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;
32 void *data; /* user data */
35 int sampling; /* inside an SMPI_SAMPLE_ block? */
37 } s_smpi_process_data_t;
39 static smpi_process_data_t *process_data = NULL;
40 int process_count = 0;
41 int* index_to_process_data = NULL;
44 MPI_Comm MPI_COMM_WORLD = MPI_COMM_UNINITIALIZED;
45 int MPI_UNIVERSE_SIZE;
47 MPI_Errhandler *MPI_ERRORS_RETURN = NULL;
48 MPI_Errhandler *MPI_ERRORS_ARE_FATAL = NULL;
49 MPI_Errhandler *MPI_ERRHANDLER_NULL = NULL;
51 #define MAILBOX_NAME_MAXLEN (5 + sizeof(int) * 2 + 1)
53 static char *get_mailbox_name(char *str, int index)
55 snprintf(str, MAILBOX_NAME_MAXLEN, "SMPI-%0*x", (int) (sizeof(int) * 2),
60 static char *get_mailbox_name_small(char *str, int index)
62 snprintf(str, MAILBOX_NAME_MAXLEN, "small%0*x", (int) (sizeof(int) * 2),
67 void smpi_process_init(int *argc, char ***argv)
70 smpi_process_data_t data;
74 proc = SIMIX_process_self();
75 //FIXME: dirty cleanup method to avoid using msg cleanup functions on these processes when using MSG+SMPI
76 proc->context->cleanup_func=SIMIX_process_cleanup;
77 char* instance_id = (*argv)[1];
78 int rank = atoi((*argv)[2]);
79 index = SIMIX_process_get_PID(proc) -1;
82 smpi_current_rank = index;
84 if(!index_to_process_data){
85 index_to_process_data=(int*)xbt_malloc(SIMIX_process_count()*sizeof(int));
87 MPI_Comm* temp_comm_world = smpi_deployment_register_process(instance_id, rank, index);
88 data = smpi_process_remote_data(index);
89 data->comm_world = temp_comm_world;
91 data->instance_id = instance_id;
92 simcall_process_set_data(proc, data);
96 memmove(&(*argv)[1], &(*argv)[3], sizeof(char *) * (*argc - 3));
97 (*argv)[(*argc) - 1] = NULL;
98 (*argv)[(*argc) - 2] = NULL;
103 // set the process attached to the mailbox
104 simcall_rdv_set_receiver(data->mailbox_small, proc);
105 XBT_DEBUG("<%d> New process in the game: %p", index, proc);
107 if(smpi_privatize_global_variables){
108 switch_data_segment(index);
112 if (smpi_process_data() == NULL)
113 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.");
116 void smpi_process_destroy(void)
118 int index = smpi_process_index();
119 if(smpi_privatize_global_variables){
120 switch_data_segment(index);
122 process_data[index_to_process_data[index]]->state = SMPI_FINALIZED;
123 XBT_DEBUG("<%d> Process left the game", index);
127 * @brief Prepares the current process for termination.
129 void smpi_process_finalize(void)
132 // wait for all pending asynchronous comms to finish
133 while (SIMIX_process_has_pending_comms(SIMIX_process_self())) {
134 simcall_process_sleep(0.01);
138 int size = smpi_comm_size(MPI_COMM_WORLD);
139 int rank = smpi_comm_rank(MPI_COMM_WORLD);
140 /* All non-root send & receive zero-length message. */
142 smpi_mpi_ssend (NULL, 0, MPI_BYTE, 0,
145 smpi_mpi_recv (NULL, 0, MPI_BYTE, 0,
147 MPI_COMM_WORLD, MPI_STATUS_IGNORE);
149 /* The root collects and broadcasts the messages. */
151 MPI_Request* requests;
152 requests = (MPI_Request*)malloc( size * sizeof(MPI_Request) );
153 for (i = 1; i < size; ++i) {
154 requests[i] = smpi_mpi_irecv(NULL, 0, MPI_BYTE, MPI_ANY_SOURCE,
155 COLL_TAG_BARRIER, MPI_COMM_WORLD
158 smpi_mpi_waitall( size-1, requests+1, MPI_STATUSES_IGNORE );
159 for (i = 1; i < size; ++i) {
160 requests[i] = smpi_mpi_issend(NULL, 0, MPI_BYTE, i,
165 smpi_mpi_waitall( size-1, requests+1, MPI_STATUSES_IGNORE );
172 * @brief Check if a process is finalized
174 int smpi_process_finalized()
176 int index = smpi_process_index();
177 if (index != MPI_UNDEFINED)
