malleabilityManager.c 26.5 KB
Newer Older
1
2
3
4
#include <pthread.h>
#include "malleabilityManager.h"
#include "malleabilityStates.h"
#include "malleabilityTypes.h"
iker_martin's avatar
iker_martin committed
5
#include "malleabilityZombies.h"
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
#include "ProcessDist.h"
#include "CommDist.h"

#define MALLEABILITY_ROOT 0
#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
22
int shrink_redistribution();
23
24
25
26
27
28
29
30

int thread_creation();
int thread_check();
void* thread_async_work(void* void_arg);

typedef struct {
  int spawn_type;
  int spawn_dist;
31
  int spawn_is_single;
32
33
34
35
  int spawn_threaded;
  int comm_type;
  int comm_threaded;

36
  int grp;
37
38
39
40
  configuration *config_file;
  results_data *results;
} malleability_config_t;

iker_martin's avatar
iker_martin committed
41
typedef struct { //FIXME numC_spawned no se esta usando
42
  int myId, numP, numC, numC_spawned, root, root_parents;
43
44
45
  pthread_t async_thread;
  MPI_Comm comm, thread_comm;
  MPI_Comm intercomm;
46
  MPI_Comm user_comm;
47
  
48
49
  char *name_exec, *nodelist;
  int num_cpus, num_nodes;
50
51
52
53
54
55
56
57
58
59
60
61
} malleability_t;

int state = MAL_UNRESERVED; //FIXME Mover a otro lado

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;

62
/*
63
64
65
66
67
68
69
70
 * 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.
71
 */
72
int init_malleability(int myId, int numP, int root, MPI_Comm comm, char *name_exec, char *nodelist, int num_cpus, int num_nodes) {
73
74
75
76
77
78
79
80
81
  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));

82
83
  MPI_Comm_dup(comm, &dup_comm);
  MPI_Comm_dup(comm, &thread_comm);
84
85
86
87

  mall->myId = myId;
  mall->numP = numP;
  mall->root = root;
88
  mall->comm = dup_comm;
89
90
  mall->thread_comm = thread_comm; // TODO Refactor -- Crear solo si es necesario?
  mall->user_comm = comm;
91

92
  mall->name_exec = name_exec;
93
94
95
  mall->nodelist = nodelist;
  mall->num_cpus = num_cpus;
  mall->num_nodes = num_nodes;
96
97
98
99
100
101
102
103
104
105
106
107

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

  state = MAL_NOT_STARTED;

  // 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();
108
    return MALLEABILITY_CHILDREN;
109
  }
iker_martin's avatar
iker_martin committed
110
111

  zombies_service_init();
112
  return MALLEABILITY_NOT_CHILDREN;
113
114
}

115
116
117
118
119
/*
 * Elimina toda la memoria reservado por el modulo
 * de maleabilidad y asegura que los zombies
 * despierten si los hubiese.
 */
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
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);

  //MPI_Comm_free(&(mall->comm)); // TODO Revisar si hace falta?
  //MPI_Comm_free(&(mall->thread_comm));
  free(mall);
  free(mall_conf);
iker_martin's avatar
iker_martin committed
135
136
137
138

  zombies_awake();
  zombies_service_free();

139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
  state = MAL_UNRESERVED;
}

/*
 * 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() {
  
  if(state == MAL_UNRESERVED) return MAL_UNRESERVED;

  if(state == MAL_NOT_STARTED) {
    // Comprobar si se tiene que realizar un redimensionado
    //if(CHECK_RMS()) {return MAL_DENIED;}
iker_martin's avatar
iker_martin committed
164
    
165
166
167
168
169
170
    state = spawn_step();

    if (state == MAL_SPAWN_COMPLETED){
      state = start_redistribution();
    }

171
  } else if(state == MAL_SPAWN_PENDING || state == MAL_SPAWN_SINGLE_PENDING) { // Comprueba si el spawn ha terminado y comienza la redistribucion
172
173
    double end_real_time;

