1 /* Copyright (c) 2010-2014. The SimGrid Team.
2 * All rights reserved. */
4 /* This program is free software; you can redistribute it and/or modify it
5 * under the terms of the license (GNU LGPL) which comes with this package. */
10 #include "smpi_mpi_dt_private.h"
12 #include "simix/smx_private.h"
13 #include "colls/colls.h"
16 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_comm, smpi,
17 "Logging specific to SMPI (comm)");
21 /* Support for cartesian topology was added, but there are 2 other types of
22 * topology, graph et dist graph. In order to support them, we have to add a
23 * field MPIR_Topo_type, and replace the MPI_Topology field by an union. */
25 typedef struct s_smpi_mpi_communicator {
27 MPIR_Topo_type topoType;
28 MPI_Topology topo; // to be replaced by an union
30 MPI_Comm leaders_comm;//inter-node communicator
31 MPI_Comm intra_comm;//intra-node communicator . For MPI_COMM_WORLD this can't be used, as var is global.
32 //use an intracomm stored in the process data instead
33 int* leaders_map; //who is the leader of each process
35 int* non_uniform_map; //set if smp nodes have a different number of processes allocated
36 int is_blocked;// are ranks allocated on the same smp node contiguous ?
37 } s_smpi_mpi_communicator_t;
39 static int smpi_compare_rankmap(const void *a, const void *b)
41 const int* x = (const int*)a;
42 const int* y = (const int*)b;
59 MPI_Comm smpi_comm_new(MPI_Group group, MPI_Topology topo)
63 comm = xbt_new(s_smpi_mpi_communicator_t, 1);
65 smpi_group_use(comm->group);
69 comm->intra_comm = MPI_COMM_NULL;
70 comm->leaders_comm = MPI_COMM_NULL;
72 comm->non_uniform_map = NULL;
73 comm->leaders_map = NULL;
78 void smpi_comm_destroy(MPI_Comm comm)
80 if (comm == MPI_COMM_UNINITIALIZED)
81 comm = smpi_process_comm_world();
82 smpi_group_unuse(comm->group);
83 smpi_topo_destroy(comm->topo); // there's no use count on topos
84 smpi_comm_unuse(comm);
87 MPI_Group smpi_comm_group(MPI_Comm comm)
89 if (comm == MPI_COMM_UNINITIALIZED)
90 comm = smpi_process_comm_world();
95 MPI_Topology smpi_comm_topo(MPI_Comm comm) {
96 if (comm != MPI_COMM_NULL)
101 int smpi_comm_size(MPI_Comm comm)
103 if (comm == MPI_COMM_UNINITIALIZED)
104 comm = smpi_process_comm_world();
106 return smpi_group_size(smpi_comm_group(comm));
109 int smpi_comm_rank(MPI_Comm comm)
111 if (comm == MPI_COMM_UNINITIALIZED)
112 comm = smpi_process_comm_world();
113 return smpi_group_rank(smpi_comm_group(comm), smpi_process_index());
116 void smpi_comm_get_name (MPI_Comm comm, char* name, int* len)
118 if (comm == MPI_COMM_UNINITIALIZED)
119 comm = smpi_process_comm_world();
120 if(comm == MPI_COMM_WORLD) {
121 strcpy(name, "WORLD");
124 *len = snprintf(name, MPI_MAX_NAME_STRING, "%p", comm);
128 void smpi_comm_set_leaders_comm(MPI_Comm comm, MPI_Comm leaders){
129 if (comm == MPI_COMM_UNINITIALIZED)
130 comm = smpi_process_comm_world();
131 comm->leaders_comm=leaders;
134 void smpi_comm_set_intra_comm(MPI_Comm comm, MPI_Comm leaders){
135 comm->intra_comm=leaders;
138 int* smpi_comm_get_non_uniform_map(MPI_Comm comm){
139 if (comm == MPI_COMM_UNINITIALIZED)
140 comm = smpi_process_comm_world();
141 return comm->non_uniform_map;
144 int* smpi_comm_get_leaders_map(MPI_Comm comm){
145 if (comm == MPI_COMM_UNINITIALIZED)
146 comm = smpi_process_comm_world();
147 return comm->leaders_map;
150 MPI_Comm smpi_comm_get_leaders_comm(MPI_Comm comm){
151 if (comm == MPI_COMM_UNINITIALIZED)
152 comm = smpi_process_comm_world();
153 return comm->leaders_comm;
156 MPI_Comm smpi_comm_get_intra_comm(MPI_Comm comm){
157 if (comm == MPI_COMM_UNINITIALIZED || comm==MPI_COMM_WORLD)
158 return smpi_process_get_comm_intra();
159 else return comm->intra_comm;
162 int smpi_comm_is_uniform(MPI_Comm comm){
163 if (comm == MPI_COMM_UNINITIALIZED)
164 comm = smpi_process_comm_world();
165 return comm->is_uniform;
168 int smpi_comm_is_blocked(MPI_Comm comm){
169 if (comm == MPI_COMM_UNINITIALIZED)
170 comm = smpi_process_comm_world();
171 return comm->is_blocked;
