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 "mc/mc_record.h"
11 #include "xbt/replay.h"
12 #include "surf/surf.h"
13 #include "simix/smx_private.h"
14 #include "simgrid/sg_config.h"
16 #include <float.h> /* DBL_MAX */
21 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_kernel, smpi,
22 "Logging specific to SMPI (kernel)");
24 typedef struct s_smpi_process_data {
29 smx_rdv_t mailbox_small;
30 xbt_mutex_t mailboxes_mutex;
35 void *data; /* user data */
38 int sampling; /* inside an SMPI_SAMPLE_ block? */
40 int replaying; /* is the process replaying a trace */
41 xbt_bar_t finalization_barrier;
42 } s_smpi_process_data_t;
44 static smpi_process_data_t *process_data = NULL;
45 int process_count = 0;
46 int* index_to_process_data = NULL;
47 extern double smpi_total_benched_time;
48 xbt_os_timer_t global_timer;
49 MPI_Comm MPI_COMM_WORLD = MPI_COMM_UNINITIALIZED;
50 int MPI_UNIVERSE_SIZE;
52 MPI_Errhandler *MPI_ERRORS_RETURN = NULL;
53 MPI_Errhandler *MPI_ERRORS_ARE_FATAL = NULL;
54 MPI_Errhandler *MPI_ERRHANDLER_NULL = NULL;
56 #define MAILBOX_NAME_MAXLEN (5 + sizeof(int) * 2 + 1)
58 static char *get_mailbox_name(char *str, int index)
60 snprintf(str, MAILBOX_NAME_MAXLEN, "SMPI-%0*x", (int) (sizeof(int) * 2),
65 static char *get_mailbox_name_small(char *str, int index)
67 snprintf(str, MAILBOX_NAME_MAXLEN, "small%0*x", (int) (sizeof(int) * 2),
72 void smpi_process_init(int *argc, char ***argv)
75 smpi_process_data_t data;
79 proc = SIMIX_process_self();
80 //FIXME: dirty cleanup method to avoid using msg cleanup functions on these processes when using MSG+SMPI
81 proc->context->cleanup_func=SIMIX_process_cleanup;
82 char* instance_id = (*argv)[1];
83 int rank = atoi((*argv)[2]);
84 index = smpi_process_index_of_smx_process(proc);
86 if(!index_to_process_data){
87 index_to_process_data=(int*)xbt_malloc(SIMIX_process_count()*sizeof(int));
89 MPI_Comm* temp_comm_world;
91 smpi_deployment_register_process(instance_id, rank, index, &temp_comm_world ,&temp_bar);
92 data = smpi_process_remote_data(index);
93 data->comm_world = temp_comm_world;
94 if(temp_bar != NULL) data->finalization_barrier = temp_bar;
96 data->instance_id = instance_id;
98 xbt_free(simcall_process_get_data(proc));
99 simcall_process_set_data(proc, data);
102 memmove(&(*argv)[0], &(*argv)[2], sizeof(char *) * (*argc - 2));
103 (*argv)[(*argc) - 1] = NULL;
104 (*argv)[(*argc) - 2] = NULL;
109 // set the process attached to the mailbox
110 simcall_rdv_set_receiver(data->mailbox_small, proc);
111 XBT_DEBUG("<%d> New process in the game: %p", index, proc);
113 if(smpi_privatize_global_variables){
114 smpi_switch_data_segment(index);
118 if (smpi_process_data() == NULL)
119 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.");
122 void smpi_process_destroy(void)
124 int index = smpi_process_index();
125 if(smpi_privatize_global_variables){
126 smpi_switch_data_segment(index);
128 process_data[index_to_process_data[index]]->state = SMPI_FINALIZED;
129 XBT_DEBUG("<%d> Process left the game", index);
133 * @brief Prepares the current process for termination.
