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;
34 void *data; /* user data */
37 int sampling; /* inside an SMPI_SAMPLE_ block? */
39 int replaying; /* is the process replaying a trace */
40 xbt_bar_t finalization_barrier;
41 } s_smpi_process_data_t;
43 static smpi_process_data_t *process_data = NULL;
44 int process_count = 0;
45 int* index_to_process_data = NULL;
46 extern double smpi_total_benched_time;
47 xbt_os_timer_t global_timer;
48 MPI_Comm MPI_COMM_WORLD = MPI_COMM_UNINITIALIZED;
49 int MPI_UNIVERSE_SIZE;
51 MPI_Errhandler *MPI_ERRORS_RETURN = NULL;
52 MPI_Errhandler *MPI_ERRORS_ARE_FATAL = NULL;
53 MPI_Errhandler *MPI_ERRHANDLER_NULL = NULL;
55 #define MAILBOX_NAME_MAXLEN (5 + sizeof(int) * 2 + 1)
57 static char *get_mailbox_name(char *str, int index)
59 snprintf(str, MAILBOX_NAME_MAXLEN, "SMPI-%0*x", (int) (sizeof(int) * 2),
64 static char *get_mailbox_name_small(char *str, int index)
66 snprintf(str, MAILBOX_NAME_MAXLEN, "small%0*x", (int) (sizeof(int) * 2),
71 void smpi_process_init(int *argc, char ***argv)
74 smpi_process_data_t data;
78 proc = SIMIX_process_self();
79 //FIXME: dirty cleanup method to avoid using msg cleanup functions on these processes when using MSG+SMPI
80 proc->context->cleanup_func=SIMIX_process_cleanup;
81 char* instance_id = (*argv)[1];
82 int rank = atoi((*argv)[2]);
83 index = smpi_process_index_of_smx_process(proc);
85 if(!index_to_process_data){
86 index_to_process_data=(int*)xbt_malloc(SIMIX_process_count()*sizeof(int));
88 MPI_Comm* temp_comm_world;
90 smpi_deployment_register_process(instance_id, rank, index, &temp_comm_world ,&temp_bar);
91 data = smpi_process_remote_data(index);
92 data->comm_world = temp_comm_world;
93 if(temp_bar != NULL) data->finalization_barrier = temp_bar;
95 data->instance_id = instance_id;
97 xbt_free(simcall_process_get_data(proc));
98 simcall_process_set_data(proc, data);
101 memmove(&(*argv)[0], &(*argv)[2], sizeof(char *) * (*argc - 2));
102 (*argv)[(*argc) - 1] = NULL;
103 (*argv)[(*argc) - 2] = NULL;
108 // set the process attached to the mailbox
109 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:
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 smx_rdv_t smpi_process_remote_mailbox(int index)
260 smpi_process_data_t data = smpi_process_remote_data(index);
261 return data->mailbox;
265 smx_rdv_t smpi_process_remote_mailbox_small(int index)
267 smpi_process_data_t data = smpi_process_remote_data(index);
268 return data->mailbox_small;
271 xbt_os_timer_t smpi_process_timer(void)
273 smpi_process_data_t data = smpi_process_data();
277 void smpi_process_simulated_start(void)
279 smpi_process_data_t data = smpi_process_data();
280 data->simulated = SIMIX_get_clock();
283 double smpi_process_simulated_elapsed(void)
285 smpi_process_data_t data = smpi_process_data();
286 return SIMIX_get_clock() - data->simulated;
289 MPI_Comm smpi_process_comm_self(void)
291 smpi_process_data_t data = smpi_process_data();
292 if(data->comm_self==MPI_COMM_NULL){
293 MPI_Group group = smpi_group_new(1);
294 data->comm_self = smpi_comm_new(group, NULL);
295 smpi_group_set_mapping(group, smpi_process_index(), 0);
298 return data->comm_self;
301 MPI_Comm smpi_process_get_comm_intra(void)
303 smpi_process_data_t data = smpi_process_data();
304 return data->comm_intra;
307 void smpi_process_set_comm_intra(MPI_Comm comm)
309 smpi_process_data_t data = smpi_process_data();
310 data->comm_intra = comm;
313 void smpi_process_set_sampling(int s)
315 smpi_process_data_t data = smpi_process_data();
319 int smpi_process_get_sampling(void)
321 smpi_process_data_t data = smpi_process_data();
322 return data->sampling;
326 void print_request(const char *message, MPI_Request request)
329 ("%s request %p [buf = %p, size = %zu, src = %d, dst = %d, tag = %d, flags = %x]",
330 message, request, request->buf, request->size, request->src,
331 request->dst, request->tag, request->flags);
334 void smpi_comm_copy_buffer_callback(smx_synchro_t comm,
335 void *buff, size_t buff_size)
337 XBT_DEBUG("Copy the data over");
340 if((smpi_privatize_global_variables)
341 && ((char*)buff >= start_data_exe)
342 && ((char*)buff < start_data_exe + size_data_exe )
344 XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !");
345 smpi_switch_data_segment(((smpi_process_data_t)SIMIX_process_get_data(comm->comm.src_proc))->index);
346 tmpbuff = (void*)xbt_malloc(buff_size);
347 memcpy(tmpbuff, buff, buff_size);
351 if((smpi_privatize_global_variables)
352 && ((char*)comm->comm.dst_buff >= start_data_exe)
353 && ((char*)comm->comm.dst_buff < start_data_exe + size_data_exe )
355 XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment");
356 smpi_switch_data_segment(((smpi_process_data_t)SIMIX_process_get_data(comm->comm.dst_proc))->index);
360 memcpy(comm->comm.dst_buff, tmpbuff, buff_size);
361 if (comm->comm.detached) {
362 // if this is a detached send, the source buffer was duplicated by SMPI
363 // sender to make the original buffer available to the application ASAP
365 //It seems that the request is used after the call there this should
366 //be free somewhereelse but where???
