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;
32 void *data; /* user data */
35 int sampling; /* inside an SMPI_SAMPLE_ block? */
37 xbt_bar_t finalization_barrier;
38 } s_smpi_process_data_t;
40 static smpi_process_data_t *process_data = NULL;
41 int process_count = 0;
42 int* index_to_process_data = NULL;
45 MPI_Comm MPI_COMM_WORLD = MPI_COMM_UNINITIALIZED;
46 int MPI_UNIVERSE_SIZE;
48 MPI_Errhandler *MPI_ERRORS_RETURN = NULL;
49 MPI_Errhandler *MPI_ERRORS_ARE_FATAL = NULL;
50 MPI_Errhandler *MPI_ERRHANDLER_NULL = NULL;
52 #define MAILBOX_NAME_MAXLEN (5 + sizeof(int) * 2 + 1)
54 static char *get_mailbox_name(char *str, int index)
56 snprintf(str, MAILBOX_NAME_MAXLEN, "SMPI-%0*x", (int) (sizeof(int) * 2),
61 static char *get_mailbox_name_small(char *str, int index)
63 snprintf(str, MAILBOX_NAME_MAXLEN, "small%0*x", (int) (sizeof(int) * 2),
68 void smpi_process_init(int *argc, char ***argv)
71 smpi_process_data_t data;
75 proc = SIMIX_process_self();
76 //FIXME: dirty cleanup method to avoid using msg cleanup functions on these processes when using MSG+SMPI
77 proc->context->cleanup_func=SIMIX_process_cleanup;
78 char* instance_id = (*argv)[1];
79 int rank = atoi((*argv)[2]);
80 index = SIMIX_process_get_PID(proc) -1;
83 smpi_current_rank = index;
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;
96 xbt_free(simcall_process_get_data(proc));
97 simcall_process_set_data(proc, data);
101 memmove(&(*argv)[1], &(*argv)[3], sizeof(char *) * (*argc - 3));
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);
110 XBT_DEBUG("<%d> New process in the game: %p", index, proc);
112 if(smpi_privatize_global_variables){
113 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 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;
180 int smpi_process_argc(void)
182 smpi_process_data_t data = smpi_process_data();
183 return data->argc ? *(data->argc) - 1 : 0;
186 int smpi_process_getarg(integer * index, char *dst, ftnlen len)
188 smpi_process_data_t data = smpi_process_data();
192 if (!data->argc || !data->argv || *index < 1 || *index >= *(data->argc)) {
195 arg = (*data->argv)[*index];
196 for (i = 0; i < len && arg[i] != '\0'; i++) {
199 for (; i < len; i++) {
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 return SIMIX_process_self_get_data(SIMIX_process_self());
223 smpi_process_data_t smpi_process_remote_data(int index)
225 return process_data[index_to_process_data[index]];
228 void smpi_process_set_user_data(void *data)
230 smpi_process_data_t process_data = smpi_process_data();
231 process_data->data = data;
234 void *smpi_process_get_user_data()
236 smpi_process_data_t process_data = smpi_process_data();
237 return process_data->data;
240 int smpi_process_count(void)
242 return process_count;
245 int smpi_process_index(void)
247 smpi_process_data_t data = smpi_process_data();
248 //return -1 if not initialized
249 return data ? data->index : MPI_UNDEFINED;
252 MPI_Comm smpi_process_comm_world(void)
254 smpi_process_data_t data = smpi_process_data();
255 //return MPI_COMM_NULL if not initialized
256 return data ? *data->comm_world : MPI_COMM_NULL;
259 smx_rdv_t smpi_process_mailbox(void)
261 smpi_process_data_t data = smpi_process_data();
262 return data->mailbox;
265 smx_rdv_t smpi_process_mailbox_small(void)
267 smpi_process_data_t data = smpi_process_data();
268 return data->mailbox_small;
271 smx_rdv_t smpi_process_remote_mailbox(int index)
273 smpi_process_data_t data = smpi_process_remote_data(index);
274 return data->mailbox;
278 smx_rdv_t smpi_process_remote_mailbox_small(int index)
280 smpi_process_data_t data = smpi_process_remote_data(index);
281 return data->mailbox_small;
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 void smpi_process_set_sampling(int s)
316 smpi_process_data_t data = smpi_process_data();
320 int smpi_process_get_sampling(void)
322 smpi_process_data_t data = smpi_process_data();
323 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_action_t comm,
335 void *buff, size_t buff_size)
337 XBT_DEBUG("Copy the data over");
338 if(_xbt_replay_is_active()) return;
341 if((smpi_privatize_global_variables)
342 && ((char*)buff >= start_data_exe)
343 && ((char*)buff < start_data_exe + size_data_exe )
345 XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !");
346 switch_data_segment(((smpi_process_data_t)SIMIX_process_get_data(comm->comm.src_proc))->index);
347 tmpbuff = (void*)xbt_malloc(buff_size);
348 memcpy(tmpbuff, buff, buff_size);
352 if((smpi_privatize_global_variables)
353 && ((char*)comm->comm.dst_buff >= start_data_exe)
354 && ((char*)comm->comm.dst_buff < start_data_exe + size_data_exe )
356 XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment");
357 switch_data_segment(((smpi_process_data_t)SIMIX_process_get_data(comm->comm.dst_proc))->index);
361 memcpy(comm->comm.dst_buff, tmpbuff, buff_size);
362 if (comm->comm.detached) {
363 // if this is a detached send, the source buffer was duplicated by SMPI
364 // sender to make the original buffer available to the application ASAP
366 //It seems that the request is used after the call there this should
367 //be free somewhereelse but where???