174 MPI_Comm smpi_comm_split(MPI_Comm comm, int color, int key)
176 if (comm == MPI_COMM_UNINITIALIZED)
177 comm = smpi_process_comm_world();
178 int system_tag = 123;
179 int index, rank, size, i, j, count, reqs;
183 MPI_Group group, group_root, group_out;
184 MPI_Request* requests;
186 group_root = group_out = NULL;
187 group = smpi_comm_group(comm);
188 rank = smpi_comm_rank(comm);
189 size = smpi_comm_size(comm);
190 /* Gather all colors and keys on rank 0 */
191 sendbuf = xbt_new(int, 2);
195 recvbuf = xbt_new(int, 2 * size);
199 smpi_mpi_gather(sendbuf, 2, MPI_INT, recvbuf, 2, MPI_INT, 0, comm);
201 /* Do the actual job */
203 rankmap = xbt_new(int, 2 * size);
204 for(i = 0; i < size; i++) {
205 if(recvbuf[2 * i] == MPI_UNDEFINED) {
209 for(j = i + 1; j < size; j++) {
210 if(recvbuf[2 * i] == recvbuf[2 * j]) {
211 recvbuf[2 * j] = MPI_UNDEFINED;
212 rankmap[2 * count] = j;
213 rankmap[2 * count + 1] = recvbuf[2 * j + 1];
217 /* Add self in the group */
218 recvbuf[2 * i] = MPI_UNDEFINED;
219 rankmap[2 * count] = i;
220 rankmap[2 * count + 1] = recvbuf[2 * i + 1];
222 qsort(rankmap, count, 2 * sizeof(int), &smpi_compare_rankmap);
223 group_out = smpi_group_new(count);
225 group_root = group_out; /* Save root's group */
227 for(j = 0; j < count; j++) {
228 //increment refcounter in order to avoid freeing the group too quick before copy
229 index = smpi_group_index(group, rankmap[2 * j]);
230 smpi_group_set_mapping(group_out, index, j);
232 requests = xbt_new(MPI_Request, count);
234 for(j = 0; j < count; j++) {
235 if(rankmap[2 * j] != 0) {
236 requests[reqs] = smpi_isend_init(&group_out, 1, MPI_PTR, rankmap[2 * j], system_tag, comm);
240 smpi_mpi_startall(reqs, requests);
241 smpi_mpi_waitall(reqs, requests, MPI_STATUS_IGNORE);
245 group_out = group_root; /* exit with root's group */
247 if(color != MPI_UNDEFINED) {
248 smpi_mpi_recv(&group_out, 1, MPI_PTR, 0, system_tag, comm, MPI_STATUS_IGNORE);
250 group_out=smpi_group_copy(group_out);
252 } /* otherwise, exit with group_out == NULL */
254 return group_out ? smpi_comm_new(group_out, NULL) : MPI_COMM_NULL;
257 void smpi_comm_use(MPI_Comm comm){
258 if (comm == MPI_COMM_UNINITIALIZED)
259 comm = smpi_process_comm_world();
263 void smpi_comm_unuse(MPI_Comm comm){
264 if (comm == MPI_COMM_UNINITIALIZED)
265 comm = smpi_process_comm_world();
267 if(comm->refcount==0){
268 if(comm->intra_comm != MPI_COMM_NULL)
269 smpi_comm_unuse(comm->intra_comm);
270 if(comm->leaders_comm != MPI_COMM_NULL)
271 smpi_comm_unuse(comm->leaders_comm);
272 if(comm->non_uniform_map !=NULL)
273 xbt_free(comm->non_uniform_map);
274 if(comm->leaders_map !=NULL)
275 xbt_free(comm->leaders_map);
281 compare_ints (const void *a, const void *b)
283 const int *da = (const int *) a;
284 const int *db = (const int *) b;
286 return (*da > *db) - (*da < *db);
289 void smpi_comm_init_smp(MPI_Comm comm){
292 if (comm == MPI_COMM_UNINITIALIZED)
293 comm = smpi_process_comm_world();
295 int comm_size =smpi_comm_size(comm);
297 // If we are in replay - perform an ugly hack
298 // say to SimGrid that we are not in replay for a while, because we need
299 // the buffers to be copied for the following calls
300 int replaying = 0; //cache data to set it back again after
301 if(smpi_process_get_replaying()){
303 smpi_process_set_replaying(0);
306 if(smpi_privatize_global_variables){ //we need to switch here, as the called function may silently touch global variables
307 smpi_switch_data_segment(smpi_process_index());
309 //identify neighbours in comm
310 //get the indexes of all processes sharing the same simix host
311 xbt_swag_t process_list = simcall_host_get_process_list(SIMIX_host_self());
312 int intra_comm_size = 0;
313 //only one process/node, disable SMP support and return
314 // if(intra_comm_size==1){
315 // smpi_comm_set_intra_comm(comm, MPI_COMM_SELF);
316 // //smpi_comm_set_leaders_comm(comm, comm);
317 // smpi_process_set_comm_intra(MPI_COMM_SELF);
323 int min_index=INT_MAX;//the minimum index will be the leader
