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 int index = smpi_process_index();
134 // wait for all pending asynchronous comms to finish
135 xbt_barrier_wait(process_data[index_to_process_data[index]]->finalization_barrier);
140 * @brief Check if a process is finalized
142 int smpi_process_finalized()
144 int index = smpi_process_index();
145 if (index != MPI_UNDEFINED)
146 return (process_data[index_to_process_data[index]]->state == SMPI_FINALIZED);
152 * @brief Check if a process is initialized
154 int smpi_process_initialized(void)
156 int index = smpi_process_index();
157 return ( (index != MPI_UNDEFINED)
158 && (process_data[index_to_process_data[index]]->state == SMPI_INITIALIZED));
162 * @brief Mark a process as initialized (=MPI_Init called)
164 void smpi_process_mark_as_initialized(void)
166 int index = smpi_process_index();
167 if ((index != MPI_UNDEFINED) && (!process_data[index_to_process_data[index]]->state != SMPI_FINALIZED))
168 process_data[index_to_process_data[index]]->state = SMPI_INITIALIZED;
172 int smpi_global_size(void)
174 char *value = getenv("SMPI_GLOBAL_SIZE");
178 "Please set env var SMPI_GLOBAL_SIZE to expected number of processes.\n");
184 smpi_process_data_t smpi_process_data(void)
186 return SIMIX_process_self_get_data(SIMIX_process_self());
189 smpi_process_data_t smpi_process_remote_data(int index)
191 return process_data[index_to_process_data[index]];
194 void smpi_process_set_user_data(void *data)
196 smpi_process_data_t process_data = smpi_process_data();
197 process_data->data = data;
200 void *smpi_process_get_user_data()
202 smpi_process_data_t process_data = smpi_process_data();
203 return process_data->data;
206 int smpi_process_count(void)
208 return process_count;
211 int smpi_process_index(void)
213 smpi_process_data_t data = smpi_process_data();
214 //return -1 if not initialized
215 return data ? data->index : MPI_UNDEFINED;
218 MPI_Comm smpi_process_comm_world(void)
220 smpi_process_data_t data = smpi_process_data();
221 //return MPI_COMM_NULL if not initialized
222 return data ? *data->comm_world : MPI_COMM_NULL;
225 smx_rdv_t smpi_process_mailbox(void)
227 smpi_process_data_t data = smpi_process_data();
228 return data->mailbox;
231 smx_rdv_t smpi_process_mailbox_small(void)
233 smpi_process_data_t data = smpi_process_data();
234 return data->mailbox_small;
237 smx_rdv_t smpi_process_remote_mailbox(int index)
239 smpi_process_data_t data = smpi_process_remote_data(index);
240 return data->mailbox;
244 smx_rdv_t smpi_process_remote_mailbox_small(int index)
246 smpi_process_data_t data = smpi_process_remote_data(index);
247 return data->mailbox_small;
250 xbt_os_timer_t smpi_process_timer(void)
252 smpi_process_data_t data = smpi_process_data();
256 void smpi_process_simulated_start(void)
258 smpi_process_data_t data = smpi_process_data();
259 data->simulated = SIMIX_get_clock();
262 double smpi_process_simulated_elapsed(void)
264 smpi_process_data_t data = smpi_process_data();
265 return SIMIX_get_clock() - data->simulated;
268 MPI_Comm smpi_process_comm_self(void)
270 smpi_process_data_t data = smpi_process_data();
271 if(data->comm_self==MPI_COMM_NULL){
272 MPI_Group group = smpi_group_new(1);
273 data->comm_self = smpi_comm_new(group, NULL);
274 smpi_group_set_mapping(group, smpi_process_index(), 0);
277 return data->comm_self;
280 void smpi_process_set_sampling(int s)
282 smpi_process_data_t data = smpi_process_data();
286 int smpi_process_get_sampling(void)
288 smpi_process_data_t data = smpi_process_data();
289 return data->sampling;
292 void print_request(const char *message, MPI_Request request)
295 ("%s request %p [buf = %p, size = %zu, src = %d, dst = %d, tag = %d, flags = %x]",
296 message, request, request->buf, request->size, request->src,
297 request->dst, request->tag, request->flags);
300 void smpi_comm_copy_buffer_callback(smx_action_t comm,
301 void *buff, size_t buff_size)
303 XBT_DEBUG("Copy the data over");
304 if(_xbt_replay_is_active()) return;
307 if((smpi_privatize_global_variables)
308 && ((char*)buff >= start_data_exe)
309 && ((char*)buff < start_data_exe + size_data_exe )
311 XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !");
312 switch_data_segment(((smpi_process_data_t)SIMIX_process_get_data(comm->comm.src_proc))->index);
313 tmpbuff = (void*)xbt_malloc(buff_size);
314 memcpy(tmpbuff, buff, buff_size);
318 if((smpi_privatize_global_variables)
319 && ((char*)comm->comm.dst_buff >= start_data_exe)
320 && ((char*)comm->comm.dst_buff < start_data_exe + size_data_exe )
322 XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment");
323 switch_data_segment(((smpi_process_data_t)SIMIX_process_get_data(comm->comm.dst_proc))->index);
327 memcpy(comm->comm.dst_buff, tmpbuff, buff_size);
328 if (comm->comm.detached) {
329 // if this is a detached send, the source buffer was duplicated by SMPI
330 // sender to make the original buffer available to the application ASAP
332 //It seems that the request is used after the call there this should
333 //be free somewhereelse but where???
