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 = SIMIX_process_get_PID(proc) -1;
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 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 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 int smpi_process_index_of_smx_process(smx_process_t process) {
225 smpi_process_data_t data = (smpi_process_data_t) SIMIX_process_get_data(process);
226 return data ? data->index : MPI_UNDEFINED;
229 MPI_Comm smpi_process_comm_world(void)
231 smpi_process_data_t data = smpi_process_data();
232 //return MPI_COMM_NULL if not initialized
233 return data ? *data->comm_world : MPI_COMM_NULL;
236 smx_rdv_t smpi_process_mailbox(void)
238 smpi_process_data_t data = smpi_process_data();
239 return data->mailbox;
242 smx_rdv_t smpi_process_mailbox_small(void)
244 smpi_process_data_t data = smpi_process_data();
245 return data->mailbox_small;
248 smx_rdv_t smpi_process_remote_mailbox(int index)
250 smpi_process_data_t data = smpi_process_remote_data(index);
251 return data->mailbox;
255 smx_rdv_t smpi_process_remote_mailbox_small(int index)
257 smpi_process_data_t data = smpi_process_remote_data(index);
258 return data->mailbox_small;
261 xbt_os_timer_t smpi_process_timer(void)
263 smpi_process_data_t data = smpi_process_data();
267 void smpi_process_simulated_start(void)
269 smpi_process_data_t data = smpi_process_data();
270 data->simulated = SIMIX_get_clock();
273 double smpi_process_simulated_elapsed(void)
275 smpi_process_data_t data = smpi_process_data();
276 return SIMIX_get_clock() - data->simulated;
279 MPI_Comm smpi_process_comm_self(void)
281 smpi_process_data_t data = smpi_process_data();
282 if(data->comm_self==MPI_COMM_NULL){
283 MPI_Group group = smpi_group_new(1);
284 data->comm_self = smpi_comm_new(group, NULL);
285 smpi_group_set_mapping(group, smpi_process_index(), 0);
288 return data->comm_self;
291 MPI_Comm smpi_process_get_comm_intra(void)
293 smpi_process_data_t data = smpi_process_data();
294 return data->comm_intra;
297 void smpi_process_set_comm_intra(MPI_Comm comm)
299 smpi_process_data_t data = smpi_process_data();
300 data->comm_intra = comm;
303 void smpi_process_set_sampling(int s)
305 smpi_process_data_t data = smpi_process_data();
309 int smpi_process_get_sampling(void)
311 smpi_process_data_t data = smpi_process_data();
312 return data->sampling;
316 void print_request(const char *message, MPI_Request request)
319 ("%s request %p [buf = %p, size = %zu, src = %d, dst = %d, tag = %d, flags = %x]",
320 message, request, request->buf, request->size, request->src,
321 request->dst, request->tag, request->flags);
324 void smpi_comm_copy_buffer_callback(smx_action_t comm,
325 void *buff, size_t buff_size)
327 XBT_DEBUG("Copy the data over");
328 if(_xbt_replay_is_active()) return;
331 if((smpi_privatize_global_variables)
332 && ((char*)buff >= start_data_exe)
333 && ((char*)buff < start_data_exe + size_data_exe )
335 XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !");
336 switch_data_segment(((smpi_process_data_t)SIMIX_process_get_data(comm->comm.src_proc))->index);
337 tmpbuff = (void*)xbt_malloc(buff_size);
338 memcpy(tmpbuff, buff, buff_size);
342 if((smpi_privatize_global_variables)
343 && ((char*)comm->comm.dst_buff >= start_data_exe)
344 && ((char*)comm->comm.dst_buff < start_data_exe + size_data_exe )
346 XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment");
347 switch_data_segment(((smpi_process_data_t)SIMIX_process_get_data(comm->comm.dst_proc))->index);
351 memcpy(comm->comm.dst_buff, tmpbuff, buff_size);
352 if (comm->comm.detached) {
353 // if this is a detached send, the source buffer was duplicated by SMPI
354 // sender to make the original buffer available to the application ASAP
356 //It seems that the request is used after the call there this should
357 //be free somewhereelse but where???
