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;
33 void *data; /* user data */
36 int sampling; /* inside an SMPI_SAMPLE_ block? */
38 xbt_bar_t finalization_barrier;
39 } s_smpi_process_data_t;
41 static smpi_process_data_t *process_data = NULL;
42 int process_count = 0;
43 int* index_to_process_data = NULL;
46 MPI_Comm MPI_COMM_WORLD = MPI_COMM_UNINITIALIZED;
47 int MPI_UNIVERSE_SIZE;
49 MPI_Errhandler *MPI_ERRORS_RETURN = NULL;
50 MPI_Errhandler *MPI_ERRORS_ARE_FATAL = NULL;
51 MPI_Errhandler *MPI_ERRHANDLER_NULL = NULL;
53 #define MAILBOX_NAME_MAXLEN (5 + sizeof(int) * 2 + 1)
55 static char *get_mailbox_name(char *str, int index)
57 snprintf(str, MAILBOX_NAME_MAXLEN, "SMPI-%0*x", (int) (sizeof(int) * 2),
62 static char *get_mailbox_name_small(char *str, int index)
64 snprintf(str, MAILBOX_NAME_MAXLEN, "small%0*x", (int) (sizeof(int) * 2),
69 void smpi_process_init(int *argc, char ***argv)
72 smpi_process_data_t data;
76 proc = SIMIX_process_self();
77 //FIXME: dirty cleanup method to avoid using msg cleanup functions on these processes when using MSG+SMPI
78 proc->context->cleanup_func=SIMIX_process_cleanup;
79 char* instance_id = (*argv)[1];
80 int rank = atoi((*argv)[2]);
81 index = smpi_process_index_of_smx_process(proc);
83 if(!index_to_process_data){
84 index_to_process_data=(int*)xbt_malloc(SIMIX_process_count()*sizeof(int));
86 MPI_Comm* temp_comm_world;
88 smpi_deployment_register_process(instance_id, rank, index, &temp_comm_world ,&temp_bar);
89 data = smpi_process_remote_data(index);
90 data->comm_world = temp_comm_world;
91 if(temp_bar != NULL) data->finalization_barrier = temp_bar;
93 data->instance_id = instance_id;
94 xbt_free(simcall_process_get_data(proc));
95 simcall_process_set_data(proc, data);
99 memmove(&(*argv)[1], &(*argv)[3], sizeof(char *) * (*argc - 3));
100 (*argv)[(*argc) - 1] = NULL;
101 (*argv)[(*argc) - 2] = NULL;
106 // set the process attached to the mailbox
107 simcall_rdv_set_receiver(data->mailbox_small, proc);
109 XBT_DEBUG("<%d> New process in the game: %p", index, proc);
111 if(smpi_privatize_global_variables){
112 smpi_switch_data_segment(index);
116 if (smpi_process_data() == NULL)
117 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.");
120 void smpi_process_destroy(void)
122 int index = smpi_process_index();
123 if(smpi_privatize_global_variables){
124 smpi_switch_data_segment(index);
126 process_data[index_to_process_data[index]]->state = SMPI_FINALIZED;
127 XBT_DEBUG("<%d> Process left the game", index);
131 * @brief Prepares the current process for termination.
