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 "simgrid/s4u/Host.hpp"
9 #include "src/simix/smx_private.h"
11 #include "private.hpp"
12 #include "smpi_comm.hpp"
13 #include "smpi_coll.hpp"
14 #include "smpi_datatype.hpp"
15 #include "smpi_process.hpp"
16 #include "smpi_request.hpp"
17 #include "smpi_status.hpp"
18 #include "smpi_win.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. */
28 static int smpi_compare_rankmap(const void *a, const void *b)
30 const int* x = static_cast<const int*>(a);
31 const int* y = static_cast<const int*>(b);
51 std::unordered_map<int, smpi_key_elem> Comm::keyvals_;
52 int Comm::keyval_id_=0;
54 Comm::Comm(MPI_Group group, MPI_Topology topo) : group_(group), topo_(topo)
57 topoType_ = MPI_INVALID_TOPO;
58 intra_comm_ = MPI_COMM_NULL;
59 leaders_comm_ = MPI_COMM_NULL;
61 non_uniform_map_ = nullptr;
62 leaders_map_ = nullptr;
66 void Comm::destroy(Comm* comm)
68 if (comm == MPI_COMM_UNINITIALIZED){
69 Comm::destroy(smpi_process()->comm_world());
72 delete comm->topo_; // there's no use count on topos
76 int Comm::dup(MPI_Comm* newcomm){
77 if(smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP){ //we need to switch as the called function may silently touch global variables
78 smpi_switch_data_segment(smpi_process()->index());
80 MPI_Group cp = new Group(this->group());
81 (*newcomm) = new Comm(cp, this->topo());
82 int ret = MPI_SUCCESS;
84 if (not attributes()->empty()) {
87 for(auto it : *attributes()){
88 smpi_key_elem elem = keyvals_.at(it.first);
89 if (elem != nullptr && elem->copy_fn.comm_copy_fn != MPI_NULL_COPY_FN) {
90 ret = elem->copy_fn.comm_copy_fn(this, it.first, nullptr, it.second, &value_out, &flag);
91 if (ret != MPI_SUCCESS) {
92 Comm::destroy(*newcomm);
93 *newcomm = MPI_COMM_NULL;
98 (*newcomm)->attributes()->insert({it.first, value_out});
106 MPI_Group Comm::group()
108 if (this == MPI_COMM_UNINITIALIZED)
109 return smpi_process()->comm_world()->group();
113 MPI_Topology Comm::topo() {
119 if (this == MPI_COMM_UNINITIALIZED)
120 return smpi_process()->comm_world()->size();
121 return group_->size();
126 if (this == MPI_COMM_UNINITIALIZED)
127 return smpi_process()->comm_world()->rank();
128 return group_->rank(smpi_process()->index());
131 void Comm::get_name (char* name, int* len)
133 if (this == MPI_COMM_UNINITIALIZED){
134 smpi_process()->comm_world()->get_name(name, len);
137 if(this == MPI_COMM_WORLD) {
138 strncpy(name, "WORLD",5);
141 *len = snprintf(name, MPI_MAX_NAME_STRING, "%p", this);
145 void Comm::set_leaders_comm(MPI_Comm leaders){
146 if (this == MPI_COMM_UNINITIALIZED){
147 smpi_process()->comm_world()->set_leaders_comm(leaders);
150 leaders_comm_=leaders;
153 void Comm::set_intra_comm(MPI_Comm leaders){
157 int* Comm::get_non_uniform_map(){
158 if (this == MPI_COMM_UNINITIALIZED)
159 return smpi_process()->comm_world()->get_non_uniform_map();
160 return non_uniform_map_;
163 int* Comm::get_leaders_map(){
164 if (this == MPI_COMM_UNINITIALIZED)
165 return smpi_process()->comm_world()->get_leaders_map();
169 MPI_Comm Comm::get_leaders_comm(){
170 if (this == MPI_COMM_UNINITIALIZED)
171 return smpi_process()->comm_world()->get_leaders_comm();
172 return leaders_comm_;
175 MPI_Comm Comm::get_intra_comm(){
176 if (this == MPI_COMM_UNINITIALIZED || this==MPI_COMM_WORLD)
177 return smpi_process()->comm_intra();
178 else return intra_comm_;
181 int Comm::is_uniform(){
182 if (this == MPI_COMM_UNINITIALIZED)
183 return smpi_process()->comm_world()->is_uniform();
187 int Comm::is_blocked(){
188 if (this == MPI_COMM_UNINITIALIZED)
189 return smpi_process()->comm_world()->is_blocked();
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 rank = 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 int* rankmap = xbt_new(int, 2 * size);
220 for (int i = 0; i < size; i++) {
221 if (recvbuf[2 * i] != MPI_UNDEFINED) {
223 for (int j = i + 1; j < size; j++) {
