1 /* Copyright (c) 2007-2015. 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 "src/mc/mc_record.h"
11 #include "xbt/replay.h"
12 #include "surf/surf.h"
13 #include "src/simix/smx_private.h"
14 #include "simgrid/sg_config.h"
15 #include "src/mc/mc_replay.h"
16 #include "src/msg/msg_private.h"
18 #include <float.h> /* DBL_MAX */
23 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_kernel, smpi,
24 "Logging specific to SMPI (kernel)");
26 typedef struct s_smpi_process_data {
31 smx_rdv_t mailbox_small;
32 xbt_mutex_t mailboxes_mutex;
37 void *data; /* user data */
40 int sampling; /* inside an SMPI_SAMPLE_ block? */
42 int replaying; /* is the process replaying a trace */
43 xbt_bar_t finalization_barrier;
44 } s_smpi_process_data_t;
46 static smpi_process_data_t *process_data = NULL;
47 int process_count = 0;
48 int smpi_universe_size = 0;
49 int* index_to_process_data = NULL;
50 extern double smpi_total_benched_time;
51 xbt_os_timer_t global_timer;
52 MPI_Comm MPI_COMM_WORLD = MPI_COMM_UNINITIALIZED;
54 MPI_Errhandler *MPI_ERRORS_RETURN = NULL;
55 MPI_Errhandler *MPI_ERRORS_ARE_FATAL = NULL;
56 MPI_Errhandler *MPI_ERRHANDLER_NULL = NULL;
58 #define MAILBOX_NAME_MAXLEN (5 + sizeof(int) * 2 + 1)
60 static char *get_mailbox_name(char *str, int index)
62 snprintf(str, MAILBOX_NAME_MAXLEN, "SMPI-%0*x", (int) (sizeof(int) * 2),
67 static char *get_mailbox_name_small(char *str, int index)
69 snprintf(str, MAILBOX_NAME_MAXLEN, "small%0*x", (int) (sizeof(int) * 2),
74 void smpi_process_init(int *argc, char ***argv)
77 smpi_process_data_t data;
81 proc = SIMIX_process_self();
82 //FIXME: dirty cleanup method to avoid using msg cleanup functions on these processes when using MSG+SMPI
83 SIMIX_process_set_cleanup_function(proc, SIMIX_process_cleanup);
84 char* instance_id = (*argv)[1];
85 int rank = atoi((*argv)[2]);
86 /* Now using segment index of the process */
87 index = proc->segment_index;
89 if(!index_to_process_data){
90 index_to_process_data=(int*)xbt_malloc(SIMIX_process_count()*sizeof(int));
93 if(smpi_privatize_global_variables){
94 /* Done at the process's creation */
95 SMPI_switch_data_segment(index);
98 MPI_Comm* temp_comm_world;
100 smpi_deployment_register_process(instance_id, rank, index, &temp_comm_world ,&temp_bar);
101 data = smpi_process_remote_data(index);
102 data->comm_world = temp_comm_world;
103 if(temp_bar != NULL) data->finalization_barrier = temp_bar;
105 data->instance_id = instance_id;
107 //xbt_free(simcall_process_get_data(proc));
109 simdata_process_t simdata = simcall_process_get_data(proc);
110 simdata->data = data;
114 memmove(&(*argv)[0], &(*argv)[2], sizeof(char *) * (*argc - 2));
115 (*argv)[(*argc) - 1] = NULL;
116 (*argv)[(*argc) - 2] = NULL;
121 // set the process attached to the mailbox
122 simcall_rdv_set_receiver(data->mailbox_small, proc);
123 XBT_DEBUG("<%d> New process in the game: %p", index, proc);
126 if (smpi_process_data() == NULL)
127 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.");
130 void smpi_process_destroy(void)
132 int index = smpi_process_index();
133 if(smpi_privatize_global_variables){
134 smpi_switch_data_segment(index);
136 process_data[index_to_process_data[index]]->state = SMPI_FINALIZED;
137 XBT_DEBUG("<%d> Process left the game", index);
141 * @brief Prepares the current process for termination.
