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;
33 void *data; /* user data */
36 int sampling; /* inside an SMPI_SAMPLE_ block? */
38 int replaying; /* is the process replaying a trace */
39 xbt_bar_t finalization_barrier;
40 } s_smpi_process_data_t;
42 static smpi_process_data_t *process_data = NULL;
43 int process_count = 0;
44 int* index_to_process_data = NULL;
45 extern double smpi_total_benched_time;
46 xbt_os_timer_t global_timer;
47 MPI_Comm MPI_COMM_WORLD = MPI_COMM_UNINITIALIZED;
48 int MPI_UNIVERSE_SIZE;
50 MPI_Errhandler *MPI_ERRORS_RETURN = NULL;
51 MPI_Errhandler *MPI_ERRORS_ARE_FATAL = NULL;
52 MPI_Errhandler *MPI_ERRHANDLER_NULL = NULL;
54 #define MAILBOX_NAME_MAXLEN (5 + sizeof(int) * 2 + 1)
56 static char *get_mailbox_name(char *str, int index)
58 snprintf(str, MAILBOX_NAME_MAXLEN, "SMPI-%0*x", (int) (sizeof(int) * 2),
63 static char *get_mailbox_name_small(char *str, int index)
65 snprintf(str, MAILBOX_NAME_MAXLEN, "small%0*x", (int) (sizeof(int) * 2),
70 void smpi_process_init(int *argc, char ***argv)
73 smpi_process_data_t data;
77 proc = SIMIX_process_self();
78 //FIXME: dirty cleanup method to avoid using msg cleanup functions on these processes when using MSG+SMPI
79 proc->context->cleanup_func=SIMIX_process_cleanup;
80 char* instance_id = (*argv)[1];
81 int rank = atoi((*argv)[2]);
82 index = smpi_process_index_of_smx_process(proc);
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;
89 smpi_deployment_register_process(instance_id, rank, index, &temp_comm_world ,&temp_bar);
90 data = smpi_process_remote_data(index);
91 data->comm_world = temp_comm_world;
92 if(temp_bar != NULL) data->finalization_barrier = temp_bar;
94 data->instance_id = instance_id;
96 xbt_free(simcall_process_get_data(proc));
97 simcall_process_set_data(proc, data);
100 memmove(&(*argv)[0], &(*argv)[2], sizeof(char *) * (*argc - 2));
101 (*argv)[(*argc) - 1] = NULL;
102 (*argv)[(*argc) - 2] = NULL;
107 // set the process attached to the mailbox
108 simcall_rdv_set_receiver(data->mailbox_small, proc);
110 XBT_DEBUG("<%d> New process in the game: %p", index, proc);
112 if(smpi_privatize_global_variables){
113 smpi_switch_data_segment(index);
117 if (smpi_process_data() == NULL)
118 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.");
121 void smpi_process_destroy(void)
123 int index = smpi_process_index();
124 if(smpi_privatize_global_variables){
125 smpi_switch_data_segment(index);
127 process_data[index_to_process_data[index]]->state = SMPI_FINALIZED;
128 XBT_DEBUG("<%d> Process left the game", index);
132 * @brief Prepares the current process for termination.
