1 /* smpi_coll.c -- various optimized routing for collectives */
3 /* Copyright (c) 2009-2015. 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 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_coll, smpi,
18 "Logging specific to SMPI (coll)");
20 s_mpi_coll_description_t mpi_coll_gather_description[] = {
22 "gather default collective",
23 (void*)smpi_mpi_gather},
24 COLL_GATHERS(COLL_DESCRIPTION, COLL_COMMA),
25 {NULL, NULL, NULL} /* this array must be NULL terminated */
29 s_mpi_coll_description_t mpi_coll_allgather_description[] = {
31 "allgather default collective",
32 (void*)smpi_mpi_allgather},
33 COLL_ALLGATHERS(COLL_DESCRIPTION, COLL_COMMA),
34 {NULL, NULL, NULL} /* this array must be NULL terminated */
37 s_mpi_coll_description_t mpi_coll_allgatherv_description[] = {
39 "allgatherv default collective",
40 (void*)smpi_mpi_allgatherv},
41 COLL_ALLGATHERVS(COLL_DESCRIPTION, COLL_COMMA),
42 {NULL, NULL, NULL} /* this array must be NULL terminated */
45 s_mpi_coll_description_t mpi_coll_allreduce_description[] = {
47 "allreduce default collective",
48 (void*)smpi_mpi_allreduce},
49 COLL_ALLREDUCES(COLL_DESCRIPTION, COLL_COMMA),
50 {NULL, NULL, NULL} /* this array must be NULL terminated */
53 s_mpi_coll_description_t mpi_coll_reduce_scatter_description[] = {
55 "reduce_scatter default collective",
56 (void*)smpi_mpi_reduce_scatter},
57 COLL_REDUCE_SCATTERS(COLL_DESCRIPTION, COLL_COMMA),
58 {NULL, NULL, NULL} /* this array must be NULL terminated */
61 s_mpi_coll_description_t mpi_coll_scatter_description[] = {
63 "scatter default collective",
64 (void*)smpi_mpi_scatter},
65 COLL_SCATTERS(COLL_DESCRIPTION, COLL_COMMA),
66 {NULL, NULL, NULL} /* this array must be NULL terminated */
69 s_mpi_coll_description_t mpi_coll_barrier_description[] = {
71 "barrier default collective",
72 (void*)smpi_mpi_barrier},
73 COLL_BARRIERS(COLL_DESCRIPTION, COLL_COMMA),
74 {NULL, NULL, NULL} /* this array must be NULL terminated */
76 s_mpi_coll_description_t mpi_coll_alltoall_description[] = {
78 "Ompi alltoall default collective",
79 (void*)smpi_coll_tuned_alltoall_ompi2},
80 COLL_ALLTOALLS(COLL_DESCRIPTION, COLL_COMMA),
82 "Alltoall Bruck (SG) collective",
83 (void*)smpi_coll_tuned_alltoall_bruck},
85 "Alltoall basic linear (SG) collective",
86 (void*)smpi_coll_tuned_alltoall_basic_linear},
87 {NULL, NULL, NULL} /* this array must be NULL terminated */
90 s_mpi_coll_description_t mpi_coll_alltoallv_description[] = {
92 "Ompi alltoallv default collective",
93 (void*)smpi_coll_basic_alltoallv},
94 COLL_ALLTOALLVS(COLL_DESCRIPTION, COLL_COMMA),
95 {NULL, NULL, NULL} /* this array must be NULL terminated */
98 s_mpi_coll_description_t mpi_coll_bcast_description[] = {
100 "bcast default collective ",
101 (void*)smpi_mpi_bcast},
102 COLL_BCASTS(COLL_DESCRIPTION, COLL_COMMA),
103 {NULL, NULL, NULL} /* this array must be NULL terminated */
106 s_mpi_coll_description_t mpi_coll_reduce_description[] = {
108 "reduce default collective",
109 (void*)smpi_mpi_reduce},
110 COLL_REDUCES(COLL_DESCRIPTION, COLL_COMMA),
111 {NULL, NULL, NULL} /* this array must be NULL terminated */
116 /** Displays the long description of all registered models, and quit */
117 void coll_help(const char *category, s_mpi_coll_description_t * table)
120 printf("Long description of the %s models accepted by this simulator:\n",
122 for (i = 0; table[i].name; i++)
123 printf(" %s: %s\n", table[i].name, table[i].description);
126 int find_coll_description(s_mpi_coll_description_t * table,
127 char *name, const char *desc)
130 char *name_list = NULL;
132 if(name==NULL){//no argument provided, use active selector's algorithm
133 name=(char*)sg_cfg_get_string("smpi/coll_selector");
136 for (i = 0; table[i].name; i++)
137 if (!strcmp(name, table[i].name)) {
138 if (strcmp(table[i].name,"default"))
139 XBT_INFO("Switch to algorithm %s for collective %s",table[i].name,desc);
144 // collective seems not handled by the active selector, try with default one
