malleabilityManager.c 29.8 KB
Newer Older
1
#include <pthread.h>
2
#include <string.h>
3
4
#include "malleabilityManager.h"
#include "malleabilityStates.h"
5
#include "malleabilityDataStructures.h"
6
#include "malleabilityTypes.h"
iker_martin's avatar
iker_martin committed
7
#include "malleabilityZombies.h"
8
#include "spawn_methods/GenericSpawn.h"
9
10
11
12
13
14
15
16
17
18
19
20
21
22
#include "CommDist.h"

#define MALLEABILITY_USE_SYNCHRONOUS 0
#define MALLEABILITY_USE_ASYNCHRONOUS 1


void send_data(int numP_children, malleability_data_t *data_struct, int is_asynchronous);
void recv_data(int numP_parents, malleability_data_t *data_struct, int is_asynchronous);

void Children_init();
int spawn_step();
int start_redistribution();
int check_redistribution();
int end_redistribution();
iker_martin's avatar
iker_martin committed
23
int shrink_redistribution();
24

25
26
27
void comm_node_data(int rootBcast, int is_child_group);
void def_nodeinfo_type(MPI_Datatype *node_type);

28
29
int thread_creation();
int thread_check();
30
void* thread_async_work();
31

32
33
void print_comms_state();

34
typedef struct {
35
  int spawn_method;
36
  int spawn_dist;
37
38
39
  int spawn_strategies;
  //int spawn_is_single;
  //int spawn_threaded;
40
41
42
  int comm_type;
  int comm_threaded;

43
  int grp;
44
45
46
47
  configuration *config_file;
  results_data *results;
} malleability_config_t;

iker_martin's avatar
iker_martin committed
48
typedef struct { //FIXME numC_spawned no se esta usando
49
  int myId, numP, numC, numC_spawned, root, root_parents;
50
51
52
  pthread_t async_thread;
  MPI_Comm comm, thread_comm;
  MPI_Comm intercomm;
53
  MPI_Comm user_comm;
54
  
55
  char *name_exec, *nodelist;
56
  int num_cpus, num_nodes, nodelist_len;
57
58
} malleability_t;

59
int state = MALL_UNRESERVED; //FIXME Mover a otro lado
60
61
62
63
64
65
66
67
68

malleability_config_t *mall_conf;
malleability_t *mall;

malleability_data_t *rep_s_data;
malleability_data_t *dist_s_data;
malleability_data_t *rep_a_data;
malleability_data_t *dist_a_data;

69
/*
70
71
72
73
74
75
76
77
 * Inicializa la reserva de memoria para el modulo de maleabilidad
 * creando todas las estructuras necesarias y copias de comunicadores
 * para no interferir en la aplicación.
 *
 * Si es llamada por un grupo de procesos creados de forma dinámica,
 * inicializan la comunicacion con sus padres. En este caso, al terminar 
 * la comunicacion los procesos hijo estan preparados para ejecutar la
 * aplicacion.
78
 */
79
int init_malleability(int myId, int numP, int root, MPI_Comm comm, char *name_exec, char *nodelist, int num_cpus, int num_nodes) {
80
81
82
83
84
85
86
87
88
  MPI_Comm dup_comm, thread_comm;

  mall_conf = (malleability_config_t *) malloc(sizeof(malleability_config_t));
  mall = (malleability_t *) malloc(sizeof(malleability_t));
  rep_s_data = (malleability_data_t *) malloc(sizeof(malleability_data_t));
  dist_s_data = (malleability_data_t *) malloc(sizeof(malleability_data_t));
  rep_a_data = (malleability_data_t *) malloc(sizeof(malleability_data_t));
  dist_a_data = (malleability_data_t *) malloc(sizeof(malleability_data_t));

89
90
  MPI_Comm_dup(comm, &dup_comm);
  MPI_Comm_dup(comm, &thread_comm);
91
92
  MPI_Comm_set_name(dup_comm, "MPI_COMM_MALL");
  MPI_Comm_set_name(thread_comm, "MPI_COMM_MALL_THREAD");
93
94
95
96

  mall->myId = myId;
  mall->numP = numP;
  mall->root = root;
97
  mall->comm = dup_comm;
98
  mall->thread_comm = thread_comm;
99
  mall->user_comm = comm;
100

101
  mall->name_exec = name_exec;
102
103
104
  mall->nodelist = nodelist;
  mall->num_cpus = num_cpus;
  mall->num_nodes = num_nodes;
105
106
107
108
109
110

  rep_s_data->entries = 0;
  rep_a_data->entries = 0;
  dist_s_data->entries = 0;
  dist_a_data->entries = 0;

111
  state = MALL_NOT_STARTED;
112
113
114
115
116

  // Si son el primer grupo de procesos, obtienen los datos de los padres
  MPI_Comm_get_parent(&(mall->intercomm));
  if(mall->intercomm != MPI_COMM_NULL ) { 
    Children_init();
117
    return MALLEABILITY_CHILDREN;
118
  }
iker_martin's avatar
iker_martin committed
119

