_intra_comm(MPI_Comm leaders){
137 int* Comm::get_non_uniform_map(){
138 if (this == MPI_COMM_UNINITIALIZED)
139 return smpi_process()->comm_world()->get_non_uniform_map();
140 return non_uniform_map_;
143 int* Comm::get_leaders_map(){
144 if (this == MPI_COMM_UNINITIALIZED)
145 return smpi_process()->comm_world()->get_leaders_map();
149 MPI_Comm Comm::get_leaders_comm(){
150 if (this == MPI_COMM_UNINITIALIZED)
151 return smpi_process()->comm_world()->get_leaders_comm();
152 return leaders_comm_;
155 MPI_Comm Comm::get_intra_comm(){
156 if (this == MPI_COMM_UNINITIALIZED || this==MPI_COMM_WORLD)
157 return smpi_process()->comm_intra();
158 else return intra_comm_;
161 int Comm::is_uniform(){
162 if (this == MPI_COMM_UNINITIALIZED)
163 return smpi_process()->comm_world()->is_uniform();
167 int Comm::is_blocked(){
168 if (this == MPI_COMM_UNINITIALIZED)
169 return smpi_process()->comm_world()->is_blocked();
173 MPI_Comm Comm::split(int color, int key)
175 if (this == MPI_COMM_UNINITIALIZED)
176 return smpi_process()->comm_world()->split(color, key);
177 int system_tag = 123;
180 MPI_Group group_root = nullptr;
181 MPI_Group group_out = nullptr;
182 MPI_Group group = this->group();
183 int rank = this->rank();
184 int size = this->size();
185 /* Gather all colors and keys on rank 0 */
186 int* sendbuf = xbt_new(int, 2);
190 recvbuf = xbt_new(int, 2 * size);
194 Coll_gather_default::gather(sendbuf, 2, MPI_INT, recvbuf, 2, MPI_INT, 0, this);
196 /* Do the actual job */
198 MPI_Group* group_snd = xbt_new(MPI_Group, size);
199 std::vector<std::pair<int, int>> rankmap;
200 rankmap.reserve(size);
201 for (int i = 0; i < size; i++) {
202 if (recvbuf[2 * i] != MPI_UNDEFINED) {
204 for (int j = i + 1; j < size; j++) {
205 if(recvbuf[2 * i] == recvbuf[2 * j]) {
206 recvbuf[2 * j] = MPI_UNDEFINED;
207 rankmap.push_back({recvbuf[2 * j + 1], j});
210 /* Add self in the group */
211 recvbuf[2 * i] = MPI_UNDEFINED;
212 rankmap.push_back({recvbuf[2 * i + 1], i});
213 std::sort(begin(rankmap), end(rankmap));
214 group_out = new Group(rankmap.size());
216 group_root = group_out; /* Save root's group */
218 for (unsigned j = 0; j < rankmap.size(); j++) {
219 int index = group->index(rankmap[j].second);
220 group_out->set_mapping(index, j);
222 MPI_Request* requests = xbt_new(MPI_Request, rankmap.size());
224 for (auto const& rank : rankmap) {
225 if (rank.second != 0) {
226 group_snd[reqs]=new Group(group_out);
227 requests[reqs] = Request::isend(&(group_snd[reqs]), 1, MPI_PTR, rank.second, system_tag, this);
231 if(i != 0 && group_out != MPI_COMM_WORLD->group() && group_out != MPI_GROUP_EMPTY)
232 Group::unref(group_out);
234 Request::waitall(reqs, requests, MPI_STATUS_IGNORE);
240 group_out = group_root; /* exit with root's group */
242 if(color != MPI_UNDEFINED) {
243 Request::recv(&group_out, 1, MPI_PTR, 0, system_tag, this, MPI_STATUS_IGNORE);
244 } /* otherwise, exit with group_out == nullptr */
246 return group_out!=nullptr ? new Comm(group_out, nullptr) : MPI_COMM_NULL;
250 if (this == MPI_COMM_UNINITIALIZED){
251 smpi_process()->comm_world()->ref();
258 void Comm::cleanup_smp(){
