malleabilityManager.c 27.3 KB
Newer Older
1
2
3
#include <pthread.h>
#include "malleabilityManager.h"
#include "malleabilityStates.h"
4
#include "malleabilityDataStructures.h"
5
#include "malleabilityTypes.h"
iker_martin's avatar
iker_martin committed
6
#include "malleabilityZombies.h"
7
#include "spawn_methods/GenericSpawn.h"
8
9
10
11
12
13
14
15
16
17
18
19
20
21
#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
22
int shrink_redistribution();
23
24
25

int thread_creation();
int thread_check();
26
void* thread_async_work();
27
28

typedef struct {
29
  int spawn_method;
30
  int spawn_dist;
31
32
33
  int spawn_strategies;
  //int spawn_is_single;
  //int spawn_threaded;
34
35
36
  int comm_type;
  int comm_threaded;

37
  int grp;
38
39
40
41
  configuration *config_file;
  results_data *results;
} malleability_config_t;

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

53
int state = MALL_UNRESERVED; //FIXME Mover a otro lado
54
55
56
57
58
59
60
61
62

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;

63
/*
64
65
66
67
68
69
70
71
 * 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.
72
 */
73
int init_malleability(int myId, int numP, int root, MPI_Comm comm, char *name_exec, char *nodelist, int num_cpus, int num_nodes) {
74
75
76
77
78
79
80
81
82
  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));

83
84
  MPI_Comm_dup(comm, &dup_comm);
  MPI_Comm_dup(comm, &thread_comm);
85
86
87
88

  mall->myId = myId;
  mall->numP = numP;
  mall->root = root;
89
  mall->comm = dup_comm;
90
91
  mall->thread_comm = thread_comm; // TODO Refactor -- Crear solo si es necesario?
  mall->user_comm = comm;
92

93
  mall->name_exec = name_exec;
94
95
96
  mall->nodelist = nodelist;
  mall->num_cpus = num_cpus;
  mall->num_nodes = num_nodes;
97
98
99
100
101
102

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

103
  state = MALL_NOT_STARTED;
104
105
106
107
108

  // 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();
109
    return MALLEABILITY_CHILDREN;
110
  }
iker_martin's avatar
iker_martin committed
111
112

  zombies_service_init();
113
  return MALLEABILITY_NOT_CHILDREN;
114
115
}

116
117
118
119
120
/*
 * Elimina toda la memoria reservado por el modulo
 * de maleabilidad y asegura que los zombies
 * despierten si los hubiese.
 */
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
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
136
137
138
139

  zombies_awake();
  zombies_service_free();

140
  state = MALL_UNRESERVED;
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
}

/*
 * 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() {
  
160
  if(state == MALL_UNRESERVED) return MALL_UNRESERVED;
161

162
  if(state == MALL_NOT_STARTED) {
163
    // Comprobar si se tiene que realizar un redimensionado
164
    //if(CHECK_RMS()) {return MALL_DENIED;}
iker_martin's avatar
iker_martin committed
165
    
166
167
    state = spawn_step();

168
    if (state == MALL_SPAWN_COMPLETED){
169
170
171
      state = start_redistribution();
    }

172
  } else if(state == MALL_SPAWN_PENDING || state == MALL_SPAWN_SINGLE_PENDING) { // Comprueba si el spawn ha terminado y comienza la redistribucion
173
174
    double end_real_time;

175
    /*if(mall_conf->spawn_type == MALL_SPAWN_MERGE && mall_conf->spawn_type == MALL_SPAWN_PTHREAD && mall->numP > mall->numC) {
176
177
178
      state = shrink_redistribution(); //TODO REFACTOR

    } else {
179
180
181
182
    */
      //state = check_slurm_comm(mall->myId, mall->root, mall->numP, &(mall->intercomm), mall->comm, mall->thread_comm, &end_real_time); //FIXMENOW
      state = check_spawn_state(&(mall->intercomm), mall->comm, MALL_DIST_PENDING, &end_real_time); //FIXME 3 argumento depende de la distribucion
      if (state == MALL_SPAWN_COMPLETED || state == MALL_DIST_ADAPTED) {  //FIXME MALL_DIST_ADAPTED tiene que recoger los tiempos
183
        mall_conf->results->spawn_time[mall_conf->grp] = MPI_Wtime() - mall_conf->results->spawn_start;
184
185

        if(malleability_spawn_contains_strat(mall_conf->spawn_strategies, MALL_SPAWN_PTHREAD, NULL)) {
186
187
          mall_conf->results->spawn_real_time[mall_conf->grp] = end_real_time - mall_conf->results->spawn_start;
        }
188
        //TODO Si es MERGE, metodo diferente de redistribucion de datos
189
190
        state = start_redistribution();
      }
191
    //}
192

193
  } else if(state == MALL_DIST_PENDING) {
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
    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;
}

214
void get_benchmark_configuration(configuration **config_file) {
215
216
217
218
219
220
221
  *config_file = mall_conf->config_file;
}

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

222
void get_benchmark_results(results_data **results) {
223
224
225
226
  *results = mall_conf->results;
}
//-------------------------------------------------------------------------------------------------------------

