1 /* Copyright (c) 2007-2014. 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. */
8 #include "smpi_mpi_dt_private.h"
10 #include "xbt/replay.h"
11 #include "surf/surf.h"
12 #include "simix/smx_private.h"
13 #include "simgrid/sg_config.h"
15 #include <float.h> /* DBL_MAX */
20 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_kernel, smpi,
21 "Logging specific to SMPI (kernel)");
23 typedef struct s_smpi_process_data {
28 smx_rdv_t mailbox_small;
31 void *data; /* user data */
34 int sampling; /* inside an SMPI_SAMPLE_ block? */
35 } s_smpi_process_data_t;
37 static smpi_process_data_t *process_data = NULL;
38 static int process_count = 0;
40 MPI_Comm MPI_COMM_WORLD = MPI_COMM_NULL;
41 int MPI_UNIVERSE_SIZE;
43 MPI_Errhandler *MPI_ERRORS_RETURN = NULL;
44 MPI_Errhandler *MPI_ERRORS_ARE_FATAL = NULL;
45 MPI_Errhandler *MPI_ERRHANDLER_NULL = NULL;
47 #define MAILBOX_NAME_MAXLEN (5 + sizeof(int) * 2 + 1)
49 static char *get_mailbox_name(char *str, int index)
51 snprintf(str, MAILBOX_NAME_MAXLEN, "SMPI-%0*x", (int) (sizeof(int) * 2),
56 static char *get_mailbox_name_small(char *str, int index)
58 snprintf(str, MAILBOX_NAME_MAXLEN, "small%0*x", (int) (sizeof(int) * 2),
63 void smpi_process_init(int *argc, char ***argv)
66 smpi_process_data_t data;
70 proc = SIMIX_process_self();
71 index = atoi((*argv)[1]);
73 smpi_current_rank = index;
75 data = smpi_process_remote_data(index);
76 simcall_process_set_data(proc, data);
79 memmove(&(*argv)[1], &(*argv)[2], sizeof(char *) * (*argc - 2));
80 (*argv)[(*argc) - 1] = NULL;
85 // set the process attached to the mailbox
86 simcall_rdv_set_receiver(data->mailbox_small, proc);
87 XBT_DEBUG("<%d> New process in the game: %p", index, proc);
89 if (smpi_process_data() == NULL)
90 xbt_die("smpi_process_data() returned NULL. You probably gave a NULL parameter to MPI_Init. Although it's required by MPI-2, this is currently not supported by SMPI.");
93 void smpi_process_destroy(void)
95 int index = smpi_process_index();
96 process_data[index]->state = SMPI_FINALIZED;
97 XBT_DEBUG("<%d> Process left the game", index);
101 * @brief Prepares the current process for termination.
103 void smpi_process_finalize(void)
106 int size = smpi_comm_size(MPI_COMM_WORLD);
107 int rank = smpi_comm_rank(MPI_COMM_WORLD);
108 /* All non-root send & receive zero-length message. */
110 smpi_mpi_ssend (NULL, 0, MPI_BYTE, 0,
113 smpi_mpi_recv (NULL, 0, MPI_BYTE, 0,
115 MPI_COMM_WORLD, MPI_STATUS_IGNORE);
117 /* The root collects and broadcasts the messages. */
119 MPI_Request* requests;
120 requests = (MPI_Request*)malloc( size * sizeof(MPI_Request) );
121 for (i = 1; i < size; ++i) {
122 requests[i] = smpi_mpi_irecv(NULL, 0, MPI_BYTE, MPI_ANY_SOURCE,
123 COLL_TAG_BARRIER, MPI_COMM_WORLD
126 smpi_mpi_waitall( size-1, requests+1, MPI_STATUSES_IGNORE );
127 for (i = 1; i < size; ++i) {
128 requests[i] = smpi_mpi_issend(NULL, 0, MPI_BYTE, i,
133 smpi_mpi_waitall( size-1, requests+1, MPI_STATUSES_IGNORE );
140 * @brief Check if a process is finalized
142 int smpi_process_finalized()
144 int index = smpi_process_index();
145 if (index != MPI_UNDEFINED)
146 return (process_data[index]->state == SMPI_FINALIZED);
152 * @brief Check if a process is initialized
154 int smpi_process_initialized(void)
156 int index = smpi_process_index();
157 return ( (index != MPI_UNDEFINED)
158 && (process_data[index]->state == SMPI_INITIALIZED));
162 * @brief Mark a process as initialized (=MPI_Init called)
164 void smpi_process_mark_as_initialized(void)
166 int index = smpi_process_index();
167 if ((index != MPI_UNDEFINED) && (!process_data[index]->state != SMPI_FINALIZED))