174
175
176
177
178
179
180
181
182
183
184
185
186
    if(mall_conf->spawn_type == COMM_SPAWN_MERGE_PTHREAD && mall->numP > mall->numC) {
      state = shrink_redistribution(); //TODO REFACTOR

    } else {
      state = check_slurm_comm(mall->myId, mall->root, mall->numP, &(mall->intercomm), mall->comm, mall->thread_comm, &end_real_time);
      if (state == MAL_SPAWN_COMPLETED) {  
        mall_conf->results->spawn_time[mall_conf->grp] = MPI_Wtime() - mall_conf->results->spawn_start;
        if(mall_conf->spawn_type == COMM_SPAWN_PTHREAD || mall_conf->spawn_type == COMM_SPAWN_MERGE_PTHREAD) {
          mall_conf->results->spawn_real_time[mall_conf->grp] = end_real_time - mall_conf->results->spawn_start;
        }
        //TODO Si es MERGE SHRINK, metodo diferente de redistribucion de datos
        state = start_redistribution();
      }
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
    }

  } else if(state == MAL_DIST_PENDING) {
    if(mall_conf->comm_type == MAL_USE_THREAD) {
      state = thread_check();
    } else {
      state = check_redistribution();
    }
  }

  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;
}

210
void get_benchmark_configuration(configuration **config_file) {
211
212
213
214
215
216
217
  *config_file = mall_conf->config_file;
}

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

218
void get_benchmark_results(results_data **results) {
219
220
221
222
  *results = mall_conf->results;
}
//-------------------------------------------------------------------------------------------------------------

223
void set_malleability_configuration(int spawn_type, int spawn_is_single, int spawn_dist, int spawn_threaded, int comm_type, int comm_threaded) {
224
  mall_conf->spawn_type = spawn_type;
225
  mall_conf->spawn_is_single = spawn_is_single;
226
227
228
229
230
231
232
233
  mall_conf->spawn_dist = spawn_dist;
  mall_conf->spawn_threaded = spawn_threaded;
  mall_conf->comm_type = comm_type;
  mall_conf->comm_threaded = comm_threaded;
}

/*
 * To be deprecated
234
 * Tiene que ser llamado despues de setear la config
235
236
 */
void set_children_number(int numC){
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
  if((mall_conf->spawn_type == COMM_SPAWN_MERGE || mall_conf->spawn_type == COMM_SPAWN_MERGE_PTHREAD) && (numC - mall->numP >= 0)) {
    mall->numC = numC;
    mall->numC_spawned = numC - mall->numP;

    if(numC == mall->numP) { // Migrar
      mall->numC_spawned = numC;
      if(mall_conf->spawn_type == COMM_SPAWN_MERGE)
        mall_conf->spawn_type = COMM_SPAWN_SERIAL;
      else
	mall_conf->spawn_type = COMM_SPAWN_PTHREAD;
    }
  } else {
    mall->numC = numC;
    mall->numC_spawned = numC;
  }
}

/*
 * TODO
 */
void get_malleability_user_comm(MPI_Comm *comm) {
  *comm = mall->user_comm;
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
}

/*
 * 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".
 */
void malleability_add_data(void *data, int total_qty, int type, int is_replicated, int is_constant) {

  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 {
      int 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;
      }
      
      add_data(data, total_qty, type, total_reqs, dist_a_data);
    }
  }
}

/*
 * Devuelve el numero de entradas para la estructura de descripcion de 
 * datos elegida.
 */
void malleability_get_entries(int *entries, int is_replicated, int is_constant){
  
  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
 */
324
void malleability_get_data(void **data, int index, int is_replicated, int is_constant) {
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
  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;
    }
  }

341
  *data = data_struct->arrays[index];
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
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
}


//======================================================||
//================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) {
  int i;
  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
      send_async(aux, data_struct->qty[i], mall->myId, mall->numP, mall->root, mall->intercomm, numP_children, data_struct->requests, mall_conf->comm_type);
    }
  } else {
    for(i=0; i < data_struct->entries; i++) {
      aux = (char *) data_struct->arrays[i]; //TODO Comprobar que realmente es un char
      send_sync(aux, data_struct->qty[i], mall->myId, mall->numP, mall->root, mall->intercomm, numP_children);
    }
  }
}