120
121
122
123
124
125
126
  if(nodelist != NULL) { //TODO To be deprecated by using Slurm or else statement
    mall->nodelist_len = strlen(nodelist);
  } else { // If no nodelist is detected, get it from the actual run
    mall->nodelist = malloc(MPI_MAX_PROCESSOR_NAME * sizeof(char));
    MPI_Get_processor_name(mall->nodelist, &mall->nodelist_len);
    //TODO Get name of each process and create real nodelist
  }
127

iker_martin's avatar
iker_martin committed
128
  zombies_service_init();
129
  return MALLEABILITY_NOT_CHILDREN;
130
131
}

132
133
134
135
136
/*
 * Elimina toda la memoria reservado por el modulo
 * de maleabilidad y asegura que los zombies
 * despierten si los hubiese.
 */
137
138
139
140
141
142
143
144
145
146
147
void free_malleability() {	
  free_malleability_data_struct(rep_s_data);
  free_malleability_data_struct(rep_a_data);
  free_malleability_data_struct(dist_s_data);
  free_malleability_data_struct(dist_a_data);

  free(rep_s_data);
  free(rep_a_data);
  free(dist_s_data);
  free(dist_a_data);

148
149
  if(mall->comm != MPI_COMM_WORLD) MPI_Comm_free(&(mall->comm));
  if(mall->thread_comm != MPI_COMM_WORLD) MPI_Comm_free(&(mall->thread_comm));
150
151
  free(mall);
  free(mall_conf);
iker_martin's avatar
iker_martin committed
152
153
154
155

  zombies_awake();
  zombies_service_free();

156
  state = MALL_UNRESERVED;
157
158
}

159
160
/* 
 * TODO Reescribir
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
 * Se realiza el redimensionado de procesos por parte de los padres.
 *
 * Se crean los nuevos procesos con la distribucion fisica elegida y
 * a continuacion se transmite la informacion a los mismos.
 *
 * Si hay datos asincronos a transmitir, primero se comienza a
 * transmitir estos y se termina la funcion. Se tiene que comprobar con
 * llamando a la función de nuevo que se han terminado de enviar
 *
 * Si hay ademas datos sincronos a enviar, no se envian aun.
 *
 * Si solo hay datos sincronos se envian tras la creacion de los procesos
 * y finalmente se desconectan los dos grupos de procesos.
 */
int malleability_checkpoint() {
176
177
178
179
180
181
182
183
  double end_real_time;

  switch(state) {
    case MALL_UNRESERVED:
      break;
    case MALL_NOT_STARTED:
      // Comprobar si se tiene que realizar un redimensionado
      //if(CHECK_RMS()) {return MALL_DENIED;}
184

185
186
187
188
189
190
      state = spawn_step();

      if (state == MALL_SPAWN_COMPLETED || state == MALL_SPAWN_ADAPT_POSTPONE){
        malleability_checkpoint();
      }
      break;
191

192
193
194
195
196
197
    case MALL_SPAWN_PENDING: // Comprueba si el spawn ha terminado y comienza la redistribucion
    case MALL_SPAWN_SINGLE_PENDING:
      state = check_spawn_state(&(mall->intercomm), mall->comm, &end_real_time);
      if (state == MALL_SPAWN_COMPLETED || state == MALL_SPAWN_ADAPTED) {
        mall_conf->results->spawn_time[mall_conf->grp] = MPI_Wtime() - mall_conf->results->spawn_start;
        mall_conf->results->spawn_real_time[mall_conf->grp] = end_real_time - mall_conf->results->spawn_start;
198

199
200
201
        malleability_checkpoint();
      }
      break;
202

203
204
205
206
207
    case MALL_SPAWN_ADAPT_POSTPONE:
    case MALL_SPAWN_COMPLETED:
      state = start_redistribution();
      malleability_checkpoint();
      break;
208

209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
    case MALL_DIST_PENDING:
      if(mall_conf->comm_type == MAL_USE_THREAD) {
        state = thread_check();
      } else {
        state = check_redistribution();
      }
      if(state != MALL_DIST_PENDING) {
        malleability_checkpoint();
      }
      break;

    case MALL_SPAWN_ADAPT_PENDING:
      mall_conf->results->spawn_start = MPI_Wtime();
      unset_spawn_postpone_flag(state);
      state = check_spawn_state(&(mall->intercomm), mall->comm, &end_real_time);
224

225
      if(!malleability_spawn_contains_strat(mall_conf->spawn_strategies, MALL_SPAWN_PTHREAD, NULL)) {
226
227
        mall_conf->results->spawn_time[mall_conf->grp] = MPI_Wtime() - mall_conf->results->spawn_start;
      }
228
      break;
229

