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 int smpi_process_index_of_smx_process(smx_process_t process) {
223 smpi_process_data_t data = SIMIX_process_get_data(process);
224 return data ? data->index : MPI_UNDEFINED;
227 MPI_Comm smpi_process_comm_world(void)
229 smpi_process_data_t data = smpi_process_data();
230 //return MPI_COMM_NULL if not initialized
231 return data ? *data->comm_world : MPI_COMM_NULL;
234 smx_rdv_t smpi_process_mailbox(void)
236 smpi_process_data_t data = smpi_process_data();
237 return data->mailbox;
240 smx_rdv_t smpi_process_mailbox_small(void)
242 smpi_process_data_t data = smpi_process_data();
243 return data->mailbox_small;
246 smx_rdv_t smpi_process_remote_mailbox(int index)
248 smpi_process_data_t data = smpi_process_remote_data(index);
249 return data->mailbox;
253 smx_rdv_t smpi_process_remote_mailbox_small(int index)
255 smpi_process_data_t data = smpi_process_remote_data(index);
256 return data->mailbox_small;
259 xbt_os_timer_t smpi_process_timer(void)
261 smpi_process_data_t data = smpi_process_data();
265 void smpi_process_simulated_start(void)
267 smpi_process_data_t data = smpi_process_data();
268 data->simulated = SIMIX_get_clock();
271 double smpi_process_simulated_elapsed(void)
273 smpi_process_data_t data = smpi_process_data();
274 return SIMIX_get_clock() - data->simulated;
277 MPI_Comm smpi_process_comm_self(void)
279 smpi_process_data_t data = smpi_process_data();
280 if(data->comm_self==MPI_COMM_NULL){
281 MPI_Group group = smpi_group_new(1);
282 data->comm_self = smpi_comm_new(group, NULL);
283 smpi_group_set_mapping(group, smpi_process_index(), 0);
286 return data->comm_self;
289 void smpi_process_set_sampling(int s)
291 smpi_process_data_t data = smpi_process_data();
295 int smpi_process_get_sampling(void)
297 smpi_process_data_t data = smpi_process_data();
298 return data->sampling;
301 void print_request(const char *message, MPI_Request request)
304 ("%s request %p [buf = %p, size = %zu, src = %d, dst = %d, tag = %d, flags = %x]",
305 message, request, request->buf, request->size, request->src,
306 request->dst, request->tag, request->flags);
309 void smpi_comm_copy_buffer_callback(smx_action_t comm,
310 void *buff, size_t buff_size)
312 XBT_DEBUG("Copy the data over");
313 if(_xbt_replay_is_active()) return;
316 if((smpi_privatize_global_variables)
317 && ((char*)buff >= start_data_exe)
318 && ((char*)buff < start_data_exe + size_data_exe )
320 XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !");
321 switch_data_segment(((smpi_process_data_t)SIMIX_process_get_data(comm->comm.src_proc))->index);
322 tmpbuff = (void*)xbt_malloc(buff_size);
323 memcpy(tmpbuff, buff, buff_size);
327 if((smpi_privatize_global_variables)
328 && ((char*)comm->comm.dst_buff >= start_data_exe)
329 && ((char*)comm->comm.dst_buff < start_data_exe + size_data_exe )
331 XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment");
332 switch_data_segment(((smpi_process_data_t)SIMIX_process_get_data(comm->comm.dst_proc))->index);
336 memcpy(comm->comm.dst_buff, tmpbuff, buff_size);
337 if (comm->comm.detached) {
338 // if this is a detached send, the source buffer was duplicated by SMPI
339 // sender to make the original buffer available to the application ASAP
341 //It seems that the request is used after the call there this should
342 //be free somewhereelse but where???
343 //xbt_free(comm->comm.src_data);// inside SMPI the request is keep
344 //inside the user data and should be free
345 comm->comm.src_buff = NULL;
348 if(tmpbuff!=buff)xbt_free(tmpbuff);
352 static void smpi_check_options(){
353 //check correctness of MPI parameters
355 xbt_assert(sg_cfg_get_int("smpi/async_small_thres") <=
356 sg_cfg_get_int("smpi/send_is_detached_thres"));
358 if (sg_cfg_is_default_value("smpi/running_power")) {
359 XBT_INFO("You did not set the power of the host running the simulation. "
360 "The timings will certainly not be accurate. "
361 "Use the option \"--cfg=smpi/running_power:<flops>\" to set its value."
