1 /* smpi_coll.c -- various optimized routing for collectives */
3 /* Copyright (c) 2009-2013. The SimGrid Team.
4 * All rights reserved. */
6 /* This program is free software; you can redistribute it and/or modify it
7 * under the terms of the license (GNU LGPL) which comes with this package. */
14 #include "colls/colls.h"
15 #include "simgrid/sg_config.h"
17 s_mpi_coll_description_t mpi_coll_gather_description[] = {
19 "gather default collective",
21 COLL_GATHERS(COLL_DESCRIPTION, COLL_COMMA),
22 {NULL, NULL, NULL} /* this array must be NULL terminated */
26 s_mpi_coll_description_t mpi_coll_allgather_description[] = {
28 "allgather default collective",
30 COLL_ALLGATHERS(COLL_DESCRIPTION, COLL_COMMA),
31 {NULL, NULL, NULL} /* this array must be NULL terminated */
34 s_mpi_coll_description_t mpi_coll_allgatherv_description[] = {
36 "allgatherv default collective",
38 COLL_ALLGATHERVS(COLL_DESCRIPTION, COLL_COMMA),
39 {NULL, NULL, NULL} /* this array must be NULL terminated */
42 s_mpi_coll_description_t mpi_coll_allreduce_description[] = {
44 "allreduce default collective",
46 COLL_ALLREDUCES(COLL_DESCRIPTION, COLL_COMMA),
47 {NULL, NULL, NULL} /* this array must be NULL terminated */
50 s_mpi_coll_description_t mpi_coll_reduce_scatter_description[] = {
52 "reduce_scatter default collective",
53 smpi_mpi_reduce_scatter},
54 COLL_REDUCE_SCATTERS(COLL_DESCRIPTION, COLL_COMMA),
55 {NULL, NULL, NULL} /* this array must be NULL terminated */
58 s_mpi_coll_description_t mpi_coll_scatter_description[] = {
60 "scatter default collective",
62 COLL_SCATTERS(COLL_DESCRIPTION, COLL_COMMA),
63 {NULL, NULL, NULL} /* this array must be NULL terminated */
66 s_mpi_coll_description_t mpi_coll_barrier_description[] = {
68 "barrier default collective",
70 COLL_BARRIERS(COLL_DESCRIPTION, COLL_COMMA),
71 {NULL, NULL, NULL} /* this array must be NULL terminated */
73 s_mpi_coll_description_t mpi_coll_alltoall_description[] = {
75 "Ompi alltoall default collective",
76 smpi_coll_tuned_alltoall_ompi2},
77 COLL_ALLTOALLS(COLL_DESCRIPTION, COLL_COMMA),
79 "Alltoall Bruck (SG) collective",
80 smpi_coll_tuned_alltoall_bruck},
82 "Alltoall basic linear (SG) collective",
83 smpi_coll_tuned_alltoall_basic_linear},
84 {NULL, NULL, NULL} /* this array must be NULL terminated */
87 s_mpi_coll_description_t mpi_coll_alltoallv_description[] = {
89 "Ompi alltoallv default collective",
90 smpi_coll_basic_alltoallv},
91 COLL_ALLTOALLVS(COLL_DESCRIPTION, COLL_COMMA),
92 {NULL, NULL, NULL} /* this array must be NULL terminated */
95 s_mpi_coll_description_t mpi_coll_bcast_description[] = {
97 "bcast default collective",
99 COLL_BCASTS(COLL_DESCRIPTION, COLL_COMMA),
100 {NULL, NULL, NULL} /* this array must be NULL terminated */
103 s_mpi_coll_description_t mpi_coll_reduce_description[] = {
105 "reduce default collective",
107 COLL_REDUCES(COLL_DESCRIPTION, COLL_COMMA),
108 {NULL, NULL, NULL} /* this array must be NULL terminated */
113 /** Displays the long description of all registered models, and quit */
114 void coll_help(const char *category, s_mpi_coll_description_t * table)
117 printf("Long description of the %s models accepted by this simulator:\n",
119 for (i = 0; table[i].name; i++)
120 printf(" %s: %s\n", table[i].name, table[i].description);
123 int find_coll_description(s_mpi_coll_description_t * table,
127 char *name_list = NULL;
129 if(name==NULL){//no argument provided, use active selector's algorithm
130 name=(char*)sg_cfg_get_string("smpi/coll_selector");
133 for (i = 0; table[i].name; i++)
134 if (!strcmp(name, table[i].name)) {
139 // collective seems not handled by the active selector, try with default one
140 name=(char*)"default";
141 for (i = 0; table[i].name; i++)
142 if (!strcmp(name, table[i].name)) {
146 name_list = strdup(table[0].name);
147 for (i = 1; table[i].name; i++) {