230
    case MALL_SPAWN_ADAPTED:
231
      state = shrink_redistribution();
232
      malleability_checkpoint();
233
      break;
234

235
236
237
238
    case MALL_DIST_COMPLETED: //TODO No es esto muy feo?
      state = MALL_COMPLETED;
      break;
  }
239
240
241
242
243
244
245
246
247
248
249
250
251
  return state;
}

// Funciones solo necesarias por el benchmark
//-------------------------------------------------------------------------------------------------------------
void set_benchmark_grp(int grp) {
  mall_conf->grp = grp;
}

void set_benchmark_configuration(configuration *config_file) {
  mall_conf->config_file = config_file;
}

252
void get_benchmark_configuration(configuration **config_file) {
253
254
255
256
257
258
259
  *config_file = mall_conf->config_file;
}

void set_benchmark_results(results_data *results) {
  mall_conf->results = results;
}

260
void get_benchmark_results(results_data **results) {
261
262
263
264
  *results = mall_conf->results;
}
//-------------------------------------------------------------------------------------------------------------

265
266
267
void set_malleability_configuration(int spawn_method, int spawn_strategies, int spawn_dist, int comm_type, int comm_threaded) {
  mall_conf->spawn_method = spawn_method;
  mall_conf->spawn_strategies = spawn_strategies;
268
269
270
271
272
273
274
  mall_conf->spawn_dist = spawn_dist;
  mall_conf->comm_type = comm_type;
  mall_conf->comm_threaded = comm_threaded;
}

/*
 * To be deprecated
275
 * Tiene que ser llamado despues de setear la config
276
277
 */
void set_children_number(int numC){
278
  if((mall_conf->spawn_method == MALL_SPAWN_MERGE) && (numC >= mall->numP)) {
279
280
281
282
283
    mall->numC = numC;
    mall->numC_spawned = numC - mall->numP;

    if(numC == mall->numP) { // Migrar
      mall->numC_spawned = numC;
284
      mall_conf->spawn_method = MALL_SPAWN_BASELINE;
285
286
287
288
289
290
291
292
293
294
295
296
    }
  } else {
    mall->numC = numC;
    mall->numC_spawned = numC;
  }
}

/*
 * TODO
 */
void get_malleability_user_comm(MPI_Comm *comm) {
  *comm = mall->user_comm;
297
298
299
300
301
302
303
304
305
306
}

/*
 * Anyade a la estructura concreta de datos elegida
 * el nuevo set de datos "data" de un total de "total_qty" elementos.
 *
 * Los datos variables se tienen que anyadir cuando quieran ser mandados, no antes
 *
 * Mas informacion en la funcion "add_data".
 */
307
void malleability_add_data(void *data, size_t total_qty, int type, int is_replicated, int is_constant) {
308
309
310
311
312
313
314
315
316
317
318

  if(is_constant) {
    if(is_replicated) {
      add_data(data, total_qty, type, 0, rep_s_data); //FIXME Numero magico
    } else {
      add_data(data, total_qty, type, 0, dist_s_data); //FIXME Numero magico
    }
  } else {
    if(is_replicated) {
      add_data(data, total_qty, type, 0, rep_a_data); //FIXME Numero magico || Un request?
    } else {
319
      size_t total_reqs = 0;
320
321
322
323
324
325
      
      if(mall_conf->comm_type  == MAL_USE_NORMAL) {
        total_reqs = 1;
      } else if(mall_conf->comm_type  == MAL_USE_IBARRIER) {
        total_reqs = 2;
      } else if(mall_conf->comm_type  == MAL_USE_POINT) {
326
        total_reqs = mall->numC;
327
328
329
330
331
332
333
      }
      
      add_data(data, total_qty, type, total_reqs, dist_a_data);
    }
  }
}

334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
/*
 * Modifica en la estructura concreta de datos elegida en el indice "index"
 * con el set de datos "data" de un total de "total_qty" elementos.
 *
 * Los datos variables se tienen que modificar cuando quieran ser mandados, no antes
 *
 * Mas informacion en la funcion "modify_data".
 */
void malleability_modify_data(void *data, size_t index, size_t total_qty, int type, int is_replicated, int is_constant) {
  if(is_constant) {
    if(is_replicated) {
      modify_data(data, index, total_qty, type, 0, rep_s_data); //FIXME Numero magico
    } else {
      modify_data(data, index, total_qty, type, 0, dist_s_data); //FIXME Numero magico
    }
  } else {
    if(is_replicated) {
      modify_data(data, index, total_qty, type, 0, rep_a_data); //FIXME Numero magico || UN request?
    } else {
      size_t total_reqs = 0;
      
      if(mall_conf->comm_type  == MAL_USE_NORMAL) {
        total_reqs = 1;
      } else if(mall_conf->comm_type  == MAL_USE_IBARRIER) {
        total_reqs = 2;
      } else if(mall_conf->comm_type  == MAL_USE_POINT) {
        total_reqs = mall->numC;
      }
      
      modify_data(data, index, total_qty, type, total_reqs, dist_a_data); //FIXME Numero magico
    }
  }
}