367 //xbt_free(comm->comm.src_data);// inside SMPI the request is keep
368 //inside the user data and should be free
369 comm->comm.src_buff = NULL;
372 if(tmpbuff!=buff)xbt_free(tmpbuff);
377 void smpi_comm_null_copy_buffer_callback(smx_synchro_t comm,
378 void *buff, size_t buff_size)
383 static void smpi_check_options(){
384 //check correctness of MPI parameters
386 xbt_assert(sg_cfg_get_int("smpi/async_small_thres") <=
387 sg_cfg_get_int("smpi/send_is_detached_thres"));
389 if (sg_cfg_is_default_value("smpi/running_power")) {
390 XBT_INFO("You did not set the power of the host running the simulation. "
391 "The timings will certainly not be accurate. "
392 "Use the option \"--cfg=smpi/running_power:<flops>\" to set its value."
393 "Check http://simgrid.org/simgrid/latest/doc/options.html#options_smpi_bench for more information.");
397 int smpi_enabled(void) {
398 return process_data != NULL;
401 void smpi_global_init(void)
405 char name[MAILBOX_NAME_MAXLEN];
408 if (!MC_is_active()) {
409 global_timer = xbt_os_timer_new();
410 xbt_os_walltimer_start(global_timer);
412 if (process_count == 0){
413 process_count = SIMIX_process_count();
416 process_data = xbt_new0(smpi_process_data_t, process_count);
417 for (i = 0; i < process_count; i++) {
418 process_data[i] = xbt_new(s_smpi_process_data_t, 1);
419 //process_data[i]->index = i;
420 process_data[i]->argc = NULL;
421 process_data[i]->argv = NULL;
422 process_data[i]->mailbox = simcall_rdv_create(get_mailbox_name(name, i));
423 process_data[i]->mailbox_small =
424 simcall_rdv_create(get_mailbox_name_small(name, i));
425 process_data[i]->timer = xbt_os_timer_new();
427 MC_ignore_heap(process_data[i]->timer, xbt_os_timer_size());
428 process_data[i]->comm_self = MPI_COMM_NULL;
429 process_data[i]->comm_intra = MPI_COMM_NULL;
430 process_data[i]->comm_world = NULL;
431 process_data[i]->state = SMPI_UNINITIALIZED;
432 process_data[i]->sampling = 0;
433 process_data[i]->finalization_barrier = NULL;
435 //if the process was launched through smpirun script
436 //we generate a global mpi_comm_world
437 //if not, we let MPI_COMM_NULL, and the comm world
438 //will be private to each mpi instance
440 group = smpi_group_new(process_count);
441 MPI_COMM_WORLD = smpi_comm_new(group, NULL);
442 xbt_bar_t bar=xbt_barrier_init(process_count);
444 MPI_UNIVERSE_SIZE = smpi_comm_size(MPI_COMM_WORLD);
445 for (i = 0; i < process_count; i++) {
446 smpi_group_set_mapping(group, i, i);
447 process_data[i]->finalization_barrier = bar;
452 void smpi_global_destroy(void)
454 int count = smpi_process_count();
457 smpi_bench_destroy();
458 if (MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED){
459 while (smpi_group_unuse(smpi_comm_group(MPI_COMM_WORLD)) > 0);
460 xbt_free(MPI_COMM_WORLD);
461 xbt_barrier_destroy(process_data[0]->finalization_barrier);
463 smpi_deployment_cleanup_instances();
465 MPI_COMM_WORLD = MPI_COMM_NULL;
466 for (i = 0; i < count; i++) {
467 if(process_data[i]->comm_self!=MPI_COMM_NULL){
468 smpi_group_unuse(smpi_comm_group(process_data[i]->comm_self));
469 smpi_comm_destroy(process_data[i]->comm_self);
471 if(process_data[i]->comm_intra!=MPI_COMM_NULL){
472 smpi_group_unuse(smpi_comm_group(process_data[i]->comm_intra));
473 smpi_comm_destroy(process_data[i]->comm_intra);