143 void smpi_process_finalize(void)
145 // This leads to an explosion of the search graph
146 // which cannot be reduced:
147 if(MC_is_active() || MC_record_replay_is_active())
150 int index = smpi_process_index();
151 // wait for all pending asynchronous comms to finish
152 xbt_barrier_wait(process_data[index_to_process_data[index]]->finalization_barrier);
156 * @brief Check if a process is finalized
158 int smpi_process_finalized()
160 int index = smpi_process_index();
161 if (index != MPI_UNDEFINED)
162 return (process_data[index_to_process_data[index]]->state == SMPI_FINALIZED);
168 * @brief Check if a process is initialized
170 int smpi_process_initialized(void)
172 if (!index_to_process_data){
176 int index = smpi_process_index();
177 return ( (index != MPI_UNDEFINED)
178 && (process_data[index_to_process_data[index]]->state == SMPI_INITIALIZED));
183 * @brief Mark a process as initialized (=MPI_Init called)
185 void smpi_process_mark_as_initialized(void)
187 int index = smpi_process_index();
188 if ((index != MPI_UNDEFINED) && (process_data[index_to_process_data[index]]->state != SMPI_FINALIZED))
189 process_data[index_to_process_data[index]]->state = SMPI_INITIALIZED;
192 void smpi_process_set_replaying(int value){
193 int index = smpi_process_index();
194 if ((index != MPI_UNDEFINED) && (process_data[index_to_process_data[index]]->state != SMPI_FINALIZED))
195 process_data[index_to_process_data[index]]->replaying = value;
198 int smpi_process_get_replaying(){
199 int index = smpi_process_index();
200 if (index != MPI_UNDEFINED)
201 return process_data[index_to_process_data[index]]->replaying;
202 else return _xbt_replay_is_active();
206 int smpi_global_size(void)
208 char *value = getenv("SMPI_GLOBAL_SIZE");
212 "Please set env var SMPI_GLOBAL_SIZE to expected number of processes.\n");
218 smpi_process_data_t smpi_process_data(void)
220 simdata_process_t simdata = SIMIX_process_self_get_data(SIMIX_process_self());
221 return simdata->data;
224 smpi_process_data_t smpi_process_remote_data(int index)
226 return process_data[index_to_process_data[index]];
229 void smpi_process_set_user_data(void *data)
231 smpi_process_data_t process_data = smpi_process_data();
232 process_data->data = data;
235 void *smpi_process_get_user_data()
237 smpi_process_data_t process_data = smpi_process_data();
238 return process_data->data;
241 int smpi_process_count(void)
243 return process_count;
246 int smpi_process_index(void)
248 smpi_process_data_t data = smpi_process_data();
249 //return -1 if not initialized
250 return data ? data->index : MPI_UNDEFINED;
253 MPI_Comm smpi_process_comm_world(void)
255 smpi_process_data_t data = smpi_process_data();
256 //return MPI_COMM_NULL if not initialized
257 return data ? *data->comm_world : MPI_COMM_NULL;
260 smx_rdv_t smpi_process_mailbox(void)
262 smpi_process_data_t data = smpi_process_data();
263 return data->mailbox;
266 smx_rdv_t smpi_process_mailbox_small(void)
268 smpi_process_data_t data = smpi_process_data();
269 return data->mailbox_small;
272 xbt_mutex_t smpi_process_mailboxes_mutex(void)
274 smpi_process_data_t data = smpi_process_data();
275 return data->mailboxes_mutex;