368
369
370
371
/*
 * Devuelve el numero de entradas para la estructura de descripcion de 
 * datos elegida.
 */
372
void malleability_get_entries(size_t *entries, int is_replicated, int is_constant){
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
  
  if(is_constant) {
    if(is_replicated) {
      *entries = rep_s_data->entries;
    } else {
      *entries = dist_s_data->entries;
    }
  } else {
    if(is_replicated) {
      *entries = rep_a_data->entries;
    } else {
      *entries = dist_a_data->entries;
    }
  }
}

/*
 * Devuelve el elemento de la lista "index" al usuario.
 * La devolución es en el mismo orden que lo han metido los padres
 * con la funcion "malleability_add_data()".
 * Es tarea del usuario saber el tipo de esos datos.
 * TODO Refactor a que sea automatico
 */
396
void malleability_get_data(void **data, int index, int is_replicated, int is_constant) {
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
  malleability_data_t *data_struct;

  if(is_constant) {
    if(is_replicated) {
      data_struct = rep_s_data;
    } else {
      data_struct = dist_s_data;
    }
  } else {
    if(is_replicated) {
      data_struct = rep_a_data;
    } else {
      data_struct = dist_a_data;
    }
  }

413
  *data = data_struct->arrays[index];
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
}


//======================================================||
//================PRIVATE FUNCTIONS=====================||
//================DATA COMMUNICATION====================||
//======================================================||
//======================================================||


/*
 * Funcion generalizada para enviar datos desde los hijos.
 * La asincronizidad se refiere a si el hilo padre e hijo lo hacen
 * de forma bloqueante o no. El padre puede tener varios hilos.
 */
void send_data(int numP_children, malleability_data_t *data_struct, int is_asynchronous) {
430
  size_t i;
431
432
433
434
435
  char *aux;

  if(is_asynchronous) {
    for(i=0; i < data_struct->entries; i++) {
      aux = (char *) data_struct->arrays[i]; //TODO Comprobar que realmente es un char
436
      send_async(aux, data_struct->qty[i], mall->myId, mall->numP, mall->intercomm, numP_children, data_struct->requests, mall_conf->comm_type);
437
438
439
440
    }
  } else {
    for(i=0; i < data_struct->entries; i++) {
      aux = (char *) data_struct->arrays[i]; //TODO Comprobar que realmente es un char
441
      send_sync(aux, data_struct->qty[i], mall->myId, mall->numP, mall->intercomm, numP_children);
442
443
444
445
446
447
448
449
450
451
    }
  }
}

/*
 * Funcion generalizada para recibir datos desde los hijos.
 * La asincronizidad se refiere a si el hilo padre e hijo lo hacen
 * de forma bloqueante o no. El padre puede tener varios hilos.
 */
void recv_data(int numP_parents, malleability_data_t *data_struct, int is_asynchronous) {
452
  size_t i;
453
454
455
456
457
  char *aux;

  if(is_asynchronous) {
    for(i=0; i < data_struct->entries; i++) {
      aux = (char *) data_struct->arrays[i]; //TODO Comprobar que realmente es un char
458
      recv_async(&aux, data_struct->qty[i], mall->myId, mall->numP, mall->intercomm, numP_parents, mall_conf->comm_type);
459
460
461
462
463
      data_struct->arrays[i] = (void *) aux;
    }
  } else {
    for(i=0; i < data_struct->entries; i++) {
      aux = (char *) data_struct->arrays[i]; //TODO Comprobar que realmente es un char
464
      recv_sync(&aux, data_struct->qty[i], mall->myId, mall->numP, mall->intercomm, numP_parents);
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
      data_struct->arrays[i] = (void *) aux;
    }
  }
}

//======================================================||
//================PRIVATE FUNCTIONS=====================||
//=====================CHILDREN=========================||
//======================================================||
//======================================================||

/*
 * Inicializacion de los datos de los hijos.
 * En la misma se reciben datos de los padres: La configuracion
 * de la ejecucion a realizar; y los datos a recibir de los padres
 * ya sea de forma sincrona, asincrona o ambas.
 */
void Children_init() {
483
484
  size_t i;
  int numP_parents, root_parents;
485
  int is_intercomm;
486

487
488
  malleability_connect_children(mall->myId, mall->numP, mall->root, mall->comm, &numP_parents, &root_parents, &(mall->intercomm));
  MPI_Comm_test_inter(mall->intercomm, &is_intercomm);
489
  // TODO A partir de este punto tener en cuenta si es BASELINE o MERGE
490