178 return (process_data[index_to_process_data[index]]->state == SMPI_FINALIZED);
184 * @brief Check if a process is initialized
186 int smpi_process_initialized(void)
188 int index = smpi_process_index();
189 return ( (index != MPI_UNDEFINED)
190 && (process_data[index_to_process_data[index]]->state == SMPI_INITIALIZED));
194 * @brief Mark a process as initialized (=MPI_Init called)
196 void smpi_process_mark_as_initialized(void)
198 int index = smpi_process_index();
199 if ((index != MPI_UNDEFINED) && (!process_data[index_to_process_data[index]]->state != SMPI_FINALIZED))
200 process_data[index_to_process_data[index]]->state = SMPI_INITIALIZED;
205 int smpi_process_argc(void)
207 smpi_process_data_t data = smpi_process_data();
208 return data->argc ? *(data->argc) - 1 : 0;
211 int smpi_process_getarg(integer * index, char *dst, ftnlen len)
213 smpi_process_data_t data = smpi_process_data();
217 if (!data->argc || !data->argv || *index < 1 || *index >= *(data->argc)) {
220 arg = (*data->argv)[*index];
221 for (i = 0; i < len && arg[i] != '\0'; i++) {
224 for (; i < len; i++) {
231 int smpi_global_size(void)
233 char *value = getenv("SMPI_GLOBAL_SIZE");
237 "Please set env var SMPI_GLOBAL_SIZE to expected number of processes.\n");
243 smpi_process_data_t smpi_process_data(void)
245 return SIMIX_process_self_get_data(SIMIX_process_self());
248 smpi_process_data_t smpi_process_remote_data(int index)
250 return process_data[index_to_process_data[index]];
253 void smpi_process_set_user_data(void *data)
255 smpi_process_data_t process_data = smpi_process_data();
256 process_data->data = data;
259 void *smpi_process_get_user_data()
261 smpi_process_data_t process_data = smpi_process_data();
262 return process_data->data;
265 int smpi_process_count(void)
267 return process_count;
270 int smpi_process_index(void)
272 smpi_process_data_t data = smpi_process_data();
273 //return -1 if not initialized
274 return data ? data->index : MPI_UNDEFINED;
277 MPI_Comm smpi_process_comm_world(void)
279 smpi_process_data_t data = smpi_process_data();
280 //return MPI_COMM_NULL if not initialized
281 return data ? *data->comm_world : MPI_COMM_NULL;
284 smx_rdv_t smpi_process_mailbox(void)
286 smpi_process_data_t data = smpi_process_data();
287 return data->mailbox;
290 smx_rdv_t smpi_process_mailbox_small(void)
292 smpi_process_data_t data = smpi_process_data();
293 return data->mailbox_small;
296 smx_rdv_t smpi_process_remote_mailbox(int index)
298 smpi_process_data_t data = smpi_process_remote_data(index);
299 return data->mailbox;
303 smx_rdv_t smpi_process_remote_mailbox_small(int index)
305 smpi_process_data_t data = smpi_process_remote_data(index);
306 return data->mailbox_small;
309 xbt_os_timer_t smpi_process_timer(void)
311 smpi_process_data_t data = smpi_process_data();
315 void smpi_process_simulated_start(void)
317 smpi_process_data_t data = smpi_process_data();
318 data->simulated = SIMIX_get_clock();
321 double smpi_process_simulated_elapsed(void)
323 smpi_process_data_t data = smpi_process_data();
324 return SIMIX_get_clock() - data->simulated;
327 MPI_Comm smpi_process_comm_self(void)
329 smpi_process_data_t data = smpi_process_data();
330 if(data->comm_self==MPI_COMM_NULL){
331 MPI_Group group = smpi_group_new(1);
332 data->comm_self = smpi_comm_new(group, NULL);
333 smpi_group_set_mapping(group, smpi_process_index(), 0);
336 return data->comm_self;
339 void smpi_process_set_sampling(int s)
341 smpi_process_data_t data = smpi_process_data();
345 int smpi_process_get_sampling(void)
347 smpi_process_data_t data = smpi_process_data();