149 xbt_realloc(name_list,
150 strlen(name_list) + strlen(table[i].name) + 3);
151 strcat(name_list, ", ");
152 strcat(name_list, table[i].name);
154 xbt_die("Model '%s' is invalid! Valid models are: %s.", name, name_list);
158 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_coll, smpi,
159 "Logging specific to SMPI (coll)");
161 int (*mpi_coll_gather_fun)(void *, int, MPI_Datatype, void*, int, MPI_Datatype, int root, MPI_Comm);
162 int (*mpi_coll_allgather_fun)(void *, int, MPI_Datatype, void*, int, MPI_Datatype, MPI_Comm);
163 int (*mpi_coll_allgatherv_fun)(void *, int, MPI_Datatype, void*, int*, int*, MPI_Datatype, MPI_Comm);
164 int (*mpi_coll_allreduce_fun)(void *sbuf, void *rbuf, int rcount, MPI_Datatype dtype, MPI_Op op, MPI_Comm comm);
165 int (*mpi_coll_alltoall_fun)(void *, int, MPI_Datatype, void*, int, MPI_Datatype, MPI_Comm);
166 int (*mpi_coll_alltoallv_fun)(void *, int*, int*, MPI_Datatype, void*, int*, int*, MPI_Datatype, MPI_Comm);
167 int (*mpi_coll_bcast_fun)(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm com);
168 int (*mpi_coll_reduce_fun)(void *buf, void *rbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm);
169 int (*mpi_coll_reduce_scatter_fun)(void *sbuf, void *rbuf, int *rcounts,MPI_Datatype dtype,MPI_Op op,MPI_Comm comm);
170 int (*mpi_coll_scatter_fun)(void *sendbuf, int sendcount, MPI_Datatype sendtype,void *recvbuf, int recvcount, MPI_Datatype recvtype,int root, MPI_Comm comm);
171 int (*mpi_coll_barrier_fun)(MPI_Comm comm);
181 typedef struct s_proc_tree *proc_tree_t;
186 static proc_tree_t alloc_tree(int arity)
191 tree = xbt_new(struct s_proc_tree, 1);
192 tree->PROCTREE_A = arity;
194 tree->numChildren = 0;
195 tree->child = xbt_new(int, arity);
196 for (i = 0; i < arity; i++) {
207 static void free_tree(proc_tree_t tree)
209 xbt_free(tree->child);
214 * Build the tree depending on a process rank (index) and the group size (extent)
215 * @param root the rank of the tree root
216 * @param rank the rank of the calling process
217 * @param size the total number of processes
219 static void build_tree(int root, int rank, int size, proc_tree_t * tree)
221 int index = (rank - root + size) % size;
222 int firstChildIdx = index * (*tree)->PROCTREE_A + 1;
226 (*tree)->root = root;
228 for (i = 0; i < (*tree)->PROCTREE_A && firstChildIdx + i < size; i++) {
229 (*tree)->child[i] = (firstChildIdx + i + root) % size;
230 (*tree)->numChildren++;
236 (*tree)->parent = (((index - 1) / (*tree)->PROCTREE_A) + root) % size;
243 static void tree_bcast(void *buf, int count, MPI_Datatype datatype,
244 MPI_Comm comm, proc_tree_t tree)
246 int system_tag = COLL_TAG_BCAST;
248 MPI_Request *requests;
250 rank = smpi_comm_rank(comm);
251 /* wait for data from my parent in the tree */
253 XBT_DEBUG("<%d> tree_bcast(): i am not root: recv from %d, tag=%d)",
254 rank, tree->parent, system_tag + rank);
255 smpi_mpi_recv(buf, count, datatype, tree->parent, system_tag + rank,
256 comm, MPI_STATUS_IGNORE);
258 requests = xbt_new(MPI_Request, tree->numChildren);
259 XBT_DEBUG("<%d> creates %d requests (1 per child)", rank,
261 /* iniates sends to ranks lower in the tree */
262 for (i = 0; i < tree->numChildren; i++) {
263 if (tree->child[i] == -1) {
264 requests[i] = MPI_REQUEST_NULL;
266 XBT_DEBUG("<%d> send to <%d>, tag=%d", rank, tree->child[i],
267 system_tag + tree->child[i]);
269 smpi_isend_init(buf, count, datatype, tree->child[i],
270 system_tag + tree->child[i], comm);
273 smpi_mpi_startall(tree->numChildren, requests);
274 smpi_mpi_waitall(tree->numChildren, requests, MPI_STATUS_IGNORE);
281 static void tree_antibcast(void *buf, int count, MPI_Datatype datatype,
282 MPI_Comm comm, proc_tree_t tree)
284 int system_tag = COLL_TAG_BCAST;
286 MPI_Request *requests;
288 rank = smpi_comm_rank(comm);
289 // everyone sends to its parent, except root.
