1 /* Copyright (c) 2009, 2010. 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. */
9 #include "msg/msg.h" /* Yeah! If you want to use msg, you need to include msg/msg.h */
10 #include "simgrid/simix.h" /* semaphores for the barrier */
11 #include "xbt.h" /* calloc, printf */
12 #include "instr/instr_private.h"
13 #include <xbt/replay.h>
15 /** @addtogroup MSG_examples
17 * @section MSG_ex_actions Trace driven simulations
19 * The <b>actions/actions.c</b> example demonstrates how to run trace-driven simulations. It
20 * is very handy when you want to test an algorithm or protocol that
21 * does nothing unless it receives some events from outside. For
22 * example, a P2P protocol reacts to requests from the user, but
23 * does nothing if there is no such event.
25 * In such situations, SimGrid allows to write your protocol in your
26 * C file, and the events to react to in a separate text file.
27 * Declare a function handling each of the events that you want to
28 * accept in your trace files, register them using \ref
29 * xbt_replay_action_register in your main, and then use \ref
30 * MSG_action_trace_run to launch the simulation. You can either
31 * have one trace file containing all your events, or a file per
32 * simulated process. Check the tesh files in the example directory
33 * for details on how to do it.
35 * This example uses this approach to replay MPI-like traces. It
36 * comes with a set of event handlers reproducing MPI events. This
37 * is somehow similar to SMPI, yet differently implemented. This
38 * code should probably be changed to use SMPI internals instead,
43 XBT_LOG_NEW_DEFAULT_CATEGORY(actions,
44 "Messages specific for this msg example");
45 int communicator_size = 0;
47 static void action_Isend(const char *const *action);
50 int last_Irecv_sender_id;
53 int allReduce_counter;
54 xbt_dynar_t isends; /* of msg_comm_t */
55 /* Used to implement irecv+wait */
56 xbt_dynar_t irecvs; /* of msg_comm_t */
57 xbt_dynar_t tasks; /* of msg_task_t */
58 } s_process_globals_t, *process_globals_t;
61 static double parse_double(const char *string)
66 value = strtod(string, &endptr);
68 THROWF(unknown_error, 0, "%s is not a double", string);
72 static int get_rank (const char *process_name)
74 return atoi(&(process_name[1]));
77 static void asynchronous_cleanup(void) {
78 process_globals_t globals = (process_globals_t) MSG_process_get_data(MSG_process_self());
80 /* Destroy any isend which correspond to completed communications */
83 while ((found = MSG_comm_testany(globals->isends)) != -1) {
84 xbt_dynar_remove_at(globals->isends,found,&comm);
85 MSG_comm_destroy(comm);
90 static void action_send(const char *const *action)
94 const char *size_str = action[3];
95 double size=parse_double(size_str);
96 double clock = MSG_get_clock(); /* this "call" is free thanks to inlining */
98 sprintf(to, "%s_%s", MSG_process_get_name(MSG_process_self()),action[2]);
100 if (XBT_LOG_ISENABLED(actions, xbt_log_priority_verbose))
101 name = xbt_str_join_array(action, " ");
104 int rank = get_rank(MSG_process_get_name(MSG_process_self()));
105 int dst_traced = get_rank(action[2]);
106 TRACE_smpi_ptp_in(rank, rank, dst_traced, "send");
107 TRACE_smpi_send(rank, rank, dst_traced);
110 XBT_DEBUG("Entering Send: %s (size: %lg)", name, size);
112 action_Isend(action);
114 MSG_task_send(MSG_task_create(name, 0, size, NULL), to);
117 XBT_VERB("%s %f", name, MSG_get_clock() - clock);
122 TRACE_smpi_ptp_out(rank, rank, dst_traced, "send");