358 //xbt_free(comm->comm.src_data);// inside SMPI the request is keep
359 //inside the user data and should be free
360 comm->comm.src_buff = NULL;
363 if(tmpbuff!=buff)xbt_free(tmpbuff);
367 static void smpi_check_options(){
368 //check correctness of MPI parameters
370 xbt_assert(sg_cfg_get_int("smpi/async_small_thres") <=
371 sg_cfg_get_int("smpi/send_is_detached_thres"));
373 if (sg_cfg_is_default_value("smpi/running_power")) {
374 XBT_INFO("You did not set the power of the host running the simulation. "
375 "The timings will certainly not be accurate. "
376 "Use the option \"--cfg=smpi/running_power:<flops>\" to set its value."
377 "Check http://simgrid.org/simgrid/latest/doc/options.html#options_smpi_bench for more information.");
381 int smpi_enabled(void) {
382 return process_data != NULL;
385 void smpi_global_init(void)
389 char name[MAILBOX_NAME_MAXLEN];
393 if (process_count == 0){
394 process_count = SIMIX_process_count();
397 process_data = xbt_new0(smpi_process_data_t, process_count);
398 for (i = 0; i < process_count; i++) {
399 process_data[i] = xbt_new(s_smpi_process_data_t, 1);
400 //process_data[i]->index = i;
401 process_data[i]->argc = NULL;
402 process_data[i]->argv = NULL;
403 process_data[i]->mailbox = simcall_rdv_create(get_mailbox_name(name, i));
404 process_data[i]->mailbox_small =
405 simcall_rdv_create(get_mailbox_name_small(name, i));
406 process_data[i]->timer = xbt_os_timer_new();
408 MC_ignore_heap(process_data[i]->timer, xbt_os_timer_size());
409 process_data[i]->comm_self = MPI_COMM_NULL;
410 process_data[i]->comm_intra = MPI_COMM_NULL;
411 process_data[i]->comm_world = NULL;
412 process_data[i]->state = SMPI_UNINITIALIZED;
413 process_data[i]->sampling = 0;
414 process_data[i]->finalization_barrier = NULL;
416 //if the process was launched through smpirun script
417 //we generate a global mpi_comm_world
418 //if not, we let MPI_COMM_NULL, and the comm world
419 //will be private to each mpi instance
421 group = smpi_group_new(process_count);
422 MPI_COMM_WORLD = smpi_comm_new(group, NULL);
423 xbt_bar_t bar=xbt_barrier_init(process_count);
425 MPI_UNIVERSE_SIZE = smpi_comm_size(MPI_COMM_WORLD);
426 for (i = 0; i < process_count; i++) {
427 smpi_group_set_mapping(group, i, i);
428 process_data[i]->finalization_barrier = bar;
433 void smpi_global_destroy(void)
435 int count = smpi_process_count();
438 smpi_bench_destroy();
439 if (MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED){
440 while (smpi_group_unuse(smpi_comm_group(MPI_COMM_WORLD)) > 0);
441 xbt_free(MPI_COMM_WORLD);
442 xbt_barrier_destroy(process_data[0]->finalization_barrier);
444 smpi_deployment_cleanup_instances();
446 MPI_COMM_WORLD = MPI_COMM_NULL;
447 for (i = 0; i < count; i++) {
448 if(process_data[i]->comm_self!=MPI_COMM_NULL){
449 smpi_group_unuse(smpi_comm_group(process_data[i]->comm_self));
450 smpi_comm_destroy(process_data[i]->comm_self);
452 if(process_data[i]->comm_intra!=MPI_COMM_NULL){
453 smpi_group_unuse(smpi_comm_group(process_data[i]->comm_intra));
454 smpi_comm_destroy(process_data[i]->comm_intra);
456 xbt_os_timer_free(process_data[i]->timer);
