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"
15 #include "mc/mc_replay.h"
17 #include <float.h> /* DBL_MAX */
22 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_kernel, smpi,
23 "Logging specific to SMPI (kernel)");
25 typedef struct s_smpi_process_data {
30 smx_rdv_t mailbox_small;
31 xbt_mutex_t mailboxes_mutex;
36 void *data; /* user data */
39 int sampling; /* inside an SMPI_SAMPLE_ block? */
41 int replaying; /* is the process replaying a trace */
42 xbt_bar_t finalization_barrier;
43 } s_smpi_process_data_t;
45 static smpi_process_data_t *process_data = NULL;
46 int process_count = 0;
47 int smpi_universe_size = 0;
48 int* index_to_process_data = NULL;
49 extern double smpi_total_benched_time;
50 xbt_os_timer_t global_timer;
51 MPI_Comm MPI_COMM_WORLD = MPI_COMM_UNINITIALIZED;
53 MPI_Errhandler *MPI_ERRORS_RETURN = NULL;
54 MPI_Errhandler *MPI_ERRORS_ARE_FATAL = NULL;
55 MPI_Errhandler *MPI_ERRHANDLER_NULL = NULL;
57 #define MAILBOX_NAME_MAXLEN (5 + sizeof(int) * 2 + 1)
59 static char *get_mailbox_name(char *str, int index)
61 snprintf(str, MAILBOX_NAME_MAXLEN, "SMPI-%0*x", (int) (sizeof(int) * 2),
66 static char *get_mailbox_name_small(char *str, int index)
68 snprintf(str, MAILBOX_NAME_MAXLEN, "small%0*x", (int) (sizeof(int) * 2),
73 void smpi_process_init(int *argc, char ***argv)
76 smpi_process_data_t data;
80 proc = SIMIX_process_self();
81 //FIXME: dirty cleanup method to avoid using msg cleanup functions on these processes when using MSG+SMPI
82 proc->context->cleanup_func=SIMIX_process_cleanup;
83 char* instance_id = (*argv)[1];
84 int rank = atoi((*argv)[2]);
85 index = smpi_process_index_of_smx_process(proc);
87 if(!index_to_process_data){
88 index_to_process_data=(int*)xbt_malloc(SIMIX_process_count()*sizeof(int));
90 MPI_Comm* temp_comm_world;
92 smpi_deployment_register_process(instance_id, rank, index, &temp_comm_world ,&temp_bar);
93 data = smpi_process_remote_data(index);
94 data->comm_world = temp_comm_world;
95 if(temp_bar != NULL) data->finalization_barrier = temp_bar;
97 data->instance_id = instance_id;
99 xbt_free(simcall_process_get_data(proc));
100 simcall_process_set_data(proc, data);
103 memmove(&(*argv)[0], &(*argv)[2], sizeof(char *) * (*argc - 2));
104 (*argv)[(*argc) - 1] = NULL;
105 (*argv)[(*argc) - 2] = NULL;
110 // set the process attached to the mailbox
111 simcall_rdv_set_receiver(data->mailbox_small, proc);
112 XBT_DEBUG("<%d> New process in the game: %p", index, proc);
114 if(smpi_privatize_global_variables){
115 smpi_switch_data_segment(index);
119 if (smpi_process_data() == NULL)
120 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.");
123 void smpi_process_destroy(void)
125 int index = smpi_process_index();
126 if(smpi_privatize_global_variables){
127 smpi_switch_data_segment(index);
129 process_data[index_to_process_data[index]]->state = SMPI_FINALIZED;
130 XBT_DEBUG("<%d> Process left the game", index);
134 * @brief Prepares the current process for termination.
