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"
15 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_comm, smpi,
16 "Logging specific to SMPI (comm)");
20 /* Support for cartesian topology was added, but there are 2 other types of
21 * topology, graph et dist graph. In order to support them, we have to add a
22 * field MPIR_Topo_type, and replace the MPI_Topology field by an union. */
24 typedef struct s_smpi_mpi_communicator {
26 MPIR_Topo_type topoType;
27 MPI_Topology topo; // to be replaced by an union
29 MPI_Comm leaders_comm;//inter-node communicator
30 MPI_Comm intra_comm;//intra-node communicator . For MPI_COMM_WORLD this can't be used, as var is global.
31 //use an intracomm stored in the process data instead
32 int* leaders_map; //who is the leader of each process
34 int* non_uniform_map; //set if smp nodes have a different number of processes allocated
35 int is_blocked;// are ranks allocated on the same smp node contiguous ?
36 } s_smpi_mpi_communicator_t;
38 static int smpi_compare_rankmap(const void *a, const void *b)
40 const int* x = (const int*)a;
41 const int* y = (const int*)b;
58 MPI_Comm smpi_comm_new(MPI_Group group, MPI_Topology topo)
62 comm = xbt_new(s_smpi_mpi_communicator_t, 1);
64 smpi_group_use(comm->group);
68 comm->intra_comm = MPI_COMM_NULL;
69 comm->leaders_comm = MPI_COMM_NULL;
71 comm->non_uniform_map = NULL;
72 comm->leaders_map = NULL;
77 void smpi_comm_destroy(MPI_Comm comm)
79 if (comm == MPI_COMM_UNINITIALIZED)
80 comm = smpi_process_comm_world();
81 smpi_group_unuse(comm->group);
82 smpi_topo_destroy(comm->topo); // there's no use count on topos
83 smpi_comm_unuse(comm);
86 MPI_Group smpi_comm_group(MPI_Comm comm)
88 if (comm == MPI_COMM_UNINITIALIZED)
89 comm = smpi_process_comm_world();
94 MPI_Topology smpi_comm_topo(MPI_Comm comm) {
95 if (comm != MPI_COMM_NULL)
100 int smpi_comm_size(MPI_Comm comm)
102 if (comm == MPI_COMM_UNINITIALIZED)
103 comm = smpi_process_comm_world();
105 return smpi_group_size(smpi_comm_group(comm));
108 int smpi_comm_rank(MPI_Comm comm)
110 if (comm == MPI_COMM_UNINITIALIZED)
111 comm = smpi_process_comm_world();
112 return smpi_group_rank(smpi_comm_group(comm), smpi_process_index());
115 void smpi_comm_get_name (MPI_Comm comm, char* name, int* len)
117 if (comm == MPI_COMM_UNINITIALIZED)
118 comm = smpi_process_comm_world();
119 if(comm == MPI_COMM_WORLD) {
120 strcpy(name, "WORLD");
123 *len = snprintf(name, MPI_MAX_NAME_STRING, "%p", comm);
127 void smpi_comm_set_leaders_comm(MPI_Comm comm, MPI_Comm leaders){
128 comm->leaders_comm=leaders;
131 void smpi_comm_set_intra_comm(MPI_Comm comm, MPI_Comm leaders){
132 comm->intra_comm=leaders;
135 int* smpi_comm_get_non_uniform_map(MPI_Comm comm){
136 return comm->non_uniform_map;
139 int* smpi_comm_get_leaders_map(MPI_Comm comm){
140 return comm->leaders_map;
143 MPI_Comm smpi_comm_get_leaders_comm(MPI_Comm comm){
144 return comm->leaders_comm;
147 MPI_Comm smpi_comm_get_intra_comm(MPI_Comm comm){
148 if(comm==MPI_COMM_WORLD) return smpi_process_get_comm_intra();
149 else return comm->intra_comm;
152 int smpi_comm_is_uniform(MPI_Comm comm){
153 return comm->is_uniform;
156 int smpi_comm_is_blocked(MPI_Comm comm){
157 return comm->is_blocked;
160 MPI_Comm smpi_comm_split(MPI_Comm comm, int color, int key)
162 if (comm == MPI_COMM_UNINITIALIZED)
163 comm = smpi_process_comm_world();
164 int system_tag = 123;
165 int index, rank, size, i, j, count, reqs;