457 simcall_rdv_destroy(process_data[i]->mailbox);
458 simcall_rdv_destroy(process_data[i]->mailbox_small);
459 xbt_free(process_data[i]);
461 xbt_free(process_data);
464 xbt_free(index_to_process_data);
465 if(smpi_privatize_global_variables)
466 smpi_destroy_global_memory_segments();
471 void __attribute__ ((weak)) user_main_()
473 xbt_die("Should not be in this smpi_simulated_main");
477 int __attribute__ ((weak)) smpi_simulated_main_(int argc, char **argv)
479 smpi_process_init(&argc, &argv);
484 int __attribute__ ((weak)) main(int argc, char **argv)
486 return smpi_main(smpi_simulated_main_, argc, argv);
491 static void smpi_init_logs(){
493 /* Connect log categories. See xbt/log.c */
494 XBT_LOG_CONNECT(smpi); /* Keep this line as soon as possible in this
495 function: xbt_log_appender_file.c depends on it
496 DO NOT connect this in XBT or so, or it will be
497 useless to xbt_log_appender_file.c */
499 XBT_LOG_CONNECT(instr_smpi);
501 XBT_LOG_CONNECT(smpi_base);
502 XBT_LOG_CONNECT(smpi_bench);
503 XBT_LOG_CONNECT(smpi_coll);
504 XBT_LOG_CONNECT(smpi_colls);
505 XBT_LOG_CONNECT(smpi_comm);
506 XBT_LOG_CONNECT(smpi_dvfs);
507 XBT_LOG_CONNECT(smpi_group);
508 XBT_LOG_CONNECT(smpi_kernel);
509 XBT_LOG_CONNECT(smpi_mpi);
510 XBT_LOG_CONNECT(smpi_mpi_dt);
511 XBT_LOG_CONNECT(smpi_pmpi);
512 XBT_LOG_CONNECT(smpi_replay);
513 XBT_LOG_CONNECT(smpi_rma);
518 static void smpi_init_options(){
519 int gather_id = find_coll_description(mpi_coll_gather_description,
520 sg_cfg_get_string("smpi/gather"));
521 mpi_coll_gather_fun = (int (*)(void *, int, MPI_Datatype,
522 void *, int, MPI_Datatype, int, MPI_Comm))
523 mpi_coll_gather_description[gather_id].coll;
525 int allgather_id = find_coll_description(mpi_coll_allgather_description,
526 sg_cfg_get_string("smpi/allgather"));
527 mpi_coll_allgather_fun = (int (*)(void *, int, MPI_Datatype,
528 void *, int, MPI_Datatype, MPI_Comm))
529 mpi_coll_allgather_description[allgather_id].coll;
531 int allgatherv_id = find_coll_description(mpi_coll_allgatherv_description,
532 sg_cfg_get_string("smpi/allgatherv"));
533 mpi_coll_allgatherv_fun = (int (*)(void *, int, MPI_Datatype, void *, int *,
534 int *, MPI_Datatype, MPI_Comm))
535 mpi_coll_allgatherv_description[allgatherv_id].coll;
537 int allreduce_id = find_coll_description(mpi_coll_allreduce_description,
538 sg_cfg_get_string("smpi/allreduce"));
539 mpi_coll_allreduce_fun = (int (*)(void *sbuf, void *rbuf, int rcount,
540 MPI_Datatype dtype, MPI_Op op,
542 mpi_coll_allreduce_description[allreduce_id].coll;
544 int alltoall_id = find_coll_description(mpi_coll_alltoall_description,
545 sg_cfg_get_string("smpi/alltoall"));
546 mpi_coll_alltoall_fun = (int (*)(void *, int, MPI_Datatype,
547 void *, int, MPI_Datatype, MPI_Comm))
548 mpi_coll_alltoall_description[alltoall_id].coll;
550 int alltoallv_id = find_coll_description(mpi_coll_alltoallv_description,
551 sg_cfg_get_string("smpi/alltoallv"));
552 mpi_coll_alltoallv_fun = (int (*)(void *, int *, int *, MPI_Datatype,
553 void *, int *, int *, MPI_Datatype,
555 mpi_coll_alltoallv_description[alltoallv_id].coll;