259 if (intra_comm_ != MPI_COMM_NULL)
260 Comm::unref(intra_comm_);
261 if (leaders_comm_ != MPI_COMM_NULL)
262 Comm::unref(leaders_comm_);
263 if (non_uniform_map_ != nullptr)
264 xbt_free(non_uniform_map_);
265 if (leaders_map_ != nullptr)
266 xbt_free(leaders_map_);
269 void Comm::unref(Comm* comm){
270 if (comm == MPI_COMM_UNINITIALIZED){
271 Comm::unref(smpi_process()->comm_world());
275 Group::unref(comm->group_);
277 if(comm->refcount_==0){
279 comm->cleanup_attr<Comm>();
284 void Comm::init_smp(){
287 if (this == MPI_COMM_UNINITIALIZED)
288 smpi_process()->comm_world()->init_smp();
290 int comm_size = this->size();
292 // If we are in replay - perform an ugly hack
293 // tell SimGrid we are not in replay for a while, because we need the buffers to be copied for the following calls
294 bool replaying = false; //cache data to set it back again after
295 if(smpi_process()->replaying()){
297 smpi_process()->set_replaying(false);
300 if(smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP){ //we need to switch as the called function may silently touch global variables
301 smpi_switch_data_segment(smpi_process()->index());
303 //identify neighbours in comm
304 //get the indexes of all processes sharing the same simix host
305 xbt_swag_t process_list = sg_host_self()->extension<simgrid::simix::Host>()->process_list;
306 int intra_comm_size = 0;
307 int min_index = INT_MAX; // the minimum index will be the leader
308 smx_actor_t actor = nullptr;
309 xbt_swag_foreach(actor, process_list)
311 int index = actor->pid - 1;
313 if (this->group()->rank(index) != MPI_UNDEFINED) {
315 // the process is in the comm
316 if (index < min_index)
320 XBT_DEBUG("number of processes deployed on my node : %d", intra_comm_size);
321 MPI_Group group_intra = new Group(intra_comm_size);
324 xbt_swag_foreach(actor, process_list) {
325 int index = actor->pid -1;
326 if(this->group()->rank(index)!=MPI_UNDEFINED){
327 group_intra->set_mapping(index, i);
332 MPI_Comm comm_intra = new Comm(group_intra, nullptr);
335 int * leaders_map= static_cast<int*>(xbt_malloc0(sizeof(int)*comm_size));
336 int * leader_list= static_cast<int*>(xbt_malloc0(sizeof(int)*comm_size));
337 for(i=0; i<comm_size; i++){
341 Coll_allgather_mpich::allgather(&leader, 1, MPI_INT , leaders_map, 1, MPI_INT, this);
343 if(smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP){ //we need to switch as the called function may silently touch global variables
344 smpi_switch_data_segment(smpi_process()->index());
347 if(leaders_map_==nullptr){
348 leaders_map_= leaders_map;
350 xbt_free(leaders_map);
353 int leader_group_size = 0;
354 for(i=0; i<comm_size; i++){
355 int already_done = 0;
356 for (j = 0; j < leader_group_size; j++) {
357 if (leaders_map_[i] == leader_list[j]) {
361 if (already_done == 0) {
362 leader_list[leader_group_size] = leaders_map_[i];
366 std::sort(leader_list, leader_list + leader_group_size);
368 MPI_Group leaders_group = new Group(leader_group_size);
370 MPI_Comm leader_comm = MPI_COMM_NULL;
371 if(MPI_COMM_WORLD!=MPI_COMM_UNINITIALIZED && this!=MPI_COMM_WORLD){
372 //create leader_communicator