145 name=(char*)"default";
146 for (i = 0; table[i].name; i++)
147 if (!strcmp(name, table[i].name)) {
152 xbt_die("No collective is valid for '%s'! This is a bug.",name);
153 name_list = xbt_strdup(table[0].name);
154 for (i = 1; table[i].name; i++) {
155 name_list = static_cast<char*>(xbt_realloc(name_list,
156 strlen(name_list) + strlen(table[i].name) + 3));
157 strcat(name_list, ", ");
158 strcat(name_list, table[i].name);
160 xbt_die("Collective '%s' is invalid! Valid collectives are: %s.", name, name_list);
164 int (*mpi_coll_gather_fun)(void *, int, MPI_Datatype, void*, int, MPI_Datatype, int root, MPI_Comm);
165 int (*mpi_coll_allgather_fun)(void *, int, MPI_Datatype, void*, int, MPI_Datatype, MPI_Comm);
166 int (*mpi_coll_allgatherv_fun)(void *, int, MPI_Datatype, void*, int*, int*, MPI_Datatype, MPI_Comm);
167 int (*mpi_coll_allreduce_fun)(void *sbuf, void *rbuf, int rcount, MPI_Datatype dtype, MPI_Op op, MPI_Comm comm);
168 int (*mpi_coll_alltoall_fun)(void *, int, MPI_Datatype, void*, int, MPI_Datatype, MPI_Comm);
169 int (*mpi_coll_alltoallv_fun)(void *, int*, int*, MPI_Datatype, void*, int*, int*, MPI_Datatype, MPI_Comm);
170 int (*mpi_coll_bcast_fun)(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm com);
171 int (*mpi_coll_reduce_fun)(void *buf, void *rbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm);
172 int (*mpi_coll_reduce_scatter_fun)(void *sbuf, void *rbuf, int *rcounts,MPI_Datatype dtype,MPI_Op op,MPI_Comm comm);
173 int (*mpi_coll_scatter_fun)(void *sendbuf, int sendcount, MPI_Datatype sendtype,void *recvbuf, int recvcount, MPI_Datatype recvtype,int root, MPI_Comm comm);
174 int (*mpi_coll_barrier_fun)(MPI_Comm comm);
177 int smpi_coll_tuned_alltoall_ompi2(void *sendbuf, int sendcount,
178 MPI_Datatype sendtype, void *recvbuf,
179 int recvcount, MPI_Datatype recvtype,
183 size = smpi_comm_size(comm);
184 sendsize = smpi_datatype_size(sendtype) * sendcount;
185 if (sendsize < 200 && size > 12) {
187 smpi_coll_tuned_alltoall_bruck(sendbuf, sendcount, sendtype,
188 recvbuf, recvcount, recvtype,
190 } else if (sendsize < 3000) {
192 smpi_coll_tuned_alltoall_basic_linear(sendbuf, sendcount,
194 recvcount, recvtype, comm);
197 smpi_coll_tuned_alltoall_ring(sendbuf, sendcount, sendtype,
198 recvbuf, recvcount, recvtype,
206 * Openmpi calls this routine when the message size sent to each rank < 2000 bytes and size < 12
207 * FIXME: uh, check smpi_pmpi again, but this routine is called for > 12, not
210 int smpi_coll_tuned_alltoall_bruck(void *sendbuf, int sendcount,
211 MPI_Datatype sendtype, void *recvbuf,
212 int recvcount, MPI_Datatype recvtype,
215 int system_tag = 777;
216 int i, rank, size, err, count;
218 MPI_Aint sendext = 0;
219 MPI_Aint recvext = 0;
220 MPI_Request *requests;
222 // FIXME: check implementation
223 rank = smpi_comm_rank(comm);
224 size = smpi_comm_size(comm);
225 XBT_DEBUG("<%d> algorithm alltoall_bruck() called.", rank);
226 smpi_datatype_extent(sendtype, &lb, &sendext);
227 smpi_datatype_extent(recvtype, &lb, &recvext);
228 /* Local copy from self */
230 smpi_datatype_copy((char *)sendbuf + rank * sendcount * sendext,
232 (char *)recvbuf + rank * recvcount * recvext,
233 recvcount, recvtype);
234 if (err == MPI_SUCCESS && size > 1) {
235 /* Initiate all send/recv to/from others. */
236 requests = xbt_new(MPI_Request, 2 * (size - 1));
238 /* Create all receives that will be posted first */
239 for (i = 0; i < size; ++i) {
241 XBT_DEBUG("<%d> skip request creation [src = %d, recvcount = %d]",
246 smpi_irecv_init((char *)recvbuf + i * recvcount * recvext, recvcount,
247 recvtype, i, system_tag, comm);
250 /* Now create all sends */
251 for (i = 0; i < size; ++i) {
253 XBT_DEBUG("<%d> skip request creation [dst = %d, sendcount = %d]",
258 smpi_isend_init((char *)sendbuf + i * sendcount * sendext, sendcount,
259 sendtype, i, system_tag, comm);
262 /* Wait for them all. */
263 smpi_mpi_startall(count, requests);