136 void smpi_process_finalize(void)
138 // This leads to an explosion of the search graph
139 // which cannot be reduced:
140 if(MC_is_active() || MC_record_replay_is_active())
143 int index = smpi_process_index();
144 // wait for all pending asynchronous comms to finish
145 xbt_barrier_wait(process_data[index_to_process_data[index]]->finalization_barrier);
149 * @brief Check if a process is finalized
151 int smpi_process_finalized()
153 int index = smpi_process_index();
154 if (index != MPI_UNDEFINED)
155 return (process_data[index_to_process_data[index]]->state == SMPI_FINALIZED);
161 * @brief Check if a process is initialized
163 int smpi_process_initialized(void)
165 int index = smpi_process_index();
166 return ( (index != MPI_UNDEFINED)
167 && (process_data[index_to_process_data[index]]->state == SMPI_INITIALIZED));
171 * @brief Mark a process as initialized (=MPI_Init called)
173 void smpi_process_mark_as_initialized(void)
175 int index = smpi_process_index();
176 if ((index != MPI_UNDEFINED) && (process_data[index_to_process_data[index]]->state != SMPI_FINALIZED))
177 process_data[index_to_process_data[index]]->state = SMPI_INITIALIZED;
180 void smpi_process_set_replaying(int value){
181 int index = smpi_process_index();
182 if ((index != MPI_UNDEFINED) && (process_data[index_to_process_data[index]]->state != SMPI_FINALIZED))
183 process_data[index_to_process_data[index]]->replaying = value;
186 int smpi_process_get_replaying(){
187 int index = smpi_process_index();
188 if (index != MPI_UNDEFINED)
189 return process_data[index_to_process_data[index]]->replaying;
190 else return _xbt_replay_is_active();
194 int smpi_global_size(void)
196 char *value = getenv("SMPI_GLOBAL_SIZE");
200 "Please set env var SMPI_GLOBAL_SIZE to expected number of processes.\n");
206 smpi_process_data_t smpi_process_data(void)
208 return SIMIX_process_self_get_data(SIMIX_process_self());
211 smpi_process_data_t smpi_process_remote_data(int index)
213 return process_data[index_to_process_data[index]];
216 void smpi_process_set_user_data(void *data)
218 smpi_process_data_t process_data = smpi_process_data();
219 process_data->data = data;
222 void *smpi_process_get_user_data()
224 smpi_process_data_t process_data = smpi_process_data();
225 return process_data->data;
228 int smpi_process_count(void)
230 return process_count;
233 int smpi_process_index(void)
235 smpi_process_data_t data = smpi_process_data();
236 //return -1 if not initialized
237 return data ? data->index : MPI_UNDEFINED;
240 MPI_Comm smpi_process_comm_world(void)
242 smpi_process_data_t data = smpi_process_data();
243 //return MPI_COMM_NULL if not initialized
244 return data ? *data->comm_world : MPI_COMM_NULL;
247 smx_rdv_t smpi_process_mailbox(void)
249 smpi_process_data_t data = smpi_process_data();
250 return data->mailbox;
253 smx_rdv_t smpi_process_mailbox_small(void)
255 smpi_process_data_t data = smpi_process_data();
256 return data->mailbox_small;
259 xbt_mutex_t smpi_process_mailboxes_mutex(void)
261 smpi_process_data_t data = smpi_process_data();
262 return data->mailboxes_mutex;
265 smx_rdv_t smpi_process_remote_mailbox(int index)
267 smpi_process_data_t data = smpi_process_remote_data(index);
268 return data->mailbox;
272 smx_rdv_t smpi_process_remote_mailbox_small(int index)
274 smpi_process_data_t data = smpi_process_remote_data(index);
275 return data->mailbox_small;
278 xbt_mutex_t smpi_process_remote_mailboxes_mutex(int index)
