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;
82 if(!index_to_process_data){
83 index_to_process_data=(int*)xbt_malloc(SIMIX_process_count()*sizeof(int));
85 MPI_Comm* temp_comm_world;
87 smpi_deployment_register_process(instance_id, rank, index, &temp_comm_world ,&temp_bar);
88 data = smpi_process_remote_data(index);
89 data->comm_world = temp_comm_world;
90 if(temp_bar != NULL) data->finalization_barrier = temp_bar;
92 data->instance_id = instance_id;
93 xbt_free(simcall_process_get_data(proc));
94 simcall_process_set_data(proc, data);
98 memmove(&(*argv)[1], &(*argv)[3], sizeof(char *) * (*argc - 3));
99 (*argv)[(*argc) - 1] = NULL;
100 (*argv)[(*argc) - 2] = NULL;
105 // set the process attached to the mailbox
106 simcall_rdv_set_receiver(data->mailbox_small, proc);
107 XBT_DEBUG("<%d> New process in the game: %p", index, proc);
109 if(smpi_privatize_global_variables){
110 switch_data_segment(index);
114 if (smpi_process_data() == NULL)
115 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.");
118 void smpi_process_destroy(void)
120 int index = smpi_process_index();
121 if(smpi_privatize_global_variables){
122 switch_data_segment(index);
124 process_data[index_to_process_data[index]]->state = SMPI_FINALIZED;
125 XBT_DEBUG("<%d> Process left the game", index);
129 * @brief Prepares the current process for termination.
131 void smpi_process_finalize(void)
133 // This leads to an explosion of the search graph
134 // which cannot be reduced:
138 int index = smpi_process_index();
139 // wait for all pending asynchronous comms to finish
140 xbt_barrier_wait(process_data[index_to_process_data[index]]->finalization_barrier);
144 * @brief Check if a process is finalized
146 int smpi_process_finalized()
148 int index = smpi_process_index();
149 if (index != MPI_UNDEFINED)
150 return (process_data[index_to_process_data[index]]->state == SMPI_FINALIZED);
156 * @brief Check if a process is initialized
158 int smpi_process_initialized(void)
160 int index = smpi_process_index();
161 return ( (index != MPI_UNDEFINED)
162 && (process_data[index_to_process_data[index]]->state == SMPI_INITIALIZED));
166 * @brief Mark a process as initialized (=MPI_Init called)
168 void smpi_process_mark_as_initialized(void)
170 int index = smpi_process_index();
171 if ((index != MPI_UNDEFINED) && (!process_data[index_to_process_data[index]]->state != SMPI_FINALIZED))
172 process_data[index_to_process_data[index]]->state = SMPI_INITIALIZED;
176 int smpi_global_size(void)
178 char *value = getenv("SMPI_GLOBAL_SIZE");
182 "Please set env var SMPI_GLOBAL_SIZE to expected number of processes.\n");
188 smpi_process_data_t smpi_process_data(void)
190 return SIMIX_process_self_get_data(SIMIX_process_self());
193 smpi_process_data_t smpi_process_remote_data(int index)
195 return process_data[index_to_process_data[index]];
198 void smpi_process_set_user_data(void *data)
200 smpi_process_data_t process_data = smpi_process_data();
201 process_data->data = data;
204 void *smpi_process_get_user_data()
206 smpi_process_data_t process_data = smpi_process_data();
207 return process_data->data;
210 int smpi_process_count(void)
212 return process_count;
215 int smpi_process_index(void)
217 smpi_process_data_t data = smpi_process_data();
218 //return -1 if not initialized
219 return data ? data->index : MPI_UNDEFINED;
222 MPI_Comm smpi_process_comm_world(void)
224 smpi_process_data_t data = smpi_process_data();
225 //return MPI_COMM_NULL if not initialized
226 return data ? *data->comm_world : MPI_COMM_NULL;
229 smx_rdv_t smpi_process_mailbox(void)
231 smpi_process_data_t data = smpi_process_data();
232 return data->mailbox;
235 smx_rdv_t smpi_process_mailbox_small(void)
237 smpi_process_data_t data = smpi_process_data();
238 return data->mailbox_small;
241 smx_rdv_t smpi_process_remote_mailbox(int index)
243 smpi_process_data_t data = smpi_process_remote_data(index);
