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;
31 void *data; /* user data */
34 int sampling; /* inside an SMPI_SAMPLE_ block? */
35 } s_smpi_process_data_t;
37 static smpi_process_data_t *process_data = NULL;
38 static int process_count = 0;
40 MPI_Comm MPI_COMM_WORLD = MPI_COMM_NULL;
41 int MPI_UNIVERSE_SIZE;
43 MPI_Errhandler *MPI_ERRORS_RETURN = NULL;
44 MPI_Errhandler *MPI_ERRORS_ARE_FATAL = NULL;
45 MPI_Errhandler *MPI_ERRHANDLER_NULL = NULL;
47 #define MAILBOX_NAME_MAXLEN (5 + sizeof(int) * 2 + 1)
49 static char *get_mailbox_name(char *str, int index)
51 snprintf(str, MAILBOX_NAME_MAXLEN, "SMPI-%0*x", (int) (sizeof(int) * 2),
56 static char *get_mailbox_name_small(char *str, int index)
58 snprintf(str, MAILBOX_NAME_MAXLEN, "small%0*x", (int) (sizeof(int) * 2),
63 void smpi_process_init(int *argc, char ***argv)
66 smpi_process_data_t data;
70 proc = SIMIX_process_self();
71 index = atoi((*argv)[1]);
73 smpi_current_rank = index;
75 data = smpi_process_remote_data(index);
76 simcall_process_set_data(proc, data);
79 memmove(&(*argv)[1], &(*argv)[2], sizeof(char *) * (*argc - 2));
80 (*argv)[(*argc) - 1] = NULL;
85 // set the process attached to the mailbox
86 simcall_rdv_set_receiver(data->mailbox_small, proc);
87 XBT_DEBUG("<%d> New process in the game: %p", index, proc);
91 void smpi_process_destroy(void)
93 int index = smpi_process_index();
94 process_data[index]->state = SMPI_FINALIZED;
95 XBT_DEBUG("<%d> Process left the game", index);
99 * @brief Prepares the current process for termination.
101 void smpi_process_finalize(void)
104 // wait for all pending asynchronous comms to finish
105 while (SIMIX_process_has_pending_comms(SIMIX_process_self())) {
106 simcall_process_sleep(0.01);
110 int size = smpi_comm_size(MPI_COMM_WORLD);
111 int rank = smpi_comm_rank(MPI_COMM_WORLD);
112 /* All non-root send & receive zero-length message. */
114 smpi_mpi_ssend (NULL, 0, MPI_BYTE, 0,
117 smpi_mpi_recv (NULL, 0, MPI_BYTE, 0,
119 MPI_COMM_WORLD, MPI_STATUS_IGNORE);
121 /* The root collects and broadcasts the messages. */
123 MPI_Request* requests;
124 requests = (MPI_Request*)malloc( size * sizeof(MPI_Request) );
125 for (i = 1; i < size; ++i) {
126 requests[i] = smpi_mpi_irecv(NULL, 0, MPI_BYTE, MPI_ANY_SOURCE,
127 COLL_TAG_BARRIER, MPI_COMM_WORLD
130 smpi_mpi_waitall( size-1, requests+1, MPI_STATUSES_IGNORE );
131 for (i = 1; i < size; ++i) {
132 requests[i] = smpi_mpi_issend(NULL, 0, MPI_BYTE, i,
137 smpi_mpi_waitall( size-1, requests+1, MPI_STATUSES_IGNORE );
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]->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]->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]->state != SMPI_FINALIZED))
172 process_data[index]->state = SMPI_INITIALIZED;
177 int smpi_process_argc(void)
179 smpi_process_data_t data = smpi_process_data();
180 return data->argc ? *(data->argc) - 1 : 0;
183 int smpi_process_getarg(integer * index, char *dst, ftnlen len)
185 smpi_process_data_t data = smpi_process_data();
189 if (!data->argc || !data->argv || *index < 1 || *index >= *(data->argc)) {
192 arg = (*data->argv)[*index];
193 for (i = 0; i < len && arg[i] != '\0'; i++) {
196 for (; i < len; i++) {
203 int smpi_global_size(void)
205 char *value = getenv("SMPI_GLOBAL_SIZE");
209 "Please set env var SMPI_GLOBAL_SIZE to expected number of processes.\n");
215 smpi_process_data_t smpi_process_data(void)
217 return SIMIX_process_self_get_data(SIMIX_process_self());
220 smpi_process_data_t smpi_process_remote_data(int index)
222 return process_data[index];
225 void smpi_process_set_user_data(void *data)
227 smpi_process_data_t process_data = smpi_process_data();
228 process_data->data = data;
231 void *smpi_process_get_user_data()
233 smpi_process_data_t process_data = smpi_process_data();