/*
 * 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) {
  int i;
  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
      recv_async(&aux, data_struct->qty[i], mall->myId, mall->numP, mall->root, mall->intercomm, numP_parents, mall_conf->comm_type);
      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
      recv_sync(&aux, data_struct->qty[i], mall->myId, mall->numP, mall->root, mall->intercomm, numP_parents);
      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() {
  int numP_parents, root_parents, i;
412
413
414
415
416
417
418
  int spawn_is_single;
  MPI_Comm aux;

  MPI_Bcast(&spawn_is_single, 1, MPI_INT, MALLEABILITY_ROOT, mall->intercomm); 
  if(spawn_is_single) {
    malleability_establish_connection(mall->myId, MALLEABILITY_ROOT, &(mall->intercomm));
  }
419
  MPI_Bcast(&(mall_conf->spawn_type), 1, MPI_INT, MALLEABILITY_ROOT, mall->intercomm); 
420
421
422
423
424
425

  MPI_Bcast(&root_parents, 1, MPI_INT, MALLEABILITY_ROOT, mall->intercomm); 
  MPI_Bcast(&numP_parents, 1, MPI_INT, root_parents, mall->intercomm);

  mall_conf->config_file = recv_config_file(mall->root, mall->intercomm);
  mall_conf->results = (results_data *) malloc(sizeof(results_data));
426
  init_results_data(mall_conf->results, mall_conf->config_file->resizes, RESULTS_INIT_DATA_QTY);
427
428
429
430
431
432
433
434
435
436
437
438
439

  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
  }
  
440
  comm_data_info(rep_s_data, dist_s_data, MALLEABILITY_CHILDREN, mall->myId, root_parents, mall->intercomm);
441
442
443
444
445
446
447
448
  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++) {
449
450
451
452
453
454
455
      MPI_Datatype datatype;
      if(rep_s_data->types[i] == MAL_INT) {
        datatype = MPI_INT;
      } else {
        datatype = MPI_CHAR;
      }
      MPI_Bcast(rep_s_data->arrays[i], rep_s_data->qty[i], datatype, root_parents, mall->intercomm);
456
457
458
    } 
  }

459
  if(mall_conf->spawn_type == COMM_SPAWN_MERGE || mall_conf->spawn_type == COMM_SPAWN_MERGE_PTHREAD) {
460
    proc_adapt_expand(&(mall->numP), mall->numP+numP_parents, mall->intercomm, &(mall->comm), MALLEABILITY_CHILDREN);
461

462
    if(mall->thread_comm != MPI_COMM_WORLD) MPI_Comm_free(&(mall->thread_comm));
463
464
465
466

    MPI_Comm_dup(mall->comm, &aux);
    mall->thread_comm = aux;
    MPI_Comm_dup(mall->comm, &aux);
iker_martin's avatar
iker_martin committed
467
    mall->user_comm = aux;
468
469
  } 

470
471
472
  // Guardar los resultados de esta transmision
  recv_results(mall_conf->results, mall->root, mall_conf->config_file->resizes, mall->intercomm);

473
  MPI_Comm_disconnect(&(mall->intercomm));
474

475
476
477
478
479
480
481
482
483
484
485
486
487
488
}

//======================================================||
//================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();
iker_martin's avatar
iker_martin committed
489
490
491
492
493

  if((mall_conf->spawn_type == COMM_SPAWN_MERGE || mall_conf->spawn_type == COMM_SPAWN_MERGE_PTHREAD) && mall->numP > mall->numC) {
    state = shrink_redistribution();
    return state; 
  }
494
 
495
  state = init_slurm_comm(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_type, mall_conf->spawn_is_single, mall->thread_comm, &(mall->intercomm));
496