244 return data->mailbox;
248 smx_rdv_t smpi_process_remote_mailbox_small(int index)
250 smpi_process_data_t data = smpi_process_remote_data(index);
251 return data->mailbox_small;
254 xbt_os_timer_t smpi_process_timer(void)
256 smpi_process_data_t data = smpi_process_data();
260 void smpi_process_simulated_start(void)
262 smpi_process_data_t data = smpi_process_data();
263 data->simulated = SIMIX_get_clock();
266 double smpi_process_simulated_elapsed(void)
268 smpi_process_data_t data = smpi_process_data();
269 return SIMIX_get_clock() - data->simulated;
272 MPI_Comm smpi_process_comm_self(void)
274 smpi_process_data_t data = smpi_process_data();
275 if(data->comm_self==MPI_COMM_NULL){
276 MPI_Group group = smpi_group_new(1);
277 data->comm_self = smpi_comm_new(group, NULL);
278 smpi_group_set_mapping(group, smpi_process_index(), 0);
281 return data->comm_self;
284 void smpi_process_set_sampling(int s)
286 smpi_process_data_t data = smpi_process_data();
290 int smpi_process_get_sampling(void)
292 smpi_process_data_t data = smpi_process_data();
293 return data->sampling;
296 void print_request(const char *message, MPI_Request request)
299 ("%s request %p [buf = %p, size = %zu, src = %d, dst = %d, tag = %d, flags = %x]",
300 message, request, request->buf, request->size, request->src,
301 request->dst, request->tag, request->flags);
304 void smpi_comm_copy_buffer_callback(smx_action_t comm,
305 void *buff, size_t buff_size)
307 XBT_DEBUG("Copy the data over");
308 if(_xbt_replay_is_active()) return;
311 if((smpi_privatize_global_variables)
312 && ((char*)buff >= start_data_exe)
313 && ((char*)buff < start_data_exe + size_data_exe )
315 XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !");
316 switch_data_segment(((smpi_process_data_t)SIMIX_process_get_data(comm->comm.src_proc))->index);
317 tmpbuff = (void*)xbt_malloc(buff_size);
318 memcpy(tmpbuff, buff, buff_size);
322 if((smpi_privatize_global_variables)
323 && ((char*)comm->comm.dst_buff >= start_data_exe)
324 && ((char*)comm->comm.dst_buff < start_data_exe + size_data_exe )
326 XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment");
327 switch_data_segment(((smpi_process_data_t)SIMIX_process_get_data(comm->comm.dst_proc))->index);
331 memcpy(comm->comm.dst_buff, tmpbuff, buff_size);
332 if (comm->comm.detached) {
333 // if this is a detached send, the source buffer was duplicated by SMPI
334 // sender to make the original buffer available to the application ASAP
336 //It seems that the request is used after the call there this should
337 //be free somewhereelse but where???
338 //xbt_free(comm->comm.src_data);// inside SMPI the request is keep
339 //inside the user data and should be free
340 comm->comm.src_buff = NULL;
343 if(tmpbuff!=buff)xbt_free(tmpbuff);
347 static void smpi_check_options(){
348 //check correctness of MPI parameters
350 xbt_assert(sg_cfg_get_int("smpi/async_small_thres") <=
351 sg_cfg_get_int("smpi/send_is_detached_thres"));
353 if (sg_cfg_is_default_value("smpi/running_power")) {
354 XBT_INFO("You did not set the power of the host running the simulation. "
355 "The timings will certainly not be accurate. "
356 "Use the option \"--cfg=smpi/running_power:<flops>\" to set its value."
357 "Check http://simgrid.org/simgrid/latest/doc/options.html#options_smpi_bench for more information.");
361 int smpi_enabled(void) {
362 return process_data != NULL;
365 void smpi_global_init(void)
369 char name[MAILBOX_NAME_MAXLEN];
373 if (process_count == 0){
374 process_count = SIMIX_process_count();
377 process_data = xbt_new0(smpi_process_data_t, process_count);
378 for (i = 0; i < process_count; i++) {
379 process_data[i] = xbt_new(s_smpi_process_data_t, 1);
380 //process_data[i]->index = i;
381 process_data[i]->argc = NULL;
382 process_data[i]->argv = NULL;
383 process_data[i]->mailbox = simcall_rdv_create(get_mailbox_name(name, i));
384 process_data[i]->mailbox_small =