135 void smpi_process_finalize(void)
137 // This leads to an explosion of the search graph
138 // which cannot be reduced:
139 if(MC_is_active() || MC_record_replay_is_active())
142 int index = smpi_process_index();
143 // wait for all pending asynchronous comms to finish
144 xbt_barrier_wait(process_data[index_to_process_data[index]]->finalization_barrier);
148 * @brief Check if a process is finalized
150 int smpi_process_finalized()
152 int index = smpi_process_index();
153 if (index != MPI_UNDEFINED)
154 return (process_data[index_to_process_data[index]]->state == SMPI_FINALIZED);
160 * @brief Check if a process is initialized
162 int smpi_process_initialized(void)
164 int index = smpi_process_index();
165 return ( (index != MPI_UNDEFINED)
166 && (process_data[index_to_process_data[index]]->state == SMPI_INITIALIZED));
170 * @brief Mark a process as initialized (=MPI_Init called)
172 void smpi_process_mark_as_initialized(void)
174 int index = smpi_process_index();
175 if ((index != MPI_UNDEFINED) && (process_data[index_to_process_data[index]]->state != SMPI_FINALIZED))
176 process_data[index_to_process_data[index]]->state = SMPI_INITIALIZED;
179 void smpi_process_set_replaying(int value){
180 int index = smpi_process_index();
181 if ((index != MPI_UNDEFINED) && (process_data[index_to_process_data[index]]->state != SMPI_FINALIZED))
182 process_data[index_to_process_data[index]]->replaying = value;
185 int smpi_process_get_replaying(){
186 int index = smpi_process_index();
187 if (index != MPI_UNDEFINED)
188 return process_data[index_to_process_data[index]]->replaying;
189 else return _xbt_replay_is_active();
193 int smpi_global_size(void)
195 char *value = getenv("SMPI_GLOBAL_SIZE");
199 "Please set env var SMPI_GLOBAL_SIZE to expected number of processes.\n");
205 smpi_process_data_t smpi_process_data(void)
207 return SIMIX_process_self_get_data(SIMIX_process_self());
210 smpi_process_data_t smpi_process_remote_data(int index)
212 return process_data[index_to_process_data[index]];
215 void smpi_process_set_user_data(void *data)
217 smpi_process_data_t process_data = smpi_process_data();
218 process_data->data = data;
221 void *smpi_process_get_user_data()
223 smpi_process_data_t process_data = smpi_process_data();
224 return process_data->data;
227 int smpi_process_count(void)
229 return process_count;
232 int smpi_process_index(void)
234 smpi_process_data_t data = smpi_process_data();
235 //return -1 if not initialized
236 return data ? data->index : MPI_UNDEFINED;
239 MPI_Comm smpi_process_comm_world(void)
241 smpi_process_data_t data = smpi_process_data();
242 //return MPI_COMM_NULL if not initialized
243 return data ? *data->comm_world : MPI_COMM_NULL;
246 smx_rdv_t smpi_process_mailbox(void)
248 smpi_process_data_t data = smpi_process_data();
249 return data->mailbox;
252 smx_rdv_t smpi_process_mailbox_small(void)
254 smpi_process_data_t data = smpi_process_data();
255 return data->mailbox_small;
258 xbt_mutex_t smpi_process_mailboxes_mutex(void)
260 smpi_process_data_t data = smpi_process_data();
261 return data->mailboxes_mutex;
264 smx_rdv_t smpi_process_remote_mailbox(int index)
266 smpi_process_data_t data = smpi_process_remote_data(index);
267 return data->mailbox;
271 smx_rdv_t smpi_process_remote_mailbox_small(int index)
273 smpi_process_data_t data = smpi_process_remote_data(index);
274 return data->mailbox_small;
277 xbt_mutex_t smpi_process_remote_mailboxes_mutex(int index)
279 smpi_process_data_t data = smpi_process_remote_data(index);
280 return data->mailboxes_mutex;
283 xbt_os_timer_t smpi_process_timer(void)
285 smpi_process_data_t data = smpi_process_data();
289 void smpi_process_simulated_start(void)
291 smpi_process_data_t data = smpi_process_data();
292 data->simulated = SIMIX_get_clock();
295 double smpi_process_simulated_elapsed(void)
297 smpi_process_data_t data = smpi_process_data();
298 return SIMIX_get_clock() - data->simulated;
301 MPI_Comm smpi_process_comm_self(void)
303 smpi_process_data_t data = smpi_process_data();