334 //xbt_free(comm->comm.src_data);// inside SMPI the request is keep
335 //inside the user data and should be free
336 comm->comm.src_buff = NULL;
339 if(tmpbuff!=buff)xbt_free(tmpbuff);
343 static void smpi_check_options(){
344 //check correctness of MPI parameters
346 xbt_assert(sg_cfg_get_int("smpi/async_small_thres") <=
347 sg_cfg_get_int("smpi/send_is_detached_thres"));
349 if (sg_cfg_is_default_value("smpi/running_power")) {
350 XBT_INFO("You did not set the power of the host running the simulation. "
351 "The timings will certainly not be accurate. "
352 "Use the option \"--cfg=smpi/running_power:<flops>\" to set its value."
353 "Check http://simgrid.org/simgrid/latest/doc/options.html#options_smpi_bench for more information. ");
357 int smpi_enabled(void) {
358 return process_data != NULL;
361 void smpi_global_init(void)
365 char name[MAILBOX_NAME_MAXLEN];
369 if (process_count == 0){
370 process_count = SIMIX_process_count();
373 process_data = xbt_new0(smpi_process_data_t, process_count);
374 for (i = 0; i < process_count; i++) {
375 process_data[i] = xbt_new(s_smpi_process_data_t, 1);
376 //process_data[i]->index = i;
377 process_data[i]->argc = NULL;
378 process_data[i]->argv = NULL;
379 process_data[i]->mailbox = simcall_rdv_create(get_mailbox_name(name, i));
380 process_data[i]->mailbox_small =
381 simcall_rdv_create(get_mailbox_name_small(name, i));
382 process_data[i]->timer = xbt_os_timer_new();
384 MC_ignore_heap(process_data[i]->timer, xbt_os_timer_size());
385 process_data[i]->comm_self = MPI_COMM_NULL;
386 process_data[i]->comm_world = NULL;
387 process_data[i]->state = SMPI_UNINITIALIZED;
388 process_data[i]->sampling = 0;
389 process_data[i]->finalization_barrier = NULL;
391 //if the process was launched through smpirun script
392 //we generate a global mpi_comm_world
393 //if not, we let MPI_COMM_NULL, and the comm world
394 //will be private to each mpi instance
396 group = smpi_group_new(process_count);
397 MPI_COMM_WORLD = smpi_comm_new(group, NULL);
398 xbt_bar_t bar=xbt_barrier_init(process_count);
400 MPI_UNIVERSE_SIZE = smpi_comm_size(MPI_COMM_WORLD);
401 for (i = 0; i < process_count; i++) {
402 smpi_group_set_mapping(group, i, i);
403 process_data[i]->finalization_barrier = bar;
408 void smpi_global_destroy(void)
410 int count = smpi_process_count();
413 smpi_bench_destroy();
414 if (MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED){
415 while (smpi_group_unuse(smpi_comm_group(MPI_COMM_WORLD)) > 0);
416 xbt_free(MPI_COMM_WORLD);
417 xbt_barrier_destroy(process_data[0]->finalization_barrier);
419 smpi_deployment_cleanup_instances();
421 MPI_COMM_WORLD = MPI_COMM_NULL;
422 for (i = 0; i < count; i++) {
423 if(process_data[i]->comm_self!=MPI_COMM_NULL){
424 smpi_group_unuse(smpi_comm_group(process_data[i]->comm_self));
425 smpi_comm_destroy(process_data[i]->comm_self);
427 xbt_os_timer_free(process_data[i]->timer);
428 simcall_rdv_destroy(process_data[i]->mailbox);
429 simcall_rdv_destroy(process_data[i]->mailbox_small);
430 xbt_free(process_data[i]);