368 //xbt_free(comm->comm.src_data);// inside SMPI the request is keep
369 //inside the user data and should be free
370 comm->comm.src_buff = NULL;
373 if(tmpbuff!=buff)xbt_free(tmpbuff);
377 static void smpi_check_options(){
378 //check correctness of MPI parameters
380 xbt_assert(sg_cfg_get_int("smpi/async_small_thres") <=
381 sg_cfg_get_int("smpi/send_is_detached_thres"));
383 if (sg_cfg_is_default_value("smpi/running_power")) {
384 XBT_INFO("You did not set the power of the host running the simulation. "
385 "The timings will certainly not be accurate. "
386 "Use the option \"--cfg=smpi/running_power:<flops>\" to set its value."
387 "Check http://simgrid.org/simgrid/latest/doc/options.html#options_smpi_bench for more information. ");
391 void smpi_global_init(void)
395 char name[MAILBOX_NAME_MAXLEN];
399 if (process_count == 0){
400 process_count = SIMIX_process_count();
403 process_data = xbt_new0(smpi_process_data_t, process_count);
404 for (i = 0; i < process_count; i++) {
405 process_data[i] = xbt_new(s_smpi_process_data_t, 1);
406 //process_data[i]->index = i;
407 process_data[i]->argc = NULL;
408 process_data[i]->argv = NULL;
409 process_data[i]->mailbox = simcall_rdv_create(get_mailbox_name(name, i));
410 process_data[i]->mailbox_small =
411 simcall_rdv_create(get_mailbox_name_small(name, i));
412 process_data[i]->timer = xbt_os_timer_new();
414 MC_ignore_heap(process_data[i]->timer, xbt_os_timer_size());
415 process_data[i]->comm_self = MPI_COMM_NULL;
416 process_data[i]->comm_world = NULL;
417 process_data[i]->state = SMPI_UNINITIALIZED;
418 process_data[i]->sampling = 0;
419 process_data[i]->finalization_barrier = NULL;
421 //if the process was launched through smpirun script
422 //we generate a global mpi_comm_world
423 //if not, we let MPI_COMM_NULL, and the comm world
424 //will be private to each mpi instance
426 group = smpi_group_new(process_count);
427 MPI_COMM_WORLD = smpi_comm_new(group, NULL);
428 xbt_bar_t bar=xbt_barrier_init(process_count);
430 MPI_UNIVERSE_SIZE = smpi_comm_size(MPI_COMM_WORLD);
431 for (i = 0; i < process_count; i++) {
432 smpi_group_set_mapping(group, i, i);
433 process_data[i]->finalization_barrier = bar;
438 void smpi_global_destroy(void)
440 int count = smpi_process_count();
443 smpi_bench_destroy();
444 if (MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED){
445 while (smpi_group_unuse(smpi_comm_group(MPI_COMM_WORLD)) > 0);
446 xbt_free(MPI_COMM_WORLD);
447 xbt_barrier_destroy(process_data[0]->finalization_barrier);
449 smpi_deployment_cleanup_instances();
451 MPI_COMM_WORLD = MPI_COMM_NULL;
452 for (i = 0; i < count; i++) {
453 if(process_data[i]->comm_self!=MPI_COMM_NULL){
454 smpi_group_unuse(smpi_comm_group(process_data[i]->comm_self));
455 smpi_comm_destroy(process_data[i]->comm_self);
457 xbt_os_timer_free(process_data[i]->timer);
458 simcall_rdv_destroy(process_data[i]->mailbox);
459 simcall_rdv_destroy(process_data[i]->mailbox_small);
460 xbt_free(process_data[i]);
462 xbt_free(process_data);
465 xbt_free(index_to_process_data);
466 if(smpi_privatize_global_variables)
467 smpi_destroy_global_memory_segments();