234 return process_data->data;
237 int smpi_process_count(void)
239 return process_count;
242 int smpi_process_index(void)
244 smpi_process_data_t data = smpi_process_data();
245 //return -1 if not initialized
246 return data ? data->index : MPI_UNDEFINED;
249 smx_rdv_t smpi_process_mailbox(void)
251 smpi_process_data_t data = smpi_process_data();
252 return data->mailbox;
255 smx_rdv_t smpi_process_mailbox_small(void)
257 smpi_process_data_t data = smpi_process_data();
258 return data->mailbox_small;
261 smx_rdv_t smpi_process_remote_mailbox(int index)
263 smpi_process_data_t data = smpi_process_remote_data(index);
264 return data->mailbox;
268 smx_rdv_t smpi_process_remote_mailbox_small(int index)
270 smpi_process_data_t data = smpi_process_remote_data(index);
271 return data->mailbox_small;
274 xbt_os_timer_t smpi_process_timer(void)
276 smpi_process_data_t data = smpi_process_data();
280 void smpi_process_simulated_start(void)
282 smpi_process_data_t data = smpi_process_data();
283 data->simulated = SIMIX_get_clock();
286 double smpi_process_simulated_elapsed(void)
288 smpi_process_data_t data = smpi_process_data();
289 return SIMIX_get_clock() - data->simulated;
292 MPI_Comm smpi_process_comm_self(void)
294 smpi_process_data_t data = smpi_process_data();
295 if(data->comm_self==MPI_COMM_NULL){
296 MPI_Group group = smpi_group_new(1);
297 data->comm_self = smpi_comm_new(group);
298 smpi_group_set_mapping(group, smpi_process_index(), 0);
301 return data->comm_self;
304 void smpi_process_set_sampling(int s)
306 smpi_process_data_t data = smpi_process_data();
310 int smpi_process_get_sampling(void)
312 smpi_process_data_t data = smpi_process_data();
313 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 static 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;
329 memcpy(comm->comm.dst_buff, buff, buff_size);
330 if (comm->comm.detached) {
331 // if this is a detached send, the source buffer was duplicated by SMPI
332 // sender to make the original buffer available to the application ASAP
334 //It seems that the request is used after the call there this should
335 //be free somewhereelse but where???
336 //xbt_free(comm->comm.src_data);// inside SMPI the request is keep
337 //inside the user data and should be free
338 comm->comm.src_buff = NULL;
342 void smpi_global_init(void)
346 char name[MAILBOX_NAME_MAXLEN];
348 SIMIX_comm_set_copy_data_callback(&smpi_comm_copy_buffer_callback);
349 process_count = SIMIX_process_count();
350 process_data = xbt_new(smpi_process_data_t, process_count);
351 for (i = 0; i < process_count; i++) {
352 process_data[i] = xbt_new(s_smpi_process_data_t, 1);
353 process_data[i]->index = i;
354 process_data[i]->argc = NULL;
355 process_data[i]->argv = NULL;
356 process_data[i]->mailbox = simcall_rdv_create(get_mailbox_name(name, i));
357 process_data[i]->mailbox_small =
358 simcall_rdv_create(get_mailbox_name_small(name, i));
359 process_data[i]->timer = xbt_os_timer_new();
361 MC_ignore_heap(process_data[i]->timer, xbt_os_timer_size());
362 process_data[i]->comm_self = MPI_COMM_NULL;
363 process_data[i]->state = SMPI_UNINITIALIZED;
364 process_data[i]->sampling = 0;
366 group = smpi_group_new(process_count);
367 MPI_COMM_WORLD = smpi_comm_new(group);
368 MPI_UNIVERSE_SIZE = smpi_comm_size(MPI_COMM_WORLD);
369 for (i = 0; i < process_count; i++) {
370 smpi_group_set_mapping(group, i, i);
373 //check correctness of MPI parameters
375 xbt_assert(sg_cfg_get_int("smpi/async_small_thres") <=
376 sg_cfg_get_int("smpi/send_is_detached_thres"));
378 if (sg_cfg_is_default_value("smpi/running_power")) {
379 XBT_INFO("You did not set the power of the host running the simulation. "
380 "The timings will certainly not be accurate. "
381 "Use the option \"--cfg=smpi/running_power:<flops>\" to set its value."