348 return data->sampling;
351 void print_request(const char *message, MPI_Request request)
354 ("%s request %p [buf = %p, size = %zu, src = %d, dst = %d, tag = %d, flags = %x]",
355 message, request, request->buf, request->size, request->src,
356 request->dst, request->tag, request->flags);
359 void smpi_comm_copy_buffer_callback(smx_action_t comm,
360 void *buff, size_t buff_size)
362 XBT_DEBUG("Copy the data over");
363 if(_xbt_replay_is_active()) return;
366 if((smpi_privatize_global_variables)
367 && ((char*)buff >= start_data_exe)
368 && ((char*)buff < start_data_exe + size_data_exe )
370 XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !");
371 switch_data_segment(((smpi_process_data_t)SIMIX_process_get_data(comm->comm.src_proc))->index);
372 tmpbuff = (void*)xbt_malloc(buff_size);
373 memcpy(tmpbuff, buff, buff_size);
377 if((smpi_privatize_global_variables)
378 && ((char*)comm->comm.dst_buff >= start_data_exe)
379 && ((char*)comm->comm.dst_buff < start_data_exe + size_data_exe )
381 XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment");
382 switch_data_segment(((smpi_process_data_t)SIMIX_process_get_data(comm->comm.dst_proc))->index);
386 memcpy(comm->comm.dst_buff, tmpbuff, buff_size);
387 if (comm->comm.detached) {
388 // if this is a detached send, the source buffer was duplicated by SMPI
389 // sender to make the original buffer available to the application ASAP
391 //It seems that the request is used after the call there this should
392 //be free somewhereelse but where???
393 //xbt_free(comm->comm.src_data);// inside SMPI the request is keep
394 //inside the user data and should be free
395 comm->comm.src_buff = NULL;
398 if(tmpbuff!=buff)xbt_free(tmpbuff);
402 static void smpi_check_options(){
403 //check correctness of MPI parameters
405 xbt_assert(sg_cfg_get_int("smpi/async_small_thres") <=
406 sg_cfg_get_int("smpi/send_is_detached_thres"));
408 if (sg_cfg_is_default_value("smpi/running_power")) {
409 XBT_INFO("You did not set the power of the host running the simulation. "
410 "The timings will certainly not be accurate. "
411 "Use the option \"--cfg=smpi/running_power:<flops>\" to set its value."
412 "Check http://simgrid.org/simgrid/latest/doc/options.html#options_smpi_bench for more information. ");
416 void smpi_global_init(void)
420 char name[MAILBOX_NAME_MAXLEN];
424 if (process_count == 0){
425 process_count = SIMIX_process_count();
428 process_data = xbt_new0(smpi_process_data_t, process_count);
429 for (i = 0; i < process_count; i++) {
430 process_data[i] = xbt_new(s_smpi_process_data_t, 1);
431 //process_data[i]->index = i;
432 process_data[i]->argc = NULL;
433 process_data[i]->argv = NULL;
434 process_data[i]->mailbox = simcall_rdv_create(get_mailbox_name(name, i));
435 process_data[i]->mailbox_small =
436 simcall_rdv_create(get_mailbox_name_small(name, i));
437 process_data[i]->timer = xbt_os_timer_new();
439 MC_ignore_heap(process_data[i]->timer, xbt_os_timer_size());
440 process_data[i]->comm_self = MPI_COMM_NULL;
441 process_data[i]->comm_world = NULL;
442 process_data[i]->state = SMPI_UNINITIALIZED;
443 process_data[i]->sampling = 0;
445 //if the process was launched through smpirun script
446 //we generate a global mpi_comm_world
447 //if not, we let MPI_COMM_NULL, and the comm world
448 //will be private to each mpi instance
450 group = smpi_group_new(process_count);
451 MPI_COMM_WORLD = smpi_comm_new(group, NULL);
452 MPI_UNIVERSE_SIZE = smpi_comm_size(MPI_COMM_WORLD);
453 for (i = 0; i < process_count; i++) {