557 int bcast_id = find_coll_description(mpi_coll_bcast_description,
558 sg_cfg_get_string("smpi/bcast"));
559 mpi_coll_bcast_fun = (int (*)(void *buf, int count, MPI_Datatype datatype,
560 int root, MPI_Comm com))
561 mpi_coll_bcast_description[bcast_id].coll;
563 int reduce_id = find_coll_description(mpi_coll_reduce_description,
564 sg_cfg_get_string("smpi/reduce"));
565 mpi_coll_reduce_fun = (int (*)(void *buf, void *rbuf, int count,
566 MPI_Datatype datatype, MPI_Op op,
567 int root, MPI_Comm comm))
568 mpi_coll_reduce_description[reduce_id].coll;
570 int reduce_scatter_id =
571 find_coll_description(mpi_coll_reduce_scatter_description,
572 sg_cfg_get_string("smpi/reduce_scatter"));
573 mpi_coll_reduce_scatter_fun = (int (*)(void *sbuf, void *rbuf, int *rcounts,
574 MPI_Datatype dtype, MPI_Op op,
576 mpi_coll_reduce_scatter_description[reduce_scatter_id].coll;
578 int scatter_id = find_coll_description(mpi_coll_scatter_description,
579 sg_cfg_get_string("smpi/scatter"));
580 mpi_coll_scatter_fun = (int (*)(void *sendbuf, int sendcount,
581 MPI_Datatype sendtype, void *recvbuf,
582 int recvcount, MPI_Datatype recvtype,
583 int root, MPI_Comm comm))
584 mpi_coll_scatter_description[scatter_id].coll;
586 int barrier_id = find_coll_description(mpi_coll_barrier_description,
587 sg_cfg_get_string("smpi/barrier"));
588 mpi_coll_barrier_fun = (int (*)(MPI_Comm comm))
589 mpi_coll_barrier_description[barrier_id].coll;
591 smpi_cpu_threshold = sg_cfg_get_double("smpi/cpu_threshold");
592 smpi_running_power = sg_cfg_get_double("smpi/running_power");
593 smpi_privatize_global_variables = sg_cfg_get_boolean("smpi/privatize_global_variables");
594 if (smpi_cpu_threshold < 0)
595 smpi_cpu_threshold = DBL_MAX;
599 int smpi_main(int (*realmain) (int argc, char *argv[]), int argc, char *argv[])
601 srand(SMPI_RAND_SEED);
603 if (getenv("SMPI_PRETEND_CC") != NULL) {
604 /* Hack to ensure that smpicc can pretend to be a simple
605 * compiler. Particularly handy to pass it to the configuration tools */
612 TRACE_global_init(&argc, argv);
614 TRACE_add_start_function(TRACE_smpi_alloc);
615 TRACE_add_end_function(TRACE_smpi_release);
618 SIMIX_global_init(&argc, argv);
622 // parse the platform file: get the host list
623 SIMIX_create_environment(argv[1]);
624 SIMIX_comm_set_copy_data_callback(&smpi_comm_copy_buffer_callback);
625 SIMIX_function_register_default(realmain);
626 SIMIX_launch_application(argv[2]);
630 smpi_check_options();
632 if(smpi_privatize_global_variables)
633 smpi_initialize_global_memory_segments();
635 /* Clean IO before the run */
640 MC_do_the_modelcheck_for_real();
644 if (sg_cfg_get_boolean("smpi/display_timing"))
645 XBT_INFO("Simulation time: %g seconds.", SIMIX_get_clock());
647 smpi_global_destroy();
656 // This function can be called from extern file, to initialize logs, options, and processes of smpi
657 // without the need of smpirun
662 smpi_check_options();
664 if (TRACE_is_enabled() && TRACE_is_configured()) {
668 if(smpi_privatize_global_variables)
669 smpi_initialize_global_memory_segments();
672 void SMPI_finalize(){
673 smpi_global_destroy();