280 smpi_process_data_t data = smpi_process_remote_data(index);
281 return data->mailboxes_mutex;
284 xbt_os_timer_t smpi_process_timer(void)
286 smpi_process_data_t data = smpi_process_data();
290 void smpi_process_simulated_start(void)
292 smpi_process_data_t data = smpi_process_data();
293 data->simulated = SIMIX_get_clock();
296 double smpi_process_simulated_elapsed(void)
298 smpi_process_data_t data = smpi_process_data();
299 return SIMIX_get_clock() - data->simulated;
302 MPI_Comm smpi_process_comm_self(void)
304 smpi_process_data_t data = smpi_process_data();
305 if(data->comm_self==MPI_COMM_NULL){
306 MPI_Group group = smpi_group_new(1);
307 data->comm_self = smpi_comm_new(group, NULL);
308 smpi_group_set_mapping(group, smpi_process_index(), 0);
311 return data->comm_self;
314 MPI_Comm smpi_process_get_comm_intra(void)
316 smpi_process_data_t data = smpi_process_data();
317 return data->comm_intra;
320 void smpi_process_set_comm_intra(MPI_Comm comm)
322 smpi_process_data_t data = smpi_process_data();
323 data->comm_intra = comm;
326 void smpi_process_set_sampling(int s)
328 smpi_process_data_t data = smpi_process_data();
332 int smpi_process_get_sampling(void)
334 smpi_process_data_t data = smpi_process_data();
335 return data->sampling;
339 void print_request(const char *message, MPI_Request request)
342 ("%s request %p [buf = %p, size = %zu, src = %d, dst = %d, tag = %d, flags = %x]",
343 message, request, request->buf, request->size, request->src,
344 request->dst, request->tag, request->flags);
347 void smpi_comm_copy_buffer_callback(smx_synchro_t comm,
348 void *buff, size_t buff_size)
350 XBT_DEBUG("Copy the data over");
353 if((smpi_privatize_global_variables)
354 && ((char*)buff >= smpi_start_data_exe)
355 && ((char*)buff < smpi_start_data_exe + smpi_size_data_exe )
357 XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !");
358 smpi_switch_data_segment(((smpi_process_data_t)SIMIX_process_get_data(comm->comm.src_proc))->index);
359 tmpbuff = (void*)xbt_malloc(buff_size);
360 memcpy(tmpbuff, buff, buff_size);
364 if((smpi_privatize_global_variables)
365 && ((char*)comm->comm.dst_buff >= smpi_start_data_exe)
366 && ((char*)comm->comm.dst_buff < smpi_start_data_exe + smpi_size_data_exe )
368 XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment");
369 smpi_switch_data_segment(((smpi_process_data_t)SIMIX_process_get_data(comm->comm.dst_proc))->index);
373 memcpy(comm->comm.dst_buff, tmpbuff, buff_size);
374 if (comm->comm.detached) {
375 // if this is a detached send, the source buffer was duplicated by SMPI
376 // sender to make the original buffer available to the application ASAP
378 //It seems that the request is used after the call there this should
379 //be free somewhereelse but where???
380 //xbt_free(comm->comm.src_data);// inside SMPI the request is keep
381 //inside the user data and should be free
382 comm->comm.src_buff = NULL;
385 if(tmpbuff!=buff)xbt_free(tmpbuff);
390 void smpi_comm_null_copy_buffer_callback(smx_synchro_t comm,
391 void *buff, size_t buff_size)
396 static void smpi_check_options(){
397 //check correctness of MPI parameters
399 xbt_assert(sg_cfg_get_int("smpi/async_small_thres") <=
400 sg_cfg_get_int("smpi/send_is_detached_thres"));
402 if (sg_cfg_is_default_value("smpi/running_power")) {
403 XBT_INFO("You did not set the power of the host running the simulation. "
404 "The timings will certainly not be accurate. "
405 "Use the option \"--cfg=smpi/running_power:<flops>\" to set its value."