133 void smpi_process_finalize(void)
135 // This leads to an explosion of the search graph
136 // which cannot be reduced:
140 int index = smpi_process_index();
141 // wait for all pending asynchronous comms to finish
142 xbt_barrier_wait(process_data[index_to_process_data[index]]->finalization_barrier);
146 * @brief Check if a process is finalized
148 int smpi_process_finalized()
150 int index = smpi_process_index();
151 if (index != MPI_UNDEFINED)
152 return (process_data[index_to_process_data[index]]->state == SMPI_FINALIZED);
158 * @brief Check if a process is initialized
160 int smpi_process_initialized(void)
162 int index = smpi_process_index();
163 return ( (index != MPI_UNDEFINED)
164 && (process_data[index_to_process_data[index]]->state == SMPI_INITIALIZED));
168 * @brief Mark a process as initialized (=MPI_Init called)
170 void smpi_process_mark_as_initialized(void)
172 int index = smpi_process_index();
173 if ((index != MPI_UNDEFINED) && (!process_data[index_to_process_data[index]]->state != SMPI_FINALIZED))
174 process_data[index_to_process_data[index]]->state = SMPI_INITIALIZED;
178 int smpi_global_size(void)
180 char *value = getenv("SMPI_GLOBAL_SIZE");
184 "Please set env var SMPI_GLOBAL_SIZE to expected number of processes.\n");
190 smpi_process_data_t smpi_process_data(void)
192 return SIMIX_process_self_get_data(SIMIX_process_self());
195 smpi_process_data_t smpi_process_remote_data(int index)
197 return process_data[index_to_process_data[index]];
200 void smpi_process_set_user_data(void *data)
202 smpi_process_data_t process_data = smpi_process_data();
203 process_data->data = data;
206 void *smpi_process_get_user_data()
208 smpi_process_data_t process_data = smpi_process_data();
209 return process_data->data;
212 int smpi_process_count(void)
214 return process_count;
217 int smpi_process_index(void)
219 smpi_process_data_t data = smpi_process_data();
220 //return -1 if not initialized
221 return data ? data->index : MPI_UNDEFINED;
224 MPI_Comm smpi_process_comm_world(void)
226 smpi_process_data_t data = smpi_process_data();
227 //return MPI_COMM_NULL if not initialized
228 return data ? *data->comm_world : MPI_COMM_NULL;
231 smx_rdv_t smpi_process_mailbox(void)
233 smpi_process_data_t data = smpi_process_data();
234 return data->mailbox;
237 smx_rdv_t smpi_process_mailbox_small(void)
239 smpi_process_data_t data = smpi_process_data();
240 return data->mailbox_small;
243 smx_rdv_t smpi_process_remote_mailbox(int index)
245 smpi_process_data_t data = smpi_process_remote_data(index);
246 return data->mailbox;
250 smx_rdv_t smpi_process_remote_mailbox_small(int index)
252 smpi_process_data_t data = smpi_process_remote_data(index);
253 return data->mailbox_small;
256 xbt_os_timer_t smpi_process_timer(void)
258 smpi_process_data_t data = smpi_process_data();
262 void smpi_process_simulated_start(void)
264 smpi_process_data_t data = smpi_process_data();
265 data->simulated = SIMIX_get_clock();
268 double smpi_process_simulated_elapsed(void)
270 smpi_process_data_t data = smpi_process_data();
271 return SIMIX_get_clock() - data->simulated;
274 MPI_Comm smpi_process_comm_self(void)
276 smpi_process_data_t data = smpi_process_data();
277 if(data->comm_self==MPI_COMM_NULL){
278 MPI_Group group = smpi_group_new(1);
279 data->comm_self = smpi_comm_new(group, NULL);
280 smpi_group_set_mapping(group, smpi_process_index(), 0);
283 return data->comm_self;
286 MPI_Comm smpi_process_get_comm_intra(void)
288 smpi_process_data_t data = smpi_process_data();
289 return data->comm_intra;
292 void smpi_process_set_comm_intra(MPI_Comm comm)
294 smpi_process_data_t data = smpi_process_data();
295 data->comm_intra = comm;
298 void smpi_process_set_sampling(int s)
300 smpi_process_data_t data = smpi_process_data();
304 int smpi_process_get_sampling(void)
306 smpi_process_data_t data = smpi_process_data();
307 return data->sampling;
311 void print_request(const char *message, MPI_Request request)
314 ("%s request %p [buf = %p, size = %zu, src = %d, dst = %d, tag = %d, flags = %x]",
315 message, request, request->buf, request->size, request->src,
316 request->dst, request->tag, request->flags);
319 void smpi_comm_copy_buffer_callback(smx_action_t comm,
320 void *buff, size_t buff_size)
322 XBT_DEBUG("Copy the data over");
323 if(_xbt_replay_is_active()) return;
326 if((smpi_privatize_global_variables)
327 && ((char*)buff >= start_data_exe)
328 && ((char*)buff < start_data_exe + size_data_exe )
330 XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !");
331 smpi_switch_data_segment(((smpi_process_data_t)SIMIX_process_get_data(comm->comm.src_proc))->index);
332 tmpbuff = (void*)xbt_malloc(buff_size);
333 memcpy(tmpbuff, buff, buff_size);
337 if((smpi_privatize_global_variables)
338 && ((char*)comm->comm.dst_buff >= start_data_exe)
339 && ((char*)comm->comm.dst_buff < start_data_exe + size_data_exe )
341 XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment");
342 smpi_switch_data_segment(((smpi_process_data_t)SIMIX_process_get_data(comm->comm.dst_proc))->index);
346 memcpy(comm->comm.dst_buff, tmpbuff, buff_size);
347 if (comm->comm.detached) {
348 // if this is a detached send, the source buffer was duplicated by SMPI
349 // sender to make the original buffer available to the application ASAP
351 //It seems that the request is used after the call there this should
352 //be free somewhereelse but where???