491
  recv_config_file(mall->root, mall->intercomm, &(mall_conf->config_file));
492
  comm_node_data(root_parents, MALLEABILITY_CHILDREN);
493

494
  mall_conf->results = (results_data *) malloc(sizeof(results_data));
495
  init_results_data(mall_conf->results, mall_conf->config_file->n_resizes, mall_conf->config_file->n_stages, RESULTS_INIT_DATA_QTY);
496
497
498
499
500
501
502
503
504
505
506
507
508

  if(dist_a_data->entries || rep_a_data->entries) { // Recibir datos asincronos
    comm_data_info(rep_a_data, dist_a_data, MALLEABILITY_CHILDREN, mall->myId, root_parents, mall->intercomm);

    if(mall_conf->comm_type == MAL_USE_NORMAL || mall_conf->comm_type == MAL_USE_IBARRIER || mall_conf->comm_type == MAL_USE_POINT) {
      recv_data(numP_parents, dist_a_data, 1);

    } else if (mall_conf->comm_type == MAL_USE_THREAD) { //TODO Modificar uso para que tenga sentido comm_threaded
      recv_data(numP_parents, dist_a_data, 0);
    }
    mall_conf->results->async_end= MPI_Wtime(); // Obtener timestamp de cuando termina comm asincrona
  }
  
509
  comm_data_info(rep_s_data, dist_s_data, MALLEABILITY_CHILDREN, mall->myId, root_parents, mall->intercomm);
510
511
512
513
514
515
516
517
  if(dist_s_data->entries || rep_s_data->entries) { // Recibir datos sincronos
    recv_data(numP_parents, dist_s_data, 0);

    mall_conf->results->sync_end = MPI_Wtime(); // Obtener timestamp de cuando termina comm sincrona

    // TODO Crear funcion especifica y anyadir para Asinc
    // TODO Tener en cuenta el tipo y qty
    for(i=0; i<rep_s_data->entries; i++) {
518
519
520
521
522
523
      MPI_Datatype datatype;
      if(rep_s_data->types[i] == MAL_INT) {
        datatype = MPI_INT;
      } else {
        datatype = MPI_CHAR;
      }
524
      MPI_Bcast(rep_s_data->arrays[i], rep_s_data->qty[i], datatype, root_parents, mall->intercomm);
525
526
527
    } 
  }

528
  // Guardar los resultados de esta transmision
529
  comm_results(mall_conf->results, mall->root, mall_conf->config_file->n_resizes, mall->intercomm);
530
531
532
533
534
535
536
  if(!is_intercomm) {
    if(mall->thread_comm != MPI_COMM_WORLD) MPI_Comm_free(&(mall->thread_comm));
    if(mall->comm != MPI_COMM_WORLD) MPI_Comm_free(&(mall->comm));
    if(mall->user_comm != MPI_COMM_WORLD) MPI_Comm_free(&(mall->user_comm)); //TODO No es peligroso?

    MPI_Comm_dup(mall->intercomm, &(mall->thread_comm));
    MPI_Comm_dup(mall->intercomm, &(mall->comm));
537
    MPI_Comm_dup(mall->intercomm, &(mall->user_comm)); 
538
  }
539

540
  MPI_Comm_disconnect(&(mall->intercomm));
541
542
543
544
545
546
547
548
549
550
551
552
553
554
}

//======================================================||
//================PRIVATE FUNCTIONS=====================||
//=====================PARENTS==========================||
//======================================================||
//======================================================||

/*
 * Se encarga de realizar la creacion de los procesos hijos.
 * Si se pide en segundo plano devuelve el estado actual.
 */
int spawn_step(){
  mall_conf->results->spawn_start = MPI_Wtime();
555
 
556
  state = init_spawn(mall->name_exec, mall->num_cpus, mall->num_nodes, mall->nodelist, mall->myId, mall->numP, mall->numC, mall->root, mall_conf->spawn_dist, mall_conf->spawn_method, mall_conf->spawn_strategies, mall->thread_comm, &(mall->intercomm));
557

558
  if(!malleability_spawn_contains_strat(mall_conf->spawn_strategies, MALL_SPAWN_PTHREAD, NULL)) {
559
560
561
562
563
      mall_conf->results->spawn_time[mall_conf->grp] = MPI_Wtime() - mall_conf->results->spawn_start;
  }
  return state;
}

