malleabilityManager.c 33.8 KB
Newer Older
iker_martin's avatar
iker_martin committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
#include <pthread.h>
#include <string.h>
#include "malleabilityManager.h"
#include "malleabilityStates.h"
#include "malleabilityDataStructures.h"
#include "malleabilityTypes.h"
#include "malleabilityZombies.h"
#include "spawn_methods/GenericSpawn.h"
#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();
int shrink_redistribution();

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

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

void print_comms_state();
void malleability_comms_update(MPI_Comm comm);

35
36
37
void comm_results(int root, int compute); //FIXME Borrar
void def_malleability_results(MPI_Datatype *new_type); //FIXME BORRAR

iker_martin's avatar
iker_martin committed
38
39
40
41
42
43
44
45
typedef struct {
  int spawn_method;
  int spawn_dist;
  int spawn_strategies;
  int red_method;
  int red_strategies;

  int grp;
46
47
48
49
50
51

  double exec_start;
  double spawn_start, spawn_time;
  double sync_time, sync_end;
  double async_time, async_end;
  double malleability_end, malleability_time;
iker_martin's avatar
iker_martin committed
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
} malleability_config_t;

typedef struct { //FIXME numC_spawned no se esta usando
  int myId, numP, numC, numC_spawned, root, root_parents;
  pthread_t async_thread;
  MPI_Comm comm, thread_comm;
  MPI_Comm intercomm;
  MPI_Comm user_comm;
  int dup_user_comm;
  
  char *name_exec, *nodelist;
  int num_cpus, num_nodes, nodelist_len;
} malleability_t;

int state = MALL_UNRESERVED; //FIXME Mover a otro lado
67
int dep_not_send = 1; //FIXME BORRAR
iker_martin's avatar
iker_martin committed
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
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
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192

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;

/*
 * 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.
 */
int init_malleability(int myId, int numP, int root, MPI_Comm comm, char *name_exec, char *nodelist, int num_cpus, int num_nodes) {
  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));

  mall->dup_user_comm = 0;
  MPI_Comm_dup(comm, &dup_comm);
  MPI_Comm_dup(comm, &thread_comm);
  MPI_Comm_set_name(dup_comm, "MPI_COMM_MALL");
  MPI_Comm_set_name(thread_comm, "MPI_COMM_MALL_THREAD");

  mall->myId = myId;
  mall->numP = numP;
  mall->root = root;
  mall->comm = dup_comm;
  mall->thread_comm = thread_comm;
  mall->user_comm = comm;

  mall->name_exec = name_exec;
  mall->nodelist = nodelist;
  mall->num_cpus = num_cpus;
  mall->num_nodes = num_nodes;

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

  state = MALL_NOT_STARTED;

  zombies_service_init();

  // 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();
    return MALLEABILITY_CHILDREN;
  }

  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
  }

  return MALLEABILITY_NOT_CHILDREN;
}

/*
 * Elimina toda la memoria reservado por el modulo
 * de maleabilidad y asegura que los zombies
 * despierten si los hubiese.
 */
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);

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

  zombies_awake();
  zombies_service_free();

  state = MALL_UNRESERVED;
}

/* 
 * TODO Reescribir
 * 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() {
  double end_real_time;

  switch(state) {
    case MALL_UNRESERVED:
      break;
    case MALL_NOT_STARTED:
      // Comprobar si se tiene que realizar un redimensionado
193
      mall_conf->malleability_time = MPI_Wtime();
iker_martin's avatar
iker_martin committed
194
195
196
197
198
199
200
201
202
203
204
205

      state = spawn_step();