169 MPI_Group group, group_root, group_out;
170 MPI_Request* requests;
172 group_root = group_out = NULL;
173 group = smpi_comm_group(comm);
174 rank = smpi_comm_rank(comm);
175 size = smpi_comm_size(comm);
176 /* Gather all colors and keys on rank 0 */
177 sendbuf = xbt_new(int, 2);
181 recvbuf = xbt_new(int, 2 * size);
185 smpi_mpi_gather(sendbuf, 2, MPI_INT, recvbuf, 2, MPI_INT, 0, comm);
187 /* Do the actual job */
189 rankmap = xbt_new(int, 2 * size);
190 for(i = 0; i < size; i++) {
191 if(recvbuf[2 * i] == MPI_UNDEFINED) {
195 for(j = i + 1; j < size; j++) {
196 if(recvbuf[2 * i] == recvbuf[2 * j]) {
197 recvbuf[2 * j] = MPI_UNDEFINED;
198 rankmap[2 * count] = j;
199 rankmap[2 * count + 1] = recvbuf[2 * j + 1];
203 /* Add self in the group */
204 recvbuf[2 * i] = MPI_UNDEFINED;
205 rankmap[2 * count] = i;
206 rankmap[2 * count + 1] = recvbuf[2 * i + 1];
208 qsort(rankmap, count, 2 * sizeof(int), &smpi_compare_rankmap);
209 group_out = smpi_group_new(count);
211 group_root = group_out; /* Save root's group */
213 for(j = 0; j < count; j++) {
214 //increment refcounter in order to avoid freeing the group too quick before copy
215 index = smpi_group_index(group, rankmap[2 * j]);
216 smpi_group_set_mapping(group_out, index, j);
218 requests = xbt_new(MPI_Request, count);
220 for(j = 0; j < count; j++) {
221 if(rankmap[2 * j] != 0) {
222 requests[reqs] = smpi_isend_init(&group_out, 1, MPI_PTR, rankmap[2 * j], system_tag, comm);
226 smpi_mpi_startall(reqs, requests);
227 smpi_mpi_waitall(reqs, requests, MPI_STATUS_IGNORE);
231 group_out = group_root; /* exit with root's group */
233 if(color != MPI_UNDEFINED) {
234 smpi_mpi_recv(&group_out, 1, MPI_PTR, 0, system_tag, comm, MPI_STATUS_IGNORE);
236 group_out=smpi_group_copy(group_out);
238 } /* otherwise, exit with group_out == NULL */
240 return group_out ? smpi_comm_new(group_out, NULL) : MPI_COMM_NULL;
243 void smpi_comm_use(MPI_Comm comm){
244 if (comm == MPI_COMM_UNINITIALIZED)
245 comm = smpi_process_comm_world();
249 void smpi_comm_unuse(MPI_Comm comm){
250 if (comm == MPI_COMM_UNINITIALIZED)
251 comm = smpi_process_comm_world();
253 if(comm->refcount==0){
254 if(comm->intra_comm != MPI_COMM_NULL)
255 smpi_comm_unuse(comm->intra_comm);
256 if(comm->leaders_comm != MPI_COMM_NULL)
257 smpi_comm_unuse(comm->leaders_comm);
258 if(comm->non_uniform_map !=NULL)
259 xbt_free(comm->non_uniform_map);
260 if(comm->leaders_map !=NULL)
261 xbt_free(comm->leaders_map);
267 compare_ints (const void *a, const void *b)
269 const int *da = (const int *) a;
270 const int *db = (const int *) b;
272 return (*da > *db) - (*da < *db);
275 void smpi_comm_init_smp(MPI_Comm comm){
277 int comm_size =smpi_comm_size(comm);
279 if(smpi_privatize_global_variables){ //we need to switch here, as the called function may silently touch global variables
280 switch_data_segment(smpi_process_index());
282 //identify neighbours in comm
283 //get the indexes of all processes sharing the same simix host
284 xbt_swag_t process_list = simcall_host_get_process_list(SIMIX_host_self());
285 int intra_comm_size = 0;
286 //only one process/node, disable SMP support and return
287 // if(intra_comm_size==1){
288 // smpi_comm_set_intra_comm(comm, MPI_COMM_SELF);
289 // //smpi_comm_set_leaders_comm(comm, comm);
290 // smpi_process_set_comm_intra(MPI_COMM_SELF);
293 XBT_DEBUG("number of processes deployed on my node : %d", intra_comm_size);
296 int min_index=INT_MAX;//the minimum index will be the leader