304 if(data->comm_self==MPI_COMM_NULL){
305 MPI_Group group = smpi_group_new(1);
306 data->comm_self = smpi_comm_new(group, NULL);
307 smpi_group_set_mapping(group, smpi_process_index(), 0);
310 return data->comm_self;
313 MPI_Comm smpi_process_get_comm_intra(void)
315 smpi_process_data_t data = smpi_process_data();
316 return data->comm_intra;
319 void smpi_process_set_comm_intra(MPI_Comm comm)
321 smpi_process_data_t data = smpi_process_data();
322 data->comm_intra = comm;
325 void smpi_process_set_sampling(int s)
327 smpi_process_data_t data = smpi_process_data();
331 int smpi_process_get_sampling(void)
333 smpi_process_data_t data = smpi_process_data();
334 return data->sampling;
338 void print_request(const char *message, MPI_Request request)
341 ("%s request %p [buf = %p, size = %zu, src = %d, dst = %d, tag = %d, flags = %x]",
342 message, request, request->buf, request->size, request->src,
343 request->dst, request->tag, request->flags);
346 void smpi_comm_copy_buffer_callback(smx_synchro_t comm,
347 void *buff, size_t buff_size)
349 XBT_DEBUG("Copy the data over");
352 if((smpi_privatize_global_variables)
353 && ((char*)buff >= smpi_start_data_exe)
354 && ((char*)buff < smpi_start_data_exe + smpi_size_data_exe )
356 XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !");
357 smpi_switch_data_segment(((smpi_process_data_t)SIMIX_process_get_data(comm->comm.src_proc))->index);
358 tmpbuff = (void*)xbt_malloc(buff_size);
359 memcpy(tmpbuff, buff, buff_size);
363 if((smpi_privatize_global_variables)
364 && ((char*)comm->comm.dst_buff >= smpi_start_data_exe)
365 && ((char*)comm->comm.dst_buff < smpi_start_data_exe + smpi_size_data_exe )
367 XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment");
368 smpi_switch_data_segment(((smpi_process_data_t)SIMIX_process_get_data(comm->comm.dst_proc))->index);
372 memcpy(comm->comm.dst_buff, tmpbuff, buff_size);
373 if (comm->comm.detached) {
374 // if this is a detached send, the source buffer was duplicated by SMPI
375 // sender to make the original buffer available to the application ASAP
377 //It seems that the request is used after the call there this should
378 //be free somewhereelse but where???
379 //xbt_free(comm->comm.src_data);// inside SMPI the request is keep
380 //inside the user data and should be free
381 comm->comm.src_buff = NULL;
384 if(tmpbuff!=buff)xbt_free(tmpbuff);
389 void smpi_comm_null_copy_buffer_callback(smx_synchro_t comm,
390 void *buff, size_t buff_size)
395 static void smpi_check_options(){
396 //check correctness of MPI parameters
398 xbt_assert(sg_cfg_get_int("smpi/async_small_thres") <=
399 sg_cfg_get_int("smpi/send_is_detached_thres"));
401 if (sg_cfg_is_default_value("smpi/running_power")) {
402 XBT_INFO("You did not set the power of the host running the simulation. "
403 "The timings will certainly not be accurate. "
404 "Use the option \"--cfg=smpi/running_power:<flops>\" to set its value."
405 "Check http://simgrid.org/simgrid/latest/doc/options.html#options_smpi_bench for more information.");
409 int smpi_enabled(void) {
410 return process_data != NULL;
413 void smpi_global_init(void)
417 char name[MAILBOX_NAME_MAXLEN];
420 if (!MC_is_active()) {
421 global_timer = xbt_os_timer_new();
422 xbt_os_walltimer_start(global_timer);
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]->mailboxes_mutex=xbt_mutex_init();
438 process_data[i]->timer = xbt_os_timer_new();
440 MC_ignore_heap(process_data[i]->timer, xbt_os_timer_size());
441 process_data[i]->comm_self = MPI_COMM_NULL;
442 process_data[i]->comm_intra = MPI_COMM_NULL;
443 process_data[i]->comm_world = NULL;
444 process_data[i]->state = SMPI_UNINITIALIZED;
445 process_data[i]->sampling = 0;
446 process_data[i]->finalization_barrier = NULL;
448 //if the process was launched through smpirun script
449 //we generate a global mpi_comm_world
450 //if not, we let MPI_COMM_NULL, and the comm world
451 //will be private to each mpi instance
453 group = smpi_group_new(process_count);