385 simcall_rdv_create(get_mailbox_name_small(name, i));
386 process_data[i]->timer = xbt_os_timer_new();
388 MC_ignore_heap(process_data[i]->timer, xbt_os_timer_size());
389 process_data[i]->comm_self = MPI_COMM_NULL;
390 process_data[i]->comm_world = NULL;
391 process_data[i]->state = SMPI_UNINITIALIZED;
392 process_data[i]->sampling = 0;
393 process_data[i]->finalization_barrier = NULL;
395 //if the process was launched through smpirun script
396 //we generate a global mpi_comm_world
397 //if not, we let MPI_COMM_NULL, and the comm world
398 //will be private to each mpi instance
400 group = smpi_group_new(process_count);
401 MPI_COMM_WORLD = smpi_comm_new(group, NULL);
402 xbt_bar_t bar=xbt_barrier_init(process_count);
404 MPI_UNIVERSE_SIZE = smpi_comm_size(MPI_COMM_WORLD);
405 for (i = 0; i < process_count; i++) {
406 smpi_group_set_mapping(group, i, i);
407 process_data[i]->finalization_barrier = bar;
412 void smpi_global_destroy(void)
414 int count = smpi_process_count();
417 smpi_bench_destroy();
418 if (MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED){
419 while (smpi_group_unuse(smpi_comm_group(MPI_COMM_WORLD)) > 0);
420 xbt_free(MPI_COMM_WORLD);
421 xbt_barrier_destroy(process_data[0]->finalization_barrier);
423 smpi_deployment_cleanup_instances();
425 MPI_COMM_WORLD = MPI_COMM_NULL;
426 for (i = 0; i < count; i++) {
427 if(process_data[i]->comm_self!=MPI_COMM_NULL){
428 smpi_group_unuse(smpi_comm_group(process_data[i]->comm_self));
429 smpi_comm_destroy(process_data[i]->comm_self);
431 xbt_os_timer_free(process_data[i]->timer);
432 simcall_rdv_destroy(process_data[i]->mailbox);
433 simcall_rdv_destroy(process_data[i]->mailbox_small);
434 xbt_free(process_data[i]);
436 xbt_free(process_data);
439 xbt_free(index_to_process_data);
440 if(smpi_privatize_global_variables)
441 smpi_destroy_global_memory_segments();
446 void __attribute__ ((weak)) user_main_()
448 xbt_die("Should not be in this smpi_simulated_main");
452 int __attribute__ ((weak)) smpi_simulated_main_(int argc, char **argv)
454 smpi_process_init(&argc, &argv);
459 int __attribute__ ((weak)) main(int argc, char **argv)
461 return smpi_main(smpi_simulated_main_, argc, argv);
466 static void smpi_init_logs(){
468 /* Connect log categories. See xbt/log.c */
469 XBT_LOG_CONNECT(smpi); /* Keep this line as soon as possible in this
470 function: xbt_log_appender_file.c depends on it
471 DO NOT connect this in XBT or so, or it will be
472 useless to xbt_log_appender_file.c */
474 XBT_LOG_CONNECT(instr_smpi);
476 XBT_LOG_CONNECT(smpi_base);
477 XBT_LOG_CONNECT(smpi_bench);
478 XBT_LOG_CONNECT(smpi_coll);
479 XBT_LOG_CONNECT(smpi_colls);
480 XBT_LOG_CONNECT(smpi_comm);
481 XBT_LOG_CONNECT(smpi_dvfs);
482 XBT_LOG_CONNECT(smpi_group);
483 XBT_LOG_CONNECT(smpi_kernel);
484 XBT_LOG_CONNECT(smpi_mpi);
485 XBT_LOG_CONNECT(smpi_mpi_dt);
486 XBT_LOG_CONNECT(smpi_pmpi);
487 XBT_LOG_CONNECT(smpi_replay);
488 XBT_LOG_CONNECT(smpi_rma);
493 static void smpi_init_options(){
494 int gather_id = find_coll_description(mpi_coll_gather_description,
495 sg_cfg_get_string("smpi/gather"));
496 mpi_coll_gather_fun = (int (*)(void *, int, MPI_Datatype,
497 void *, int, MPI_Datatype, int, MPI_Comm))
498 mpi_coll_gather_description[gather_id].coll;
500 int allgather_id = find_coll_description(mpi_coll_allgather_description,
501 sg_cfg_get_string("smpi/allgather"));
502 mpi_coll_allgather_fun = (int (*)(void *, int, MPI_Datatype,
503 void *, int, MPI_Datatype, MPI_Comm))
504 mpi_coll_allgather_description[allgather_id].coll;
506 int allgatherv_id = find_coll_description(mpi_coll_allgatherv_description,
507 sg_cfg_get_string("smpi/allgatherv"));
508 mpi_coll_allgatherv_fun = (int (*)(void *, int, MPI_Datatype, void *, int *,
509 int *, MPI_Datatype, MPI_Comm))
510 mpi_coll_allgatherv_description[allgatherv_id].coll;
512 int allreduce_id = find_coll_description(mpi_coll_allreduce_description,