168 process_data[index]->state = SMPI_INITIALIZED;
173 int smpi_process_argc(void)
175 smpi_process_data_t data = smpi_process_data();
176 return data->argc ? *(data->argc) - 1 : 0;
179 int smpi_process_getarg(integer * index, char *dst, ftnlen len)
181 smpi_process_data_t data = smpi_process_data();
185 if (!data->argc || !data->argv || *index < 1 || *index >= *(data->argc)) {
188 arg = (*data->argv)[*index];
189 for (i = 0; i < len && arg[i] != '\0'; i++) {
192 for (; i < len; i++) {
199 int smpi_global_size(void)
201 char *value = getenv("SMPI_GLOBAL_SIZE");
205 "Please set env var SMPI_GLOBAL_SIZE to expected number of processes.\n");
211 smpi_process_data_t smpi_process_data(void)
213 return SIMIX_process_self_get_data(SIMIX_process_self());
216 smpi_process_data_t smpi_process_remote_data(int index)
218 return process_data[index];
221 void smpi_process_set_user_data(void *data)
223 smpi_process_data_t process_data = smpi_process_data();
224 process_data->data = data;
227 void *smpi_process_get_user_data()
229 smpi_process_data_t process_data = smpi_process_data();
230 return process_data->data;
233 int smpi_process_count(void)
235 return process_count;
238 int smpi_process_index(void)
240 smpi_process_data_t data = smpi_process_data();
241 //return -1 if not initialized
242 return data ? data->index : MPI_UNDEFINED;
245 smx_rdv_t smpi_process_mailbox(void)
247 smpi_process_data_t data = smpi_process_data();
248 return data->mailbox;
251 smx_rdv_t smpi_process_mailbox_small(void)
253 smpi_process_data_t data = smpi_process_data();
254 return data->mailbox_small;
257 smx_rdv_t smpi_process_remote_mailbox(int index)
259 smpi_process_data_t data = smpi_process_remote_data(index);
260 return data->mailbox;
264 smx_rdv_t smpi_process_remote_mailbox_small(int index)
266 smpi_process_data_t data = smpi_process_remote_data(index);
267 return data->mailbox_small;
270 xbt_os_timer_t smpi_process_timer(void)
272 smpi_process_data_t data = smpi_process_data();
276 void smpi_process_simulated_start(void)
278 smpi_process_data_t data = smpi_process_data();
279 data->simulated = SIMIX_get_clock();
282 double smpi_process_simulated_elapsed(void)
284 smpi_process_data_t data = smpi_process_data();
285 return SIMIX_get_clock() - data->simulated;
288 MPI_Comm smpi_process_comm_self(void)
290 smpi_process_data_t data = smpi_process_data();
291 if(data->comm_self==MPI_COMM_NULL){
292 MPI_Group group = smpi_group_new(1);
293 data->comm_self = smpi_comm_new(group, NULL);
294 smpi_group_set_mapping(group, smpi_process_index(), 0);
297 return data->comm_self;
300 void smpi_process_set_sampling(int s)
302 smpi_process_data_t data = smpi_process_data();
306 int smpi_process_get_sampling(void)
308 smpi_process_data_t data = smpi_process_data();
309 return data->sampling;
312 void print_request(const char *message, MPI_Request request)
315 ("%s request %p [buf = %p, size = %zu, src = %d, dst = %d, tag = %d, flags = %x]",
316 message, request, request->buf, request->size, request->src,
317 request->dst, request->tag, request->flags);
320 static void smpi_comm_copy_buffer_callback(smx_action_t comm,
321 void *buff, size_t buff_size)
323 XBT_DEBUG("Copy the data over");
324 if(_xbt_replay_is_active()) return;
325 memcpy(comm->comm.dst_buff, buff, buff_size);
326 if (comm->comm.detached) {
327 // if this is a detached send, the source buffer was duplicated by SMPI
328 // sender to make the original buffer available to the application ASAP
330 //It seems that the request is used after the call there this should
331 //be free somewhereelse but where???