125 asynchronous_cleanup();
128 static void action_Isend(const char *const *action)
131 const char *size = action[3];
132 double clock = MSG_get_clock();
133 process_globals_t globals = (process_globals_t) MSG_process_get_data(MSG_process_self());
136 sprintf(to, "%s_%s", MSG_process_get_name(MSG_process_self()),action[2]);
138 MSG_task_isend( MSG_task_create(to,0,parse_double(size),NULL), to);
139 xbt_dynar_push(globals->isends,&comm);
141 XBT_DEBUG("Isend on %s", MSG_process_get_name(MSG_process_self()));
142 XBT_VERB("%s %f", xbt_str_join_array(action, " "), MSG_get_clock() - clock);
144 asynchronous_cleanup();
148 static void action_recv(const char *const *action)
151 char mailbox_name[250];
152 msg_task_t task = NULL;
153 double clock = MSG_get_clock();
155 sprintf(mailbox_name, "%s_%s", action[2],
156 MSG_process_get_name(MSG_process_self()));
158 if (XBT_LOG_ISENABLED(actions, xbt_log_priority_verbose))
159 name = xbt_str_join_array(action, " ");
162 int rank = get_rank(MSG_process_get_name(MSG_process_self()));
163 int src_traced = get_rank(action[2]);
164 TRACE_smpi_ptp_in(rank, src_traced, rank, "recv");
167 XBT_DEBUG("Receiving: %s", name);
168 msg_error_t res = MSG_task_receive(&task, mailbox_name);
169 // MSG_task_receive(&task, MSG_process_get_name(MSG_process_self()));
170 XBT_VERB("%s %f", name, MSG_get_clock() - clock);
173 MSG_task_destroy(task);
178 TRACE_smpi_ptp_out(rank, src_traced, rank, "recv");
179 TRACE_smpi_recv(rank, src_traced, rank);
182 asynchronous_cleanup();
185 static void action_Irecv(const char *const *action)
188 double clock = MSG_get_clock();
189 process_globals_t globals = (process_globals_t) MSG_process_get_data(MSG_process_self());
191 XBT_DEBUG("Irecv on %s", MSG_process_get_name(MSG_process_self()));
193 int rank = get_rank(MSG_process_get_name(MSG_process_self()));
194 int src_traced = get_rank(action[2]);
195 globals->last_Irecv_sender_id = src_traced;
196 MSG_process_set_data(MSG_process_self(), (void *) globals);
198 TRACE_smpi_ptp_in(rank, src_traced, rank, "Irecv");
201 sprintf(mailbox, "%s_%s", action[2],
202 MSG_process_get_name(MSG_process_self()));
204 xbt_dynar_push(globals->tasks,&t);
207 xbt_dynar_get_ptr(globals->tasks, xbt_dynar_length(globals->tasks)-1),
209 xbt_dynar_push(globals->irecvs,&c);
211 XBT_VERB("%s %f", xbt_str_join_array(action, " "), MSG_get_clock() - clock);
214 TRACE_smpi_ptp_out(rank, src_traced, rank, "Irecv");
217 asynchronous_cleanup();
221 static void action_wait(const char *const *action)
224 msg_task_t task = NULL;
226 double clock = MSG_get_clock();
227 process_globals_t globals = (process_globals_t) MSG_process_get_data(MSG_process_self());
229 xbt_assert(xbt_dynar_length(globals->irecvs),
230 "action wait not preceded by any irecv: %s", xbt_str_join_array(action," "));
232 if (XBT_LOG_ISENABLED(actions, xbt_log_priority_verbose))
233 name = xbt_str_join_array(action, " ");
235 process_globals_t counters = (process_globals_t) MSG_process_get_data(MSG_process_self());
236 int src_traced = counters->last_Irecv_sender_id;
237 int rank = get_rank(MSG_process_get_name(MSG_process_self()));
238 TRACE_smpi_ptp_in(rank, src_traced, rank, "wait");
241 XBT_DEBUG("Entering %s", name);
242 comm = xbt_dynar_pop_as(globals->irecvs,msg_comm_t);
243 MSG_comm_wait(comm,-1);
244 task = xbt_dynar_pop_as(globals->tasks,msg_task_t);
245 MSG_comm_destroy(comm);
246 MSG_task_destroy(task);
248 XBT_VERB("%s %f", name, MSG_get_clock() - clock);
251 TRACE_smpi_ptp_out(rank, src_traced, rank, "wait");