134 void smpi_process_finalize(void)
136 // This leads to an explosion of the search graph
137 // which cannot be reduced:
141 int index = smpi_process_index();
142 // wait for all pending asynchronous comms to finish
143 xbt_barrier_wait(process_data[index_to_process_data[index]]->finalization_barrier);
147 * @brief Check if a process is finalized
149 int smpi_process_finalized()
151 int index = smpi_process_index();
152 if (index != MPI_UNDEFINED)
153 return (process_data[index_to_process_data[index]]->state == SMPI_FINALIZED);
159 * @brief Check if a process is initialized
161 int smpi_process_initialized(void)
163 int index = smpi_process_index();
164 return ( (index != MPI_UNDEFINED)
165 && (process_data[index_to_process_data[index]]->state == SMPI_INITIALIZED));
169 * @brief Mark a process as initialized (=MPI_Init called)
171 void smpi_process_mark_as_initialized(void)
173 int index = smpi_process_index();
174 if ((index != MPI_UNDEFINED) && (process_data[index_to_process_data[index]]->state != SMPI_FINALIZED))
175 process_data[index_to_process_data[index]]->state = SMPI_INITIALIZED;
178 void smpi_process_set_replaying(int value){
179 int index = smpi_process_index();
180 if ((index != MPI_UNDEFINED) && (process_data[index_to_process_data[index]]->state != SMPI_FINALIZED))
181 process_data[index_to_process_data[index]]->replaying = value;
184 int smpi_process_get_replaying(){
185 int index = smpi_process_index();
186 if (index != MPI_UNDEFINED)
187 return process_data[index_to_process_data[index]]->replaying;
188 else return _xbt_replay_is_active();
192 int smpi_global_size(void)
194 char *value = getenv("SMPI_GLOBAL_SIZE");
198 "Please set env var SMPI_GLOBAL_SIZE to expected number of processes.\n");
204 smpi_process_data_t smpi_process_data(void)
206 return SIMIX_process_self_get_data(SIMIX_process_self());
209 smpi_process_data_t smpi_process_remote_data(int index)
211 return process_data[index_to_process_data[index]];
214 void smpi_process_set_user_data(void *data)
216 smpi_process_data_t process_data = smpi_process_data();
217 process_data->data = data;
220 void *smpi_process_get_user_data()
222 smpi_process_data_t process_data = smpi_process_data();
223 return process_data->data;
226 int smpi_process_count(void)
228 return process_count;
231 int smpi_process_index(void)
233 smpi_process_data_t data = smpi_process_data();
234 //return -1 if not initialized
235 return data ? data->index : MPI_UNDEFINED;
238 MPI_Comm smpi_process_comm_world(void)
240 smpi_process_data_t data = smpi_process_data();
241 //return MPI_COMM_NULL if not initialized
242 return data ? *data->comm_world : MPI_COMM_NULL;
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 MPI_Comm smpi_process_get_comm_intra(void)
302 smpi_process_data_t data = smpi_process_data();
303 return data->comm_intra;
306 void smpi_process_set_comm_intra(MPI_Comm comm)
308 smpi_process_data_t data = smpi_process_data();
309 data->comm_intra = comm;
312 void smpi_process_set_sampling(int s)
314 smpi_process_data_t data = smpi_process_data();
318 int smpi_process_get_sampling(void)
320 smpi_process_data_t data = smpi_process_data();
321 return data->sampling;
325 void print_request(const char *message, MPI_Request request)
328 ("%s request %p [buf = %p, size = %zu, src = %d, dst = %d, tag = %d, flags = %x]",
329 message, request, request->buf, request->size, request->src,
330 request->dst, request->tag, request->flags);
333 void smpi_comm_copy_buffer_callback(smx_synchro_t comm,
334 void *buff, size_t buff_size)
336 XBT_DEBUG("Copy the data over");
339 if((smpi_privatize_global_variables)
340 && ((char*)buff >= start_data_exe)
341 && ((char*)buff < start_data_exe + size_data_exe )
343 XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !");
344 smpi_switch_data_segment(((smpi_process_data_t)SIMIX_process_get_data(comm->comm.src_proc))->index);
345 tmpbuff = (void*)xbt_malloc(buff_size);
346 memcpy(tmpbuff, buff, buff_size);
350 if((smpi_privatize_global_variables)
351 && ((char*)comm->comm.dst_buff >= start_data_exe)
352 && ((char*)comm->comm.dst_buff < start_data_exe + size_data_exe )
354 XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment");
355 smpi_switch_data_segment(((smpi_process_data_t)SIMIX_process_get_data(comm->comm.dst_proc))->index);
359 memcpy(comm->comm.dst_buff, tmpbuff, buff_size);
360 if (comm->comm.detached) {
361 // if this is a detached send, the source buffer was duplicated by SMPI
362 // sender to make the original buffer available to the application ASAP
364 //It seems that the request is used after the call there this should
365 //be free somewhereelse but where???