227
228
229
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;
230
231
232
233
234
235
236
  mall_conf->spawn_dist = spawn_dist;
  mall_conf->comm_type = comm_type;
  mall_conf->comm_threaded = comm_threaded;
}

/*
 * To be deprecated
237
 * Tiene que ser llamado despues de setear la config
238
239
 */
void set_children_number(int numC){
240
  if((mall_conf->spawn_method == MALL_SPAWN_MERGE) && (numC >= mall->numP)) {
241
242
243
244
245
    mall->numC = numC;
    mall->numC_spawned = numC - mall->numP;

    if(numC == mall->numP) { // Migrar
      mall->numC_spawned = numC;
246
      mall_conf->spawn_method = MALL_SPAWN_BASELINE;
247
248
249
250
251
252
253
254
255
256
257
258
    }
  } 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
  int is_intercomm;
  //MPI_Comm aux;
414

415
416
  malleability_connect_children(mall->myId, mall->numP, mall->root, mall->comm, &numP_parents, &root_parents, &(mall->intercomm));
  /*MPI_Bcast(&spawn_is_single, 1, MPI_INT, MALLEABILITY_ROOT, mall->intercomm); 
417
  if(spawn_is_single) {
418
    malleability_establish_connection(mall->myId, MALLEABILITY_ROOT, &(mall->intercomm)); //FIXMENOW
419
  }
420
  MPI_Bcast(&(mall_conf->spawn_type), 1, MPI_INT, MALLEABILITY_ROOT, mall->intercomm); 
421
422
  MPI_Bcast(&root_parents, 1, MPI_INT, MALLEABILITY_ROOT, mall->intercomm); 
  MPI_Bcast(&numP_parents, 1, MPI_INT, root_parents, mall->intercomm);
423
424
425
  */
  // TODO A partir de este punto tener en cuenta si es BASELINE o MERGE
  MPI_Comm_test_inter(mall->intercomm, &is_intercomm);
426

427
428
  recv_config_file(mall->root, mall->intercomm, &(mall_conf->config_file));

429
  mall_conf->results = (results_data *) malloc(sizeof(results_data));
430
  init_results_data(mall_conf->results, mall_conf->config_file->n_resizes, mall_conf->config_file->n_stages, RESULTS_INIT_DATA_QTY);
431
432
433
434
435
436
437
438
439
440
441
442
443

  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
  }
  
444
  comm_data_info(rep_s_data, dist_s_data, MALLEABILITY_CHILDREN, mall->myId, root_parents, mall->intercomm);
445
446
447
448
449
450
451
452
  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++) {
453
454
455
456
457
458
459
      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);
460
461
462
    } 
  }

463
464
465
  /*
  if(mall_conf->spawn_type == MALL_SPAWN_MERGE || mall_conf->spawn_type == MALL_SPAWN_MERGE_PTHREAD) {
    proc_adapt_expand(&(mall->numP), mall->numP+numP_parents, mall->intercomm, &(mall->comm), MALLEABILITY_CHILDREN); //FIXMENOW
466

467
    if(mall->thread_comm != MPI_COMM_WORLD) MPI_Comm_free(&(mall->thread_comm));
468
469
470
471

    MPI_Comm_dup(mall->comm, &aux);
    mall->thread_comm = aux;
    MPI_Comm_dup(mall->comm, &aux);
iker_martin's avatar
iker_martin committed
472
    mall->user_comm = aux;
473
  } 
474
  */
475

476
  // Guardar los resultados de esta transmision
477
  recv_results(mall_conf->results, mall->root, mall_conf->config_file->n_resizes, mall->intercomm);
478
  MPI_Comm_disconnect(&(mall->intercomm));
479
480
481
482
483
484
485
486
487
488
489
490
491
492
}

//======================================================||
//================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
493

494
495
  /* FIXME Mantener funcion de shrink_redistribuition
  if((mall_conf->spawn_type == MALL_SPAWN_MERGE) && mall->numP > mall->numC) {
iker_martin's avatar
iker_martin committed
496
497
498
    state = shrink_redistribution();
    return state; 
  }
499
  */
500
 
501
502
  //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)); //FIXMENOW
  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));
503

504
  if(!malleability_spawn_contains_strat(mall_conf->spawn_strategies, MALL_SPAWN_PTHREAD, NULL)) {
505
      mall_conf->results->spawn_time[mall_conf->grp] = MPI_Wtime() - mall_conf->results->spawn_start;
506
  } else {
507
508
      //mall_conf->results->spawn_thread_time[mall_conf->grp] = MPI_Wtime() - mall_conf->results->spawn_start;
      //mall_conf->results->spawn_start = MPI_Wtime();
509
510
511
512
  }
  return state;
}

513

514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
/*
 * 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;

532
  /*
533
  MPI_Bcast(&(mall_conf->spawn_type), 1, MPI_INT, rootBcast, mall->intercomm);
534
535
  MPI_Bcast(&(mall->root), 1, MPI_INT, rootBcast, mall->intercomm);
  MPI_Bcast(&(mall->numP), 1, MPI_INT, rootBcast, mall->intercomm);
536
  */
537