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

    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) {
206
        mall_conf->spawn_time = MPI_Wtime() - mall_conf->spawn_start;
iker_martin's avatar
iker_martin committed
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229

        malleability_checkpoint();
      }
      break;

    case MALL_SPAWN_ADAPT_POSTPONE:
    case MALL_SPAWN_COMPLETED:
      state = start_redistribution();
      malleability_checkpoint();
      break;

    case MALL_DIST_PENDING:
      if(malleability_red_contains_strat(mall_conf->red_strategies, MALL_RED_THREAD, NULL)) {
        state = thread_check();
      } else {
        state = check_redistribution();
      }
      if(state != MALL_DIST_PENDING) { 
        malleability_checkpoint();
      }
      break;

    case MALL_SPAWN_ADAPT_PENDING:
230
      mall_conf->spawn_start = MPI_Wtime();
iker_martin's avatar
iker_martin committed
231
232
233
234
      unset_spawn_postpone_flag(state);
      state = check_spawn_state(&(mall->intercomm), mall->comm, &end_real_time);

      if(!malleability_spawn_contains_strat(mall_conf->spawn_strategies, MALL_SPAWN_PTHREAD, NULL)) {
235
        mall_conf->spawn_time = MPI_Wtime() - mall_conf->spawn_start;
iker_martin's avatar
iker_martin committed
236
237
238
239
240
241
242
243
244
245
	malleability_checkpoint();
      }
      break;

    case MALL_SPAWN_ADAPTED:
      state = shrink_redistribution();
      malleability_checkpoint();
      break;

    case MALL_DIST_COMPLETED: //TODO No es esto muy feo?
246
      mall_conf->malleability_end = MPI_Wtime();
iker_martin's avatar
iker_martin committed
247
248
249
250
251
252
253
254
255
256
257
258
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
      state = MALL_COMPLETED;
      break;
  }
  return state;
}

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

void set_malleability_configuration(int spawn_method, int spawn_strategies, int spawn_dist, int red_method, int red_strategies) {
  mall_conf->spawn_method = spawn_method;
  mall_conf->spawn_strategies = spawn_strategies;
  mall_conf->spawn_dist = spawn_dist;
  mall_conf->red_method = red_method;
  mall_conf->red_strategies = red_strategies;
}

/*
 * To be deprecated
 * Tiene que ser llamado despues de setear la config
 */
void set_children_number(int numC){
  if((mall_conf->spawn_method == MALL_SPAWN_MERGE) && (numC >= mall->numP)) {
    mall->numC = numC;
    mall->numC_spawned = numC - mall->numP;

    if(numC == mall->numP) { // Migrar
      mall->numC_spawned = numC;
      mall_conf->spawn_method = MALL_SPAWN_BASELINE;
    }
  } else {
    mall->numC = numC;
    mall->numC_spawned = numC;
  }
}

/*
 * TODO
 */
void get_malleability_user_comm(MPI_Comm *comm) {
  if(mall->dup_user_comm) {
    if(mall->user_comm != MPI_COMM_WORLD) MPI_Comm_free(&(mall->user_comm));
    MPI_Comm_dup(mall->comm, &(mall->user_comm));
    MPI_Comm_set_name(mall->user_comm, "MPI_COMM_MALL_USER");
    mall->dup_user_comm = 0;
  }
  *comm = mall->user_comm;
}

/*
 * 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".
 *
 * //FIXME Si es constante se debería ir a asincrono, no sincrono
 */
310
void malleability_add_data(void *data, size_t total_qty, int type, int dependency, int is_replicated, int is_constant) {
iker_martin's avatar
iker_martin committed
311
312
313
314
  size_t total_reqs = 0;

  if(is_constant) {
    if(is_replicated) {
315
      add_data(data, total_qty, type, dependency, total_reqs, rep_s_data);
iker_martin's avatar
iker_martin committed
316
    } else {
317
      add_data(data, total_qty, type, dependency, total_reqs, dist_s_data);
iker_martin's avatar
iker_martin committed
318
319
320
    }
  } else {
    if(is_replicated) {
321
      add_data(data, total_qty, type, dependency, total_reqs, rep_a_data); //FIXME total_reqs==0 ??? 
iker_martin's avatar
iker_martin committed
322
323
324
325
326
327
328
329
330
    } else {
      if(mall_conf->red_method  == MALL_RED_BASELINE) {
        total_reqs = 1;
      } else if(mall_conf->red_method  == MALL_RED_IBARRIER) { //TODO This is a strategy, not a method
        total_reqs = 2;
      } else if(mall_conf->red_method  == MALL_RED_POINT) {
        total_reqs = mall->numC;
      }
      