362 "Check http://simgrid.org/simgrid/latest/doc/options.html#options_smpi_bench for more information.");
366 int smpi_enabled(void) {
367 return process_data != NULL;
370 void smpi_global_init(void)
374 char name[MAILBOX_NAME_MAXLEN];
378 if (process_count == 0){
379 process_count = SIMIX_process_count();
382 process_data = xbt_new0(smpi_process_data_t, process_count);
383 for (i = 0; i < process_count; i++) {
384 process_data[i] = xbt_new(s_smpi_process_data_t, 1);
385 //process_data[i]->index = i;
386 process_data[i]->argc = NULL;
387 process_data[i]->argv = NULL;
388 process_data[i]->mailbox = simcall_rdv_create(get_mailbox_name(name, i));
389 process_data[i]->mailbox_small =
390 simcall_rdv_create(get_mailbox_name_small(name, i));
391 process_data[i]->timer = xbt_os_timer_new();
393 MC_ignore_heap(process_data[i]->timer, xbt_os_timer_size());
394 process_data[i]->comm_self = MPI_COMM_NULL;
395 process_data[i]->comm_world = NULL;
396 process_data[i]->state = SMPI_UNINITIALIZED;
397 process_data[i]->sampling = 0;
398 process_data[i]->finalization_barrier = NULL;
400 //if the process was launched through smpirun script
401 //we generate a global mpi_comm_world
402 //if not, we let MPI_COMM_NULL, and the comm world
403 //will be private to each mpi instance
405 group = smpi_group_new(process_count);
406 MPI_COMM_WORLD = smpi_comm_new(group, NULL);
407 xbt_bar_t bar=xbt_barrier_init(process_count);
409 MPI_UNIVERSE_SIZE = smpi_comm_size(MPI_COMM_WORLD);
410 for (i = 0; i < process_count; i++) {
411 smpi_group_set_mapping(group, i, i);
412 process_data[i]->finalization_barrier = bar;
417 void smpi_global_destroy(void)
419 int count = smpi_process_count();
422 smpi_bench_destroy();
423 if (MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED){
424 while (smpi_group_unuse(smpi_comm_group(MPI_COMM_WORLD)) > 0);
425 xbt_free(MPI_COMM_WORLD);
426 xbt_barrier_destroy(process_data[0]->finalization_barrier);
428 smpi_deployment_cleanup_instances();
430 MPI_COMM_WORLD = MPI_COMM_NULL;
431 for (i = 0; i < count; i++) {
432 if(process_data[i]->comm_self!=MPI_COMM_NULL){
433 smpi_group_unuse(smpi_comm_group(process_data[i]->comm_self));
434 smpi_comm_destroy(process_data[i]->comm_self);
436 xbt_os_timer_free(process_data[i]->timer);
437 simcall_rdv_destroy(process_data[i]->mailbox);
438 simcall_rdv_destroy(process_data[i]->mailbox_small);
439 xbt_free(process_data[i]);
441 xbt_free(process_data);
444 xbt_free(index_to_process_data);
445 if(smpi_privatize_global_variables)
446 smpi_destroy_global_memory_segments();
451 void __attribute__ ((weak)) user_main_()
453 xbt_die("Should not be in this smpi_simulated_main");
457 int __attribute__ ((weak)) smpi_simulated_main_(int argc, char **argv)
459 smpi_process_init(&argc, &argv);
464 int __attribute__ ((weak)) main(int argc, char **argv)
466 return smpi_main(smpi_simulated_main_, argc, argv);
471 static void smpi_init_logs(){
473 /* Connect log categories. See xbt/log.c */
474 XBT_LOG_CONNECT(smpi); /* Keep this line as soon as possible in this
475 function: xbt_log_appender_file.c depends on it
476 DO NOT connect this in XBT or so, or it will be
477 useless to xbt_log_appender_file.c */
479 XBT_LOG_CONNECT(instr_smpi);
481 XBT_LOG_CONNECT(smpi_base);
482 XBT_LOG_CONNECT(smpi_bench);
483 XBT_LOG_CONNECT(smpi_coll);
484 XBT_LOG_CONNECT(smpi_colls);
485 XBT_LOG_CONNECT(smpi_comm);
486 XBT_LOG_CONNECT(smpi_dvfs);
487 XBT_LOG_CONNECT(smpi_group);
488 XBT_LOG_CONNECT(smpi_kernel);
489 XBT_LOG_CONNECT(smpi_mpi);
490 XBT_LOG_CONNECT(smpi_mpi_dt);
491 XBT_LOG_CONNECT(smpi_pmpi);
492 XBT_LOG_CONNECT(smpi_replay);
493 XBT_LOG_CONNECT(smpi_rma);
498 static void smpi_init_options(){
499 int gather_id = find_coll_description(mpi_coll_gather_description,
500 sg_cfg_get_string("smpi/gather"));
501 mpi_coll_gather_fun = (int (*)(void *, int, MPI_Datatype,
502 void *, int, MPI_Datatype, int, MPI_Comm))
503 mpi_coll_gather_description[gather_id].coll;
505 int allgather_id = find_coll_description(mpi_coll_allgather_description,
506 sg_cfg_get_string("smpi/allgather"));
507 mpi_coll_allgather_fun = (int (*)(void *, int, MPI_Datatype,
508 void *, int, MPI_Datatype, MPI_Comm))
509 mpi_coll_allgather_description[allgather_id].coll;