252 TRACE_smpi_recv(rank, src_traced, rank);
257 /* FIXME: that's a poor man's implementation: we should take the message exchanges into account */
258 static void action_barrier(const char *const *action)
261 static smx_mutex_t mutex = NULL;
262 static smx_cond_t cond = NULL;
263 static int processes_arrived_sofar=0;
265 if (XBT_LOG_ISENABLED(actions, xbt_log_priority_verbose))
266 name = xbt_str_join_array(action, " ");
268 if (mutex == NULL) { // first arriving on the barrier
269 mutex = simcall_mutex_init();
270 cond = simcall_cond_init();
271 processes_arrived_sofar=0;
273 XBT_DEBUG("Entering barrier: %s (%d already there)", name,processes_arrived_sofar);
275 simcall_mutex_lock(mutex);
276 if (++processes_arrived_sofar == communicator_size) {
277 simcall_cond_broadcast(cond);
278 simcall_mutex_unlock(mutex);
280 simcall_cond_wait(cond,mutex);
281 simcall_mutex_unlock(mutex);
284 XBT_DEBUG("Exiting barrier: %s", name);
286 processes_arrived_sofar--;
287 if (!processes_arrived_sofar) {
288 simcall_cond_destroy(cond);
289 simcall_mutex_destroy(mutex);
297 static void action_reduce(const char *const *action)
300 char *reduce_identifier;
302 double comm_size = parse_double(action[2]);
303 double comp_size = parse_double(action[3]);
304 msg_task_t comp_task = NULL;
305 const char *process_name;
306 double clock = MSG_get_clock();
308 process_globals_t counters = (process_globals_t) MSG_process_get_data(MSG_process_self());
310 xbt_assert(communicator_size, "Size of Communicator is not defined, "
311 "can't use collective operations");
313 process_name = MSG_process_get_name(MSG_process_self());
315 reduce_identifier = bprintf("reduce_%d", counters->reduce_counter++);
317 if (!strcmp(process_name, "p0")) {
318 XBT_DEBUG("%s: %s is the Root", reduce_identifier, process_name);
320 msg_comm_t *comms = xbt_new0(msg_comm_t,communicator_size-1);
321 msg_task_t *tasks = xbt_new0(msg_task_t,communicator_size-1);
322 for (i = 1; i < communicator_size; i++) {
323 sprintf(mailbox, "%s_p%d_p0", reduce_identifier, i);
324 comms[i-1] = MSG_task_irecv(&(tasks[i-1]),mailbox);
326 MSG_comm_waitall(comms,communicator_size-1,-1);
327 for (i = 1; i < communicator_size; i++) {
328 MSG_comm_destroy(comms[i-1]);
329 MSG_task_destroy(tasks[i-1]);
333 comp_task = MSG_task_create("reduce_comp", comp_size, 0, NULL);
334 XBT_DEBUG("%s: computing 'reduce_comp'", reduce_identifier);
335 MSG_task_execute(comp_task);
336 MSG_task_destroy(comp_task);
337 XBT_DEBUG("%s: computed", reduce_identifier);
340 XBT_DEBUG("%s: %s sends", reduce_identifier, process_name);
341 sprintf(mailbox, "%s_%s_p0", reduce_identifier, process_name);
342 XBT_DEBUG("put on %s", mailbox);
343 MSG_task_send(MSG_task_create(reduce_identifier, 0, comm_size, NULL),
347 XBT_VERB("%s %f", xbt_str_join_array(action, " "), MSG_get_clock() - clock);
348 free(reduce_identifier);
351 static void action_bcast(const char *const *action)
354 char *bcast_identifier;
356 double comm_size = parse_double(action[2]);
357 msg_task_t task = NULL;
358 const char *process_name;
359 double clock = MSG_get_clock();
361 process_globals_t counters = (process_globals_t) MSG_process_get_data(MSG_process_self());
363 xbt_assert(communicator_size, "Size of Communicator is not defined, "
364 "can't use collective operations");
366 process_name = MSG_process_get_name(MSG_process_self());
368 bcast_identifier = bprintf("bcast_%d", counters->bcast_counter++);
370 if (!strcmp(process_name, "p0")) {