291 XBT_DEBUG("<%d> tree_antibcast(): i am not root: send to %d, tag=%d)",
292 rank, tree->parent, system_tag + rank);
293 smpi_mpi_send(buf, count, datatype, tree->parent, system_tag + rank,
296 //every one receives as many messages as it has children
297 requests = xbt_new(MPI_Request, tree->numChildren);
298 XBT_DEBUG("<%d> creates %d requests (1 per child)", rank,
300 for (i = 0; i < tree->numChildren; i++) {
301 if (tree->child[i] == -1) {
302 requests[i] = MPI_REQUEST_NULL;
304 XBT_DEBUG("<%d> recv from <%d>, tag=%d", rank, tree->child[i],
305 system_tag + tree->child[i]);
307 smpi_irecv_init(buf, count, datatype, tree->child[i],
308 system_tag + tree->child[i], comm);
311 smpi_mpi_startall(tree->numChildren, requests);
312 smpi_mpi_waitall(tree->numChildren, requests, MPI_STATUS_IGNORE);
317 * bcast with a binary, ternary, or whatever tree ..
319 void nary_tree_bcast(void *buf, int count, MPI_Datatype datatype, int root,
320 MPI_Comm comm, int arity)
322 proc_tree_t tree = alloc_tree(arity);
325 rank = smpi_comm_rank(comm);
326 size = smpi_comm_size(comm);
327 build_tree(root, rank, size, &tree);
328 tree_bcast(buf, count, datatype, comm, tree);
333 * barrier with a binary, ternary, or whatever tree ..
335 void nary_tree_barrier(MPI_Comm comm, int arity)
337 proc_tree_t tree = alloc_tree(arity);
341 rank = smpi_comm_rank(comm);
342 size = smpi_comm_size(comm);
343 build_tree(0, rank, size, &tree);
344 tree_antibcast(&dummy, 1, MPI_CHAR, comm, tree);
345 tree_bcast(&dummy, 1, MPI_CHAR, comm, tree);
349 int smpi_coll_tuned_alltoall_ompi2(void *sendbuf, int sendcount,
350 MPI_Datatype sendtype, void *recvbuf,
351 int recvcount, MPI_Datatype recvtype,
355 size = smpi_comm_size(comm);
356 sendsize = smpi_datatype_size(sendtype) * sendcount;
357 if (sendsize < 200 && size > 12) {
359 smpi_coll_tuned_alltoall_bruck(sendbuf, sendcount, sendtype,
360 recvbuf, recvcount, recvtype,
362 } else if (sendsize < 3000) {
364 smpi_coll_tuned_alltoall_basic_linear(sendbuf, sendcount,
366 recvcount, recvtype, comm);
369 smpi_coll_tuned_alltoall_ring(sendbuf, sendcount, sendtype,
370 recvbuf, recvcount, recvtype,
378 * Openmpi calls this routine when the message size sent to each rank < 2000 bytes and size < 12
379 * FIXME: uh, check smpi_pmpi again, but this routine is called for > 12, not
382 int smpi_coll_tuned_alltoall_bruck(void *sendbuf, int sendcount,
383 MPI_Datatype sendtype, void *recvbuf,
384 int recvcount, MPI_Datatype recvtype,
387 int system_tag = 777;
388 int i, rank, size, err, count;
390 MPI_Aint sendext = 0;
391 MPI_Aint recvext = 0;
392 MPI_Request *requests;
394 // FIXME: check implementation
395 rank = smpi_comm_rank(comm);
396 size = smpi_comm_size(comm);
397 XBT_DEBUG("<%d> algorithm alltoall_bruck() called.", rank);
398 err = smpi_datatype_extent(sendtype, &lb, &sendext);
399 err = smpi_datatype_extent(recvtype, &lb, &recvext);
400 /* Local copy from self */
402 smpi_datatype_copy((char *)sendbuf + rank * sendcount * sendext,
404 (char *)recvbuf + rank * recvcount * recvext,
405 recvcount, recvtype);
406 if (err == MPI_SUCCESS && size > 1) {
407 /* Initiate all send/recv to/from others. */
408 requests = xbt_new(MPI_Request, 2 * (size - 1));
410 /* Create all receives that will be posted first */
411 for (i = 0; i < size; ++i) {
413 XBT_DEBUG("<%d> skip request creation [src = %d, recvcount = %d]",
418 smpi_irecv_init((char *)recvbuf + i * recvcount * recvext, recvcount,
419 recvtype, i, system_tag, comm);
422 /* Now create all sends */
423 for (i = 0; i < size; ++i) {
425 XBT_DEBUG("<%d> skip request creation [dst = %d, sendcount = %d]",
430 smpi_isend_init((char *)sendbuf + i * sendcount * sendext, sendcount,
431 sendtype, i, system_tag, comm);