331
      add_data(data, total_qty, type, dependency, total_reqs, dist_a_data);
iker_martin's avatar
iker_martin committed
332
333
334
335
336
337
338
339
340
341
342
343
344
    }
  }
}

/*
 * 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".
 * //FIXME Si es constante se debería ir a asincrono, no sincrono
 */
345
void malleability_modify_data(void *data, size_t index, size_t total_qty, int type, int dependency, int is_replicated, int is_constant) {
iker_martin's avatar
iker_martin committed
346
347
348
349
  size_t total_reqs = 0;

  if(is_constant) {
    if(is_replicated) {
350
      modify_data(data, index, total_qty, type, dependency, total_reqs, rep_s_data);
iker_martin's avatar
iker_martin committed
351
    } else {
352
      modify_data(data, index, total_qty, type, dependency, total_reqs, dist_s_data);
iker_martin's avatar
iker_martin committed
353
354
355
    }
  } else {
    if(is_replicated) {
356
      modify_data(data, index, total_qty, type, dependency, total_reqs, rep_a_data); //FIXME total_reqs==0 ??? 
iker_martin's avatar
iker_martin committed
357
358
359
360
361
362
363
364
365
    } else {    
      if(mall_conf->red_method  == MALL_RED_BASELINE) {
        total_reqs = 1;
      } else if(mall_conf->red_method  == MALL_RED_IBARRIER) { //TODO This is a strategy, not a method
        total_reqs = 2;
      } else if(mall_conf->red_method  == MALL_RED_POINT) {
        total_reqs = mall->numC;
      }
      
366
      modify_data(data, index, total_qty, type, dependency, total_reqs, dist_a_data);
iker_martin's avatar
iker_martin committed
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
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
    }
  }
}

/*
 * Devuelve el numero de entradas para la estructura de descripcion de 
 * datos elegida.
 * //FIXME Si es constante se debería ir a asincrono, no sincrono
 */
void malleability_get_entries(size_t *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
 * //FIXME Si es constante se debería ir a asincrono, no sincrono
 */
void malleability_get_data(void **data, size_t index, int is_replicated, int is_constant) {
  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;
    }
  }

  *data = data_struct->arrays[index];
}


//======================================================||
//================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) {
  size_t i;
436
  void *aux_send, *aux_recv;
iker_martin's avatar
iker_martin committed
437
438

  if(is_asynchronous) {
439
440
    i= dep_not_send ? 0 : data_struct->entries - 2; //FIXME BORRAR
    for(; i < data_struct->entries; i++) {
441
      if(data_struct->dependencies[i] == 1+MAL_DATA_DEPENDENT && dep_not_send && mall_conf->spawn_method == MALL_SPAWN_MERGE) break; //FIXME BORRAR dep_not_send
442
      aux_send = data_struct->arrays[i];
iker_martin's avatar
iker_martin committed
443
      aux_recv = NULL;
444
      async_communication(aux_send, &aux_recv, data_struct->qty[i], data_struct->types[i], data_struct->dependencies[i], mall->myId, mall->numP, numP_children, MALLEABILITY_NOT_CHILDREN, mall_conf->red_method, mall_conf->red_strategies, mall->intercomm, 
iker_martin's avatar
iker_martin committed
445
		      &(data_struct->requests[i]), &(data_struct->request_qty[i]));
446
      if(aux_recv != NULL) data_struct->arrays[i] = aux_recv;
iker_martin's avatar
iker_martin committed
447
448
449
    }
  } else {
    for(i=0; i < data_struct->entries; i++) {
450
      aux_send = data_struct->arrays[i];
iker_martin's avatar
iker_martin committed
451
      aux_recv = NULL;
452
453
      sync_communication(aux_send, &aux_recv, data_struct->qty[i], data_struct->types[i], data_struct->dependencies[i], mall->myId, mall->numP, numP_children, MALLEABILITY_NOT_CHILDREN, mall_conf->red_method, mall->intercomm);
      if(aux_recv != NULL) data_struct->arrays[i] = aux_recv;
iker_martin's avatar
iker_martin committed
454
455
456
457
458
459
460
461
462
463
464
    }
  }
}