366 //xbt_free(comm->comm.src_data);// inside SMPI the request is keep
367 //inside the user data and should be free
368 comm->comm.src_buff = NULL;
371 if(tmpbuff!=buff)xbt_free(tmpbuff);
376 void smpi_comm_null_copy_buffer_callback(smx_synchro_t comm,
377 void *buff, size_t buff_size)
382 static void smpi_check_options(){
383 //check correctness of MPI parameters
385 xbt_assert(sg_cfg_get_int("smpi/async_small_thres") <=
386 sg_cfg_get_int("smpi/send_is_detached_thres"));
388 if (sg_cfg_is_default_value("smpi/running_power")) {
389 XBT_INFO("You did not set the power of the host running the simulation. "
390 "The timings will certainly not be accurate. "
391 "Use the option \"--cfg=smpi/running_power:<flops>\" to set its value."
392 "Check http://simgrid.org/simgrid/latest/doc/options.html#options_smpi_bench for more information.");
396 int smpi_enabled(void) {
397 return process_data != NULL;
400 void smpi_global_init(void)
404 char name[MAILBOX_NAME_MAXLEN];
407 if (!MC_is_active()) {
408 global_timer = xbt_os_timer_new();
409 xbt_os_walltimer_start(global_timer);
411 if (process_count == 0){
412 process_count = SIMIX_process_count();
415 process_data = xbt_new0(smpi_process_data_t, process_count);
416 for (i = 0; i < process_count; i++) {
417 process_data[i] = xbt_new(s_smpi_process_data_t, 1);
418 //process_data[i]->index = i;
419 process_data[i]->argc = NULL;
420 process_data[i]->argv = NULL;
421 process_data[i]->mailbox = simcall_rdv_create(get_mailbox_name(name, i));
422 process_data[i]->mailbox_small =
423 simcall_rdv_create(get_mailbox_name_small(name, i));
424 process_data[i]->timer = xbt_os_timer_new();
426 MC_ignore_heap(process_data[i]->timer, xbt_os_timer_size());
427 process_data[i]->comm_self = MPI_COMM_NULL;
428 process_data[i]->comm_intra = MPI_COMM_NULL;
429 process_data[i]->comm_world = NULL;
430 process_data[i]->state = SMPI_UNINITIALIZED;
431 process_data[i]->sampling = 0;
432 process_data[i]->finalization_barrier = NULL;
434 //if the process was launched through smpirun script
435 //we generate a global mpi_comm_world
436 //if not, we let MPI_COMM_NULL, and the comm world
437 //will be private to each mpi instance
439 group = smpi_group_new(process_count);
440 MPI_COMM_WORLD = smpi_comm_new(group, NULL);
441 xbt_bar_t bar=xbt_barrier_init(process_count);
443 MPI_UNIVERSE_SIZE = smpi_comm_size(MPI_COMM_WORLD);
444 for (i = 0; i < process_count; i++) {
445 smpi_group_set_mapping(group, i, i);
446 process_data[i]->finalization_barrier = bar;
451 void smpi_global_destroy(void)
453 int count = smpi_process_count();
456 smpi_bench_destroy();
457 if (MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED){
458 while (smpi_group_unuse(smpi_comm_group(MPI_COMM_WORLD)) > 0);
459 xbt_free(MPI_COMM_WORLD);
460 xbt_barrier_destroy(process_data[0]->finalization_barrier);
462 smpi_deployment_cleanup_instances();
464 MPI_COMM_WORLD = MPI_COMM_NULL;
465 for (i = 0; i < count; i++) {
466 if(process_data[i]->comm_self!=MPI_COMM_NULL){
467 smpi_group_unuse(smpi_comm_group(process_data[i]->comm_self));
468 smpi_comm_destroy(process_data[i]->comm_self);
470 if(process_data[i]->comm_intra!=MPI_COMM_NULL){
471 smpi_group_unuse(smpi_comm_group(process_data[i]->comm_intra));
472 smpi_comm_destroy(process_data[i]->comm_intra);
474 xbt_os_timer_free(process_data[i]->timer);
475 simcall_rdv_destroy(process_data[i]->mailbox);