406 "Check http://simgrid.org/simgrid/latest/doc/options.html#options_smpi_bench for more information.");
410 int smpi_enabled(void) {
411 return process_data != NULL;
414 void smpi_global_init(void)
418 char name[MAILBOX_NAME_MAXLEN];
421 if (!MC_is_active()) {
422 global_timer = xbt_os_timer_new();
423 xbt_os_walltimer_start(global_timer);
425 if (process_count == 0){
426 process_count = SIMIX_process_count();
429 smpi_universe_size = process_count;
430 process_data = xbt_new0(smpi_process_data_t, process_count);
431 for (i = 0; i < process_count; i++) {
432 process_data[i] = xbt_new(s_smpi_process_data_t, 1);
433 //process_data[i]->index = i;
434 process_data[i]->argc = NULL;
435 process_data[i]->argv = NULL;
436 process_data[i]->mailbox = simcall_rdv_create(get_mailbox_name(name, i));
437 process_data[i]->mailbox_small =
438 simcall_rdv_create(get_mailbox_name_small(name, i));
439 process_data[i]->mailboxes_mutex=xbt_mutex_init();
440 process_data[i]->timer = xbt_os_timer_new();
442 MC_ignore_heap(process_data[i]->timer, xbt_os_timer_size());
443 process_data[i]->comm_self = MPI_COMM_NULL;
444 process_data[i]->comm_intra = MPI_COMM_NULL;
445 process_data[i]->comm_world = NULL;
446 process_data[i]->state = SMPI_UNINITIALIZED;
447 process_data[i]->sampling = 0;
448 process_data[i]->finalization_barrier = NULL;
450 //if the process was launched through smpirun script
451 //we generate a global mpi_comm_world
452 //if not, we let MPI_COMM_NULL, and the comm world
453 //will be private to each mpi instance
455 group = smpi_group_new(process_count);
456 MPI_COMM_WORLD = smpi_comm_new(group, NULL);
457 MPI_Attr_put(MPI_COMM_WORLD, MPI_UNIVERSE_SIZE, (void *)(MPI_Aint)process_count);
458 xbt_bar_t bar=xbt_barrier_init(process_count);
460 for (i = 0; i < process_count; i++) {
461 smpi_group_set_mapping(group, i, i);
462 process_data[i]->finalization_barrier = bar;
467 void smpi_global_destroy(void)
469 int count = smpi_process_count();
472 smpi_bench_destroy();
473 if (MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED){
474 while (smpi_group_unuse(smpi_comm_group(MPI_COMM_WORLD)) > 0);
475 xbt_free(MPI_COMM_WORLD);
476 xbt_barrier_destroy(process_data[0]->finalization_barrier);
478 smpi_deployment_cleanup_instances();
480 MPI_COMM_WORLD = MPI_COMM_NULL;
481 for (i = 0; i < count; i++) {
482 if(process_data[i]->comm_self!=MPI_COMM_NULL){
483 smpi_group_unuse(smpi_comm_group(process_data[i]->comm_self));
484 smpi_comm_destroy(process_data[i]->comm_self);
486 if(process_data[i]->comm_intra!=MPI_COMM_NULL){
487 smpi_group_unuse(smpi_comm_group(process_data[i]->comm_intra));
488 smpi_comm_destroy(process_data[i]->comm_intra);
490 xbt_os_timer_free(process_data[i]->timer);
491 simcall_rdv_destroy(process_data[i]->mailbox);
492 simcall_rdv_destroy(process_data[i]->mailbox_small);
493 xbt_mutex_destroy(process_data[i]->mailboxes_mutex);
494 xbt_free(process_data[i]);
496 xbt_free(process_data);
499 xbt_free(index_to_process_data);
500 if(smpi_privatize_global_variables)
501 smpi_destroy_global_memory_segments();
506 void __attribute__ ((weak)) user_main_()
508 xbt_die("Should not be in this smpi_simulated_main");