/*
 * 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) {
  size_t i;
465
  void *aux, *aux_s = NULL;
iker_martin's avatar
iker_martin committed
466
467
468

  if(is_asynchronous) {
    for(i=0; i < data_struct->entries; i++) {
469
470
      aux = data_struct->arrays[i];
      async_communication(aux_s, &aux, data_struct->qty[i], data_struct->types[i], data_struct->dependencies[i], mall->myId, mall->numP, numP_parents, MALLEABILITY_CHILDREN, mall_conf->red_method, mall_conf->red_strategies, mall->intercomm, 
iker_martin's avatar
iker_martin committed
471
		      &(data_struct->requests[i]), &(data_struct->request_qty[i]));
472
      data_struct->arrays[i] = aux;
iker_martin's avatar
iker_martin committed
473
474
475
    }
  } else {
    for(i=0; i < data_struct->entries; i++) {
476
477
478
      aux = data_struct->arrays[i];
      sync_communication(aux_s, &aux, data_struct->qty[i], data_struct->types[i], data_struct->dependencies[i], mall->myId, mall->numP, numP_parents, MALLEABILITY_CHILDREN, mall_conf->red_method, mall->intercomm);
      data_struct->arrays[i] = aux;
iker_martin's avatar
iker_martin committed
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
    }
  }
}

//======================================================||
//================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() {
  size_t i;
  int numP_parents, root_parents;
  int is_intercomm;

  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);
  if(!is_intercomm) { // For intracommunicators, these processes will be added
    MPI_Comm_rank(mall->intercomm, &(mall->myId));
    MPI_Comm_size(mall->intercomm, &(mall->numP));
  }

  comm_node_data(root_parents, MALLEABILITY_CHILDREN);
  MPI_Bcast(&(mall_conf->red_method), 1, MPI_INT, root_parents, mall->intercomm);
  MPI_Bcast(&(mall_conf->red_strategies), 1, MPI_INT, root_parents, mall->intercomm);

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

    if(malleability_red_contains_strat(mall_conf->red_strategies, MALL_RED_THREAD, NULL)) {
      recv_data(numP_parents, dist_a_data, MALLEABILITY_USE_SYNCHRONOUS);
    } else {
      recv_data(numP_parents, dist_a_data, MALLEABILITY_USE_ASYNCHRONOUS);
    }

519
    mall_conf->async_end= MPI_Wtime(); // Obtener timestamp de cuando termina comm asincrona
iker_martin's avatar
iker_martin committed
520
521
522
523
524
  }

  comm_data_info(rep_s_data, dist_s_data, MALLEABILITY_CHILDREN, mall->myId, root_parents, mall->intercomm);
  if(dist_s_data->entries || rep_s_data->entries) { // Recibir datos sincronos
    recv_data(numP_parents, dist_s_data, MALLEABILITY_USE_SYNCHRONOUS);
525
    mall_conf->sync_end = MPI_Wtime(); // Obtener timestamp de cuando termina comm sincrona
iker_martin's avatar
iker_martin committed
526
527
528
529
530
531
532

    // 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++) {
      MPI_Datatype datatype;
      if(rep_s_data->types[i] == MAL_INT) {
        datatype = MPI_INT;
533
534
535
      } else if(rep_s_data->types[i] == MAL_DOUBLE) {
        datatype = MPI_DOUBLE;
      } else if(rep_s_data->types[i] == MAL_CHAR) {
iker_martin's avatar
iker_martin committed
536
        datatype = MPI_CHAR;
537
538
539
      } else {
	printf("Malleability -- Redistribution recv type not recognised\n");
	MPI_Abort(MPI_COMM_WORLD, -1);
iker_martin's avatar
iker_martin committed
540
541
542
543
      }
      MPI_Bcast(rep_s_data->arrays[i], rep_s_data->qty[i], datatype, root_parents, mall->intercomm);
    } 
  }
544
  mall_conf->malleability_end = MPI_Wtime(); // Obtener timestamp de cuando termina maleabilidad
iker_martin's avatar
iker_martin committed
545
  
546
  comm_results(root_parents, 1);
iker_martin's avatar
iker_martin committed
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
  if(!is_intercomm) {
    malleability_comms_update(mall->intercomm);
  }