497
  if(mall_conf->spawn_type == COMM_SPAWN_SERIAL || mall_conf->spawn_type == COMM_SPAWN_MERGE)
498
      mall_conf->results->spawn_time[mall_conf->grp] = MPI_Wtime() - mall_conf->results->spawn_start;
499
  else if(mall_conf->spawn_type == COMM_SPAWN_PTHREAD || mall_conf->spawn_type == COMM_SPAWN_MERGE_PTHREAD) {
500
501
      //mall_conf->results->spawn_thread_time[mall_conf->grp] = MPI_Wtime() - mall_conf->results->spawn_start;
      //mall_conf->results->spawn_start = MPI_Wtime();
502
503
504
505
  }
  return state;
}

506

507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
/*
 * 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() {
  int rootBcast = MPI_PROC_NULL;
  if(mall->myId == mall->root) rootBcast = MPI_ROOT;

525
  MPI_Bcast(&(mall_conf->spawn_type), 1, MPI_INT, rootBcast, mall->intercomm);
526
527
  MPI_Bcast(&(mall->root), 1, MPI_INT, rootBcast, mall->intercomm);
  MPI_Bcast(&(mall->numP), 1, MPI_INT, rootBcast, mall->intercomm);
528

529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
  send_config_file(mall_conf->config_file, rootBcast, mall->intercomm);

  if(dist_a_data->entries || rep_a_data->entries) { // Recibir datos asincronos
    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);
      return MAL_DIST_PENDING;
    }
  } 
  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);
  if(!all_completed) return MAL_DIST_PENDING; // Continue only if asynchronous send has ended 
  

  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() {
603
604
  int result, i, rootBcast = MPI_PROC_NULL;
  MPI_Comm aux;
605
606
  if(mall->myId == mall->root) rootBcast = MPI_ROOT;

607
  if(dist_s_data->entries || rep_s_data->entries) { // Enviar datos sincronos
608
609
610
611
    comm_data_info(rep_s_data, dist_s_data, MALLEABILITY_NOT_CHILDREN, mall->myId, mall->root, mall->intercomm);
    send_data(mall->numC, dist_s_data, MALLEABILITY_USE_SYNCHRONOUS);

    // TODO Crear funcion especifica y anyadir para Asinc
612
    // TODO Tener en cuenta el tipo
613
    for(i=0; i<rep_s_data->entries; i++) {
614
615
616
617
618
619
620
      MPI_Datatype datatype;
      if(rep_s_data->types[i] == MAL_INT) {
        datatype = MPI_INT;
      } else {
        datatype = MPI_CHAR;
      }
      MPI_Bcast(rep_s_data->arrays[i], rep_s_data->qty[i], datatype, rootBcast, mall->intercomm);
621
622
    } 
  }
623
    
624
625
  if(mall_conf->spawn_type == COMM_SPAWN_MERGE || mall_conf->spawn_type == COMM_SPAWN_MERGE_PTHREAD) {
    double time_adapt = MPI_Wtime();
iker_martin's avatar
iker_martin committed
626
627
628
629
630
631
632
633
634
635

    proc_adapt_expand(&(mall->numP), mall->numC, mall->intercomm, &(mall->comm), MALLEABILITY_NOT_CHILDREN);

    if(mall->thread_comm != MPI_COMM_WORLD) MPI_Comm_free(&(mall->thread_comm));

    MPI_Comm_dup(mall->comm, &aux);
    mall->thread_comm = aux;
    MPI_Comm_dup(mall->comm, &aux);
    mall->user_comm = aux;
    mall_conf->results->spawn_time[mall_conf->grp] += MPI_Wtime() - time_adapt;
636
	
iker_martin's avatar
iker_martin committed
637
    
638
//    result = MAL_DIST_ADAPTED;
639
  }
640
641

  send_results(mall_conf->results, rootBcast, mall_conf->config_file->resizes, mall->intercomm);
642
  result = MAL_DIST_COMPLETED;
643

644
645
  MPI_Comm_disconnect(&(mall->intercomm));
  state = MAL_NOT_STARTED;
646
  return result;
647
648
}