432 xbt_free(process_data);
435 xbt_free(index_to_process_data);
436 if(smpi_privatize_global_variables)
437 smpi_destroy_global_memory_segments();
442 void __attribute__ ((weak)) user_main_()
444 xbt_die("Should not be in this smpi_simulated_main");
448 int __attribute__ ((weak)) smpi_simulated_main_(int argc, char **argv)
450 smpi_process_init(&argc, &argv);
455 int __attribute__ ((weak)) main(int argc, char **argv)
457 return smpi_main(smpi_simulated_main_, argc, argv);
462 static void smpi_init_logs(){
464 /* Connect log categories. See xbt/log.c */
465 XBT_LOG_CONNECT(smpi); /* Keep this line as soon as possible in this
466 function: xbt_log_appender_file.c depends on it
467 DO NOT connect this in XBT or so, or it will be
468 useless to xbt_log_appender_file.c */
470 XBT_LOG_CONNECT(instr_smpi);
472 XBT_LOG_CONNECT(smpi_base);
473 XBT_LOG_CONNECT(smpi_bench);
474 XBT_LOG_CONNECT(smpi_coll);
475 XBT_LOG_CONNECT(smpi_colls);
476 XBT_LOG_CONNECT(smpi_comm);
477 XBT_LOG_CONNECT(smpi_dvfs);
478 XBT_LOG_CONNECT(smpi_group);
479 XBT_LOG_CONNECT(smpi_kernel);
480 XBT_LOG_CONNECT(smpi_mpi);
481 XBT_LOG_CONNECT(smpi_mpi_dt);
482 XBT_LOG_CONNECT(smpi_pmpi);
483 XBT_LOG_CONNECT(smpi_replay);
488 static void smpi_init_options(){
489 int gather_id = find_coll_description(mpi_coll_gather_description,
490 sg_cfg_get_string("smpi/gather"));
491 mpi_coll_gather_fun = (int (*)(void *, int, MPI_Datatype,
492 void *, int, MPI_Datatype, int, MPI_Comm))
493 mpi_coll_gather_description[gather_id].coll;
495 int allgather_id = find_coll_description(mpi_coll_allgather_description,
496 sg_cfg_get_string("smpi/allgather"));
497 mpi_coll_allgather_fun = (int (*)(void *, int, MPI_Datatype,
498 void *, int, MPI_Datatype, MPI_Comm))
499 mpi_coll_allgather_description[allgather_id].coll;
501 int allgatherv_id = find_coll_description(mpi_coll_allgatherv_description,
502 sg_cfg_get_string("smpi/allgatherv"));
503 mpi_coll_allgatherv_fun = (int (*)(void *, int, MPI_Datatype, void *, int *,
504 int *, MPI_Datatype, MPI_Comm))
505 mpi_coll_allgatherv_description[allgatherv_id].coll;
507 int allreduce_id = find_coll_description(mpi_coll_allreduce_description,
508 sg_cfg_get_string("smpi/allreduce"));
509 mpi_coll_allreduce_fun = (int (*)(void *sbuf, void *rbuf, int rcount,
510 MPI_Datatype dtype, MPI_Op op,
512 mpi_coll_allreduce_description[allreduce_id].coll;
514 int alltoall_id = find_coll_description(mpi_coll_alltoall_description,
515 sg_cfg_get_string("smpi/alltoall"));
516 mpi_coll_alltoall_fun = (int (*)(void *, int, MPI_Datatype,
517 void *, int, MPI_Datatype, MPI_Comm))
518 mpi_coll_alltoall_description[alltoall_id].coll;
520 int alltoallv_id = find_coll_description(mpi_coll_alltoallv_description,
521 sg_cfg_get_string("smpi/alltoallv"));
522 mpi_coll_alltoallv_fun = (int (*)(void *, int *, int *, MPI_Datatype,
523 void *, int *, int *, MPI_Datatype,
525 mpi_coll_alltoallv_description[alltoallv_id].coll;
527 int bcast_id = find_coll_description(mpi_coll_bcast_description,