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);
483 xbt_free(index_to_process_data);
484 if(smpi_privatize_global_variables)
485 smpi_destroy_global_memory_segments();
490 void __attribute__ ((weak)) user_main_()
492 xbt_die("Should not be in this smpi_simulated_main");
496 int __attribute__ ((weak)) smpi_simulated_main_(int argc, char **argv)
498 smpi_process_init(&argc, &argv);
503 int __attribute__ ((weak)) main(int argc, char **argv)
505 return smpi_main(smpi_simulated_main_, argc, argv);
510 static void smpi_init_logs(){
512 /* Connect log categories. See xbt/log.c */
513 XBT_LOG_CONNECT(smpi); /* Keep this line as soon as possible in this
514 function: xbt_log_appender_file.c depends on it
515 DO NOT connect this in XBT or so, or it will be
516 useless to xbt_log_appender_file.c */
518 XBT_LOG_CONNECT(instr_smpi);
520 XBT_LOG_CONNECT(smpi_base);
521 XBT_LOG_CONNECT(smpi_bench);
522 XBT_LOG_CONNECT(smpi_coll);
523 XBT_LOG_CONNECT(smpi_colls);
524 XBT_LOG_CONNECT(smpi_comm);
525 XBT_LOG_CONNECT(smpi_dvfs);
526 XBT_LOG_CONNECT(smpi_group);
527 XBT_LOG_CONNECT(smpi_kernel);
528 XBT_LOG_CONNECT(smpi_mpi);
529 XBT_LOG_CONNECT(smpi_mpi_dt);
530 XBT_LOG_CONNECT(smpi_pmpi);
531 XBT_LOG_CONNECT(smpi_replay);
532 XBT_LOG_CONNECT(smpi_rma);
537 static void smpi_init_options(){
538 int gather_id = find_coll_description(mpi_coll_gather_description,
539 sg_cfg_get_string("smpi/gather"));
540 mpi_coll_gather_fun = (int (*)(void *, int, MPI_Datatype,
541 void *, int, MPI_Datatype, int, MPI_Comm))
542 mpi_coll_gather_description[gather_id].coll;
544 int allgather_id = find_coll_description(mpi_coll_allgather_description,
545 sg_cfg_get_string("smpi/allgather"));
546 mpi_coll_allgather_fun = (int (*)(void *, int, MPI_Datatype,
547 void *, int, MPI_Datatype, MPI_Comm))
548 mpi_coll_allgather_description[allgather_id].coll;
550 int allgatherv_id = find_coll_description(mpi_coll_allgatherv_description,
551 sg_cfg_get_string("smpi/allgatherv"));
552 mpi_coll_allgatherv_fun = (int (*)(void *, int, MPI_Datatype, void *, int *,
553 int *, MPI_Datatype, MPI_Comm))
554 mpi_coll_allgatherv_description[allgatherv_id].coll;
556 int allreduce_id = find_coll_description(mpi_coll_allreduce_description,
557 sg_cfg_get_string("smpi/allreduce"));
558 mpi_coll_allreduce_fun = (int (*)(void *sbuf, void *rbuf, int rcount,
559 MPI_Datatype dtype, MPI_Op op,
561 mpi_coll_allreduce_description[allreduce_id].coll;
563 int alltoall_id = find_coll_description(mpi_coll_alltoall_description,
564 sg_cfg_get_string("smpi/alltoall"));
565 mpi_coll_alltoall_fun = (int (*)(void *, int, MPI_Datatype,
566 void *, int, MPI_Datatype, MPI_Comm))
567 mpi_coll_alltoall_description[alltoall_id].coll;
569 int alltoallv_id = find_coll_description(mpi_coll_alltoallv_description,
570 sg_cfg_get_string("smpi/alltoallv"));
571 mpi_coll_alltoallv_fun = (int (*)(void *, int *, int *, MPI_Datatype,
572 void *, int *, int *, MPI_Datatype,
574 mpi_coll_alltoallv_description[alltoallv_id].coll;
576 int bcast_id = find_coll_description(mpi_coll_bcast_description,
577 sg_cfg_get_string("smpi/bcast"));
578 mpi_coll_bcast_fun = (int (*)(void *buf, int count, MPI_Datatype datatype,
579 int root, MPI_Comm com))
580 mpi_coll_bcast_description[bcast_id].coll;
582 int reduce_id = find_coll_description(mpi_coll_reduce_description,