471 /* Fortran specific stuff */
472 /* With smpicc, the following weak symbols are used */
473 /* With smpiff, the following weak symbols are replaced by those in libf2c */
474 int __attribute__ ((weak)) xargc;
475 char ** __attribute__ ((weak)) xargv;
478 void __attribute__ ((weak)) user_main_()
480 xbt_die("Should not be in this smpi_simulated_main");
484 int __attribute__ ((weak)) smpi_simulated_main_(int argc, char **argv)
486 smpi_process_init(&argc, &argv);
491 int __attribute__ ((weak)) main(int argc, char **argv)
493 return smpi_main(smpi_simulated_main_, argc, argv);
496 int __attribute__ ((weak)) MAIN__()
498 return smpi_main(smpi_simulated_main_, xargc, xargv);
502 static void smpi_init_logs(){
504 /* Connect log categories. See xbt/log.c */
505 XBT_LOG_CONNECT(smpi); /* Keep this line as soon as possible in this
506 function: xbt_log_appender_file.c depends on it
507 DO NOT connect this in XBT or so, or it will be
508 useless to xbt_log_appender_file.c */
510 XBT_LOG_CONNECT(instr_smpi);
512 XBT_LOG_CONNECT(smpi_base);
513 XBT_LOG_CONNECT(smpi_bench);
514 XBT_LOG_CONNECT(smpi_coll);
515 XBT_LOG_CONNECT(smpi_colls);
516 XBT_LOG_CONNECT(smpi_comm);
517 XBT_LOG_CONNECT(smpi_dvfs);
518 XBT_LOG_CONNECT(smpi_group);
519 XBT_LOG_CONNECT(smpi_kernel);
520 XBT_LOG_CONNECT(smpi_mpi);
521 XBT_LOG_CONNECT(smpi_mpi_dt);
522 XBT_LOG_CONNECT(smpi_pmpi);
523 XBT_LOG_CONNECT(smpi_replay);
528 static void smpi_init_options(){
529 int gather_id = find_coll_description(mpi_coll_gather_description,
530 sg_cfg_get_string("smpi/gather"));
531 mpi_coll_gather_fun = (int (*)(void *, int, MPI_Datatype,
532 void *, int, MPI_Datatype, int, MPI_Comm))
533 mpi_coll_gather_description[gather_id].coll;
535 int allgather_id = find_coll_description(mpi_coll_allgather_description,
536 sg_cfg_get_string("smpi/allgather"));
537 mpi_coll_allgather_fun = (int (*)(void *, int, MPI_Datatype,
538 void *, int, MPI_Datatype, MPI_Comm))
539 mpi_coll_allgather_description[allgather_id].coll;
541 int allgatherv_id = find_coll_description(mpi_coll_allgatherv_description,
542 sg_cfg_get_string("smpi/allgatherv"));
543 mpi_coll_allgatherv_fun = (int (*)(void *, int, MPI_Datatype, void *, int *,
544 int *, MPI_Datatype, MPI_Comm))
545 mpi_coll_allgatherv_description[allgatherv_id].coll;
547 int allreduce_id = find_coll_description(mpi_coll_allreduce_description,
548 sg_cfg_get_string("smpi/allreduce"));
549 mpi_coll_allreduce_fun = (int (*)(void *sbuf, void *rbuf, int rcount,
550 MPI_Datatype dtype, MPI_Op op,
552 mpi_coll_allreduce_description[allreduce_id].coll;
554 int alltoall_id = find_coll_description(mpi_coll_alltoall_description,
555 sg_cfg_get_string("smpi/alltoall"));
556 mpi_coll_alltoall_fun = (int (*)(void *, int, MPI_Datatype,
557 void *, int, MPI_Datatype, MPI_Comm))
558 mpi_coll_alltoall_description[alltoall_id].coll;
560 int alltoallv_id = find_coll_description(mpi_coll_alltoallv_description,
561 sg_cfg_get_string("smpi/alltoallv"));
562 mpi_coll_alltoallv_fun = (int (*)(void *, int *, int *, MPI_Datatype,
563 void *, int *, int *, MPI_Datatype,
565 mpi_coll_alltoallv_description[alltoallv_id].coll;