511 int allgatherv_id = find_coll_description(mpi_coll_allgatherv_description,
512 sg_cfg_get_string("smpi/allgatherv"));
513 mpi_coll_allgatherv_fun = (int (*)(void *, int, MPI_Datatype, void *, int *,
514 int *, MPI_Datatype, MPI_Comm))
515 mpi_coll_allgatherv_description[allgatherv_id].coll;
517 int allreduce_id = find_coll_description(mpi_coll_allreduce_description,
518 sg_cfg_get_string("smpi/allreduce"));
519 mpi_coll_allreduce_fun = (int (*)(void *sbuf, void *rbuf, int rcount,
520 MPI_Datatype dtype, MPI_Op op,
522 mpi_coll_allreduce_description[allreduce_id].coll;
524 int alltoall_id = find_coll_description(mpi_coll_alltoall_description,
525 sg_cfg_get_string("smpi/alltoall"));
526 mpi_coll_alltoall_fun = (int (*)(void *, int, MPI_Datatype,
527 void *, int, MPI_Datatype, MPI_Comm))
528 mpi_coll_alltoall_description[alltoall_id].coll;
530 int alltoallv_id = find_coll_description(mpi_coll_alltoallv_description,
531 sg_cfg_get_string("smpi/alltoallv"));
532 mpi_coll_alltoallv_fun = (int (*)(void *, int *, int *, MPI_Datatype,
533 void *, int *, int *, MPI_Datatype,
535 mpi_coll_alltoallv_description[alltoallv_id].coll;
537 int bcast_id = find_coll_description(mpi_coll_bcast_description,
538 sg_cfg_get_string("smpi/bcast"));
539 mpi_coll_bcast_fun = (int (*)(void *buf, int count, MPI_Datatype datatype,
540 int root, MPI_Comm com))
541 mpi_coll_bcast_description[bcast_id].coll;
543 int reduce_id = find_coll_description(mpi_coll_reduce_description,
544 sg_cfg_get_string("smpi/reduce"));
545 mpi_coll_reduce_fun = (int (*)(void *buf, void *rbuf, int count,
546 MPI_Datatype datatype, MPI_Op op,
547 int root, MPI_Comm comm))
548 mpi_coll_reduce_description[reduce_id].coll;
550 int reduce_scatter_id =
551 find_coll_description(mpi_coll_reduce_scatter_description,
552 sg_cfg_get_string("smpi/reduce_scatter"));
553 mpi_coll_reduce_scatter_fun = (int (*)(void *sbuf, void *rbuf, int *rcounts,
554 MPI_Datatype dtype, MPI_Op op,
556 mpi_coll_reduce_scatter_description[reduce_scatter_id].coll;
558 int scatter_id = find_coll_description(mpi_coll_scatter_description,
559 sg_cfg_get_string("smpi/scatter"));
560 mpi_coll_scatter_fun = (int (*)(void *sendbuf, int sendcount,
561 MPI_Datatype sendtype, void *recvbuf,
562 int recvcount, MPI_Datatype recvtype,
563 int root, MPI_Comm comm))
564 mpi_coll_scatter_description[scatter_id].coll;
566 int barrier_id = find_coll_description(mpi_coll_barrier_description,
567 sg_cfg_get_string("smpi/barrier"));
568 mpi_coll_barrier_fun = (int (*)(MPI_Comm comm))
569 mpi_coll_barrier_description[barrier_id].coll;
571 smpi_cpu_threshold = sg_cfg_get_double("smpi/cpu_threshold");
572 smpi_running_power = sg_cfg_get_double("smpi/running_power");
573 smpi_privatize_global_variables = sg_cfg_get_boolean("smpi/privatize_global_variables");
574 if (smpi_cpu_threshold < 0)
575 smpi_cpu_threshold = DBL_MAX;
579 int smpi_main(int (*realmain) (int argc, char *argv[]), int argc, char *argv[])
581 srand(SMPI_RAND_SEED);
583 if (getenv("SMPI_PRETEND_CC") != NULL) {
584 /* Hack to ensure that smpicc can pretend to be a simple
585 * compiler. Particularly handy to pass it to the configuration tools */
592 TRACE_global_init(&argc, argv);
594 TRACE_add_start_function(TRACE_smpi_alloc);
595 TRACE_add_end_function(TRACE_smpi_release);
598 SIMIX_global_init(&argc, argv);
600 // parse the platform file: get the host list
601 SIMIX_create_environment(argv[1]);
602 SIMIX_comm_set_copy_data_callback(&smpi_comm_copy_buffer_callback);
603 SIMIX_function_register_default(realmain);
604 SIMIX_launch_application(argv[2]);
610 smpi_check_options();
612 if(smpi_privatize_global_variables)
613 smpi_initialize_global_memory_segments();
615 /* Clean IO before the run */
620 MC_do_the_modelcheck_for_real();
624 if (sg_cfg_get_boolean("smpi/display_timing"))
625 XBT_INFO("Simulation time: %g seconds.", SIMIX_get_clock());
627 smpi_global_destroy();
636 // This function can be called from extern file, to initialize logs, options, and processes of smpi
637 // without the need of smpirun
642 smpi_check_options();
644 if (TRACE_is_enabled() && TRACE_is_configured()) {
648 if(smpi_privatize_global_variables)
649 smpi_initialize_global_memory_segments();
652 void SMPI_finalize(){
653 smpi_global_destroy();