583 sg_cfg_get_string("smpi/reduce"));
584 mpi_coll_reduce_fun = (int (*)(void *buf, void *rbuf, int count,
585 MPI_Datatype datatype, MPI_Op op,
586 int root, MPI_Comm comm))
587 mpi_coll_reduce_description[reduce_id].coll;
589 int reduce_scatter_id =
590 find_coll_description(mpi_coll_reduce_scatter_description,
591 sg_cfg_get_string("smpi/reduce_scatter"));
592 mpi_coll_reduce_scatter_fun = (int (*)(void *sbuf, void *rbuf, int *rcounts,
593 MPI_Datatype dtype, MPI_Op op,
595 mpi_coll_reduce_scatter_description[reduce_scatter_id].coll;
597 int scatter_id = find_coll_description(mpi_coll_scatter_description,
598 sg_cfg_get_string("smpi/scatter"));
599 mpi_coll_scatter_fun = (int (*)(void *sendbuf, int sendcount,
600 MPI_Datatype sendtype, void *recvbuf,
601 int recvcount, MPI_Datatype recvtype,
602 int root, MPI_Comm comm))
603 mpi_coll_scatter_description[scatter_id].coll;
605 int barrier_id = find_coll_description(mpi_coll_barrier_description,
606 sg_cfg_get_string("smpi/barrier"));
607 mpi_coll_barrier_fun = (int (*)(MPI_Comm comm))
608 mpi_coll_barrier_description[barrier_id].coll;
610 smpi_cpu_threshold = sg_cfg_get_double("smpi/cpu_threshold");
611 smpi_running_power = sg_cfg_get_double("smpi/running_power");
612 smpi_privatize_global_variables = sg_cfg_get_boolean("smpi/privatize_global_variables");
613 if (smpi_cpu_threshold < 0)
614 smpi_cpu_threshold = DBL_MAX;
618 int smpi_main(int (*realmain) (int argc, char *argv[]), int argc, char *argv[])
620 srand(SMPI_RAND_SEED);
622 if (getenv("SMPI_PRETEND_CC") != NULL) {
623 /* Hack to ensure that smpicc can pretend to be a simple
624 * compiler. Particularly handy to pass it to the configuration tools */
631 TRACE_global_init(&argc, argv);
633 TRACE_add_start_function(TRACE_smpi_alloc);
634 TRACE_add_end_function(TRACE_smpi_release);
637 SIMIX_global_init(&argc, argv);
641 // parse the platform file: get the host list
642 SIMIX_create_environment(argv[1]);
643 SIMIX_comm_set_copy_data_callback(&smpi_comm_copy_buffer_callback);
644 SIMIX_function_register_default(realmain);
645 SIMIX_launch_application(argv[2]);
649 smpi_check_options();
651 if(smpi_privatize_global_variables)
652 smpi_initialize_global_memory_segments();
654 /* Clean IO before the run */
658 if (MC_is_active()) {
659 MC_do_the_modelcheck_for_real();
664 xbt_os_walltimer_stop(global_timer);
665 if (sg_cfg_get_boolean("smpi/display_timing")){
666 double global_time = xbt_os_timer_elapsed(global_timer);
667 XBT_INFO("Simulated time: %g seconds. \n "
668 "The simulation took %g seconds (after parsing and platform setup)\n"
669 "%g seconds were actual computation of the application"
670 , SIMIX_get_clock(), global_time , smpi_total_benched_time);
672 if (smpi_total_benched_time/global_time>=0.75)
673 XBT_INFO("More than 75%% of the time was spent inside the application code.\n"
674 "You may want to use sampling functions or trace replay to reduce this.");
678 smpi_global_destroy();
687 // This function can be called from extern file, to initialize logs, options, and processes of smpi
688 // without the need of smpirun
693 smpi_check_options();
695 if (TRACE_is_enabled() && TRACE_is_configured()) {
699 if(smpi_privatize_global_variables)
700 smpi_initialize_global_memory_segments();
703 void SMPI_finalize(){
704 smpi_global_destroy();