278 smx_rdv_t smpi_process_remote_mailbox(int index)
280 smpi_process_data_t data = smpi_process_remote_data(index);
281 return data->mailbox;
285 smx_rdv_t smpi_process_remote_mailbox_small(int index)
287 smpi_process_data_t data = smpi_process_remote_data(index);
288 return data->mailbox_small;
291 xbt_mutex_t smpi_process_remote_mailboxes_mutex(int index)
293 smpi_process_data_t data = smpi_process_remote_data(index);
294 return data->mailboxes_mutex;
297 xbt_os_timer_t smpi_process_timer(void)
299 smpi_process_data_t data = smpi_process_data();
303 void smpi_process_simulated_start(void)
305 smpi_process_data_t data = smpi_process_data();
306 data->simulated = SIMIX_get_clock();
309 double smpi_process_simulated_elapsed(void)
311 smpi_process_data_t data = smpi_process_data();
312 return SIMIX_get_clock() - data->simulated;
315 MPI_Comm smpi_process_comm_self(void)
317 smpi_process_data_t data = smpi_process_data();
318 if(data->comm_self==MPI_COMM_NULL){
319 MPI_Group group = smpi_group_new(1);
320 data->comm_self = smpi_comm_new(group, NULL);
321 smpi_group_set_mapping(group, smpi_process_index(), 0);
324 return data->comm_self;
327 MPI_Comm smpi_process_get_comm_intra(void)
329 smpi_process_data_t data = smpi_process_data();
330 return data->comm_intra;
333 void smpi_process_set_comm_intra(MPI_Comm comm)
335 smpi_process_data_t data = smpi_process_data();
336 data->comm_intra = comm;
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;
352 void print_request(const char *message, MPI_Request request)
355 ("%s request %p [buf = %p, size = %zu, src = %d, dst = %d, tag = %d, flags = %x]",
356 message, request, request->buf, request->size, request->src,
357 request->dst, request->tag, request->flags);
360 void smpi_comm_copy_buffer_callback(smx_synchro_t comm,
361 void *buff, size_t buff_size)
363 XBT_DEBUG("Copy the data over");
366 if((smpi_privatize_global_variables)
367 && ((char*)buff >= smpi_start_data_exe)
368 && ((char*)buff < smpi_start_data_exe + smpi_size_data_exe )
370 XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !");
371 smpi_switch_data_segment(((smpi_process_data_t)(((simdata_process_t)SIMIX_process_get_data(comm->comm.src_proc))->data))->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 >= smpi_start_data_exe)
379 && ((char*)comm->comm.dst_buff < smpi_start_data_exe + smpi_size_data_exe )
381 XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment");
382 smpi_switch_data_segment(((smpi_process_data_t)(((simdata_process_t)SIMIX_process_get_data(comm->comm.dst_proc))->data))->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);
403 void smpi_comm_null_copy_buffer_callback(smx_synchro_t comm,
404 void *buff, size_t buff_size)
409 static void smpi_check_options(){
410 //check correctness of MPI parameters
412 xbt_assert(sg_cfg_get_int("smpi/async_small_thresh") <=
413 sg_cfg_get_int("smpi/send_is_detached_thres"));
415 if (sg_cfg_is_default_value("smpi/running_power")) {
416 XBT_INFO("You did not set the power of the host running the simulation. "
417 "The timings will certainly not be accurate. "
418 "Use the option \"--cfg=smpi/running_power:<flops>\" to set its value."