454 MPI_COMM_WORLD = smpi_comm_new(group, NULL);
455 xbt_bar_t bar=xbt_barrier_init(process_count);
457 MPI_UNIVERSE_SIZE = smpi_comm_size(MPI_COMM_WORLD);
458 for (i = 0; i < process_count; i++) {
459 smpi_group_set_mapping(group, i, i);
460 process_data[i]->finalization_barrier = bar;
465 void smpi_global_destroy(void)
467 int count = smpi_process_count();
470 smpi_bench_destroy();
471 if (MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED){
472 while (smpi_group_unuse(smpi_comm_group(MPI_COMM_WORLD)) > 0);
473 xbt_free(MPI_COMM_WORLD);
474 xbt_barrier_destroy(process_data[0]->finalization_barrier);
476 smpi_deployment_cleanup_instances();
478 MPI_COMM_WORLD = MPI_COMM_NULL;
479 for (i = 0; i < count; i++) {
480 if(process_data[i]->comm_self!=MPI_COMM_NULL){
481 smpi_group_unuse(smpi_comm_group(process_data[i]->comm_self));
482 smpi_comm_destroy(process_data[i]->comm_self);
484 if(process_data[i]->comm_intra!=MPI_COMM_NULL){
485 smpi_group_unuse(smpi_comm_group(process_data[i]->comm_intra));
486 smpi_comm_destroy(process_data[i]->comm_intra);
488 xbt_os_timer_free(process_data[i]->timer);
489 simcall_rdv_destroy(process_data[i]->mailbox);
490 simcall_rdv_destroy(process_data[i]->mailbox_small);
491 xbt_mutex_destroy(process_data[i]->mailboxes_mutex);
492 xbt_free(process_data[i]);
494 xbt_free(process_data);
497 xbt_free(index_to_process_data);
498 if(smpi_privatize_global_variables)
499 smpi_destroy_global_memory_segments();
504 void __attribute__ ((weak)) user_main_()
506 xbt_die("Should not be in this smpi_simulated_main");
510 int __attribute__ ((weak)) smpi_simulated_main_(int argc, char **argv)
512 smpi_process_init(&argc, &argv);
517 int __attribute__ ((weak)) main(int argc, char **argv)
519 return smpi_main(smpi_simulated_main_, argc, argv);
524 static void smpi_init_logs(){
526 /* Connect log categories. See xbt/log.c */
527 XBT_LOG_CONNECT(smpi); /* Keep this line as soon as possible in this
528 function: xbt_log_appender_file.c depends on it
529 DO NOT connect this in XBT or so, or it will be
530 useless to xbt_log_appender_file.c */
531 XBT_LOG_CONNECT(instr_smpi);
532 XBT_LOG_CONNECT(smpi_base);
533 XBT_LOG_CONNECT(smpi_bench);
534 XBT_LOG_CONNECT(smpi_coll);
535 XBT_LOG_CONNECT(smpi_colls);
536 XBT_LOG_CONNECT(smpi_comm);
537 XBT_LOG_CONNECT(smpi_dvfs);
538 XBT_LOG_CONNECT(smpi_group);
539 XBT_LOG_CONNECT(smpi_kernel);
540 XBT_LOG_CONNECT(smpi_mpi);
541 XBT_LOG_CONNECT(smpi_mpi_dt);
542 XBT_LOG_CONNECT(smpi_pmpi);
543 XBT_LOG_CONNECT(smpi_replay);
544 XBT_LOG_CONNECT(smpi_rma);
549 static void smpi_init_options(){
550 int gather_id = find_coll_description(mpi_coll_gather_description,
551 sg_cfg_get_string("smpi/gather"),"gather");
552 mpi_coll_gather_fun = (int (*)(void *, int, MPI_Datatype,
553 void *, int, MPI_Datatype, int, MPI_Comm))
554 mpi_coll_gather_description[gather_id].coll;
556 int allgather_id = find_coll_description(mpi_coll_allgather_description,
557 sg_cfg_get_string("smpi/allgather"),"allgather");
558 mpi_coll_allgather_fun = (int (*)(void *, int, MPI_Datatype,
559 void *, int, MPI_Datatype, MPI_Comm))
560 mpi_coll_allgather_description[allgather_id].coll;
562 int allgatherv_id = find_coll_description(mpi_coll_allgatherv_description,
563 sg_cfg_get_string("smpi/allgatherv"),"allgatherv");
564 mpi_coll_allgatherv_fun = (int (*)(void *, int, MPI_Datatype, void *, int *,
565 int *, MPI_Datatype, MPI_Comm))
566 mpi_coll_allgatherv_description[allgatherv_id].coll;
568 int allreduce_id = find_coll_description(mpi_coll_allreduce_description,
569 sg_cfg_get_string("smpi/allreduce"),"allreduce");
570 mpi_coll_allreduce_fun = (int (*)(void *sbuf, void *rbuf, int rcount,
571 MPI_Datatype dtype, MPI_Op op,
573 mpi_coll_allreduce_description[allreduce_id].coll;
575 int alltoall_id = find_coll_description(mpi_coll_alltoall_description,
576 sg_cfg_get_string("smpi/alltoall"),"alltoall");
577 mpi_coll_alltoall_fun = (int (*)(void *, int, MPI_Datatype,
578 void *, int, MPI_Datatype, MPI_Comm))
579 mpi_coll_alltoall_description[alltoall_id].coll;