649
650
651
652

///=============================================
///=============================================
///=============================================
653
double time_adapt, time_adapt_end;
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
int state_shrink=0; //TODO Refactor
pthread_t thread_shrink;
MPI_Comm comm_shrink;

int thread_shrink_creation();
void *thread_shrink_work();
/*
 * Crea una hebra para ejecutar una comunicación en segundo plano.
 */
int thread_shrink_creation() {
  if(pthread_create(&thread_shrink, NULL, thread_shrink_work, NULL)) {
    printf("Error al crear el hilo\n");
    MPI_Abort(MPI_COMM_WORLD, -1);
    return -1;
  }
  return MAL_SPAWN_PENDING;
}
void* thread_shrink_work() {
  proc_adapt_shrink(mall->numC, &comm_shrink, mall->myId);
673
  time_adapt_end = MPI_Wtime();
674
675
676
677
678
679
  state_shrink=2;
  pthread_exit(NULL);
}
///=============================================
///=============================================
///=============================================
iker_martin's avatar
iker_martin committed
680
int shrink_redistribution() {
681
    int global_state;
682
    double time_aux;
iker_martin's avatar
iker_martin committed
683
684
    MPI_Comm aux_comm;

685
686
687
688
689
690
691
692
693
694
695
    if(mall_conf->spawn_type == COMM_SPAWN_MERGE_PTHREAD) {
      if(state_shrink == 0) {
        time_adapt = MPI_Wtime();
	state_shrink = 1;
        MPI_Comm_dup(mall->comm, &comm_shrink);
        thread_shrink_creation();
	return MAL_SPAWN_PENDING;
      } else if(state_shrink>0) {
        MPI_Allreduce(&state_shrink, &global_state, 1, MPI_INT, MPI_MIN, mall->comm);

	if(global_state < 2) return MAL_SPAWN_PENDING;
696
	time_aux = MPI_Wtime();
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
        if(pthread_join(thread_shrink, NULL)) { 
          printf("Error al esperar al hilo\n");
          MPI_Abort(MPI_COMM_WORLD, -1);
          return -10;
        }
        MPI_Comm_dup(mall->comm, &aux_comm);
	mall->comm = comm_shrink;
      }

    } else {
      time_adapt = MPI_Wtime();
      MPI_Comm_dup(mall->comm, &aux_comm);
      proc_adapt_shrink( mall->numC, &(mall->comm), mall->myId);
    }

    //TODO REFACTOR -- Que solo la llamada de collect iters este fuera de los hilos
713
    zombies_collect_suspended(aux_comm, mall->myId, mall->numP, mall->numC, mall->root, (void *) mall_conf->results, mall->user_comm);
iker_martin's avatar
iker_martin committed
714
715
    
    if(mall->myId < mall->numC) {
716
      MPI_Comm_free(&aux_comm);
iker_martin's avatar
iker_martin committed
717
718
719
720
721
722
      MPI_Comm_dup(mall->comm, &aux_comm);
      mall->thread_comm = aux_comm;
      MPI_Comm_dup(mall->comm, &aux_comm);
      mall->user_comm = aux_comm;

      mall_conf->results->spawn_time[mall_conf->grp] = MPI_Wtime() - time_adapt;
723
724
725
      if(mall_conf->spawn_type == COMM_SPAWN_MERGE_PTHREAD) {
          mall_conf->results->spawn_real_time[mall_conf->grp] = time_adapt_end - time_adapt + MPI_Wtime() - time_aux;
      }
iker_martin's avatar
iker_martin committed
726
727
728
729
730
731
      return MAL_DIST_COMPLETED; //FIXME Refactor Poner a SPAWN_COMPLETED
    } else {
      return MAL_ZOMBIE;
    }
}

732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
// 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;
  }
  return MAL_DIST_PENDING;
}

/*
 * 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);
  if(all_completed != MAL_DIST_COMPLETED) return MAL_DIST_PENDING; // Continue only if asynchronous send has ended 
763
  //FIXME No se tiene en cuenta el estado MAL_APP_ENDED
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786

  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".
 */
void* thread_async_work(void* void_arg) {
  send_data(mall->numC, dist_a_data, MALLEABILITY_USE_SYNCHRONOUS);
  state = MAL_DIST_COMPLETED;
  pthread_exit(NULL);
}