353 //xbt_free(comm->comm.src_data);// inside SMPI the request is keep
354 //inside the user data and should be free
355 comm->comm.src_buff = NULL;
358 if(tmpbuff!=buff)xbt_free(tmpbuff);
362 static void smpi_check_options(){
363 //check correctness of MPI parameters
365 xbt_assert(sg_cfg_get_int("smpi/async_small_thres") <=
366 sg_cfg_get_int("smpi/send_is_detached_thres"));
368 if (sg_cfg_is_default_value("smpi/running_power")) {
369 XBT_INFO("You did not set the power of the host running the simulation. "
370 "The timings will certainly not be accurate. "
371 "Use the option \"--cfg=smpi/running_power:<flops>\" to set its value."
372 "Check http://simgrid.org/simgrid/latest/doc/options.html#options_smpi_bench for more information.");
376 int smpi_enabled(void) {
377 return process_data != NULL;
380 void smpi_global_init(void)
384 char name[MAILBOX_NAME_MAXLEN];
388 if (process_count == 0){
389 process_count = SIMIX_process_count();
392 process_data = xbt_new0(smpi_process_data_t, process_count);
393 for (i = 0; i < process_count; i++) {
394 process_data[i] = xbt_new(s_smpi_process_data_t, 1);
395 //process_data[i]->index = i;
396 process_data[i]->argc = NULL;
397 process_data[i]->argv = NULL;
398 process_data[i]->mailbox = simcall_rdv_create(get_mailbox_name(name, i));
399 process_data[i]->mailbox_small =
400 simcall_rdv_create(get_mailbox_name_small(name, i));
401 process_data[i]->timer = xbt_os_timer_new();
403 MC_ignore_heap(process_data[i]->timer, xbt_os_timer_size());
404 process_data[i]->comm_self = MPI_COMM_NULL;
405 process_data[i]->comm_intra = MPI_COMM_NULL;
406 process_data[i]->comm_world = NULL;
407 process_data[i]->state = SMPI_UNINITIALIZED;
408 process_data[i]->sampling = 0;
409 process_data[i]->finalization_barrier = NULL;
411 //if the process was launched through smpirun script
412 //we generate a global mpi_comm_world
413 //if not, we let MPI_COMM_NULL, and the comm world
414 //will be private to each mpi instance
416 group = smpi_group_new(process_count);
417 MPI_COMM_WORLD = smpi_comm_new(group, NULL);
418 xbt_bar_t bar=xbt_barrier_init(process_count);
420 MPI_UNIVERSE_SIZE = smpi_comm_size(MPI_COMM_WORLD);
421 for (i = 0; i < process_count; i++) {
422 smpi_group_set_mapping(group, i, i);
423 process_data[i]->finalization_barrier = bar;
428 void smpi_global_destroy(void)
430 int count = smpi_process_count();
433 smpi_bench_destroy();
434 if (MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED){
435 while (smpi_group_unuse(smpi_comm_group(MPI_COMM_WORLD)) > 0);
436 xbt_free(MPI_COMM_WORLD);
437 xbt_barrier_destroy(process_data[0]->finalization_barrier);
439 smpi_deployment_cleanup_instances();
441 MPI_COMM_WORLD = MPI_COMM_NULL;
442 for (i = 0; i < count; i++) {
443 if(process_data[i]->comm_self!=MPI_COMM_NULL){
444 smpi_group_unuse(smpi_comm_group(process_data[i]->comm_self));
445 smpi_comm_destroy(process_data[i]->comm_self);
447 if(process_data[i]->comm_intra!=MPI_COMM_NULL){
448 smpi_group_unuse(smpi_comm_group(process_data[i]->comm_intra));
449 smpi_comm_destroy(process_data[i]->comm_intra);
451 xbt_os_timer_free(process_data[i]->timer);