382 "Check http://simgrid.org/simgrid/latest/doc/options.html#options_smpi_bench for more information. ");
386 void smpi_global_destroy(void)
388 int count = smpi_process_count();
391 smpi_bench_destroy();
392 while (smpi_group_unuse(smpi_comm_group(MPI_COMM_WORLD)) > 0);
393 xbt_free(MPI_COMM_WORLD);
394 MPI_COMM_WORLD = MPI_COMM_NULL;
395 for (i = 0; i < count; i++) {
396 if(process_data[i]->comm_self!=MPI_COMM_NULL){
397 smpi_group_unuse(smpi_comm_group(process_data[i]->comm_self));
398 smpi_comm_destroy(process_data[i]->comm_self);
400 xbt_os_timer_free(process_data[i]->timer);
401 simcall_rdv_destroy(process_data[i]->mailbox);
402 simcall_rdv_destroy(process_data[i]->mailbox_small);
403 xbt_free(process_data[i]);
405 xbt_free(process_data);
411 /* Fortran specific stuff */
412 /* With smpicc, the following weak symbols are used */
413 /* With smpiff, the following weak symbols are replaced by those in libf2c */
414 int __attribute__ ((weak)) xargc;
415 char ** __attribute__ ((weak)) xargv;
418 void __attribute__ ((weak)) user_main_()
420 xbt_die("Should not be in this smpi_simulated_main");
424 int __attribute__ ((weak)) smpi_simulated_main_(int argc, char **argv)
426 smpi_process_init(&argc, &argv);
431 int __attribute__ ((weak)) main(int argc, char **argv)
433 return smpi_main(smpi_simulated_main_, argc, argv);
436 int __attribute__ ((weak)) MAIN__()
438 return smpi_main(smpi_simulated_main_, xargc, xargv);
442 int smpi_main(int (*realmain) (int argc, char *argv[]), int argc, char *argv[])
444 srand(SMPI_RAND_SEED);
446 if (getenv("SMPI_PRETEND_CC") != NULL) {
447 /* Hack to ensure that smpicc can pretend to be a simple
448 * compiler. Particularly handy to pass it to the configuration tools */
452 /* Connect log categories. See xbt/log.c */
453 XBT_LOG_CONNECT(smpi); /* Keep this line as soon as possible in this
454 function: xbt_log_appender_file.c depends on it
455 DO NOT connect this in XBT or so, or it will be
456 useless to xbt_log_appender_file.c */
458 XBT_LOG_CONNECT(instr_smpi);
460 XBT_LOG_CONNECT(smpi_base);
461 XBT_LOG_CONNECT(smpi_bench);
462 XBT_LOG_CONNECT(smpi_coll);
463 XBT_LOG_CONNECT(smpi_colls);
464 XBT_LOG_CONNECT(smpi_comm);
465 XBT_LOG_CONNECT(smpi_dvfs);
466 XBT_LOG_CONNECT(smpi_group);
467 XBT_LOG_CONNECT(smpi_kernel);
468 XBT_LOG_CONNECT(smpi_mpi);
469 XBT_LOG_CONNECT(smpi_mpi_dt);
470 XBT_LOG_CONNECT(smpi_pmpi);
471 XBT_LOG_CONNECT(smpi_replay);
474 TRACE_global_init(&argc, argv);
476 TRACE_add_start_function(TRACE_smpi_alloc);
477 TRACE_add_end_function(TRACE_smpi_release);
480 SIMIX_global_init(&argc, argv);
486 // parse the platform file: get the host list
487 SIMIX_create_environment(argv[1]);
489 SIMIX_function_register_default(realmain);
490 SIMIX_launch_application(argv[2]);
492 int gather_id = find_coll_description(mpi_coll_gather_description,
493 sg_cfg_get_string("smpi/gather"));
494 mpi_coll_gather_fun = (int (*)(void *, int, MPI_Datatype,
495 void *, int, MPI_Datatype, int, MPI_Comm))
496 mpi_coll_gather_description[gather_id].coll;
498 int allgather_id = find_coll_description(mpi_coll_allgather_description,
499 sg_cfg_get_string("smpi/allgather"));