419 "Check http://simgrid.org/simgrid/latest/doc/options.html#options_smpi_bench for more information.");
423 int smpi_enabled(void) {
424 return process_data != NULL;
427 void smpi_global_init(void)
431 char name[MAILBOX_NAME_MAXLEN];
434 if (!MC_is_active()) {
435 global_timer = xbt_os_timer_new();
436 xbt_os_walltimer_start(global_timer);
438 if (process_count == 0){
439 process_count = SIMIX_process_count();
442 smpi_universe_size = process_count;
443 process_data = xbt_new0(smpi_process_data_t, process_count);
444 for (i = 0; i < process_count; i++) {
445 process_data[i] = xbt_new(s_smpi_process_data_t, 1);
446 //process_data[i]->index = i;
447 process_data[i]->argc = NULL;
448 process_data[i]->argv = NULL;
449 process_data[i]->mailbox = simcall_rdv_create(get_mailbox_name(name, i));
450 process_data[i]->mailbox_small =
451 simcall_rdv_create(get_mailbox_name_small(name, i));
452 process_data[i]->mailboxes_mutex = xbt_mutex_init();
453 process_data[i]->timer = xbt_os_timer_new();
455 MC_ignore_heap(process_data[i]->timer, xbt_os_timer_size());
456 process_data[i]->comm_self = MPI_COMM_NULL;
457 process_data[i]->comm_intra = MPI_COMM_NULL;
458 process_data[i]->comm_world = NULL;
459 process_data[i]->state = SMPI_UNINITIALIZED;
460 process_data[i]->sampling = 0;
461 process_data[i]->finalization_barrier = NULL;
463 //if the process was launched through smpirun script
464 //we generate a global mpi_comm_world
465 //if not, we let MPI_COMM_NULL, and the comm world
466 //will be private to each mpi instance
468 group = smpi_group_new(process_count);
469 MPI_COMM_WORLD = smpi_comm_new(group, NULL);
470 MPI_Attr_put(MPI_COMM_WORLD, MPI_UNIVERSE_SIZE, (void *)(MPI_Aint)process_count);
471 xbt_bar_t bar=xbt_barrier_init(process_count);
473 for (i = 0; i < process_count; i++) {
474 smpi_group_set_mapping(group, i, i);
475 process_data[i]->finalization_barrier = bar;
480 void smpi_global_destroy(void)
482 int count = smpi_process_count();
485 smpi_bench_destroy();
486 if (MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED){
487 while (smpi_group_unuse(smpi_comm_group(MPI_COMM_WORLD)) > 0);
488 xbt_free(MPI_COMM_WORLD);
489 xbt_barrier_destroy(process_data[0]->finalization_barrier);
491 smpi_deployment_cleanup_instances();
493 MPI_COMM_WORLD = MPI_COMM_NULL;
494 for (i = 0; i < count; i++) {
495 if(process_data[i]->comm_self!=MPI_COMM_NULL){
496 smpi_group_unuse(smpi_comm_group(process_data[i]->comm_self));
497 smpi_comm_destroy(process_data[i]->comm_self);
499 if(process_data[i]->comm_intra!=MPI_COMM_NULL){
500 smpi_group_unuse(smpi_comm_group(process_data[i]->comm_intra));
501 smpi_comm_destroy(process_data[i]->comm_intra);
503 xbt_os_timer_free(process_data[i]->timer);
504 simcall_rdv_destroy(process_data[i]->mailbox);
505 simcall_rdv_destroy(process_data[i]->mailbox_small);
506 xbt_mutex_destroy(process_data[i]->mailboxes_mutex);
507 xbt_free(process_data[i]);
509 xbt_free(process_data);
512 xbt_free(index_to_process_data);
513 if(smpi_privatize_global_variables)
514 smpi_destroy_global_memory_segments();
519 void __attribute__ ((weak)) user_main_()
521 xbt_die("Should not be in this smpi_simulated_main");
525 int __attribute__ ((weak)) smpi_simulated_main_(int argc, char **argv)
