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 "surf/surf.h"
11 #include "simix/smx_private.h"
12 #include "simgrid/sg_config.h"
14 #include <float.h> /* DBL_MAX */
19 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_kernel, smpi,
20 "Logging specific to SMPI (kernel)");
22 typedef struct s_smpi_process_data {
27 smx_rdv_t mailbox_small;
30 void *data; /* user data */
33 int sampling; /* inside an SMPI_SAMPLE_ block? */
34 } s_smpi_process_data_t;
36 static smpi_process_data_t *process_data = NULL;
37 static int process_count = 0;
39 MPI_Comm MPI_COMM_WORLD = MPI_COMM_NULL;
40 int MPI_UNIVERSE_SIZE;
42 MPI_Errhandler *MPI_ERRORS_RETURN = NULL;
43 MPI_Errhandler *MPI_ERRORS_ARE_FATAL = NULL;
44 MPI_Errhandler *MPI_ERRHANDLER_NULL = NULL;
46 #define MAILBOX_NAME_MAXLEN (5 + sizeof(int) * 2 + 1)
48 static char *get_mailbox_name(char *str, int index)
50 snprintf(str, MAILBOX_NAME_MAXLEN, "SMPI-%0*x", (int) (sizeof(int) * 2),
55 static char *get_mailbox_name_small(char *str, int index)
57 snprintf(str, MAILBOX_NAME_MAXLEN, "small%0*x", (int) (sizeof(int) * 2),
62 void smpi_process_init(int *argc, char ***argv)
65 smpi_process_data_t data;
69 proc = SIMIX_process_self();
70 index = atoi((*argv)[1]);
72 smpi_current_rank = index;
74 data = smpi_process_remote_data(index);
75 simcall_process_set_data(proc, data);
78 memmove(&(*argv)[1], &(*argv)[2], sizeof(char *) * (*argc - 2));
79 (*argv)[(*argc) - 1] = NULL;
84 // set the process attached to the mailbox
85 simcall_rdv_set_receiver(data->mailbox_small, proc);
86 XBT_DEBUG("<%d> New process in the game: %p", index, proc);
90 void smpi_process_destroy(void)
92 int index = smpi_process_index();
93 process_data[index]->index = -100;
94 XBT_DEBUG("<%d> Process left the game", index);
98 * @brief Prepares the current process for termination.
100 void smpi_process_finalize(void)
102 // wait for all pending asynchronous comms to finish
103 while (SIMIX_process_has_pending_comms(SIMIX_process_self())) {
104 simcall_process_sleep(0.01);
109 * @brief Check if a process is finalized
111 int smpi_process_finalized()
113 return (smpi_process_index() == -100);
114 // If finalized, this value has been set to -100;
118 * @brief Check if a process is initialized
120 int smpi_process_initialized(void)
122 int index = smpi_process_index();
123 return ((index != -100) && (index != MPI_UNDEFINED)
124 && (process_data[index]->initialized));
128 * @brief Mark a process as initialized (=MPI_Init called)
130 void smpi_process_mark_as_initialized(void)
132 int index = smpi_process_index();
133 if ((index != -100) && (index != MPI_UNDEFINED))
134 process_data[index]->initialized = 1;
139 int smpi_process_argc(void)
141 smpi_process_data_t data = smpi_process_data();
142 return data->argc ? *(data->argc) - 1 : 0;
145 int smpi_process_getarg(integer * index, char *dst, ftnlen len)
147 smpi_process_data_t data = smpi_process_data();
151 if (!data->argc || !data->argv || *index < 1 || *index >= *(data->argc)) {
154 arg = (*data->argv)[*index];
155 for (i = 0; i < len && arg[i] != '\0'; i++) {
158 for (; i < len; i++) {
165 int smpi_global_size(void)
167 char *value = getenv("SMPI_GLOBAL_SIZE");
171 "Please set env var SMPI_GLOBAL_SIZE to expected number of processes.\n");
177 smpi_process_data_t smpi_process_data(void)
179 return SIMIX_process_self_get_data(SIMIX_process_self());
182 smpi_process_data_t smpi_process_remote_data(int index)
184 return process_data[index];