512 int __attribute__ ((weak)) smpi_simulated_main_(int argc, char **argv)
514 smpi_process_init(&argc, &argv);
519 int __attribute__ ((weak)) main(int argc, char **argv)
521 return smpi_main(smpi_simulated_main_, argc, argv);
526 static void smpi_init_logs(){
528 /* Connect log categories. See xbt/log.c */
529 XBT_LOG_CONNECT(smpi); /* Keep this line as soon as possible in this
530 function: xbt_log_appender_file.c depends on it
531 DO NOT connect this in XBT or so, or it will be
532 useless to xbt_log_appender_file.c */
533 XBT_LOG_CONNECT(instr_smpi);
534 XBT_LOG_CONNECT(smpi_base);
535 XBT_LOG_CONNECT(smpi_bench);
536 XBT_LOG_CONNECT(smpi_coll);
537 XBT_LOG_CONNECT(smpi_colls);
538 XBT_LOG_CONNECT(smpi_comm);
539 XBT_LOG_CONNECT(smpi_dvfs);
540 XBT_LOG_CONNECT(smpi_group);
541 XBT_LOG_CONNECT(smpi_kernel);
542 XBT_LOG_CONNECT(smpi_mpi);
543 XBT_LOG_CONNECT(smpi_mpi_dt);
544 XBT_LOG_CONNECT(smpi_pmpi);
545 XBT_LOG_CONNECT(smpi_replay);
546 XBT_LOG_CONNECT(smpi_rma);
551 static void smpi_init_options(){
552 int gather_id = find_coll_description(mpi_coll_gather_description,
553 sg_cfg_get_string("smpi/gather"),"gather");
554 mpi_coll_gather_fun = (int (*)(void *, int, MPI_Datatype,
555 void *, int, MPI_Datatype, int, MPI_Comm))
556 mpi_coll_gather_description[gather_id].coll;
558 int allgather_id = find_coll_description(mpi_coll_allgather_description,
559 sg_cfg_get_string("smpi/allgather"),"allgather");
560 mpi_coll_allgather_fun = (int (*)(void *, int, MPI_Datatype,
561 void *, int, MPI_Datatype, MPI_Comm))
562 mpi_coll_allgather_description[allgather_id].coll;
564 int allgatherv_id = find_coll_description(mpi_coll_allgatherv_description,
565 sg_cfg_get_string("smpi/allgatherv"),"allgatherv");
566 mpi_coll_allgatherv_fun = (int (*)(void *, int, MPI_Datatype, void *, int *,
567 int *, MPI_Datatype, MPI_Comm))
568 mpi_coll_allgatherv_description[allgatherv_id].coll;
570 int allreduce_id = find_coll_description(mpi_coll_allreduce_description,
571 sg_cfg_get_string("smpi/allreduce"),"allreduce");
572 mpi_coll_allreduce_fun = (int (*)(void *sbuf, void *rbuf, int rcount,
573 MPI_Datatype dtype, MPI_Op op,
575 mpi_coll_allreduce_description[allreduce_id].coll;
577 int alltoall_id = find_coll_description(mpi_coll_alltoall_description,
578 sg_cfg_get_string("smpi/alltoall"),"alltoall");
579 mpi_coll_alltoall_fun = (int (*)(void *, int, MPI_Datatype,
580 void *, int, MPI_Datatype, MPI_Comm))
581 mpi_coll_alltoall_description[alltoall_id].coll;
583 int alltoallv_id = find_coll_description(mpi_coll_alltoallv_description,
584 sg_cfg_get_string("smpi/alltoallv"),"alltoallv");
585 mpi_coll_alltoallv_fun = (int (*)(void *, int *, int *, MPI_Datatype,
586 void *, int *, int *, MPI_Datatype,
588 mpi_coll_alltoallv_description[alltoallv_id].coll;
590 int bcast_id = find_coll_description(mpi_coll_bcast_description,
591 sg_cfg_get_string("smpi/bcast"),"bcast");
592 mpi_coll_bcast_fun = (int (*)(void *buf, int count, MPI_Datatype datatype,
593 int root, MPI_Comm com))
594 mpi_coll_bcast_description[bcast_id].coll;
596 int reduce_id = find_coll_description(mpi_coll_reduce_description,
597 sg_cfg_get_string("smpi/reduce"),"reduce");
598 mpi_coll_reduce_fun = (int (*)(void *buf, void *rbuf, int count,