  MPI_Comm_disconnect(&(mall->intercomm)); //FIXME Error en OpenMPI + Merge
}

//======================================================||
//================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(){
565
  mall_conf->spawn_start = MPI_Wtime();
iker_martin's avatar
iker_martin committed
566
567
568
569
 
  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));

  if(!malleability_spawn_contains_strat(mall_conf->spawn_strategies, MALL_SPAWN_PTHREAD, NULL)) {
570
      mall_conf->spawn_time = MPI_Wtime() - mall_conf->spawn_start;
iker_martin's avatar
iker_martin committed
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
603
604
605
606
607
608
609
610
611
612
613
614
  }
  return state;
}


/*
 * 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, 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
    MPI_Comm_dup(mall->comm, &(mall->intercomm));
  }

  if(is_intercomm) {
    rootBcast = mall->myId == mall->root ? MPI_ROOT : MPI_PROC_NULL;
  } else {
    rootBcast = mall->root;
  }

  comm_node_data(rootBcast, MALLEABILITY_NOT_CHILDREN);
  MPI_Bcast(&(mall_conf->red_method), 1, MPI_INT, rootBcast, mall->intercomm);
  MPI_Bcast(&(mall_conf->red_strategies), 1, MPI_INT, rootBcast, mall->intercomm);

  comm_data_info(rep_a_data, dist_a_data, MALLEABILITY_NOT_CHILDREN, mall->myId, mall->root, mall->intercomm);
  if(dist_a_data->entries || rep_a_data->entries) { // Enviar datos asincronos
    //FIXME No se envian los datos replicados (rep_a_data)
615
    mall_conf->async_time = MPI_Wtime();
iker_martin's avatar
iker_martin committed
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
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
    if(malleability_red_contains_strat(mall_conf->red_strategies, MALL_RED_THREAD, NULL)) {
      return thread_creation();
    } else {
      send_data(mall->numC, dist_a_data, MALLEABILITY_USE_ASYNCHRONOUS);
      return MALL_DIST_PENDING; 
    }
  } 
  return end_redistribution();
}


/*
 * 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 is_intercomm, completed, local_completed, all_completed, test_err;
  size_t i, j, req_qty;
  MPI_Request *req_completed;
  local_completed = 1;
  test_err = 0;

  //FIXME Modificar para que se tenga en cuenta rep_a_data
  for(i=0; i<dist_a_data->entries; i++) {
    req_completed = dist_a_data->requests[i];
    req_qty = dist_a_data->request_qty[i];
    if(malleability_red_contains_strat(mall_conf->red_strategies, MALL_RED_IBARRIER, NULL)) { //FIXME Strategy not fully implemented
      test_err = MPI_Test(&(req_completed[req_qty-1]), &completed, MPI_STATUS_IGNORE);
      local_completed = local_completed && completed;
    } else {
      for(j=0; j<req_qty; j++) {
	test_err = MPI_Test(&(req_completed[j]), &completed, MPI_STATUS_IGNORE);
	local_completed = local_completed && completed;
      }
//      test_err = MPI_Testall(req_qty, req_completed, &completed, MPI_STATUSES_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(&local_completed, &all_completed, 1, MPI_INT, MPI_MIN, mall->comm);
  if(!all_completed) return MALL_DIST_PENDING; // Continue only if asynchronous send has ended 
  
  if(malleability_red_contains_strat(mall_conf->red_strategies, MALL_RED_IBARRIER, NULL)) { //FIXME Strategy not fully implemented
    MPI_Waitall(req_qty, req_completed, MPI_STATUSES_IGNORE);
    //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
  }
676
677
678
679
680
  if(dep_not_send && mall_conf->spawn_method == MALL_SPAWN_MERGE) { //FIXME BORRAR // A MITAD -- Algunas partes son útiles
    dep_not_send = 0;
    send_data(mall->numC, dist_a_data, MALLEABILITY_USE_ASYNCHRONOUS);
    return MALL_DIST_PENDING; 
  }
iker_martin's avatar
iker_martin committed
681
682