297 msg_process_t process = NULL;
298 xbt_swag_foreach(process, process_list) {
300 int index = SIMIX_process_get_PID(process) -1;
302 if(smpi_group_rank(smpi_comm_group(comm), index)!=MPI_UNDEFINED){
304 //the process is in the comm
305 if(index < min_index)
311 MPI_Group group_intra = smpi_group_new(intra_comm_size);
314 xbt_swag_foreach(process, process_list) {
316 int index = SIMIX_process_get_PID(process) -1;
317 if(smpi_group_rank(smpi_comm_group(comm), index)!=MPI_UNDEFINED){
318 smpi_group_set_mapping(group_intra, index, i);
324 MPI_Comm comm_intra = smpi_comm_new(group_intra, NULL);
325 //MPI_Comm shmem_comm = smpi_process_comm_intra();
326 //int intra_rank = smpi_comm_rank(shmem_comm);
329 //if(smpi_process_index()==min_index)
332 int * leaders_map= (int*)xbt_malloc0(sizeof(int)*comm_size);
333 int * leader_list= (int*)xbt_malloc0(sizeof(int)*comm_size);
334 for(i=0; i<comm_size; i++){
338 smpi_coll_tuned_allgather_mpich(&leader, 1, MPI_INT , leaders_map, 1, MPI_INT, comm);
340 if(!comm->leaders_map){
341 comm->leaders_map= leaders_map;
343 xbt_free(leaders_map);
346 int leader_group_size = 0;
347 for(i=0; i<comm_size; i++){
349 for(j=0;j<leader_group_size; j++){
350 if(comm->leaders_map[i]==leader_list[j]){
355 leader_list[leader_group_size]=comm->leaders_map[i];
359 qsort(leader_list, leader_group_size, sizeof(int),compare_ints);
361 MPI_Group leaders_group = smpi_group_new(leader_group_size);
364 MPI_Comm leader_comm = MPI_COMM_NULL;
365 if(comm!=MPI_COMM_WORLD){
366 //create leader_communicator
367 for (i=0; i< leader_group_size;i++)
368 smpi_group_set_mapping(leaders_group, leader_list[i], i);
369 leader_comm = smpi_comm_new(leaders_group, NULL);
370 smpi_comm_set_leaders_comm(comm, leader_comm);
371 smpi_comm_set_intra_comm(comm, comm_intra);
373 //create intracommunicator
374 // smpi_comm_set_intra_comm(comm, smpi_comm_split(comm, *(int*)SIMIX_host_self(), comm_rank));
376 for (i=0; i< leader_group_size;i++)
377 smpi_group_set_mapping(leaders_group, leader_list[i], i);
379 leader_comm = smpi_comm_new(leaders_group, NULL);
380 if(smpi_comm_get_leaders_comm(comm)==MPI_COMM_NULL)
381 smpi_comm_set_leaders_comm(comm, leader_comm);
382 smpi_process_set_comm_intra(comm_intra);
387 // Are the nodes uniform ? = same number of process/node
388 int my_local_size=smpi_comm_size(comm_intra);
389 if(smpi_comm_rank(comm_intra)==0) {
390 int* non_uniform_map = xbt_malloc(sizeof(int)*leader_group_size);
391 smpi_coll_tuned_allgather_mpich(&my_local_size, 1, MPI_INT,
392 non_uniform_map, 1, MPI_INT, leader_comm);
393 for(i=0; i < leader_group_size; i++) {
394 if(non_uniform_map[0] != non_uniform_map[i]) {
399 if(!is_uniform && smpi_comm_is_uniform(comm)){
400 comm->non_uniform_map= non_uniform_map;
402 xbt_free(non_uniform_map);
404 comm->is_uniform=is_uniform;
406 smpi_coll_tuned_bcast_mpich(&(comm->is_uniform),1, MPI_INT, 0, comm_intra );
409 // Are the ranks blocked ? = allocated contiguously on the SMP nodes
411 int prev=smpi_group_rank(smpi_comm_group(comm), smpi_group_index(smpi_comm_group(comm_intra), 0));
412 for (i=1; i<my_local_size; i++){
413 int this=smpi_group_rank(smpi_comm_group(comm),smpi_group_index(smpi_comm_group(comm_intra), i));
422 smpi_mpi_allreduce(&is_blocked, &(global_blocked), 1,
423 MPI_INT, MPI_LAND, comm);
425 if(comm==MPI_COMM_WORLD){
426 if(smpi_comm_rank(comm)==0){
427 comm->is_blocked=global_blocked;
430 comm->is_blocked=global_blocked;
432 xbt_free(leader_list);