527 smpi_process_init(&argc, &argv);
532 int __attribute__ ((weak)) main(int argc, char **argv)
534 return smpi_main(smpi_simulated_main_, argc, argv);
539 static void smpi_init_logs(){
541 /* Connect log categories. See xbt/log.c */
542 XBT_LOG_CONNECT(smpi); /* Keep this line as soon as possible in this
543 function: xbt_log_appender_file.c depends on it
544 DO NOT connect this in XBT or so, or it will be
545 useless to xbt_log_appender_file.c */
546 XBT_LOG_CONNECT(instr_smpi);
547 XBT_LOG_CONNECT(smpi_base);
548 XBT_LOG_CONNECT(smpi_bench);
549 XBT_LOG_CONNECT(smpi_coll);
550 XBT_LOG_CONNECT(smpi_colls);
551 XBT_LOG_CONNECT(smpi_comm);
552 XBT_LOG_CONNECT(smpi_dvfs);
553 XBT_LOG_CONNECT(smpi_group);
554 XBT_LOG_CONNECT(smpi_kernel);
555 XBT_LOG_CONNECT(smpi_mpi);
556 XBT_LOG_CONNECT(smpi_mpi_dt);
557 XBT_LOG_CONNECT(smpi_pmpi);
558 XBT_LOG_CONNECT(smpi_replay);
559 XBT_LOG_CONNECT(smpi_rma);
564 static void smpi_init_options(){
565 int gather_id = find_coll_description(mpi_coll_gather_description,
566 sg_cfg_get_string("smpi/gather"),"gather");
567 mpi_coll_gather_fun = (int (*)(void *, int, MPI_Datatype,
568 void *, int, MPI_Datatype, int, MPI_Comm))
569 mpi_coll_gather_description[gather_id].coll;
571 int allgather_id = find_coll_description(mpi_coll_allgather_description,
572 sg_cfg_get_string("smpi/allgather"),"allgather");
573 mpi_coll_allgather_fun = (int (*)(void *, int, MPI_Datatype,
574 void *, int, MPI_Datatype, MPI_Comm))
575 mpi_coll_allgather_description[allgather_id].coll;
577 int allgatherv_id = find_coll_description(mpi_coll_allgatherv_description,
578 sg_cfg_get_string("smpi/allgatherv"),"allgatherv");
579 mpi_coll_allgatherv_fun = (int (*)(void *, int, MPI_Datatype, void *, int *,
580 int *, MPI_Datatype, MPI_Comm))
581 mpi_coll_allgatherv_description[allgatherv_id].coll;
583 int allreduce_id = find_coll_description(mpi_coll_allreduce_description,
584 sg_cfg_get_string("smpi/allreduce"),"allreduce");
585 mpi_coll_allreduce_fun = (int (*)(void *sbuf, void *rbuf, int rcount,
586 MPI_Datatype dtype, MPI_Op op,
588 mpi_coll_allreduce_description[allreduce_id].coll;
590 int alltoall_id = find_coll_description(mpi_coll_alltoall_description,
591 sg_cfg_get_string("smpi/alltoall"),"alltoall");
592 mpi_coll_alltoall_fun = (int (*)(void *, int, MPI_Datatype,
593 void *, int, MPI_Datatype, MPI_Comm))
594 mpi_coll_alltoall_description[alltoall_id].coll;
596 int alltoallv_id = find_coll_description(mpi_coll_alltoallv_description,
597 sg_cfg_get_string("smpi/alltoallv"),"alltoallv");
598 mpi_coll_alltoallv_fun = (int (*)(void *, int *, int *, MPI_Datatype,
599 void *, int *, int *, MPI_Datatype,
601 mpi_coll_alltoallv_description[alltoallv_id].coll;
603 int bcast_id = find_coll_description(mpi_coll_bcast_description,
604 sg_cfg_get_string("smpi/bcast"),"bcast");
605 mpi_coll_bcast_fun = (int (*)(void *buf, int count, MPI_Datatype datatype,
606 int root, MPI_Comm com))
607 mpi_coll_bcast_description[bcast_id].coll;
609 int reduce_id = find_coll_description(mpi_coll_reduce_description,
610 sg_cfg_get_string("smpi/reduce"),"reduce");
611 mpi_coll_reduce_fun = (int (*)(void *buf, void *rbuf, int count,
612 MPI_Datatype datatype, MPI_Op op,
613 int root, MPI_Comm comm))