500 mpi_coll_allgather_fun = (int (*)(void *, int, MPI_Datatype,
501 void *, int, MPI_Datatype, MPI_Comm))
502 mpi_coll_allgather_description[allgather_id].coll;
504 int allgatherv_id = find_coll_description(mpi_coll_allgatherv_description,
505 sg_cfg_get_string("smpi/allgatherv"));
506 mpi_coll_allgatherv_fun = (int (*)(void *, int, MPI_Datatype, void *, int *,
507 int *, MPI_Datatype, MPI_Comm))
508 mpi_coll_allgatherv_description[allgatherv_id].coll;
510 int allreduce_id = find_coll_description(mpi_coll_allreduce_description,
511 sg_cfg_get_string("smpi/allreduce"));
512 mpi_coll_allreduce_fun = (int (*)(void *sbuf, void *rbuf, int rcount,
513 MPI_Datatype dtype, MPI_Op op,
515 mpi_coll_allreduce_description[allreduce_id].coll;
517 int alltoall_id = find_coll_description(mpi_coll_alltoall_description,
518 sg_cfg_get_string("smpi/alltoall"));
519 mpi_coll_alltoall_fun = (int (*)(void *, int, MPI_Datatype,
520 void *, int, MPI_Datatype, MPI_Comm))
521 mpi_coll_alltoall_description[alltoall_id].coll;
523 int alltoallv_id = find_coll_description(mpi_coll_alltoallv_description,
524 sg_cfg_get_string("smpi/alltoallv"));
525 mpi_coll_alltoallv_fun = (int (*)(void *, int *, int *, MPI_Datatype,
526 void *, int *, int *, MPI_Datatype,
528 mpi_coll_alltoallv_description[alltoallv_id].coll;
530 int bcast_id = find_coll_description(mpi_coll_bcast_description,
531 sg_cfg_get_string("smpi/bcast"));
532 mpi_coll_bcast_fun = (int (*)(void *buf, int count, MPI_Datatype datatype,
533 int root, MPI_Comm com))
534 mpi_coll_bcast_description[bcast_id].coll;
536 int reduce_id = find_coll_description(mpi_coll_reduce_description,
537 sg_cfg_get_string("smpi/reduce"));
538 mpi_coll_reduce_fun = (int (*)(void *buf, void *rbuf, int count,
539 MPI_Datatype datatype, MPI_Op op,
540 int root, MPI_Comm comm))
541 mpi_coll_reduce_description[reduce_id].coll;
543 int reduce_scatter_id =
544 find_coll_description(mpi_coll_reduce_scatter_description,
545 sg_cfg_get_string("smpi/reduce_scatter"));
546 mpi_coll_reduce_scatter_fun = (int (*)(void *sbuf, void *rbuf, int *rcounts,
547 MPI_Datatype dtype, MPI_Op op,
549 mpi_coll_reduce_scatter_description[reduce_scatter_id].coll;
551 int scatter_id = find_coll_description(mpi_coll_scatter_description,
552 sg_cfg_get_string("smpi/scatter"));
553 mpi_coll_scatter_fun = (int (*)(void *sendbuf, int sendcount,
554 MPI_Datatype sendtype, void *recvbuf,
555 int recvcount, MPI_Datatype recvtype,
556 int root, MPI_Comm comm))
557 mpi_coll_scatter_description[scatter_id].coll;
559 int barrier_id = find_coll_description(mpi_coll_barrier_description,
560 sg_cfg_get_string("smpi/barrier"));
561 mpi_coll_barrier_fun = (int (*)(MPI_Comm comm))
562 mpi_coll_barrier_description[barrier_id].coll;
564 smpi_cpu_threshold = sg_cfg_get_double("smpi/cpu_threshold");
565 smpi_running_power = sg_cfg_get_double("smpi/running_power");
566 if (smpi_cpu_threshold < 0)
567 smpi_cpu_threshold = DBL_MAX;
571 /* Clean IO before the run */
576 MC_do_the_modelcheck_for_real();
580 if (sg_cfg_get_boolean("smpi/display_timing"))
581 XBT_INFO("Simulation time: %g seconds.", SIMIX_get_clock());
583 smpi_global_destroy();