371 XBT_DEBUG("%s: %s is the Root", bcast_identifier, process_name);
373 msg_comm_t *comms = xbt_new0(msg_comm_t,communicator_size-1);
375 for (i = 1; i < communicator_size; i++) {
376 sprintf(mailbox, "%s_p0_p%d", bcast_identifier, i);
378 MSG_task_isend(MSG_task_create(mailbox,0,comm_size,NULL),
381 MSG_comm_waitall(comms,communicator_size-1,-1);
382 for (i = 1; i < communicator_size; i++)
383 MSG_comm_destroy(comms[i-1]);
386 XBT_DEBUG("%s: all messages sent by %s have been received",
387 bcast_identifier, process_name);
390 sprintf(mailbox, "%s_p0_%s", bcast_identifier, process_name);
391 MSG_task_receive(&task, mailbox);
392 MSG_task_destroy(task);
393 XBT_DEBUG("%s: %s has received", bcast_identifier, process_name);
396 XBT_VERB("%s %f", xbt_str_join_array(action, " "), MSG_get_clock() - clock);
397 free(bcast_identifier);
401 static void action_sleep(const char *const *action)
404 const char *duration = action[2];
405 double clock = MSG_get_clock();
407 if (XBT_LOG_ISENABLED(actions, xbt_log_priority_verbose))
408 name = xbt_str_join_array(action, " ");
410 XBT_DEBUG("Entering %s", name);
411 MSG_process_sleep(parse_double(duration));
412 XBT_VERB("%s %f ", name, MSG_get_clock() - clock);
417 static void action_allReduce(const char *const *action) {
419 char *allreduce_identifier;
421 double comm_size = parse_double(action[2]);
422 double comp_size = parse_double(action[3]);
423 msg_task_t task = NULL, comp_task = NULL;
424 const char *process_name;
425 double clock = MSG_get_clock();
427 process_globals_t counters = (process_globals_t) MSG_process_get_data(MSG_process_self());
429 xbt_assert(communicator_size, "Size of Communicator is not defined, "
430 "can't use collective operations");
432 process_name = MSG_process_get_name(MSG_process_self());
434 allreduce_identifier = bprintf("allReduce_%d", counters->allReduce_counter++);
436 if (!strcmp(process_name, "p0")) {
437 XBT_DEBUG("%s: %s is the Root", allreduce_identifier, process_name);
439 msg_comm_t *comms = xbt_new0(msg_comm_t,communicator_size-1);
440 msg_task_t *tasks = xbt_new0(msg_task_t,communicator_size-1);
441 for (i = 1; i < communicator_size; i++) {
442 sprintf(mailbox, "%s_p%d_p0", allreduce_identifier, i);
443 comms[i-1] = MSG_task_irecv(&(tasks[i-1]),mailbox);
445 MSG_comm_waitall(comms,communicator_size-1,-1);
446 for (i = 1; i < communicator_size; i++) {
447 MSG_comm_destroy(comms[i-1]);
448 MSG_task_destroy(tasks[i-1]);
452 comp_task = MSG_task_create("allReduce_comp", comp_size, 0, NULL);
453 XBT_DEBUG("%s: computing 'reduce_comp'", allreduce_identifier);
454 MSG_task_execute(comp_task);
455 MSG_task_destroy(comp_task);
456 XBT_DEBUG("%s: computed", allreduce_identifier);
458 for (i = 1; i < communicator_size; i++) {
459 sprintf(mailbox, "%s_p0_p%d", allreduce_identifier, i);
461 MSG_task_isend(MSG_task_create(mailbox,0,comm_size,NULL),
464 MSG_comm_waitall(comms,communicator_size-1,-1);
465 for (i = 1; i < communicator_size; i++)
466 MSG_comm_destroy(comms[i-1]);
469 XBT_DEBUG("%s: all messages sent by %s have been received",
470 allreduce_identifier, process_name);
473 XBT_DEBUG("%s: %s sends", allreduce_identifier, process_name);
474 sprintf(mailbox, "%s_%s_p0", allreduce_identifier, process_name);
475 XBT_DEBUG("put on %s", mailbox);
476 MSG_task_send(MSG_task_create(allreduce_identifier, 0, comm_size, NULL),
479 sprintf(mailbox, "%s_p0_%s", allreduce_identifier, process_name);
480 MSG_task_receive(&task, mailbox);
481 MSG_task_destroy(task);