599 MPI_Datatype datatype, MPI_Op op,
600 int root, MPI_Comm comm))
601 mpi_coll_reduce_description[reduce_id].coll;
603 int reduce_scatter_id =
604 find_coll_description(mpi_coll_reduce_scatter_description,
605 sg_cfg_get_string("smpi/reduce_scatter"),"reduce_scatter");
606 mpi_coll_reduce_scatter_fun = (int (*)(void *sbuf, void *rbuf, int *rcounts,
607 MPI_Datatype dtype, MPI_Op op,
609 mpi_coll_reduce_scatter_description[reduce_scatter_id].coll;
611 int scatter_id = find_coll_description(mpi_coll_scatter_description,
612 sg_cfg_get_string("smpi/scatter"),"scatter");
613 mpi_coll_scatter_fun = (int (*)(void *sendbuf, int sendcount,
614 MPI_Datatype sendtype, void *recvbuf,
615 int recvcount, MPI_Datatype recvtype,
616 int root, MPI_Comm comm))
617 mpi_coll_scatter_description[scatter_id].coll;
619 int barrier_id = find_coll_description(mpi_coll_barrier_description,
620 sg_cfg_get_string("smpi/barrier"),"barrier");
621 mpi_coll_barrier_fun = (int (*)(MPI_Comm comm))
622 mpi_coll_barrier_description[barrier_id].coll;
624 smpi_cpu_threshold = sg_cfg_get_double("smpi/cpu_threshold");
625 smpi_running_power = sg_cfg_get_double("smpi/running_power");
626 smpi_privatize_global_variables = sg_cfg_get_boolean("smpi/privatize_global_variables");
627 if (smpi_cpu_threshold < 0)
628 smpi_cpu_threshold = DBL_MAX;
632 int smpi_main(int (*realmain) (int argc, char *argv[]), int argc, char *argv[])
634 srand(SMPI_RAND_SEED);
636 if (getenv("SMPI_PRETEND_CC") != NULL) {
637 /* Hack to ensure that smpicc can pretend to be a simple
638 * compiler. Particularly handy to pass it to the configuration tools */
644 TRACE_global_init(&argc, argv);
646 TRACE_add_start_function(TRACE_smpi_alloc);
647 TRACE_add_end_function(TRACE_smpi_release);
649 SIMIX_global_init(&argc, argv);
653 // parse the platform file: get the host list
654 SIMIX_create_environment(argv[1]);
655 SIMIX_comm_set_copy_data_callback(&smpi_comm_copy_buffer_callback);
656 SIMIX_function_register_default(realmain);
657 SIMIX_launch_application(argv[2]);
661 smpi_check_options();
663 if(smpi_privatize_global_variables)
664 smpi_initialize_global_memory_segments();
666 /* Clean IO before the run */
670 if (MC_is_active()) {
676 xbt_os_walltimer_stop(global_timer);
677 if (sg_cfg_get_boolean("smpi/display_timing")){
678 double global_time = xbt_os_timer_elapsed(global_timer);
679 XBT_INFO("Simulated time: %g seconds. \n "
680 "The simulation took %g seconds (after parsing and platform setup)\n"
681 "%g seconds were actual computation of the application"
682 , SIMIX_get_clock(), global_time , smpi_total_benched_time);
684 if (smpi_total_benched_time/global_time>=0.75)
685 XBT_INFO("More than 75%% of the time was spent inside the application code.\n"
686 "You may want to use sampling functions or trace replay to reduce this.");
690 smpi_global_destroy();
697 // This function can be called from extern file, to initialize logs, options, and processes of smpi
698 // without the need of smpirun
703 smpi_check_options();
704 if (TRACE_is_enabled() && TRACE_is_configured())
706 if(smpi_privatize_global_variables)
707 smpi_initialize_global_memory_segments();
710 void SMPI_finalize(){
711 smpi_global_destroy();