454 smpi_group_set_mapping(group, i, i);
459 void smpi_global_destroy(void)
461 int count = smpi_process_count();
464 smpi_bench_destroy();
465 if (MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED){
466 while (smpi_group_unuse(smpi_comm_group(MPI_COMM_WORLD)) > 0);
467 xbt_free(MPI_COMM_WORLD);
469 MPI_COMM_WORLD = MPI_COMM_NULL;
470 for (i = 0; i < count; i++) {
471 if(process_data[i]->comm_self!=MPI_COMM_NULL){
472 smpi_group_unuse(smpi_comm_group(process_data[i]->comm_self));
473 smpi_comm_destroy(process_data[i]->comm_self);
475 xbt_os_timer_free(process_data[i]->timer);
476 simcall_rdv_destroy(process_data[i]->mailbox);
477 simcall_rdv_destroy(process_data[i]->mailbox_small);
478 xbt_free(process_data[i]);
480 xbt_free(process_data);
482 if(smpi_privatize_global_variables)
483 smpi_destroy_global_memory_segments();
487 /* Fortran specific stuff */
488 /* With smpicc, the following weak symbols are used */
489 /* With smpiff, the following weak symbols are replaced by those in libf2c */
490 int __attribute__ ((weak)) xargc;
491 char ** __attribute__ ((weak)) xargv;
494 void __attribute__ ((weak)) user_main_()
496 xbt_die("Should not be in this smpi_simulated_main");
500 int __attribute__ ((weak)) smpi_simulated_main_(int argc, char **argv)
502 smpi_process_init(&argc, &argv);
507 int __attribute__ ((weak)) main(int argc, char **argv)
509 return smpi_main(smpi_simulated_main_, argc, argv);
512 int __attribute__ ((weak)) MAIN__()
514 return smpi_main(smpi_simulated_main_, xargc, xargv);
518 static void smpi_init_logs(){
520 /* Connect log categories. See xbt/log.c */
521 XBT_LOG_CONNECT(smpi); /* Keep this line as soon as possible in this
522 function: xbt_log_appender_file.c depends on it
523 DO NOT connect this in XBT or so, or it will be
524 useless to xbt_log_appender_file.c */
526 XBT_LOG_CONNECT(instr_smpi);
528 XBT_LOG_CONNECT(smpi_base);
529 XBT_LOG_CONNECT(smpi_bench);
530 XBT_LOG_CONNECT(smpi_coll);
531 XBT_LOG_CONNECT(smpi_colls);
532 XBT_LOG_CONNECT(smpi_comm);
533 XBT_LOG_CONNECT(smpi_dvfs);
534 XBT_LOG_CONNECT(smpi_group);
535 XBT_LOG_CONNECT(smpi_kernel);
536 XBT_LOG_CONNECT(smpi_mpi);
537 XBT_LOG_CONNECT(smpi_mpi_dt);
538 XBT_LOG_CONNECT(smpi_pmpi);
539 XBT_LOG_CONNECT(smpi_replay);
544 static void smpi_init_options(){
545 int gather_id = find_coll_description(mpi_coll_gather_description,
546 sg_cfg_get_string("smpi/gather"));
547 mpi_coll_gather_fun = (int (*)(void *, int, MPI_Datatype,
548 void *, int, MPI_Datatype, int, MPI_Comm))
549 mpi_coll_gather_description[gather_id].coll;
551 int allgather_id = find_coll_description(mpi_coll_allgather_description,
552 sg_cfg_get_string("smpi/allgather"));
553 mpi_coll_allgather_fun = (int (*)(void *, int, MPI_Datatype,
554 void *, int, MPI_Datatype, MPI_Comm))
555 mpi_coll_allgather_description[allgather_id].coll;
557 int allgatherv_id = find_coll_description(mpi_coll_allgatherv_description,
558 sg_cfg_get_string("smpi/allgatherv"));
559 mpi_coll_allgatherv_fun = (int (*)(void *, int, MPI_Datatype, void *, int *,
560 int *, MPI_Datatype, MPI_Comm))
561 mpi_coll_allgatherv_description[allgatherv_id].coll;
563 int allreduce_id = find_coll_description(mpi_coll_allreduce_description,
564 sg_cfg_get_string("smpi/allreduce"));
565 mpi_coll_allreduce_fun = (int (*)(void *sbuf, void *rbuf, int rcount,
566 MPI_Datatype dtype, MPI_Op op,
568 mpi_coll_allreduce_description[allreduce_id].coll;
570 int alltoall_id = find_coll_description(mpi_coll_alltoall_description,
571 sg_cfg_get_string("smpi/alltoall"));