482 XBT_DEBUG("%s: %s has received", allreduce_identifier, process_name);
485 XBT_VERB("%s %f", xbt_str_join_array(action, " "), MSG_get_clock() - clock);
486 free(allreduce_identifier);
489 static void action_comm_size(const char *const *action)
492 const char *size = action[2];
493 double clock = MSG_get_clock();
495 if (XBT_LOG_ISENABLED(actions, xbt_log_priority_verbose))
496 name = xbt_str_join_array(action, " ");
497 communicator_size = parse_double(size);
498 XBT_VERB("%s %f", name, MSG_get_clock() - clock);
502 static void action_compute(const char *const *action)
505 const char *amout = action[2];
506 msg_task_t task = MSG_task_create(name, parse_double(amout), 0, NULL);
507 double clock = MSG_get_clock();
509 if (XBT_LOG_ISENABLED(actions, xbt_log_priority_verbose))
510 name = xbt_str_join_array(action, " ");
511 XBT_DEBUG("Entering %s", name);
512 MSG_task_execute(task);
513 MSG_task_destroy(task);
514 XBT_VERB("%s %f", name, MSG_get_clock() - clock);
518 static void action_init(const char *const *action)
521 TRACE_smpi_init(get_rank(MSG_process_get_name(MSG_process_self())));
523 XBT_DEBUG("Initialize the counters");
524 process_globals_t globals = (process_globals_t) calloc(1, sizeof(s_process_globals_t));
525 globals->isends = xbt_dynar_new(sizeof(msg_comm_t),NULL);
526 globals->irecvs = xbt_dynar_new(sizeof(msg_comm_t),NULL);
527 globals->tasks = xbt_dynar_new(sizeof(msg_task_t),NULL);
528 MSG_process_set_data(MSG_process_self(),globals);
532 static void action_finalize(const char *const *action)
535 TRACE_smpi_finalize(get_rank(MSG_process_get_name(MSG_process_self())));
537 process_globals_t globals = (process_globals_t) MSG_process_get_data(MSG_process_self());
539 xbt_dynar_free_container(&(globals->isends));
540 xbt_dynar_free_container(&(globals->irecvs));
541 xbt_dynar_free_container(&(globals->tasks));
547 int main(int argc, char *argv[])
549 msg_error_t res = MSG_OK;
551 /* Check the given arguments */
552 MSG_init(&argc, argv);
554 printf("Usage: %s platform_file deployment_file [action_files]\n",
557 ("example: %s msg_platform.xml msg_deployment.xml actions # if all actions are in the same file\n",
560 ("example: %s msg_platform.xml msg_deployment.xml # if actions are in separate files, specified in deployment\n",
565 /* Simulation setting */
566 MSG_create_environment(argv[1]);
568 /* No need to register functions as in classical MSG programs: the actions get started anyway */
569 MSG_launch_application(argv[2]);
571 /* Action registration */
572 xbt_replay_action_register("init", action_init);
573 xbt_replay_action_register("finalize", action_finalize);
574 xbt_replay_action_register("comm_size",action_comm_size);
575 xbt_replay_action_register("send", action_send);
576 xbt_replay_action_register("Isend", action_Isend);
577 xbt_replay_action_register("recv", action_recv);
578 xbt_replay_action_register("Irecv", action_Irecv);
579 xbt_replay_action_register("wait", action_wait);
580 xbt_replay_action_register("barrier", action_barrier);
581 xbt_replay_action_register("bcast", action_bcast);
582 xbt_replay_action_register("reduce", action_reduce);
583 xbt_replay_action_register("allReduce",action_allReduce);
584 xbt_replay_action_register("sleep", action_sleep);
585 xbt_replay_action_register("compute", action_compute);
588 /* Actually do the simulation using MSG_action_trace_run */
589 res = MSG_action_trace_run(argv[3]); // it's ok to pass a NULL argument here
591 XBT_INFO("Simulation time %g", MSG_get_clock());