434 /* Wait for them all. */
435 smpi_mpi_startall(count, requests);
436 XBT_DEBUG("<%d> wait for %d requests", rank, count);
437 smpi_mpi_waitall(count, requests, MPI_STATUS_IGNORE);
444 * Alltoall basic_linear (STARMPI:alltoall-simple)
446 int smpi_coll_tuned_alltoall_basic_linear(void *sendbuf, int sendcount,
447 MPI_Datatype sendtype,
448 void *recvbuf, int recvcount,
449 MPI_Datatype recvtype,
452 int system_tag = 888;
453 int i, rank, size, err, count;
454 MPI_Aint lb = 0, sendext = 0, recvext = 0;
455 MPI_Request *requests;
458 rank = smpi_comm_rank(comm);
459 size = smpi_comm_size(comm);
460 XBT_DEBUG("<%d> algorithm alltoall_basic_linear() called.", rank);
461 err = smpi_datatype_extent(sendtype, &lb, &sendext);
462 err = smpi_datatype_extent(recvtype, &lb, &recvext);
463 /* simple optimization */
464 err = smpi_datatype_copy((char *)sendbuf + rank * sendcount * sendext,
466 (char *)recvbuf + rank * recvcount * recvext,
467 recvcount, recvtype);
468 if (err == MPI_SUCCESS && size > 1) {
469 /* Initiate all send/recv to/from others. */
470 requests = xbt_new(MPI_Request, 2 * (size - 1));
471 /* Post all receives first -- a simple optimization */
473 for (i = (rank + 1) % size; i != rank; i = (i + 1) % size) {
475 smpi_irecv_init((char *)recvbuf + i * recvcount * recvext, recvcount,
476 recvtype, i, system_tag, comm);
479 /* Now post all sends in reverse order
480 * - We would like to minimize the search time through message queue
481 * when messages actually arrive in the order in which they were posted.
482 * TODO: check the previous assertion
484 for (i = (rank + size - 1) % size; i != rank; i = (i + size - 1) % size) {
486 smpi_isend_init((char *)sendbuf + i * sendcount * sendext, sendcount,
487 sendtype, i, system_tag, comm);
490 /* Wait for them all. */
491 smpi_mpi_startall(count, requests);
492 XBT_DEBUG("<%d> wait for %d requests", rank, count);
493 smpi_mpi_waitall(count, requests, MPI_STATUS_IGNORE);
499 int smpi_coll_basic_alltoallv(void *sendbuf, int *sendcounts,
500 int *senddisps, MPI_Datatype sendtype,
501 void *recvbuf, int *recvcounts,
502 int *recvdisps, MPI_Datatype recvtype,
505 int system_tag = 889;
506 int i, rank, size, err, count;
507 MPI_Aint lb = 0, sendext = 0, recvext = 0;
508 MPI_Request *requests;
511 rank = smpi_comm_rank(comm);
512 size = smpi_comm_size(comm);
513 XBT_DEBUG("<%d> algorithm basic_alltoallv() called.", rank);
514 err = smpi_datatype_extent(sendtype, &lb, &sendext);
515 err = smpi_datatype_extent(recvtype, &lb, &recvext);
516 /* Local copy from self */
518 smpi_datatype_copy((char *)sendbuf + senddisps[rank] * sendext,
519 sendcounts[rank], sendtype,
520 (char *)recvbuf + recvdisps[rank] * recvext,
521 recvcounts[rank], recvtype);
522 if (err == MPI_SUCCESS && size > 1) {
523 /* Initiate all send/recv to/from others. */
524 requests = xbt_new(MPI_Request, 2 * (size - 1));
526 /* Create all receives that will be posted first */
527 for (i = 0; i < size; ++i) {
528 if (i == rank || recvcounts[i] == 0) {
530 ("<%d> skip request creation [src = %d, recvcounts[src] = %d]",
531 rank, i, recvcounts[i]);
535 smpi_irecv_init((char *)recvbuf + recvdisps[i] * recvext,
536 recvcounts[i], recvtype, i, system_tag, comm);
539 /* Now create all sends */
540 for (i = 0; i < size; ++i) {
541 if (i == rank || sendcounts[i] == 0) {
543 ("<%d> skip request creation [dst = %d, sendcounts[dst] = %d]",
544 rank, i, sendcounts[i]);
548 smpi_isend_init((char *)sendbuf + senddisps[i] * sendext,
549 sendcounts[i], sendtype, i, system_tag, comm);
552 /* Wait for them all. */
553 smpi_mpi_startall(count, requests);
554 XBT_DEBUG("<%d> wait for %d requests", rank, count);
555 smpi_mpi_waitall(count, requests, MPI_STATUS_IGNORE);