564

565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
/*
 * Comienza la redistribucion de los datos con el nuevo grupo de procesos.
 *
 * Primero se envia la configuracion a utilizar al nuevo grupo de procesos y a continuacion
 * se realiza el envio asincrono y/o sincrono si lo hay.
 *
 * En caso de que haya comunicacion asincrona, se comienza y se termina la funcion 
 * indicando que se ha comenzado un envio asincrono.
 *
 * Si no hay comunicacion asincrono se pasa a realizar la sincrona si la hubiese.
 *
 * Finalmente se envian datos sobre los resultados a los hijos y se desconectan ambos
 * grupos de procesos.
 */
int start_redistribution() {
580
581
582
583
584
585
586
587
  int rootBcast, is_intercomm;

  is_intercomm = 0;
  if(mall->intercomm != MPI_COMM_NULL) {
    MPI_Comm_test_inter(mall->intercomm, &is_intercomm);
  } else { 
    // Si no tiene comunicador creado, se debe a que se ha pospuesto el Spawn
    //   y se trata del spawn Merge Shrink
588
    MPI_Comm_dup(mall->comm, &(mall->intercomm));
589
  }
590

591
592
593
594
595
  if(is_intercomm) {
    rootBcast = mall->myId == mall->root ? MPI_ROOT : MPI_PROC_NULL;
  } else {
    rootBcast = mall->root;
  }
596

597
  send_config_file(mall_conf->config_file, rootBcast, mall->intercomm);
598
  comm_node_data(rootBcast, MALLEABILITY_NOT_CHILDREN);
599

600
  if(dist_a_data->entries || rep_a_data->entries) { // Enviar datos asincronos
601
602
603
604
605
606
    mall_conf->results->async_start = MPI_Wtime();
    comm_data_info(rep_a_data, dist_a_data, MALLEABILITY_NOT_CHILDREN, mall->myId, mall->root, mall->intercomm);
    if(mall_conf->comm_type == MAL_USE_THREAD) {
      return thread_creation();
    } else {
      send_data(mall->numC, dist_a_data, MALLEABILITY_USE_ASYNCHRONOUS);
607
      return MALL_DIST_PENDING;
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
    }
  } 
  return end_redistribution();
}


/*
 * @deprecated
 * Comprueba si la redistribucion asincrona ha terminado. 
 * Si no ha terminado la funcion termina indicandolo, en caso contrario,
 * se continua con la comunicacion sincrona, el envio de resultados y
 * se desconectan los grupos de procesos.
 *
 * Esta funcion permite dos modos de funcionamiento al comprobar si la
 * comunicacion asincrona ha terminado.
 * Si se utiliza el modo "MAL_USE_NORMAL" o "MAL_USE_POINT", se considera 
 * terminada cuando los padres terminan de enviar.
 * Si se utiliza el modo "MAL_USE_IBARRIER", se considera terminada cuando
 * los hijos han terminado de recibir.
 */
int check_redistribution() {
  int completed, all_completed, test_err;
  MPI_Request *req_completed;
//dist_a_data->requests[0][X] //FIXME Numero magico 0 -- Modificar para que sea un for?

  if (mall_conf->comm_type == MAL_USE_POINT) {
    test_err = MPI_Testall(mall->numC, dist_a_data->requests[0], &completed, MPI_STATUSES_IGNORE);
  } else {
    if(mall_conf->comm_type == MAL_USE_NORMAL) {
      req_completed = &(dist_a_data->requests[0][0]);
    } else if (mall_conf->comm_type == MAL_USE_IBARRIER) {
      req_completed = &(dist_a_data->requests[0][1]);
    }

    test_err = MPI_Test(req_completed, &completed, MPI_STATUS_IGNORE);
  }
 
  if (test_err != MPI_SUCCESS && test_err != MPI_ERR_PENDING) {
    printf("P%d aborting -- Test Async\n", mall->myId);
    MPI_Abort(MPI_COMM_WORLD, test_err);
  }

  MPI_Allreduce(&completed, &all_completed, 1, MPI_INT, MPI_MIN, mall->comm);
651
  if(!all_completed) return MALL_DIST_PENDING; // Continue only if asynchronous send has ended 
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
  

  if(mall_conf->comm_type == MAL_USE_IBARRIER) {
    MPI_Wait(&(dist_a_data->requests[0][0]), MPI_STATUS_IGNORE); // Indicar como completado el envio asincrono
    //Para la desconexión de ambos grupos de procesos es necesario indicar a MPI que esta comm
    //ha terminado, aunque solo se pueda llegar a este punto cuando ha terminado
  }
  return end_redistribution();
}


/*
 * Termina la redistribución de los datos con los hijos, comprobando
 * si se han realizado iteraciones con comunicaciones en segundo plano
 * y enviando cuantas iteraciones se han realizado a los hijos.
 *
 * Además se realizan las comunicaciones síncronas se las hay.
 * Finalmente termina enviando los datos temporales a los hijos.
 */ 
int end_redistribution() {
672
673
  size_t i;
  int is_intercomm, rootBcast, local_state;
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688

  is_intercomm = 0;
  if(mall->intercomm != MPI_COMM_NULL) {
    MPI_Comm_test_inter(mall->intercomm, &is_intercomm);
  } else { 
    // Si no tiene comunicador creado, se debe a que se ha pospuesto el Spawn
    //   y se trata del spawn Merge Shrink
    mall->intercomm = mall->comm;
  }
  if(is_intercomm) {
    rootBcast = mall->myId == mall->root ? MPI_ROOT : MPI_PROC_NULL;
  } else {
    rootBcast = mall->root;
  }
  