581 int alltoallv_id = find_coll_description(mpi_coll_alltoallv_description,
582 sg_cfg_get_string("smpi/alltoallv"),"alltoallv");
583 mpi_coll_alltoallv_fun = (int (*)(void *, int *, int *, MPI_Datatype,
584 void *, int *, int *, MPI_Datatype,
586 mpi_coll_alltoallv_description[alltoallv_id].coll;
588 int bcast_id = find_coll_description(mpi_coll_bcast_description,
589 sg_cfg_get_string("smpi/bcast"),"bcast");
590 mpi_coll_bcast_fun = (int (*)(void *buf, int count, MPI_Datatype datatype,
591 int root, MPI_Comm com))
592 mpi_coll_bcast_description[bcast_id].coll;
594 int reduce_id = find_coll_description(mpi_coll_reduce_description,
595 sg_cfg_get_string("smpi/reduce"),"reduce");
596 mpi_coll_reduce_fun = (int (*)(void *buf, void *rbuf, int count,
597 MPI_Datatype datatype, MPI_Op op,
598 int root, MPI_Comm comm))
599 mpi_coll_reduce_description[reduce_id].coll;
601 int reduce_scatter_id =
602 find_coll_description(mpi_coll_reduce_scatter_description,
603 sg_cfg_get_string("smpi/reduce_scatter"),"reduce_scatter");
604 mpi_coll_reduce_scatter_fun = (int (*)(void *sbuf, void *rbuf, int *rcounts,
605 MPI_Datatype dtype, MPI_Op op,
607 mpi_coll_reduce_scatter_description[reduce_scatter_id].coll;
609 int scatter_id = find_coll_description(mpi_coll_scatter_description,
610 sg_cfg_get_string("smpi/scatter"),"scatter");
611 mpi_coll_scatter_fun = (int (*)(void *sendbuf, int sendcount,
612 MPI_Datatype sendtype, void *recvbuf,
613 int recvcount, MPI_Datatype recvtype,
614 int root, MPI_Comm comm))
615 mpi_coll_scatter_description[scatter_id].coll;
617 int barrier_id = find_coll_description(mpi_coll_barrier_description,
618 sg_cfg_get_string("smpi/barrier"),"barrier");
619 mpi_coll_barrier_fun = (int (*)(MPI_Comm comm))
620 mpi_coll_barrier_description[barrier_id].coll;
622 smpi_cpu_threshold = sg_cfg_get_double("smpi/cpu_threshold");
623 smpi_running_power = sg_cfg_get_double("smpi/running_power");
624 smpi_privatize_global_variables = sg_cfg_get_boolean("smpi/privatize_global_variables");
625 if (smpi_cpu_threshold < 0)
626 smpi_cpu_threshold = DBL_MAX;
630 int smpi_main(int (*realmain) (int argc, char *argv[]), int argc, char *argv[])
632 srand(SMPI_RAND_SEED);
634 if (getenv("SMPI_PRETEND_CC") != NULL) {
635 /* Hack to ensure that smpicc can pretend to be a simple
636 * compiler. Particularly handy to pass it to the configuration tools */
642 TRACE_global_init(&argc, argv);
644 TRACE_add_start_function(TRACE_smpi_alloc);
645 TRACE_add_end_function(TRACE_smpi_release);
647 SIMIX_global_init(&argc, argv);
651 // parse the platform file: get the host list
652 SIMIX_create_environment(argv[1]);
653 SIMIX_comm_set_copy_data_callback(&smpi_comm_copy_buffer_callback);
654 SIMIX_function_register_default(realmain);
655 SIMIX_launch_application(argv[2]);
659 smpi_check_options();
661 if(smpi_privatize_global_variables)
662 smpi_initialize_global_memory_segments();
664 /* Clean IO before the run */
668 if (MC_is_active()) {
669 MC_do_the_modelcheck_for_real();
674 xbt_os_walltimer_stop(global_timer);
675 if (sg_cfg_get_boolean("smpi/display_timing")){
676 double global_time = xbt_os_timer_elapsed(global_timer);
677 XBT_INFO("Simulated time: %g seconds. \n "
678 "The simulation took %g seconds (after parsing and platform setup)\n"
679 "%g seconds were actual computation of the application"
680 , SIMIX_get_clock(), global_time , smpi_total_benched_time);
682 if (smpi_total_benched_time/global_time>=0.75)
683 XBT_INFO("More than 75%% of the time was spent inside the application code.\n"
684 "You may want to use sampling functions or trace replay to reduce this.");
688 smpi_global_destroy();
695 // This function can be called from extern file, to initialize logs, options, and processes of smpi
696 // without the need of smpirun
701 smpi_check_options();
702 if (TRACE_is_enabled() && TRACE_is_configured())
704 if(smpi_privatize_global_variables)
705 smpi_initialize_global_memory_segments();
708 void SMPI_finalize(){
709 smpi_global_destroy();