567 int bcast_id = find_coll_description(mpi_coll_bcast_description,
568 sg_cfg_get_string("smpi/bcast"));
569 mpi_coll_bcast_fun = (int (*)(void *buf, int count, MPI_Datatype datatype,
570 int root, MPI_Comm com))
571 mpi_coll_bcast_description[bcast_id].coll;
573 int reduce_id = find_coll_description(mpi_coll_reduce_description,
574 sg_cfg_get_string("smpi/reduce"));
575 mpi_coll_reduce_fun = (int (*)(void *buf, void *rbuf, int count,
576 MPI_Datatype datatype, MPI_Op op,
577 int root, MPI_Comm comm))
578 mpi_coll_reduce_description[reduce_id].coll;
580 int reduce_scatter_id =
581 find_coll_description(mpi_coll_reduce_scatter_description,
582 sg_cfg_get_string("smpi/reduce_scatter"));
583 mpi_coll_reduce_scatter_fun = (int (*)(void *sbuf, void *rbuf, int *rcounts,
584 MPI_Datatype dtype, MPI_Op op,
586 mpi_coll_reduce_scatter_description[reduce_scatter_id].coll;
588 int scatter_id = find_coll_description(mpi_coll_scatter_description,
589 sg_cfg_get_string("smpi/scatter"));
590 mpi_coll_scatter_fun = (int (*)(void *sendbuf, int sendcount,
591 MPI_Datatype sendtype, void *recvbuf,
592 int recvcount, MPI_Datatype recvtype,
593 int root, MPI_Comm comm))
594 mpi_coll_scatter_description[scatter_id].coll;
596 int barrier_id = find_coll_description(mpi_coll_barrier_description,
597 sg_cfg_get_string("smpi/barrier"));
598 mpi_coll_barrier_fun = (int (*)(MPI_Comm comm))
599 mpi_coll_barrier_description[barrier_id].coll;
601 smpi_cpu_threshold = sg_cfg_get_double("smpi/cpu_threshold");
602 smpi_running_power = sg_cfg_get_double("smpi/running_power");
603 smpi_privatize_global_variables = sg_cfg_get_boolean("smpi/privatize_global_variables");
604 if (smpi_cpu_threshold < 0)
605 smpi_cpu_threshold = DBL_MAX;
609 int smpi_main(int (*realmain) (int argc, char *argv[]), int argc, char *argv[])
611 srand(SMPI_RAND_SEED);
613 if (getenv("SMPI_PRETEND_CC") != NULL) {
614 /* Hack to ensure that smpicc can pretend to be a simple
615 * compiler. Particularly handy to pass it to the configuration tools */
622 TRACE_global_init(&argc, argv);
624 TRACE_add_start_function(TRACE_smpi_alloc);
625 TRACE_add_end_function(TRACE_smpi_release);
628 SIMIX_global_init(&argc, argv);
630 // parse the platform file: get the host list
631 SIMIX_create_environment(argv[1]);
632 SIMIX_comm_set_copy_data_callback(&smpi_comm_copy_buffer_callback);
633 SIMIX_function_register_default(realmain);
634 SIMIX_launch_application(argv[2]);
640 smpi_check_options();
642 if(smpi_privatize_global_variables)
643 smpi_initialize_global_memory_segments();
645 /* Clean IO before the run */
650 MC_do_the_modelcheck_for_real();
654 if (sg_cfg_get_boolean("smpi/display_timing"))
655 XBT_INFO("Simulation time: %g seconds.", SIMIX_get_clock());
657 smpi_global_destroy();
666 // This function can be called from extern file, to initialize logs, options, and processes of smpi
667 // without the need of smpirun
672 smpi_check_options();
674 if (TRACE_is_enabled() && TRACE_is_configured()) {
678 if(smpi_privatize_global_variables)
679 smpi_initialize_global_memory_segments();
682 void SMPI_finalize(){
683 smpi_global_destroy();