373 for (i=0; i< leader_group_size;i++)
374 leaders_group->set_mapping(leader_list[i], i);
375 leader_comm = new Comm(leaders_group, nullptr);
376 this->set_leaders_comm(leader_comm);
377 this->set_intra_comm(comm_intra);
379 // create intracommunicator
381 for (i=0; i< leader_group_size;i++)
382 leaders_group->set_mapping(leader_list[i], i);
384 if(this->get_leaders_comm()==MPI_COMM_NULL){
385 leader_comm = new Comm(leaders_group, nullptr);
386 this->set_leaders_comm(leader_comm);
388 leader_comm=this->get_leaders_comm();
389 Group::unref(leaders_group);
391 smpi_process()->set_comm_intra(comm_intra);
394 // Are the nodes uniform ? = same number of process/node
395 int my_local_size=comm_intra->size();
396 if(comm_intra->rank()==0) {
398 int* non_uniform_map = xbt_new0(int,leader_group_size);
399 Coll_allgather_mpich::allgather(&my_local_size, 1, MPI_INT,
400 non_uniform_map, 1, MPI_INT, leader_comm);
401 for(i=0; i < leader_group_size; i++) {
402 if(non_uniform_map[0] != non_uniform_map[i]) {
407 if(is_uniform==0 && this->is_uniform()!=0){
408 non_uniform_map_ = non_uniform_map;
410 xbt_free(non_uniform_map);
412 is_uniform_=is_uniform;
414 Coll_bcast_mpich::bcast(&(is_uniform_),1, MPI_INT, 0, comm_intra );
416 if(smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP){ //we need to switch as the called function may silently touch global variables
417 smpi_switch_data_segment(smpi_process()->index());
419 // Are the ranks blocked ? = allocated contiguously on the SMP nodes
421 int prev=this->group()->rank(comm_intra->group()->index(0));
422 for (i = 1; i < my_local_size; i++) {
423 int that = this->group()->rank(comm_intra->group()->index(i));
424 if (that != prev + 1) {
432 Coll_allreduce_default::allreduce(&is_blocked, &(global_blocked), 1, MPI_INT, MPI_LAND, this);
434 if(MPI_COMM_WORLD==MPI_COMM_UNINITIALIZED || this==MPI_COMM_WORLD){
436 is_blocked_ = global_blocked;
439 is_blocked_=global_blocked;
441 xbt_free(leader_list);
444 smpi_process()->set_replaying(true);
447 MPI_Comm Comm::f2c(int id) {
449 return MPI_COMM_SELF;
451 return MPI_COMM_WORLD;
452 } else if(F2C::f2c_lookup() != nullptr && id >= 0) {
454 const auto& lookup = F2C::f2c_lookup();
455 auto comm = lookup->find(get_key_id(key, id));
456 return comm == lookup->end() ? MPI_COMM_NULL : static_cast<MPI_Comm>(comm->second);
458 return MPI_COMM_NULL;
462 void Comm::free_f(int id) {
464 F2C::f2c_lookup()->erase(id == 0 ? get_key(key, id) : get_key_id(key, id));
468 if(F2C::f2c_lookup()==nullptr){
469 F2C::set_f2c_lookup(new std::unordered_map<std::string, F2C*>);
472 (*(F2C::f2c_lookup()))[this == MPI_COMM_WORLD ? get_key(key, F2C::f2c_id()) : get_key_id(key, F2C::f2c_id())] = this;
474 return F2C::f2c_id()-1;
477 void Comm::add_rma_win(MPI_Win win){
478 rma_wins_.push_back(win);
481 void Comm::remove_rma_win(MPI_Win win){
482 rma_wins_.remove(win);
485 void Comm::finish_rma_calls(){
486 for (auto const& it : rma_wins_) {
487 if(it->rank()==this->rank()){//is it ours (for MPI_COMM_WORLD)?
488 int finished = it->finish_comms();
489 XBT_DEBUG("Barrier for rank %d - Finished %d RMA calls",this->rank(), finished);