  MPI_Comm_test_inter(mall->intercomm, &is_intercomm);
683
  if(!is_intercomm) mall_conf->async_end = MPI_Wtime(); // Merge method only
iker_martin's avatar
iker_martin committed
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
  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() {
  size_t i;
  int is_intercomm, rootBcast, local_state;

  MPI_Comm_test_inter(mall->intercomm, &is_intercomm);
  if(is_intercomm) {
    rootBcast = mall->myId == mall->root ? MPI_ROOT : MPI_PROC_NULL;
  } else {
    rootBcast = mall->root;
  }
  
  comm_data_info(rep_s_data, dist_s_data, MALLEABILITY_NOT_CHILDREN, mall->myId, mall->root, mall->intercomm);
  if(dist_s_data->entries || rep_s_data->entries) { // Enviar datos sincronos
709
    mall_conf->sync_time = MPI_Wtime();
iker_martin's avatar
iker_martin committed
710
    send_data(mall->numC, dist_s_data, MALLEABILITY_USE_SYNCHRONOUS);
711
    if(!is_intercomm) mall_conf->sync_end = MPI_Wtime(); // Merge method only
iker_martin's avatar
iker_martin committed
712
713
714
715
716
717
718

    // TODO Crear funcion especifica y anyadir para Asinc
    // TODO Tener en cuenta el tipo
    for(i=0; i<rep_s_data->entries; i++) {
      MPI_Datatype datatype;
      if(rep_s_data->types[i] == MAL_INT) {
        datatype = MPI_INT;
719
720
721
      } else if(rep_s_data->types[i] == MAL_DOUBLE) {
        datatype = MPI_DOUBLE;
      } else if(rep_s_data->types[i] == MAL_CHAR) {
iker_martin's avatar
iker_martin committed
722
        datatype = MPI_CHAR;
723
724
725
      } else {
	printf("Malleability -- Redistribution data array type not recognised\n");
	MPI_Abort(MPI_COMM_WORLD, -1);
iker_martin's avatar
iker_martin committed
726
727
728
729
730
      }
      MPI_Bcast(rep_s_data->arrays[i], rep_s_data->qty[i], datatype, rootBcast, mall->intercomm);
    } 
  }

731
732
733
  int compute = mall_conf->spawn_method == MALL_SPAWN_MERGE ? 1 : 0;
  comm_results(rootBcast, compute);

iker_martin's avatar
iker_martin committed
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
763
764
765
766
767
768
769
770
771
772
773
  local_state = MALL_DIST_COMPLETED;
  if(!is_intercomm) { // Merge Spawn
    if(mall->numP < mall->numC) { // Expand
      malleability_comms_update(mall->intercomm);
    } else { // Shrink || Merge Shrink requiere de mas tareas
      local_state = MALL_SPAWN_ADAPT_PENDING;
    }
  }

  if(mall->intercomm != MPI_COMM_NULL && mall->intercomm != MPI_COMM_WORLD) {
    MPI_Comm_disconnect(&(mall->intercomm)); //FIXME Error en OpenMPI + Merge
  }

  return local_state;
}


///=============================================
///=============================================
///=============================================
//TODO Add comment
int shrink_redistribution() {
    double time_extra = MPI_Wtime();

    //TODO Create new state before collecting zombies. Processes can perform tasks before that. Then call again Malleability to commit the change
    zombies_collect_suspended(mall->user_comm, mall->myId, mall->numP, mall->numC, mall->root);
    
    if(mall->myId < mall->numC) {
      if(mall->thread_comm != MPI_COMM_WORLD) MPI_Comm_free(&(mall->thread_comm)); //FIXME Modificar a que se pida pro el usuario el cambio y se llama a comms_update
      if(mall->comm != MPI_COMM_WORLD) MPI_Comm_free(&(mall->comm));
      mall->dup_user_comm = 1;

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

      MPI_Comm_set_name(mall->thread_comm, "MPI_COMM_MALL_THREAD");
      MPI_Comm_set_name(mall->comm, "MPI_COMM_MALL");

      MPI_Comm_free(&(mall->intercomm));

774
      mall_conf->spawn_time += MPI_Wtime() - time_extra;
iker_martin's avatar
iker_martin committed
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
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
      return MALL_DIST_COMPLETED;
    } else {
      return MALL_ZOMBIE;
    }
}