332 //xbt_free(comm->comm.src_data);// inside SMPI the request is keep
333 //inside the user data and should be free
334 comm->comm.src_buff = NULL;
338 void smpi_global_init(void)
342 char name[MAILBOX_NAME_MAXLEN];
344 SIMIX_comm_set_copy_data_callback(&smpi_comm_copy_buffer_callback);
345 process_count = SIMIX_process_count();
346 process_data = xbt_new(smpi_process_data_t, process_count);
347 for (i = 0; i < process_count; i++) {
348 process_data[i] = xbt_new(s_smpi_process_data_t, 1);
349 process_data[i]->index = i;
350 process_data[i]->argc = NULL;
351 process_data[i]->argv = NULL;
352 process_data[i]->mailbox = simcall_rdv_create(get_mailbox_name(name, i));
353 process_data[i]->mailbox_small =
354 simcall_rdv_create(get_mailbox_name_small(name, i));
355 process_data[i]->timer = xbt_os_timer_new();
357 MC_ignore_heap(process_data[i]->timer, xbt_os_timer_size());
358 process_data[i]->comm_self = MPI_COMM_NULL;
359 process_data[i]->state = SMPI_UNINITIALIZED;
360 process_data[i]->sampling = 0;
362 group = smpi_group_new(process_count);
363 MPI_COMM_WORLD = smpi_comm_new(group, NULL);
364 MPI_UNIVERSE_SIZE = smpi_comm_size(MPI_COMM_WORLD);
365 for (i = 0; i < process_count; i++) {
366 smpi_group_set_mapping(group, i, i);
369 //check correctness of MPI parameters
371 xbt_assert(sg_cfg_get_int("smpi/async_small_thres") <=
372 sg_cfg_get_int("smpi/send_is_detached_thres"));
374 if (sg_cfg_is_default_value("smpi/running_power")) {
375 XBT_INFO("You did not set the power of the host running the simulation. "
376 "The timings will certainly not be accurate. "
377 "Use the option \"--cfg=smpi/running_power:<flops>\" to set its value."
378 "Check http://simgrid.org/simgrid/latest/doc/options.html#options_smpi_bench for more information. ");
382 void smpi_global_destroy(void)
384 int count = smpi_process_count();
387 smpi_bench_destroy();
388 while (smpi_group_unuse(smpi_comm_group(MPI_COMM_WORLD)) > 0);
389 xbt_free(MPI_COMM_WORLD);
390 MPI_COMM_WORLD = MPI_COMM_NULL;
391 for (i = 0; i < count; i++) {
392 if(process_data[i]->comm_self!=MPI_COMM_NULL){
393 smpi_group_unuse(smpi_comm_group(process_data[i]->comm_self));
394 smpi_comm_destroy(process_data[i]->comm_self);
396 xbt_os_timer_free(process_data[i]->timer);
397 simcall_rdv_destroy(process_data[i]->mailbox);
398 simcall_rdv_destroy(process_data[i]->mailbox_small);
399 xbt_free(process_data[i]);
401 xbt_free(process_data);
407 /* Fortran specific stuff */
408 /* With smpicc, the following weak symbols are used */
409 /* With smpiff, the following weak symbols are replaced by those in libf2c */
410 int __attribute__ ((weak)) xargc;
411 char ** __attribute__ ((weak)) xargv;
414 void __attribute__ ((weak)) user_main_()
416 xbt_die("Should not be in this smpi_simulated_main");
420 int __attribute__ ((weak)) smpi_simulated_main_(int argc, char **argv)
422 smpi_process_init(&argc, &argv);
427 int __attribute__ ((weak)) main(int argc, char **argv)
429 return smpi_main(smpi_simulated_main_, argc, argv);
432 int __attribute__ ((weak)) MAIN__()
434 return smpi_main(smpi_simulated_main_, xargc, xargv);
438 int smpi_main(int (*realmain) (int argc, char *argv[]), int argc, char *argv[])
440 srand(SMPI_RAND_SEED);
442 if (getenv("SMPI_PRETEND_CC") != NULL) {
443 /* Hack to ensure that smpicc can pretend to be a simple
444 * compiler. Particularly handy to pass it to the configuration tools */
448 /* Connect log categories. See xbt/log.c */
449 XBT_LOG_CONNECT(smpi); /* Keep this line as soon as possible in this
450 function: xbt_log_appender_file.c depends on it
451 DO NOT connect this in XBT or so, or it will be
452 useless to xbt_log_appender_file.c */
454 XBT_LOG_CONNECT(instr_smpi);
456 XBT_LOG_CONNECT(smpi_base);
457 XBT_LOG_CONNECT(smpi_bench);
458 XBT_LOG_CONNECT(smpi_coll);
459 XBT_LOG_CONNECT(smpi_colls);
460 XBT_LOG_CONNECT(smpi_comm);
461 XBT_LOG_CONNECT(smpi_dvfs);
462 XBT_LOG_CONNECT(smpi_group);
463 XBT_LOG_CONNECT(smpi_kernel);
464 XBT_LOG_CONNECT(smpi_mpi);