324 msg_process_t process = NULL;
325 xbt_swag_foreach(process, process_list) {
327 int index = SIMIX_process_get_PID(process) -1;
329 if(smpi_group_rank(smpi_comm_group(comm), index)!=MPI_UNDEFINED){
331 //the process is in the comm
332 if(index < min_index)
337 XBT_DEBUG("number of processes deployed on my node : %d", intra_comm_size);
338 MPI_Group group_intra = smpi_group_new(intra_comm_size);
341 xbt_swag_foreach(process, process_list) {
343 int index = SIMIX_process_get_PID(process) -1;
344 if(smpi_group_rank(smpi_comm_group(comm), index)!=MPI_UNDEFINED){
345 smpi_group_set_mapping(group_intra, index, i);
351 MPI_Comm comm_intra = smpi_comm_new(group_intra, NULL);
352 //MPI_Comm shmem_comm = smpi_process_comm_intra();
353 //int intra_rank = smpi_comm_rank(shmem_comm);
356 //if(smpi_process_index()==min_index)
359 int * leaders_map= (int*)xbt_malloc0(sizeof(int)*comm_size);
360 int * leader_list= (int*)xbt_malloc0(sizeof(int)*comm_size);
361 for(i=0; i<comm_size; i++){
365 smpi_coll_tuned_allgather_mpich(&leader, 1, MPI_INT , leaders_map, 1, MPI_INT, comm);
367 if(smpi_privatize_global_variables){ //we need to switch here, as the called function may silently touch global variables
368 smpi_switch_data_segment(smpi_process_index());
371 if(!comm->leaders_map){
372 comm->leaders_map= leaders_map;
374 xbt_free(leaders_map);
377 int leader_group_size = 0;
378 for(i=0; i<comm_size; i++){
380 for(j=0;j<leader_group_size; j++){
381 if(comm->leaders_map[i]==leader_list[j]){
386 leader_list[leader_group_size]=comm->leaders_map[i];
390 qsort(leader_list, leader_group_size, sizeof(int),compare_ints);
392 MPI_Group leaders_group = smpi_group_new(leader_group_size);
395 MPI_Comm leader_comm = MPI_COMM_NULL;
396 if(MPI_COMM_WORLD!=MPI_COMM_UNINITIALIZED && comm!=MPI_COMM_WORLD){
397 //create leader_communicator
398 for (i=0; i< leader_group_size;i++)
399 smpi_group_set_mapping(leaders_group, leader_list[i], i);
400 leader_comm = smpi_comm_new(leaders_group, NULL);
401 smpi_comm_set_leaders_comm(comm, leader_comm);
402 smpi_comm_set_intra_comm(comm, comm_intra);
404 //create intracommunicator
405 // smpi_comm_set_intra_comm(comm, smpi_comm_split(comm, *(int*)SIMIX_host_self(), comm_rank));
407 for (i=0; i< leader_group_size;i++)
408 smpi_group_set_mapping(leaders_group, leader_list[i], i);
410 leader_comm = smpi_comm_new(leaders_group, NULL);
411 if(smpi_comm_get_leaders_comm(comm)==MPI_COMM_NULL)
412 smpi_comm_set_leaders_comm(comm, leader_comm);
413 smpi_process_set_comm_intra(comm_intra);
418 // Are the nodes uniform ? = same number of process/node
419 int my_local_size=smpi_comm_size(comm_intra);
420 if(smpi_comm_rank(comm_intra)==0) {
421 int* non_uniform_map = xbt_malloc0(sizeof(int)*leader_group_size);
422 smpi_coll_tuned_allgather_mpich(&my_local_size, 1, MPI_INT,
423 non_uniform_map, 1, MPI_INT, leader_comm);
424 for(i=0; i < leader_group_size; i++) {
425 if(non_uniform_map[0] != non_uniform_map[i]) {
430 if(!is_uniform && smpi_comm_is_uniform(comm)){
431 comm->non_uniform_map= non_uniform_map;
433 xbt_free(non_uniform_map);
435 comm->is_uniform=is_uniform;
437 smpi_coll_tuned_bcast_mpich(&(comm->is_uniform),1, MPI_INT, 0, comm_intra );
439 if(smpi_privatize_global_variables){ //we need to switch here, as the called function may silently touch global variables
440 smpi_switch_data_segment(smpi_process_index());
442 // Are the ranks blocked ? = allocated contiguously on the SMP nodes
444 int prev=smpi_group_rank(smpi_comm_group(comm), smpi_group_index(smpi_comm_group(comm_intra), 0));
445 for (i=1; i<my_local_size; i++){
446 int this=smpi_group_rank(smpi_comm_group(comm),smpi_group_index(smpi_comm_group(comm_intra), i));
455 smpi_mpi_allreduce(&is_blocked, &(global_blocked), 1,
456 MPI_INT, MPI_LAND, comm);
458 if(MPI_COMM_WORLD==SMPI_UNINITIALIZED || comm==MPI_COMM_WORLD){
459 if(smpi_comm_rank(comm)==0){
460 comm->is_blocked=global_blocked;
463 comm->is_blocked=global_blocked;
465 xbt_free(leader_list);
468 smpi_process_set_replaying(1);