187 void smpi_process_set_user_data(void *data)
189 smpi_process_data_t process_data = smpi_process_data();
190 process_data->data = data;
193 void *smpi_process_get_user_data()
195 smpi_process_data_t process_data = smpi_process_data();
196 return process_data->data;
199 int smpi_process_count(void)
201 return process_count;
204 int smpi_process_index(void)
206 smpi_process_data_t data = smpi_process_data();
207 //return -1 if not initialized
208 return data ? data->index : MPI_UNDEFINED;
211 smx_rdv_t smpi_process_mailbox(void)
213 smpi_process_data_t data = smpi_process_data();
214 return data->mailbox;
217 smx_rdv_t smpi_process_mailbox_small(void)
219 smpi_process_data_t data = smpi_process_data();
220 return data->mailbox_small;
223 smx_rdv_t smpi_process_remote_mailbox(int index)
225 smpi_process_data_t data = smpi_process_remote_data(index);
226 return data->mailbox;
230 smx_rdv_t smpi_process_remote_mailbox_small(int index)
232 smpi_process_data_t data = smpi_process_remote_data(index);
233 return data->mailbox_small;
236 xbt_os_timer_t smpi_process_timer(void)
238 smpi_process_data_t data = smpi_process_data();
242 void smpi_process_simulated_start(void)
244 smpi_process_data_t data = smpi_process_data();
245 data->simulated = SIMIX_get_clock();
248 double smpi_process_simulated_elapsed(void)
250 smpi_process_data_t data = smpi_process_data();
251 return SIMIX_get_clock() - data->simulated;
254 MPI_Comm smpi_process_comm_self(void)
256 smpi_process_data_t data = smpi_process_data();
257 if(data->comm_self==MPI_COMM_NULL){
258 MPI_Group group = smpi_group_new(1);
259 data->comm_self = smpi_comm_new(group);
260 smpi_group_set_mapping(group, smpi_process_index(), 0);
263 return data->comm_self;
266 void smpi_process_set_sampling(int s)
268 smpi_process_data_t data = smpi_process_data();
272 int smpi_process_get_sampling(void)
274 smpi_process_data_t data = smpi_process_data();
275 return data->sampling;
278 void print_request(const char *message, MPI_Request request)
281 ("%s request %p [buf = %p, size = %zu, src = %d, dst = %d, tag = %d, flags = %x]",
282 message, request, request->buf, request->size, request->src,
283 request->dst, request->tag, request->flags);
286 static void smpi_comm_copy_buffer_callback(smx_action_t comm,
287 void *buff, size_t buff_size)
289 XBT_DEBUG("Copy the data over");
290 memcpy(comm->comm.dst_buff, buff, buff_size);
291 if (comm->comm.detached) {
292 // if this is a detached send, the source buffer was duplicated by SMPI
293 // sender to make the original buffer available to the application ASAP
295 //It seems that the request is used after the call there this should
296 //be free somewhereelse but where???
297 //xbt_free(comm->comm.src_data);// inside SMPI the request is keep
298 //inside the user data and should be free
299 comm->comm.src_buff = NULL;
303 void smpi_global_init(void)
307 char name[MAILBOX_NAME_MAXLEN];
309 SIMIX_comm_set_copy_data_callback(&smpi_comm_copy_buffer_callback);
310 process_count = SIMIX_process_count();
311 process_data = xbt_new(smpi_process_data_t, process_count);
312 for (i = 0; i < process_count; i++) {
313 process_data[i] = xbt_new(s_smpi_process_data_t, 1);
314 process_data[i]->index = i;
315 process_data[i]->argc = NULL;
316 process_data[i]->argv = NULL;
317 process_data[i]->mailbox = simcall_rdv_create(get_mailbox_name(name, i));
318 process_data[i]->mailbox_small =
319 simcall_rdv_create(get_mailbox_name_small(name, i));