452 simcall_rdv_destroy(process_data[i]->mailbox);
453 simcall_rdv_destroy(process_data[i]->mailbox_small);
454 xbt_free(process_data[i]);
456 xbt_free(process_data);
459 xbt_free(index_to_process_data);
460 if(smpi_privatize_global_variables)
461 smpi_destroy_global_memory_segments();
466 void __attribute__ ((weak)) user_main_()
468 xbt_die("Should not be in this smpi_simulated_main");
472 int __attribute__ ((weak)) smpi_simulated_main_(int argc, char **argv)
474 smpi_process_init(&argc, &argv);
479 int __attribute__ ((weak)) main(int argc, char **argv)
481 return smpi_main(smpi_simulated_main_, argc, argv);
486 static void smpi_init_logs(){
488 /* Connect log categories. See xbt/log.c */
489 XBT_LOG_CONNECT(smpi); /* Keep this line as soon as possible in this
490 function: xbt_log_appender_file.c depends on it
491 DO NOT connect this in XBT or so, or it will be
492 useless to xbt_log_appender_file.c */
494 XBT_LOG_CONNECT(instr_smpi);
496 XBT_LOG_CONNECT(smpi_base);
497 XBT_LOG_CONNECT(smpi_bench);
498 XBT_LOG_CONNECT(smpi_coll);
499 XBT_LOG_CONNECT(smpi_colls);
500 XBT_LOG_CONNECT(smpi_comm);
501 XBT_LOG_CONNECT(smpi_dvfs);
502 XBT_LOG_CONNECT(smpi_group);
503 XBT_LOG_CONNECT(smpi_kernel);
504 XBT_LOG_CONNECT(smpi_mpi);
505 XBT_LOG_CONNECT(smpi_mpi_dt);
506 XBT_LOG_CONNECT(smpi_pmpi);
507 XBT_LOG_CONNECT(smpi_replay);
508 XBT_LOG_CONNECT(smpi_rma);
513 static void smpi_init_options(){
514 int gather_id = find_coll_description(mpi_coll_gather_description,
515 sg_cfg_get_string("smpi/gather"));
516 mpi_coll_gather_fun = (int (*)(void *, int, MPI_Datatype,
517 void *, int, MPI_Datatype, int, MPI_Comm))
518 mpi_coll_gather_description[gather_id].coll;
520 int allgather_id = find_coll_description(mpi_coll_allgather_description,
521 sg_cfg_get_string("smpi/allgather"));
522 mpi_coll_allgather_fun = (int (*)(void *, int, MPI_Datatype,
523 void *, int, MPI_Datatype, MPI_Comm))
524 mpi_coll_allgather_description[allgather_id].coll;
526 int allgatherv_id = find_coll_description(mpi_coll_allgatherv_description,
527 sg_cfg_get_string("smpi/allgatherv"));
528 mpi_coll_allgatherv_fun = (int (*)(void *, int, MPI_Datatype, void *, int *,
529 int *, MPI_Datatype, MPI_Comm))
530 mpi_coll_allgatherv_description[allgatherv_id].coll;
532 int allreduce_id = find_coll_description(mpi_coll_allreduce_description,
533 sg_cfg_get_string("smpi/allreduce"));
534 mpi_coll_allreduce_fun = (int (*)(void *sbuf, void *rbuf, int rcount,
535 MPI_Datatype dtype, MPI_Op op,
537 mpi_coll_allreduce_description[allreduce_id].coll;
539 int alltoall_id = find_coll_description(mpi_coll_alltoall_description,
540 sg_cfg_get_string("smpi/alltoall"));
541 mpi_coll_alltoall_fun = (int (*)(void *, int, MPI_Datatype,
542 void *, int, MPI_Datatype, MPI_Comm))
543 mpi_coll_alltoall_description[alltoall_id].coll;
545 int alltoallv_id = find_coll_description(mpi_coll_alltoallv_description,
546 sg_cfg_get_string("smpi/alltoallv"));
547 mpi_coll_alltoallv_fun = (int (*)(void *, int *, int *, MPI_Datatype,
548 void *, int *, int *, MPI_Datatype,
550 mpi_coll_alltoallv_description[alltoallv_id].coll;