264 XBT_DEBUG("<%d> wait for %d requests", rank, count);
265 smpi_mpi_waitall(count, requests, MPI_STATUS_IGNORE);
266 for(i = 0; i < count; i++) {
267 if(requests[i]!=MPI_REQUEST_NULL) smpi_mpi_request_free(&requests[i]);
275 * Alltoall basic_linear (STARMPI:alltoall-simple)
277 int smpi_coll_tuned_alltoall_basic_linear(void *sendbuf, int sendcount,
278 MPI_Datatype sendtype,
279 void *recvbuf, int recvcount,
280 MPI_Datatype recvtype,
283 int system_tag = 888;
284 int i, rank, size, err, count;
285 MPI_Aint lb = 0, sendext = 0, recvext = 0;
286 MPI_Request *requests;
289 rank = smpi_comm_rank(comm);
290 size = smpi_comm_size(comm);
291 XBT_DEBUG("<%d> algorithm alltoall_basic_linear() called.", rank);
292 smpi_datatype_extent(sendtype, &lb, &sendext);
293 smpi_datatype_extent(recvtype, &lb, &recvext);
294 /* simple optimization */
295 err = smpi_datatype_copy((char *)sendbuf + rank * sendcount * sendext,
297 (char *)recvbuf + rank * recvcount * recvext,
298 recvcount, recvtype);
299 if (err == MPI_SUCCESS && size > 1) {
300 /* Initiate all send/recv to/from others. */
301 requests = xbt_new(MPI_Request, 2 * (size - 1));
302 /* Post all receives first -- a simple optimization */
304 for (i = (rank + 1) % size; i != rank; i = (i + 1) % size) {
306 smpi_irecv_init((char *)recvbuf + i * recvcount * recvext, recvcount,
307 recvtype, i, system_tag, comm);
310 /* Now post all sends in reverse order
311 * - We would like to minimize the search time through message queue
312 * when messages actually arrive in the order in which they were posted.
313 * TODO: check the previous assertion
315 for (i = (rank + size - 1) % size; i != rank; i = (i + size - 1) % size) {
317 smpi_isend_init((char *)sendbuf + i * sendcount * sendext, sendcount,
318 sendtype, i, system_tag, comm);
321 /* Wait for them all. */
322 smpi_mpi_startall(count, requests);
323 XBT_DEBUG("<%d> wait for %d requests", rank, count);
324 smpi_mpi_waitall(count, requests, MPI_STATUS_IGNORE);
325 for(i = 0; i < count; i++) {
326 if(requests[i]!=MPI_REQUEST_NULL) smpi_mpi_request_free(&requests[i]);
333 int smpi_coll_basic_alltoallv(void *sendbuf, int *sendcounts,
334 int *senddisps, MPI_Datatype sendtype,
335 void *recvbuf, int *recvcounts,
336 int *recvdisps, MPI_Datatype recvtype,
339 int system_tag = 889;
340 int i, rank, size, err, count;
341 MPI_Aint lb = 0, sendext = 0, recvext = 0;
342 MPI_Request *requests;
345 rank = smpi_comm_rank(comm);
346 size = smpi_comm_size(comm);
347 XBT_DEBUG("<%d> algorithm basic_alltoallv() called.", rank);
348 smpi_datatype_extent(sendtype, &lb, &sendext);
349 smpi_datatype_extent(recvtype, &lb, &recvext);
350 /* Local copy from self */
352 smpi_datatype_copy((char *)sendbuf + senddisps[rank] * sendext,
353 sendcounts[rank], sendtype,
354 (char *)recvbuf + recvdisps[rank] * recvext,
355 recvcounts[rank], recvtype);
356 if (err == MPI_SUCCESS && size > 1) {
357 /* Initiate all send/recv to/from others. */
358 requests = xbt_new(MPI_Request, 2 * (size - 1));
360 /* Create all receives that will be posted first */
361 for (i = 0; i < size; ++i) {
362 if (i == rank || recvcounts[i] == 0) {
364 ("<%d> skip request creation [src = %d, recvcounts[src] = %d]",
365 rank, i, recvcounts[i]);
369 smpi_irecv_init((char *)recvbuf + recvdisps[i] * recvext,
370 recvcounts[i], recvtype, i, system_tag, comm);
373 /* Now create all sends */
374 for (i = 0; i < size; ++i) {
375 if (i == rank || sendcounts[i] == 0) {
377 ("<%d> skip request creation [dst = %d, sendcounts[dst] = %d]",
378 rank, i, sendcounts[i]);
382 smpi_isend_init((char *)sendbuf + senddisps[i] * sendext,
383 sendcounts[i], sendtype, i, system_tag, comm);
386 /* Wait for them all. */
387 smpi_mpi_startall(count, requests);
388 XBT_DEBUG("<%d> wait for %d requests", rank, count);
389 smpi_mpi_waitall(count, requests, MPI_STATUS_IGNORE);
390 for(i = 0; i < count; i++) {
391 if(requests[i]!=MPI_REQUEST_NULL) smpi_mpi_request_free(&requests[i]);