538
539
540
541
542
543
544
545
546
  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);
547
      return MALL_DIST_PENDING;
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
    }
  } 
  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);
591
  if(!all_completed) return MALL_DIST_PENDING; // Continue only if asynchronous send has ended 
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
  

  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() {
612
  int result, i, rootBcast = MPI_PROC_NULL;
613
  //MPI_Comm aux;
614
615
  if(mall->myId == mall->root) rootBcast = MPI_ROOT;

616
  if(dist_s_data->entries || rep_s_data->entries) { // Enviar datos sincronos
617
618
619
620
    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
621
    // TODO Tener en cuenta el tipo
622
    for(i=0; i<rep_s_data->entries; i++) {
623
624
625
626
627
628
629
      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);
630
631
    } 
  }
632
633
634
   
 /* 
  if(mall_conf->spawn_type == MALL_SPAWN_MERGE) {
635
    double time_adapt = MPI_Wtime();
iker_martin's avatar
iker_martin committed
636

637
    proc_adapt_expand(&(mall->numP), mall->numC, mall->intercomm, &(mall->comm), MALLEABILITY_NOT_CHILDREN); //FIXMENOW
iker_martin's avatar
iker_martin committed
638
639
640
641
642
643
644
645

    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;
646
	
iker_martin's avatar
iker_martin committed
647
    
648
//    result = MAL_DIST_ADAPTED;
649
  }
650
  */
651

652
  send_results(mall_conf->results, rootBcast, mall_conf->config_file->n_resizes, mall->intercomm);
653
  result = MALL_DIST_COMPLETED;
654

655
  MPI_Comm_disconnect(&(mall->intercomm));
656
  state = MALL_NOT_STARTED;
657
  return result;
658
659
}

660
661
662
663

///=============================================
///=============================================
///=============================================
664
double time_adapt, time_adapt_end;
665
/*
666
667
668
669
670
671
int state_shrink=0; //TODO Refactor
pthread_t thread_shrink;
MPI_Comm comm_shrink;

int thread_shrink_creation();
void *thread_shrink_work();
672
673


674
675
676
677
678
679
680
681
682
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() {
683
  proc_adapt_shrink(mall->numC, &comm_shrink, mall->myId); //FIXMENOW
684
  time_adapt_end = MPI_Wtime();
685
686
687
  state_shrink=2;
  pthread_exit(NULL);
}
688
*/
689
690
691
///=============================================
///=============================================
///=============================================
iker_martin's avatar
iker_martin committed
692
int shrink_redistribution() {
693
    //int global_state;
694
    double time_aux;
iker_martin's avatar
iker_martin committed
695
696
    MPI_Comm aux_comm;

697
698
    /*
    if(mall_conf->spawn_type == MALL_SPAWN_MERGE && mall_conf->spawn_type == MALL_SPAWN_PTHREAD) {
699
700
701
702
703
704
705
706
707
708
      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;
709
	time_aux = MPI_Wtime();
710
711
712
713
714
715
716
717
718
719
720
721
        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);
722
      proc_adapt_shrink( mall->numC, &(mall->comm), mall->myId); //FIXMENOW
723
    }
724
725
726
    */
    MPI_Comm_dup(mall->comm, &aux_comm);

727
728

    //TODO REFACTOR -- Que solo la llamada de collect iters este fuera de los hilos
729
    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
730
731
    
    if(mall->myId < mall->numC) {
732
      MPI_Comm_free(&aux_comm);
733
      MPI_Comm_dup(mall->comm, &aux_comm); // FIXME Que pasa con los comunicadores Thread_comm y User_comm
iker_martin's avatar
iker_martin committed
734
735
736
737
      mall->thread_comm = aux_comm;
      MPI_Comm_dup(mall->comm, &aux_comm);
      mall->user_comm = aux_comm;

738
739
740
      mall_conf->results->spawn_time[mall_conf->grp] = MPI_Wtime() - time_adapt; //FIXME Error
      if(mall_conf->spawn_method == MALL_SPAWN_MERGE && malleability_spawn_contains_strat(mall_conf->spawn_strategies,MALL_SPAWN_PTHREAD, NULL)) {
          mall_conf->results->spawn_real_time[mall_conf->grp] = time_adapt_end - time_adapt + MPI_Wtime() - time_aux; //FIXME Error
741
      }
742
      return MALL_DIST_COMPLETED; //FIXME Refactor Poner a SPAWN_COMPLETED
iker_martin's avatar
iker_martin committed
743
    } else {
744
      return MALL_ZOMBIE;
iker_martin's avatar
iker_martin committed
745
746
747
    }
}

748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
// 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;
  }
764
  return MALL_DIST_PENDING;
765
766
767
768
769
770
771
772
773
774
775
776
777
}

/*
 * 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);
778
779
  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
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797

  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".
 */
798
void* thread_async_work() {
799
  send_data(mall->numC, dist_a_data, MALLEABILITY_USE_SYNCHRONOUS);
800
  state = MALL_DIST_COMPLETED;
801
802
  pthread_exit(NULL);
}