224 if(recvbuf[2 * i] == recvbuf[2 * j]) {
225 recvbuf[2 * j] = MPI_UNDEFINED;
226 rankmap[2 * count] = j;
227 rankmap[2 * count + 1] = recvbuf[2 * j + 1];
231 /* Add self in the group */
232 recvbuf[2 * i] = MPI_UNDEFINED;
233 rankmap[2 * count] = i;
234 rankmap[2 * count + 1] = recvbuf[2 * i + 1];
236 qsort(rankmap, count, 2 * sizeof(int), &smpi_compare_rankmap);
237 group_out = new Group(count);
239 group_root = group_out; /* Save root's group */
241 for (int j = 0; j < count; j++) {
242 int index = group->index(rankmap[2 * j]);
243 group_out->set_mapping(index, j);
245 MPI_Request* requests = xbt_new(MPI_Request, count);
247 for (int j = 0; j < count; j++) {
248 if(rankmap[2 * j] != 0) {
249 group_snd[reqs]=new Group(group_out);
250 requests[reqs] = Request::isend(&(group_snd[reqs]), 1, MPI_PTR, rankmap[2 * j], system_tag, this);
254 if(i != 0 && group_out != MPI_COMM_WORLD->group() && group_out != MPI_GROUP_EMPTY)
255 Group::unref(group_out);
257 Request::waitall(reqs, requests, MPI_STATUS_IGNORE);
264 group_out = group_root; /* exit with root's group */
266 if(color != MPI_UNDEFINED) {
267 Request::recv(&group_out, 1, MPI_PTR, 0, system_tag, this, MPI_STATUS_IGNORE);
268 } /* otherwise, exit with group_out == nullptr */
270 return group_out!=nullptr ? new Comm(group_out, nullptr) : MPI_COMM_NULL;
274 if (this == MPI_COMM_UNINITIALIZED){
275 smpi_process()->comm_world()->ref();
282 void Comm::cleanup_smp(){
283 if (intra_comm_ != MPI_COMM_NULL)
284 Comm::unref(intra_comm_);
285 if (leaders_comm_ != MPI_COMM_NULL)
286 Comm::unref(leaders_comm_);
287 if (non_uniform_map_ != nullptr)
288 xbt_free(non_uniform_map_);
289 if (leaders_map_ != nullptr)
290 xbt_free(leaders_map_);
293 void Comm::unref(Comm* comm){
294 if (comm == MPI_COMM_UNINITIALIZED){
295 Comm::unref(smpi_process()->comm_world());
299 Group::unref(comm->group_);
301 if(comm->refcount_==0){
303 comm->cleanup_attr<Comm>();
308 static int compare_ints (const void *a, const void *b)
310 const int *da = static_cast<const int *>(a);
311 const int *db = static_cast<const int *>(b);
313 return static_cast<int>(*da > *db) - static_cast<int>(*da < *db);
316 void Comm::init_smp(){
319 if (this == MPI_COMM_UNINITIALIZED)
320 smpi_process()->comm_world()->init_smp();
322 int comm_size = this->size();
324 // If we are in replay - perform an ugly hack
325 // tell SimGrid we are not in replay for a while, because we need the buffers to be copied for the following calls
326 bool replaying = false; //cache data to set it back again after
327 if(smpi_process()->replaying()){
329 smpi_process()->set_replaying(false);
332 if(smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP){ //we need to switch as the called function may silently touch global variables
333 smpi_switch_data_segment(smpi_process()->index());
335 //identify neighbours in comm
336 //get the indexes of all processes sharing the same simix host
337 xbt_swag_t process_list = sg_host_self()->extension<simgrid::simix::Host>()->process_list;
338 int intra_comm_size = 0;
339 int min_index = INT_MAX; // the minimum index will be the leader
340 smx_actor_t actor = nullptr;
341 xbt_swag_foreach(actor, process_list)
343 int index = actor->pid - 1;
345 if (this->group()->rank(index) != MPI_UNDEFINED) {
347 // the process is in the comm
348 if (index < min_index)
352 XBT_DEBUG("number of processes deployed on my node : %d", intra_comm_size);
353 MPI_Group group_intra = new Group(intra_comm_size);
356 xbt_swag_foreach(actor, process_list) {
357 int index = actor->pid -1;
358 if(this->group()->rank(index)!=MPI_UNDEFINED){
359 group_intra->set_mapping(index, i);
364 MPI_Comm comm_intra = new Comm(group_intra, nullptr);
367 int * leaders_map= static_cast<int*>(xbt_malloc0(sizeof(int)*comm_size));
368 int * leader_list= static_cast<int*>(xbt_malloc0(sizeof(int)*comm_size));
369 for(i=0; i<comm_size; i++){