614 mpi_coll_reduce_description[reduce_id].coll;
616 int reduce_scatter_id =
617 find_coll_description(mpi_coll_reduce_scatter_description,
618 sg_cfg_get_string("smpi/reduce_scatter"),"reduce_scatter");
619 mpi_coll_reduce_scatter_fun = (int (*)(void *sbuf, void *rbuf, int *rcounts,
620 MPI_Datatype dtype, MPI_Op op,
622 mpi_coll_reduce_scatter_description[reduce_scatter_id].coll;
624 int scatter_id = find_coll_description(mpi_coll_scatter_description,
625 sg_cfg_get_string("smpi/scatter"),"scatter");
626 mpi_coll_scatter_fun = (int (*)(void *sendbuf, int sendcount,
627 MPI_Datatype sendtype, void *recvbuf,
628 int recvcount, MPI_Datatype recvtype,
629 int root, MPI_Comm comm))
630 mpi_coll_scatter_description[scatter_id].coll;
632 int barrier_id = find_coll_description(mpi_coll_barrier_description,
633 sg_cfg_get_string("smpi/barrier"),"barrier");
634 mpi_coll_barrier_fun = (int (*)(MPI_Comm comm))
635 mpi_coll_barrier_description[barrier_id].coll;
637 smpi_cpu_threshold = sg_cfg_get_double("smpi/cpu_threshold");
638 smpi_running_power = sg_cfg_get_double("smpi/running_power");
639 smpi_privatize_global_variables = sg_cfg_get_boolean("smpi/privatize_global_variables");
640 if (smpi_cpu_threshold < 0)
641 smpi_cpu_threshold = DBL_MAX;
645 int smpi_main(int (*realmain) (int argc, char *argv[]), int argc, char *argv[])
647 srand(SMPI_RAND_SEED);
649 if (getenv("SMPI_PRETEND_CC") != NULL) {
650 /* Hack to ensure that smpicc can pretend to be a simple
651 * compiler. Particularly handy to pass it to the configuration tools */
657 TRACE_global_init(&argc, argv);
659 TRACE_add_start_function(TRACE_smpi_alloc);
660 TRACE_add_end_function(TRACE_smpi_release);
662 SIMIX_global_init(&argc, argv);
663 MSG_init(&argc,argv);
665 SMPI_switch_data_segment = smpi_switch_data_segment;
669 // parse the platform file: get the host list
670 SIMIX_create_environment(argv[1]);
671 SIMIX_comm_set_copy_data_callback(&smpi_comm_copy_buffer_callback);
672 SIMIX_function_register_default(realmain);
673 SIMIX_launch_application(argv[2]);
677 smpi_check_options();
679 if(smpi_privatize_global_variables)
680 smpi_initialize_global_memory_segments();
682 /* Clean IO before the run */
686 if (MC_is_active()) {
692 xbt_os_walltimer_stop(global_timer);
693 if (sg_cfg_get_boolean("smpi/display_timing")){
694 double global_time = xbt_os_timer_elapsed(global_timer);
695 XBT_INFO("Simulated time: %g seconds. \n\n"
696 "The simulation took %g seconds (after parsing and platform setup)\n"
697 "%g seconds were actual computation of the application"
698 , SIMIX_get_clock(), global_time , smpi_total_benched_time);
700 if (smpi_total_benched_time/global_time>=0.75)
701 XBT_INFO("More than 75%% of the time was spent inside the application code.\n"
702 "You may want to use sampling functions or trace replay to reduce this.");
706 smpi_global_destroy();
713 // This function can be called from extern file, to initialize logs, options, and processes of smpi
714 // without the need of smpirun
719 smpi_check_options();
720 if (TRACE_is_enabled() && TRACE_is_configured())
722 if(smpi_privatize_global_variables)
723 smpi_initialize_global_memory_segments();
726 void SMPI_finalize(){
727 smpi_global_destroy();