572 mpi_coll_alltoall_fun = (int (*)(void *, int, MPI_Datatype,
573 void *, int, MPI_Datatype, MPI_Comm))
574 mpi_coll_alltoall_description[alltoall_id].coll;
576 int alltoallv_id = find_coll_description(mpi_coll_alltoallv_description,
577 sg_cfg_get_string("smpi/alltoallv"));
578 mpi_coll_alltoallv_fun = (int (*)(void *, int *, int *, MPI_Datatype,
579 void *, int *, int *, MPI_Datatype,
581 mpi_coll_alltoallv_description[alltoallv_id].coll;
583 int bcast_id = find_coll_description(mpi_coll_bcast_description,
584 sg_cfg_get_string("smpi/bcast"));
585 mpi_coll_bcast_fun = (int (*)(void *buf, int count, MPI_Datatype datatype,
586 int root, MPI_Comm com))
587 mpi_coll_bcast_description[bcast_id].coll;
589 int reduce_id = find_coll_description(mpi_coll_reduce_description,
590 sg_cfg_get_string("smpi/reduce"));
591 mpi_coll_reduce_fun = (int (*)(void *buf, void *rbuf, int count,
592 MPI_Datatype datatype, MPI_Op op,
593 int root, MPI_Comm comm))
594 mpi_coll_reduce_description[reduce_id].coll;
596 int reduce_scatter_id =
597 find_coll_description(mpi_coll_reduce_scatter_description,
598 sg_cfg_get_string("smpi/reduce_scatter"));
599 mpi_coll_reduce_scatter_fun = (int (*)(void *sbuf, void *rbuf, int *rcounts,
600 MPI_Datatype dtype, MPI_Op op,
602 mpi_coll_reduce_scatter_description[reduce_scatter_id].coll;
604 int scatter_id = find_coll_description(mpi_coll_scatter_description,
605 sg_cfg_get_string("smpi/scatter"));
606 mpi_coll_scatter_fun = (int (*)(void *sendbuf, int sendcount,
607 MPI_Datatype sendtype, void *recvbuf,
608 int recvcount, MPI_Datatype recvtype,
609 int root, MPI_Comm comm))
610 mpi_coll_scatter_description[scatter_id].coll;
612 int barrier_id = find_coll_description(mpi_coll_barrier_description,
613 sg_cfg_get_string("smpi/barrier"));
614 mpi_coll_barrier_fun = (int (*)(MPI_Comm comm))
615 mpi_coll_barrier_description[barrier_id].coll;
617 smpi_cpu_threshold = sg_cfg_get_double("smpi/cpu_threshold");
618 smpi_running_power = sg_cfg_get_double("smpi/running_power");
619 smpi_privatize_global_variables = sg_cfg_get_boolean("smpi/privatize_global_variables");
620 if (smpi_cpu_threshold < 0)
621 smpi_cpu_threshold = DBL_MAX;
625 int smpi_main(int (*realmain) (int argc, char *argv[]), int argc, char *argv[])
627 srand(SMPI_RAND_SEED);
629 if (getenv("SMPI_PRETEND_CC") != NULL) {
630 /* Hack to ensure that smpicc can pretend to be a simple
631 * compiler. Particularly handy to pass it to the configuration tools */
638 TRACE_global_init(&argc, argv);
640 TRACE_add_start_function(TRACE_smpi_alloc);
641 TRACE_add_end_function(TRACE_smpi_release);
644 SIMIX_global_init(&argc, argv);
646 // parse the platform file: get the host list
647 SIMIX_create_environment(argv[1]);
648 SIMIX_comm_set_copy_data_callback(&smpi_comm_copy_buffer_callback);
649 SIMIX_function_register_default(realmain);
650 SIMIX_launch_application(argv[2]);
656 smpi_check_options();
658 if(smpi_privatize_global_variables)
659 smpi_initialize_global_memory_segments();
661 /* Clean IO before the run */
666 MC_do_the_modelcheck_for_real();
670 if (sg_cfg_get_boolean("smpi/display_timing"))
671 XBT_INFO("Simulation time: %g seconds.", SIMIX_get_clock());
673 smpi_global_destroy();
682 // This function can be called from extern file, to initialize logs, options, and processes of smpi
683 // without the need of smpirun
688 smpi_check_options();
690 if (TRACE_is_enabled() && TRACE_is_configured()) {
694 if(smpi_privatize_global_variables)
695 smpi_initialize_global_memory_segments();
698 void SMPI_finalize(){
699 smpi_global_destroy();