689
  comm_data_info(rep_s_data, dist_s_data, MALLEABILITY_NOT_CHILDREN, mall->myId, mall->root, mall->intercomm);
690
  if(dist_s_data->entries || rep_s_data->entries) { // Enviar datos sincronos
691
692
693
    send_data(mall->numC, dist_s_data, MALLEABILITY_USE_SYNCHRONOUS);

    // TODO Crear funcion especifica y anyadir para Asinc
694
    // TODO Tener en cuenta el tipo
695
    for(i=0; i<rep_s_data->entries; i++) {
696
697
698
699
700
701
      MPI_Datatype datatype;
      if(rep_s_data->types[i] == MAL_INT) {
        datatype = MPI_INT;
      } else {
        datatype = MPI_CHAR;
      }
702
      MPI_Bcast(rep_s_data->arrays[i], rep_s_data->qty[i], datatype, rootBcast, mall->intercomm);
703
704
    } 
  }
iker_martin's avatar
iker_martin committed
705

706
  comm_results(mall_conf->results, rootBcast, mall_conf->config_file->n_resizes, mall->intercomm);
iker_martin's avatar
iker_martin committed
707

708
709
710
711
712
713
  local_state = MALL_DIST_COMPLETED;
  if(!is_intercomm) { // Merge Spawn
    if(mall->numP < mall->numC) { // Expand
      if(mall->thread_comm != MPI_COMM_WORLD) MPI_Comm_free(&(mall->thread_comm));
      if(mall->comm != MPI_COMM_WORLD) MPI_Comm_free(&(mall->comm));
      if(mall->user_comm != MPI_COMM_WORLD) MPI_Comm_free(&(mall->user_comm)); //TODO No es peligroso?
714

715
716
717
      MPI_Comm_dup(mall->intercomm, &(mall->thread_comm));
      MPI_Comm_dup(mall->intercomm, &(mall->comm));
      MPI_Comm_dup(mall->intercomm, &(mall->user_comm));
718
719
720
721

      MPI_Comm_set_name(mall->thread_comm, "MPI_COMM_MALL_THREAD");
      MPI_Comm_set_name(mall->comm, "MPI_COMM_MALL");
      MPI_Comm_set_name(mall->user_comm, "MPI_COMM_MALL_USER");
722
723
724
725
    } else { // Shrink || Merge Shrink requiere de mas tareas
      local_state = MALL_SPAWN_ADAPT_PENDING;
    }
  }
726
727


728
729
  if(mall->intercomm != MPI_COMM_NULL && mall->intercomm != MPI_COMM_WORLD) {
    MPI_Comm_disconnect(&(mall->intercomm));
730
  }
731

732
  return local_state;
733
734
}

735
736
737
738

///=============================================
///=============================================
///=============================================
739
//TODO Add comment
iker_martin's avatar
iker_martin committed
740
int shrink_redistribution() {
741
    double time_extra = MPI_Wtime();
742
743

    //TODO REFACTOR -- Que solo la llamada de collect iters este fuera de los hilos
744
    zombies_collect_suspended(mall->comm, mall->myId, mall->numP, mall->numC, mall->root, (void *) mall_conf->results, mall_conf->config_file->n_stages);
iker_martin's avatar
iker_martin committed
745
746
    
    if(mall->myId < mall->numC) {
747
748
749
750
751
752
753
754
      if(mall->thread_comm != MPI_COMM_WORLD) MPI_Comm_free(&(mall->thread_comm));
      if(mall->comm != MPI_COMM_WORLD) MPI_Comm_free(&(mall->comm));
      if(mall->user_comm != MPI_COMM_WORLD) MPI_Comm_free(&(mall->user_comm)); //TODO No es peligroso?

      MPI_Comm_dup(mall->intercomm, &(mall->thread_comm));
      MPI_Comm_dup(mall->intercomm, &(mall->comm));
      MPI_Comm_dup(mall->intercomm, &(mall->user_comm));

755
756
757
758
      MPI_Comm_set_name(mall->thread_comm, "MPI_COMM_MALL_THREAD");
      MPI_Comm_set_name(mall->comm, "MPI_COMM_MALL");
      MPI_Comm_set_name(mall->user_comm, "MPI_COMM_MALL_USER");

759
760
      MPI_Comm_free(&(mall->intercomm));