373 Coll_allgather_mpich::allgather(&leader, 1, MPI_INT , leaders_map, 1, MPI_INT, this);
375 if(smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP){ //we need to switch as the called function may silently touch global variables
376 smpi_switch_data_segment(smpi_process()->index());
379 if(leaders_map_==nullptr){
380 leaders_map_= leaders_map;
382 xbt_free(leaders_map);
385 int leader_group_size = 0;
386 for(i=0; i<comm_size; i++){
388 for(j=0;j<leader_group_size; j++){
389 if(leaders_map_[i]==leader_list[j]){
394 leader_list[leader_group_size]=leaders_map_[i];
398 qsort(leader_list, leader_group_size, sizeof(int),compare_ints);
400 MPI_Group leaders_group = new Group(leader_group_size);
402 MPI_Comm leader_comm = MPI_COMM_NULL;
403 if(MPI_COMM_WORLD!=MPI_COMM_UNINITIALIZED && this!=MPI_COMM_WORLD){
404 //create leader_communicator
405 for (i=0; i< leader_group_size;i++)
406 leaders_group->set_mapping(leader_list[i], i);
407 leader_comm = new Comm(leaders_group, nullptr);
408 this->set_leaders_comm(leader_comm);
409 this->set_intra_comm(comm_intra);
411 //create intracommunicator
413 for (i=0; i< leader_group_size;i++)
414 leaders_group->set_mapping(leader_list[i], i);
416 if(this->get_leaders_comm()==MPI_COMM_NULL){
417 leader_comm = new Comm(leaders_group, nullptr);
418 this->set_leaders_comm(leader_comm);
420 leader_comm=this->get_leaders_comm();
421 Group::unref(leaders_group);
423 smpi_process()->set_comm_intra(comm_intra);
426 // Are the nodes uniform ? = same number of process/node
427 int my_local_size=comm_intra->size();
428 if(comm_intra->rank()==0) {
430 int* non_uniform_map = xbt_new0(int,leader_group_size);
431 Coll_allgather_mpich::allgather(&my_local_size, 1, MPI_INT,
432 non_uniform_map, 1, MPI_INT, leader_comm);
433 for(i=0; i < leader_group_size; i++) {
434 if(non_uniform_map[0] != non_uniform_map[i]) {
439 if(is_uniform==0 && this->is_uniform()!=0){
440 non_uniform_map_= non_uniform_map;
442 xbt_free(non_uniform_map);
444 is_uniform_=is_uniform;
446 Coll_bcast_mpich::bcast(&(is_uniform_),1, MPI_INT, 0, comm_intra );
448 if(smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP){ //we need to switch as the called function may silently touch global variables
449 smpi_switch_data_segment(smpi_process()->index());
451 // Are the ranks blocked ? = allocated contiguously on the SMP nodes
453 int prev=this->group()->rank(comm_intra->group()->index(0));
454 for (i=1; i<my_local_size; i++){
455 int that=this->group()->rank(comm_intra->group()->index(i));
464 Coll_allreduce_default::allreduce(&is_blocked, &(global_blocked), 1, MPI_INT, MPI_LAND, this);
466 if(MPI_COMM_WORLD==MPI_COMM_UNINITIALIZED || this==MPI_COMM_WORLD){
468 is_blocked_=global_blocked;
471 is_blocked_=global_blocked;
473 xbt_free(leader_list);
476 smpi_process()->set_replaying(true);
479 MPI_Comm Comm::f2c(int id) {
481 return MPI_COMM_SELF;
483 return MPI_COMM_WORLD;
484 } else if(F2C::f2c_lookup() != nullptr && id >= 0) {
486 MPI_Comm tmp = static_cast<MPI_Comm>(xbt_dict_get_or_null(F2C::f2c_lookup(),get_key_id(key, id)));
487 return tmp != nullptr ? tmp : MPI_COMM_NULL ;
489 return MPI_COMM_NULL;
493 void Comm::free_f(int id) {
495 xbt_dict_remove(F2C::f2c_lookup(), id==0? get_key(key, id) : get_key_id(key, id));
499 if(F2C::f2c_lookup()==nullptr){
500 F2C::set_f2c_lookup(xbt_dict_new_homogeneous(nullptr));
503 xbt_dict_set(F2C::f2c_lookup(), this==MPI_COMM_WORLD? get_key(key, F2C::f2c_id()) : get_key_id(key,F2C::f2c_id()), this, nullptr);
505 return F2C::f2c_id()-1;
509 void Comm::add_rma_win(MPI_Win win){
510 rma_wins_.push_back(win);
513 void Comm::remove_rma_win(MPI_Win win){
514 rma_wins_.remove(win);
517 void Comm::finish_rma_calls(){
518 for(auto it : rma_wins_){
519 if(it->rank()==this->rank()){//is it ours (for MPI_COMM_WORLD)?
520 int finished = it->finish_comms();
521 XBT_DEBUG("Barrier for rank %d - Finished %d RMA calls",this->rank(), finished);