513 sg_cfg_get_string("smpi/allreduce"));
514 mpi_coll_allreduce_fun = (int (*)(void *sbuf, void *rbuf, int rcount,
515 MPI_Datatype dtype, MPI_Op op,
517 mpi_coll_allreduce_description[allreduce_id].coll;
519 int alltoall_id = find_coll_description(mpi_coll_alltoall_description,
520 sg_cfg_get_string("smpi/alltoall"));
521 mpi_coll_alltoall_fun = (int (*)(void *, int, MPI_Datatype,
522 void *, int, MPI_Datatype, MPI_Comm))
523 mpi_coll_alltoall_description[alltoall_id].coll;
525 int alltoallv_id = find_coll_description(mpi_coll_alltoallv_description,
526 sg_cfg_get_string("smpi/alltoallv"));
527 mpi_coll_alltoallv_fun = (int (*)(void *, int *, int *, MPI_Datatype,
528 void *, int *, int *, MPI_Datatype,
530 mpi_coll_alltoallv_description[alltoallv_id].coll;
532 int bcast_id = find_coll_description(mpi_coll_bcast_description,
533 sg_cfg_get_string("smpi/bcast"));
534 mpi_coll_bcast_fun = (int (*)(void *buf, int count, MPI_Datatype datatype,
535 int root, MPI_Comm com))
536 mpi_coll_bcast_description[bcast_id].coll;
538 int reduce_id = find_coll_description(mpi_coll_reduce_description,
539 sg_cfg_get_string("smpi/reduce"));
540 mpi_coll_reduce_fun = (int (*)(void *buf, void *rbuf, int count,
541 MPI_Datatype datatype, MPI_Op op,
542 int root, MPI_Comm comm))
543 mpi_coll_reduce_description[reduce_id].coll;
545 int reduce_scatter_id =
546 find_coll_description(mpi_coll_reduce_scatter_description,
547 sg_cfg_get_string("smpi/reduce_scatter"));
548 mpi_coll_reduce_scatter_fun = (int (*)(void *sbuf, void *rbuf, int *rcounts,
549 MPI_Datatype dtype, MPI_Op op,
551 mpi_coll_reduce_scatter_description[reduce_scatter_id].coll;
553 int scatter_id = find_coll_description(mpi_coll_scatter_description,
554 sg_cfg_get_string("smpi/scatter"));
555 mpi_coll_scatter_fun = (int (*)(void *sendbuf, int sendcount,
556 MPI_Datatype sendtype, void *recvbuf,
557 int recvcount, MPI_Datatype recvtype,
558 int root, MPI_Comm comm))
559 mpi_coll_scatter_description[scatter_id].coll;
561 int barrier_id = find_coll_description(mpi_coll_barrier_description,
562 sg_cfg_get_string("smpi/barrier"));
563 mpi_coll_barrier_fun = (int (*)(MPI_Comm comm))
564 mpi_coll_barrier_description[barrier_id].coll;
566 smpi_cpu_threshold = sg_cfg_get_double("smpi/cpu_threshold");
567 smpi_running_power = sg_cfg_get_double("smpi/running_power");
568 smpi_privatize_global_variables = sg_cfg_get_boolean("smpi/privatize_global_variables");
569 if (smpi_cpu_threshold < 0)
570 smpi_cpu_threshold = DBL_MAX;
574 int smpi_main(int (*realmain) (int argc, char *argv[]), int argc, char *argv[])
576 srand(SMPI_RAND_SEED);
578 if (getenv("SMPI_PRETEND_CC") != NULL) {
579 /* Hack to ensure that smpicc can pretend to be a simple
580 * compiler. Particularly handy to pass it to the configuration tools */
587 TRACE_global_init(&argc, argv);
589 TRACE_add_start_function(TRACE_smpi_alloc);
590 TRACE_add_end_function(TRACE_smpi_release);
593 SIMIX_global_init(&argc, argv);
595 // parse the platform file: get the host list
596 SIMIX_create_environment(argv[1]);
597 SIMIX_comm_set_copy_data_callback(&smpi_comm_copy_buffer_callback);
598 SIMIX_function_register_default(realmain);
599 SIMIX_launch_application(argv[2]);
605 smpi_check_options();
607 if(smpi_privatize_global_variables)
608 smpi_initialize_global_memory_segments();
610 /* Clean IO before the run */
615 MC_do_the_modelcheck_for_real();
619 if (sg_cfg_get_boolean("smpi/display_timing"))
620 XBT_INFO("Simulation time: %g seconds.", SIMIX_get_clock());
622 smpi_global_destroy();
631 // This function can be called from extern file, to initialize logs, options, and processes of smpi
632 // without the need of smpirun
637 smpi_check_options();
639 if (TRACE_is_enabled() && TRACE_is_configured()) {
643 if(smpi_privatize_global_variables)
644 smpi_initialize_global_memory_segments();
647 void SMPI_finalize(){
648 smpi_global_destroy();