476 simcall_rdv_destroy(process_data[i]->mailbox_small);
477 xbt_free(process_data[i]);
479 xbt_free(process_data);
482 xbt_free(index_to_process_data);
483 if(smpi_privatize_global_variables)
484 smpi_destroy_global_memory_segments();
489 void __attribute__ ((weak)) user_main_()
491 xbt_die("Should not be in this smpi_simulated_main");
495 int __attribute__ ((weak)) smpi_simulated_main_(int argc, char **argv)
497 smpi_process_init(&argc, &argv);
502 int __attribute__ ((weak)) main(int argc, char **argv)
504 return smpi_main(smpi_simulated_main_, argc, argv);
509 static void smpi_init_logs(){
511 /* Connect log categories. See xbt/log.c */
512 XBT_LOG_CONNECT(smpi); /* Keep this line as soon as possible in this
513 function: xbt_log_appender_file.c depends on it
514 DO NOT connect this in XBT or so, or it will be
515 useless to xbt_log_appender_file.c */
517 XBT_LOG_CONNECT(instr_smpi);
519 XBT_LOG_CONNECT(smpi_base);
520 XBT_LOG_CONNECT(smpi_bench);
521 XBT_LOG_CONNECT(smpi_coll);
522 XBT_LOG_CONNECT(smpi_colls);
523 XBT_LOG_CONNECT(smpi_comm);
524 XBT_LOG_CONNECT(smpi_dvfs);
525 XBT_LOG_CONNECT(smpi_group);
526 XBT_LOG_CONNECT(smpi_kernel);
527 XBT_LOG_CONNECT(smpi_mpi);
528 XBT_LOG_CONNECT(smpi_mpi_dt);
529 XBT_LOG_CONNECT(smpi_pmpi);
530 XBT_LOG_CONNECT(smpi_replay);
531 XBT_LOG_CONNECT(smpi_rma);
536 static void smpi_init_options(){
537 int gather_id = find_coll_description(mpi_coll_gather_description,
538 sg_cfg_get_string("smpi/gather"));
539 mpi_coll_gather_fun = (int (*)(void *, int, MPI_Datatype,
540 void *, int, MPI_Datatype, int, MPI_Comm))
541 mpi_coll_gather_description[gather_id].coll;
543 int allgather_id = find_coll_description(mpi_coll_allgather_description,
544 sg_cfg_get_string("smpi/allgather"));
545 mpi_coll_allgather_fun = (int (*)(void *, int, MPI_Datatype,
546 void *, int, MPI_Datatype, MPI_Comm))
547 mpi_coll_allgather_description[allgather_id].coll;
549 int allgatherv_id = find_coll_description(mpi_coll_allgatherv_description,
550 sg_cfg_get_string("smpi/allgatherv"));
551 mpi_coll_allgatherv_fun = (int (*)(void *, int, MPI_Datatype, void *, int *,
552 int *, MPI_Datatype, MPI_Comm))
553 mpi_coll_allgatherv_description[allgatherv_id].coll;
555 int allreduce_id = find_coll_description(mpi_coll_allreduce_description,
556 sg_cfg_get_string("smpi/allreduce"));
557 mpi_coll_allreduce_fun = (int (*)(void *sbuf, void *rbuf, int rcount,
558 MPI_Datatype dtype, MPI_Op op,
560 mpi_coll_allreduce_description[allreduce_id].coll;
562 int alltoall_id = find_coll_description(mpi_coll_alltoall_description,
563 sg_cfg_get_string("smpi/alltoall"));
564 mpi_coll_alltoall_fun = (int (*)(void *, int, MPI_Datatype,
565 void *, int, MPI_Datatype, MPI_Comm))
566 mpi_coll_alltoall_description[alltoall_id].coll;
568 int alltoallv_id = find_coll_description(mpi_coll_alltoallv_description,
569 sg_cfg_get_string("smpi/alltoallv"));
570 mpi_coll_alltoallv_fun = (int (*)(void *, int *, int *, MPI_Datatype,
571 void *, int *, int *, MPI_Datatype,
573 mpi_coll_alltoallv_description[alltoallv_id].coll;
575 int bcast_id = find_coll_description(mpi_coll_bcast_description,
576 sg_cfg_get_string("smpi/bcast"));
577 mpi_coll_bcast_fun = (int (*)(void *buf, int count, MPI_Datatype datatype,
578 int root, MPI_Comm com))
579 mpi_coll_bcast_description[bcast_id].coll;