552 int bcast_id = find_coll_description(mpi_coll_bcast_description,
553 sg_cfg_get_string("smpi/bcast"));
554 mpi_coll_bcast_fun = (int (*)(void *buf, int count, MPI_Datatype datatype,
555 int root, MPI_Comm com))
556 mpi_coll_bcast_description[bcast_id].coll;
558 int reduce_id = find_coll_description(mpi_coll_reduce_description,
559 sg_cfg_get_string("smpi/reduce"));
560 mpi_coll_reduce_fun = (int (*)(void *buf, void *rbuf, int count,
561 MPI_Datatype datatype, MPI_Op op,
562 int root, MPI_Comm comm))
563 mpi_coll_reduce_description[reduce_id].coll;
565 int reduce_scatter_id =
566 find_coll_description(mpi_coll_reduce_scatter_description,
567 sg_cfg_get_string("smpi/reduce_scatter"));
568 mpi_coll_reduce_scatter_fun = (int (*)(void *sbuf, void *rbuf, int *rcounts,
569 MPI_Datatype dtype, MPI_Op op,
571 mpi_coll_reduce_scatter_description[reduce_scatter_id].coll;
573 int scatter_id = find_coll_description(mpi_coll_scatter_description,
574 sg_cfg_get_string("smpi/scatter"));
575 mpi_coll_scatter_fun = (int (*)(void *sendbuf, int sendcount,
576 MPI_Datatype sendtype, void *recvbuf,
577 int recvcount, MPI_Datatype recvtype,
578 int root, MPI_Comm comm))
579 mpi_coll_scatter_description[scatter_id].coll;
581 int barrier_id = find_coll_description(mpi_coll_barrier_description,
582 sg_cfg_get_string("smpi/barrier"));
583 mpi_coll_barrier_fun = (int (*)(MPI_Comm comm))
584 mpi_coll_barrier_description[barrier_id].coll;
586 smpi_cpu_threshold = sg_cfg_get_double("smpi/cpu_threshold");
587 smpi_running_power = sg_cfg_get_double("smpi/running_power");
588 smpi_privatize_global_variables = sg_cfg_get_boolean("smpi/privatize_global_variables");
589 if (smpi_cpu_threshold < 0)
590 smpi_cpu_threshold = DBL_MAX;
594 int smpi_main(int (*realmain) (int argc, char *argv[]), int argc, char *argv[])
596 srand(SMPI_RAND_SEED);
598 if (getenv("SMPI_PRETEND_CC") != NULL) {
599 /* Hack to ensure that smpicc can pretend to be a simple
600 * compiler. Particularly handy to pass it to the configuration tools */
607 TRACE_global_init(&argc, argv);
609 TRACE_add_start_function(TRACE_smpi_alloc);
610 TRACE_add_end_function(TRACE_smpi_release);
613 SIMIX_global_init(&argc, argv);
617 // parse the platform file: get the host list
618 SIMIX_create_environment(argv[1]);
619 SIMIX_comm_set_copy_data_callback(&smpi_comm_copy_buffer_callback);
620 SIMIX_function_register_default(realmain);
621 SIMIX_launch_application(argv[2]);
625 smpi_check_options();
627 if(smpi_privatize_global_variables)
628 smpi_initialize_global_memory_segments();
630 /* Clean IO before the run */
635 MC_do_the_modelcheck_for_real();
639 if (sg_cfg_get_boolean("smpi/display_timing"))
640 XBT_INFO("Simulation time: %g seconds.", SIMIX_get_clock());
642 smpi_global_destroy();
651 // This function can be called from extern file, to initialize logs, options, and processes of smpi
652 // without the need of smpirun
657 smpi_check_options();
659 if (TRACE_is_enabled() && TRACE_is_configured()) {
663 if(smpi_privatize_global_variables)
664 smpi_initialize_global_memory_segments();
667 void SMPI_finalize(){
668 smpi_global_destroy();