465 XBT_LOG_CONNECT(smpi_mpi_dt);
466 XBT_LOG_CONNECT(smpi_pmpi);
467 XBT_LOG_CONNECT(smpi_replay);
470 TRACE_global_init(&argc, argv);
472 TRACE_add_start_function(TRACE_smpi_alloc);
473 TRACE_add_end_function(TRACE_smpi_release);
476 SIMIX_global_init(&argc, argv);
482 // parse the platform file: get the host list
483 SIMIX_create_environment(argv[1]);
485 SIMIX_function_register_default(realmain);
486 SIMIX_launch_application(argv[2]);
488 int gather_id = find_coll_description(mpi_coll_gather_description,
489 sg_cfg_get_string("smpi/gather"));
490 mpi_coll_gather_fun = (int (*)(void *, int, MPI_Datatype,
491 void *, int, MPI_Datatype, int, MPI_Comm))
492 mpi_coll_gather_description[gather_id].coll;
494 int allgather_id = find_coll_description(mpi_coll_allgather_description,
495 sg_cfg_get_string("smpi/allgather"));
496 mpi_coll_allgather_fun = (int (*)(void *, int, MPI_Datatype,
497 void *, int, MPI_Datatype, MPI_Comm))
498 mpi_coll_allgather_description[allgather_id].coll;
500 int allgatherv_id = find_coll_description(mpi_coll_allgatherv_description,
501 sg_cfg_get_string("smpi/allgatherv"));
502 mpi_coll_allgatherv_fun = (int (*)(void *, int, MPI_Datatype, void *, int *,
503 int *, MPI_Datatype, MPI_Comm))
504 mpi_coll_allgatherv_description[allgatherv_id].coll;
506 int allreduce_id = find_coll_description(mpi_coll_allreduce_description,
507 sg_cfg_get_string("smpi/allreduce"));
508 mpi_coll_allreduce_fun = (int (*)(void *sbuf, void *rbuf, int rcount,
509 MPI_Datatype dtype, MPI_Op op,
511 mpi_coll_allreduce_description[allreduce_id].coll;
513 int alltoall_id = find_coll_description(mpi_coll_alltoall_description,
514 sg_cfg_get_string("smpi/alltoall"));
515 mpi_coll_alltoall_fun = (int (*)(void *, int, MPI_Datatype,
516 void *, int, MPI_Datatype, MPI_Comm))
517 mpi_coll_alltoall_description[alltoall_id].coll;
519 int alltoallv_id = find_coll_description(mpi_coll_alltoallv_description,
520 sg_cfg_get_string("smpi/alltoallv"));
521 mpi_coll_alltoallv_fun = (int (*)(void *, int *, int *, MPI_Datatype,
522 void *, int *, int *, MPI_Datatype,
524 mpi_coll_alltoallv_description[alltoallv_id].coll;
526 int bcast_id = find_coll_description(mpi_coll_bcast_description,
527 sg_cfg_get_string("smpi/bcast"));
528 mpi_coll_bcast_fun = (int (*)(void *buf, int count, MPI_Datatype datatype,
529 int root, MPI_Comm com))
530 mpi_coll_bcast_description[bcast_id].coll;
532 int reduce_id = find_coll_description(mpi_coll_reduce_description,
533 sg_cfg_get_string("smpi/reduce"));
534 mpi_coll_reduce_fun = (int (*)(void *buf, void *rbuf, int count,
535 MPI_Datatype datatype, MPI_Op op,
536 int root, MPI_Comm comm))
537 mpi_coll_reduce_description[reduce_id].coll;
539 int reduce_scatter_id =
540 find_coll_description(mpi_coll_reduce_scatter_description,
541 sg_cfg_get_string("smpi/reduce_scatter"));
542 mpi_coll_reduce_scatter_fun = (int (*)(void *sbuf, void *rbuf, int *rcounts,
543 MPI_Datatype dtype, MPI_Op op,
545 mpi_coll_reduce_scatter_description[reduce_scatter_id].coll;
547 int scatter_id = find_coll_description(mpi_coll_scatter_description,
548 sg_cfg_get_string("smpi/scatter"));
549 mpi_coll_scatter_fun = (int (*)(void *sendbuf, int sendcount,
550 MPI_Datatype sendtype, void *recvbuf,
551 int recvcount, MPI_Datatype recvtype,
552 int root, MPI_Comm comm))
553 mpi_coll_scatter_description[scatter_id].coll;
555 int barrier_id = find_coll_description(mpi_coll_barrier_description,
556 sg_cfg_get_string("smpi/barrier"));
557 mpi_coll_barrier_fun = (int (*)(MPI_Comm comm))
558 mpi_coll_barrier_description[barrier_id].coll;
560 smpi_cpu_threshold = sg_cfg_get_double("smpi/cpu_threshold");
561 smpi_running_power = sg_cfg_get_double("smpi/running_power");
562 if (smpi_cpu_threshold < 0)
563 smpi_cpu_threshold = DBL_MAX;
567 /* Clean IO before the run */
572 MC_do_the_modelcheck_for_real();
576 if (sg_cfg_get_boolean("smpi/display_timing"))
577 XBT_INFO("Simulation time: %g seconds.", SIMIX_get_clock());
579 smpi_global_destroy();