761
762
763
      mall_conf->results->spawn_time[mall_conf->grp] += MPI_Wtime() - time_extra;
      if(malleability_spawn_contains_strat(mall_conf->spawn_strategies,MALL_SPAWN_PTHREAD, NULL)) {
          mall_conf->results->spawn_real_time[mall_conf->grp] += MPI_Wtime() - time_extra;
764
      }
765
      return MALL_DIST_COMPLETED;
iker_martin's avatar
iker_martin committed
766
    } else {
767
      return MALL_ZOMBIE;
iker_martin's avatar
iker_martin committed
768
769
770
    }
}

771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
//======================================================||
//================PRIVATE FUNCTIONS=====================||
//=================COMM NODE INFO ======================||
//======================================================||
//======================================================||
//TODO Add comment
void comm_node_data(int rootBcast, int is_child_group) {
  MPI_Datatype node_type;

  def_nodeinfo_type(&node_type);
  MPI_Bcast(mall, 1, node_type, rootBcast, mall->intercomm);

  if(is_child_group) {
    mall->nodelist = malloc((mall->nodelist_len+1) * sizeof(char));
    mall->nodelist[mall->nodelist_len] = '\0';
  }
  MPI_Bcast(mall->nodelist, mall->nodelist_len, MPI_CHAR, rootBcast, mall->intercomm);

  MPI_Type_free(&node_type);
}

//TODO Add comment
void def_nodeinfo_type(MPI_Datatype *node_type) {
  int i, counts = 3;
  int blocklengths[3] = {1, 1, 1};
  MPI_Aint displs[counts], dir;
  MPI_Datatype types[counts];

  // Rellenar vector types
  types[0] = types[1] = types[2] = MPI_INT;

  // Rellenar vector displs
  MPI_Get_address(mall, &dir);

  MPI_Get_address(&(mall->num_cpus), &displs[0]);
  MPI_Get_address(&(mall->num_nodes), &displs[1]);
  MPI_Get_address(&(mall->nodelist_len), &displs[2]);

  for(i=0;i<counts;i++) displs[i] -= dir;

  MPI_Type_create_struct(counts, blocklengths, displs, types, node_type);
  MPI_Type_commit(node_type);
}

815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
// TODO MOVER A OTRO LADO??
//======================================================||
//================PRIVATE FUNCTIONS=====================||
//===============COMM PARENTS THREADS===================||
//======================================================||
//======================================================||

/*
 * Crea una hebra para ejecutar una comunicación en segundo plano.
 */
int thread_creation() {
  if(pthread_create(&(mall->async_thread), NULL, thread_async_work, NULL)) {
    printf("Error al crear el hilo\n");
    MPI_Abort(MPI_COMM_WORLD, -1);
    return -1;
  }
831
  return MALL_DIST_PENDING;
832
833
834
835
836
837
838
839
840
841
842
843
844
}

/*
 * Comprobación por parte de una hebra maestra que indica
 * si una hebra esclava ha terminado su comunicación en segundo plano.
 *
 * El estado de la comunicación es devuelto al finalizar la función. 
 */
int thread_check() {
  int all_completed = 0;

  // Comprueba que todos los hilos han terminado la distribucion (Mismo valor en commAsync)
  MPI_Allreduce(&state, &all_completed, 1, MPI_INT, MPI_MAX, mall->comm);
845
846
  if(all_completed != MALL_DIST_COMPLETED) return MALL_DIST_PENDING; // Continue only if asynchronous send has ended 
  //FIXME No se tiene en cuenta el estado MALL_APP_ENDED
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864

  if(pthread_join(mall->async_thread, NULL)) {
    printf("Error al esperar al hilo\n");
    MPI_Abort(MPI_COMM_WORLD, -1);
    return -2;
  } 
  return end_redistribution();
}


/*
 * Función ejecutada por una hebra.
 * Ejecuta una comunicación síncrona con los hijos que
 * para el usuario se puede considerar como en segundo plano.
 *
 * Cuando termina la comunicación la hebra maestra puede comprobarlo
 * por el valor "commAsync".
 */
865
void* thread_async_work() {
866
  send_data(mall->numC, dist_a_data, MALLEABILITY_USE_SYNCHRONOUS);
867
  state = MALL_DIST_COMPLETED;
868
869
  pthread_exit(NULL);
}
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889


//==============================================================================
/*
 * Muestra por pantalla el estado actual de todos los comunicadores
 */
void print_comms_state() {
  int tester;
  char *test = malloc(MPI_MAX_OBJECT_NAME * sizeof(char));

  MPI_Comm_get_name(mall->comm, test, &tester);
  printf("P%d Comm=%d Name=%s\n", mall->myId, mall->comm, test);
  MPI_Comm_get_name(mall->user_comm, test, &tester);
  printf("P%d Comm=%d Name=%s\n", mall->myId, mall->user_comm, test);
  if(mall->intercomm != MPI_COMM_NULL) {
    MPI_Comm_get_name(mall->intercomm, test, &tester);
    printf("P%d Comm=%d Name=%s\n", mall->myId, mall->intercomm, test);
  }
  free(test);
}