528 sg_cfg_get_string("smpi/bcast"));
529 mpi_coll_bcast_fun = (int (*)(void *buf, int count, MPI_Datatype datatype,
530 int root, MPI_Comm com))
531 mpi_coll_bcast_description[bcast_id].coll;
533 int reduce_id = find_coll_description(mpi_coll_reduce_description,
534 sg_cfg_get_string("smpi/reduce"));
535 mpi_coll_reduce_fun = (int (*)(void *buf, void *rbuf, int count,
536 MPI_Datatype datatype, MPI_Op op,
537 int root, MPI_Comm comm))
538 mpi_coll_reduce_description[reduce_id].coll;
540 int reduce_scatter_id =
541 find_coll_description(mpi_coll_reduce_scatter_description,
542 sg_cfg_get_string("smpi/reduce_scatter"));
543 mpi_coll_reduce_scatter_fun = (int (*)(void *sbuf, void *rbuf, int *rcounts,
544 MPI_Datatype dtype, MPI_Op op,
546 mpi_coll_reduce_scatter_description[reduce_scatter_id].coll;
548 int scatter_id = find_coll_description(mpi_coll_scatter_description,
549 sg_cfg_get_string("smpi/scatter"));
550 mpi_coll_scatter_fun = (int (*)(void *sendbuf, int sendcount,
551 MPI_Datatype sendtype, void *recvbuf,
552 int recvcount, MPI_Datatype recvtype,
553 int root, MPI_Comm comm))
554 mpi_coll_scatter_description[scatter_id].coll;
556 int barrier_id = find_coll_description(mpi_coll_barrier_description,
557 sg_cfg_get_string("smpi/barrier"));
558 mpi_coll_barrier_fun = (int (*)(MPI_Comm comm))
559 mpi_coll_barrier_description[barrier_id].coll;
561 smpi_cpu_threshold = sg_cfg_get_double("smpi/cpu_threshold");
562 smpi_running_power = sg_cfg_get_double("smpi/running_power");
563 smpi_privatize_global_variables = sg_cfg_get_boolean("smpi/privatize_global_variables");
564 if (smpi_cpu_threshold < 0)
565 smpi_cpu_threshold = DBL_MAX;
569 int smpi_main(int (*realmain) (int argc, char *argv[]), int argc, char *argv[])
571 srand(SMPI_RAND_SEED);
573 if (getenv("SMPI_PRETEND_CC") != NULL) {
574 /* Hack to ensure that smpicc can pretend to be a simple
575 * compiler. Particularly handy to pass it to the configuration tools */
582 TRACE_global_init(&argc, argv);
584 TRACE_add_start_function(TRACE_smpi_alloc);
585 TRACE_add_end_function(TRACE_smpi_release);
588 SIMIX_global_init(&argc, argv);
590 // parse the platform file: get the host list
591 SIMIX_create_environment(argv[1]);
592 SIMIX_comm_set_copy_data_callback(&smpi_comm_copy_buffer_callback);
593 SIMIX_function_register_default(realmain);
594 SIMIX_launch_application(argv[2]);
600 smpi_check_options();
602 if(smpi_privatize_global_variables)
603 smpi_initialize_global_memory_segments();
605 /* Clean IO before the run */
610 MC_do_the_modelcheck_for_real();
614 if (sg_cfg_get_boolean("smpi/display_timing"))
615 XBT_INFO("Simulation time: %g seconds.", SIMIX_get_clock());
617 smpi_global_destroy();
626 // This function can be called from extern file, to initialize logs, options, and processes of smpi
627 // without the need of smpirun
632 smpi_check_options();
634 if (TRACE_is_enabled() && TRACE_is_configured()) {
638 if(smpi_privatize_global_variables)
639 smpi_initialize_global_memory_segments();
642 void SMPI_finalize(){
643 smpi_global_destroy();