320 process_data[i]->timer = xbt_os_timer_new();
322 MC_ignore_heap(process_data[i]->timer, xbt_os_timer_size());
323 process_data[i]->comm_self = MPI_COMM_NULL;
324 process_data[i]->initialized = 0;
325 process_data[i]->sampling = 0;
327 group = smpi_group_new(process_count);
328 MPI_COMM_WORLD = smpi_comm_new(group);
329 MPI_UNIVERSE_SIZE = smpi_comm_size(MPI_COMM_WORLD);
330 for (i = 0; i < process_count; i++) {
331 smpi_group_set_mapping(group, i, i);
334 //check correctness of MPI parameters
336 xbt_assert(sg_cfg_get_int("smpi/async_small_thres") <=
337 sg_cfg_get_int("smpi/send_is_detached_thres"));
340 void smpi_global_destroy(void)
342 int count = smpi_process_count();
345 smpi_bench_destroy();
346 while (smpi_group_unuse(smpi_comm_group(MPI_COMM_WORLD)) > 0);
347 xbt_free(MPI_COMM_WORLD);
348 MPI_COMM_WORLD = MPI_COMM_NULL;
349 for (i = 0; i < count; i++) {
350 if(process_data[i]->comm_self!=MPI_COMM_NULL){
351 smpi_group_unuse(smpi_comm_group(process_data[i]->comm_self));
352 smpi_comm_destroy(process_data[i]->comm_self);
354 xbt_os_timer_free(process_data[i]->timer);
355 simcall_rdv_destroy(process_data[i]->mailbox);
356 simcall_rdv_destroy(process_data[i]->mailbox_small);
357 xbt_free(process_data[i]);
359 xbt_free(process_data);
365 /* Fortran specific stuff */
366 /* With smpicc, the following weak symbols are used */
367 /* With smpiff, the following weak symbols are replaced by those in libf2c */
368 int __attribute__ ((weak)) xargc;
369 char ** __attribute__ ((weak)) xargv;
372 void __attribute__ ((weak)) user_main_()
374 xbt_die("Should not be in this smpi_simulated_main");
378 int __attribute__ ((weak)) smpi_simulated_main_(int argc, char **argv)
380 smpi_process_init(&argc, &argv);
382 //xbt_die("Should not be in this smpi_simulated_main");
386 int __attribute__ ((weak)) main(int argc, char **argv)
388 return smpi_main(smpi_simulated_main_, argc, argv);
391 int __attribute__ ((weak)) MAIN__()
393 return smpi_main(smpi_simulated_main_, xargc, xargv);
397 int smpi_main(int (*realmain) (int argc, char *argv[]), int argc, char *argv[])
399 srand(SMPI_RAND_SEED);
401 if (getenv("SMPI_PRETEND_CC") != NULL) {
402 /* Hack to ensure that smpicc can pretend to be a simple
403 * compiler. Particularly handy to pass it to the configuration tools */
407 /* Connect log categories. See xbt/log.c */
408 XBT_LOG_CONNECT(smpi); /* Keep this line as soon as possible in this
409 function: xbt_log_appender_file.c depends on it
410 DO NOT connect this in XBT or so, or it will be
411 useless to xbt_log_appender_file.c */
413 XBT_LOG_CONNECT(instr_smpi);
415 XBT_LOG_CONNECT(smpi_base);
416 XBT_LOG_CONNECT(smpi_bench);
417 XBT_LOG_CONNECT(smpi_coll);
418 XBT_LOG_CONNECT(smpi_colls);
419 XBT_LOG_CONNECT(smpi_comm);
420 XBT_LOG_CONNECT(smpi_dvfs);
421 XBT_LOG_CONNECT(smpi_group);
422 XBT_LOG_CONNECT(smpi_kernel);
423 XBT_LOG_CONNECT(smpi_mpi);
424 XBT_LOG_CONNECT(smpi_mpi_dt);
425 XBT_LOG_CONNECT(smpi_pmpi);
426 XBT_LOG_CONNECT(smpi_replay);
429 TRACE_global_init(&argc, argv);
431 TRACE_add_start_function(TRACE_smpi_alloc);
432 TRACE_add_end_function(TRACE_smpi_release);
435 SIMIX_global_init(&argc, argv);
441 // parse the platform file: get the host list
442 SIMIX_create_environment(argv[1]);
444 SIMIX_function_register_default(realmain);
445 SIMIX_launch_application(argv[2]);
447 int gather_id = find_coll_description(mpi_coll_gather_description,
448 sg_cfg_get_string("smpi/gather"));