581 int reduce_id = find_coll_description(mpi_coll_reduce_description,
582 sg_cfg_get_string("smpi/reduce"));
583 mpi_coll_reduce_fun = (int (*)(void *buf, void *rbuf, int count,
584 MPI_Datatype datatype, MPI_Op op,
585 int root, MPI_Comm comm))
586 mpi_coll_reduce_description[reduce_id].coll;
588 int reduce_scatter_id =
589 find_coll_description(mpi_coll_reduce_scatter_description,
590 sg_cfg_get_string("smpi/reduce_scatter"));
591 mpi_coll_reduce_scatter_fun = (int (*)(void *sbuf, void *rbuf, int *rcounts,
592 MPI_Datatype dtype, MPI_Op op,
594 mpi_coll_reduce_scatter_description[reduce_scatter_id].coll;
596 int scatter_id = find_coll_description(mpi_coll_scatter_description,
597 sg_cfg_get_string("smpi/scatter"));
598 mpi_coll_scatter_fun = (int (*)(void *sendbuf, int sendcount,
599 MPI_Datatype sendtype, void *recvbuf,
600 int recvcount, MPI_Datatype recvtype,
601 int root, MPI_Comm comm))
602 mpi_coll_scatter_description[scatter_id].coll;
604 int barrier_id = find_coll_description(mpi_coll_barrier_description,
605 sg_cfg_get_string("smpi/barrier"));
606 mpi_coll_barrier_fun = (int (*)(MPI_Comm comm))
607 mpi_coll_barrier_description[barrier_id].coll;
609 smpi_cpu_threshold = sg_cfg_get_double("smpi/cpu_threshold");
610 smpi_running_power = sg_cfg_get_double("smpi/running_power");
611 smpi_privatize_global_variables = sg_cfg_get_boolean("smpi/privatize_global_variables");
612 if (smpi_cpu_threshold < 0)
613 smpi_cpu_threshold = DBL_MAX;
617 int smpi_main(int (*realmain) (int argc, char *argv[]), int argc, char *argv[])
619 srand(SMPI_RAND_SEED);
621 if (getenv("SMPI_PRETEND_CC") != NULL) {
622 /* Hack to ensure that smpicc can pretend to be a simple
623 * compiler. Particularly handy to pass it to the configuration tools */
630 TRACE_global_init(&argc, argv);
632 TRACE_add_start_function(TRACE_smpi_alloc);
633 TRACE_add_end_function(TRACE_smpi_release);
636 SIMIX_global_init(&argc, argv);
640 // parse the platform file: get the host list
641 SIMIX_create_environment(argv[1]);
642 SIMIX_comm_set_copy_data_callback(&smpi_comm_copy_buffer_callback);
643 SIMIX_function_register_default(realmain);
644 SIMIX_launch_application(argv[2]);
648 smpi_check_options();
650 if(smpi_privatize_global_variables)
651 smpi_initialize_global_memory_segments();
653 /* Clean IO before the run */
657 if (MC_is_active()) {
658 MC_do_the_modelcheck_for_real();
663 xbt_os_walltimer_stop(global_timer);
664 if (sg_cfg_get_boolean("smpi/display_timing")){
665 double global_time = xbt_os_timer_elapsed(global_timer);
666 XBT_INFO("Simulated time: %g seconds. \n "
667 "The simulation took %g seconds (after parsing and platform setup)\n"
668 "%g seconds were actual computation of the application"
669 , SIMIX_get_clock(), global_time , smpi_total_benched_time);
671 if (smpi_total_benched_time/global_time>=0.75)
672 XBT_INFO("More than 75%% of the time was spent inside the application code.\n"
673 "You may want to use sampling functions or trace replay to reduce this.");
677 smpi_global_destroy();
686 // This function can be called from extern file, to initialize logs, options, and processes of smpi
687 // without the need of smpirun
692 smpi_check_options();
694 if (TRACE_is_enabled() && TRACE_is_configured()) {
698 if(smpi_privatize_global_variables)
699 smpi_initialize_global_memory_segments();
702 void SMPI_finalize(){
703 smpi_global_destroy();