449 mpi_coll_gather_fun = (int (*)(void *, int, MPI_Datatype,
450 void *, int, MPI_Datatype, int, MPI_Comm))
451 mpi_coll_gather_description[gather_id].coll;
453 int allgather_id = find_coll_description(mpi_coll_allgather_description,
454 sg_cfg_get_string("smpi/allgather"));
455 mpi_coll_allgather_fun = (int (*)(void *, int, MPI_Datatype,
456 void *, int, MPI_Datatype, MPI_Comm))
457 mpi_coll_allgather_description[allgather_id].coll;
459 int allgatherv_id = find_coll_description(mpi_coll_allgatherv_description,
460 sg_cfg_get_string("smpi/allgatherv"));
461 mpi_coll_allgatherv_fun = (int (*)(void *, int, MPI_Datatype, void *, int *,
462 int *, MPI_Datatype, MPI_Comm))
463 mpi_coll_allgatherv_description[allgatherv_id].coll;
465 int allreduce_id = find_coll_description(mpi_coll_allreduce_description,
466 sg_cfg_get_string("smpi/allreduce"));
467 mpi_coll_allreduce_fun = (int (*)(void *sbuf, void *rbuf, int rcount,
468 MPI_Datatype dtype, MPI_Op op,
470 mpi_coll_allreduce_description[allreduce_id].coll;
472 int alltoall_id = find_coll_description(mpi_coll_alltoall_description,
473 sg_cfg_get_string("smpi/alltoall"));
474 mpi_coll_alltoall_fun = (int (*)(void *, int, MPI_Datatype,
475 void *, int, MPI_Datatype, MPI_Comm))
476 mpi_coll_alltoall_description[alltoall_id].coll;
478 int alltoallv_id = find_coll_description(mpi_coll_alltoallv_description,
479 sg_cfg_get_string("smpi/alltoallv"));
480 mpi_coll_alltoallv_fun = (int (*)(void *, int *, int *, MPI_Datatype,
481 void *, int *, int *, MPI_Datatype,
483 mpi_coll_alltoallv_description[alltoallv_id].coll;
485 int bcast_id = find_coll_description(mpi_coll_bcast_description,
486 sg_cfg_get_string("smpi/bcast"));
487 mpi_coll_bcast_fun = (int (*)(void *buf, int count, MPI_Datatype datatype,
488 int root, MPI_Comm com))
489 mpi_coll_bcast_description[bcast_id].coll;
491 int reduce_id = find_coll_description(mpi_coll_reduce_description,
492 sg_cfg_get_string("smpi/reduce"));
493 mpi_coll_reduce_fun = (int (*)(void *buf, void *rbuf, int count,
494 MPI_Datatype datatype, MPI_Op op,
495 int root, MPI_Comm comm))
496 mpi_coll_reduce_description[reduce_id].coll;
498 int reduce_scatter_id =
499 find_coll_description(mpi_coll_reduce_scatter_description,
500 sg_cfg_get_string("smpi/reduce_scatter"));
501 mpi_coll_reduce_scatter_fun = (int (*)(void *sbuf, void *rbuf, int *rcounts,
502 MPI_Datatype dtype, MPI_Op op,
504 mpi_coll_reduce_scatter_description[reduce_scatter_id].coll;
506 int scatter_id = find_coll_description(mpi_coll_scatter_description,
507 sg_cfg_get_string("smpi/scatter"));
508 mpi_coll_scatter_fun = (int (*)(void *sendbuf, int sendcount,
509 MPI_Datatype sendtype, void *recvbuf,
510 int recvcount, MPI_Datatype recvtype,
511 int root, MPI_Comm comm))
512 mpi_coll_scatter_description[scatter_id].coll;
514 int barrier_id = find_coll_description(mpi_coll_barrier_description,
515 sg_cfg_get_string("smpi/barrier"));
516 mpi_coll_barrier_fun = (int (*)(MPI_Comm comm))
517 mpi_coll_barrier_description[barrier_id].coll;
519 smpi_cpu_threshold = sg_cfg_get_double("smpi/cpu_threshold");
520 smpi_running_power = sg_cfg_get_double("smpi/running_power");
521 if (smpi_cpu_threshold < 0)
522 smpi_cpu_threshold = DBL_MAX;
526 /* Clean IO before the run */
531 MC_do_the_modelcheck_for_real();
535 if (sg_cfg_get_boolean("smpi/display_timing"))
536 XBT_INFO("Simulation time: %g seconds.", SIMIX_get_clock());
538 smpi_global_destroy();