//======================================================||
//================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);
}

// TODO MOVER A OTRO LADO??
//======================================================||
//================PRIVATE FUNCTIONS=====================||
//===============COMM PARENTS THREADS===================||
//======================================================||
//======================================================||


int comm_state; //FIXME Usar un handler
/*
 * Crea una hebra para ejecutar una comunicación en segundo plano.
 */
int thread_creation() {
  comm_state = MALL_DIST_PENDING;
  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 comm_state;
}

/*
 * 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, is_intercomm;

  // Comprueba que todos los hilos han terminado la distribucion (Mismo valor en commAsync)
  MPI_Allreduce(&comm_state, &all_completed, 1, MPI_INT, MPI_MAX, mall->comm);
  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

  if(pthread_join(mall->async_thread, NULL)) {
    printf("Error al esperar al hilo\n");
    MPI_Abort(MPI_COMM_WORLD, -1);
    return -2;
  } 
  MPI_Comm_test_inter(mall->intercomm, &is_intercomm);
867
  if(!is_intercomm) mall_conf->async_end = MPI_Wtime(); // Merge method only
iker_martin's avatar
iker_martin committed
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
  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() {
  send_data(mall->numC, dist_a_data, MALLEABILITY_USE_SYNCHRONOUS);
  comm_state = MALL_DIST_COMPLETED;
  pthread_exit(NULL);
}


//==============================================================================
/*
 * 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);
}

void malleability_comms_update(MPI_Comm comm) {
  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(comm, &(mall->thread_comm));
  MPI_Comm_dup(comm, &(mall->comm));
  MPI_Comm_dup(comm, &(mall->user_comm)); 

  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");
}
919
920
921
922
923
924

void comm_results(int root, int compute) {
  MPI_Datatype new_type;

  def_malleability_results(&new_type);
  MPI_Bcast(mall_conf, 1, new_type, root, mall->intercomm);
iker_martin's avatar
iker_martin committed
925
/*
926
  if(compute) {
iker_martin's avatar
iker_martin committed
927
928
929
  mall_conf->sync_time = mall_conf->sync_end - mall_conf->sync_time;
  if(dist_a_data->entries) mall_conf->async_time = mall_conf->async_end - mall_conf->async_time;
  mall_conf->malleability_time = mall_conf->malleability_end - mall_conf->malleability_time;
930
  }
iker_martin's avatar
iker_martin committed
931
  */
932
933
934
935
936
937
938
939
  MPI_Type_free(&new_type);
}

void set_global_time(double ex_start) {
  mall_conf->exec_start = ex_start;
}

void retrieve_results(double *sp_time, double *sy_time, double *asy_time, double *mall_time, double *ex_start) {
iker_martin's avatar
iker_martin committed
940
941
942
  mall_conf->sync_time = mall_conf->sync_end - mall_conf->sync_time;
  if(dist_a_data->entries) mall_conf->async_time = mall_conf->async_end - mall_conf->async_time;
  mall_conf->malleability_time = mall_conf->malleability_end - mall_conf->malleability_time;
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
  *sp_time = mall_conf->spawn_time;
  *sy_time = mall_conf->sync_time;
  *asy_time = mall_conf->async_time;
  *mall_time = mall_conf->malleability_time;
  *ex_start = mall_conf->exec_start;
}

void def_malleability_results(MPI_Datatype *new_type) {
  int i, counts = 5;
  int blocklengths[counts];
  MPI_Aint displs[counts], dir;
  MPI_Datatype types[counts];

  blocklengths[0] = blocklengths[1] = blocklengths[2] = blocklengths[3] = blocklengths[4] = 1;
  types[0] = types[1] =  types[2] = types[3] = types[4] = MPI_DOUBLE;

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

  // Obtener direccion base
  MPI_Get_address(&(mall_conf->spawn_time), &displs[0]);
  MPI_Get_address(&(mall_conf->sync_time), &displs[1]);
  MPI_Get_address(&(mall_conf->async_time), &displs[2]);
  MPI_Get_address(&(mall_conf->malleability_time), &displs[3]);
  MPI_Get_address(&(mall_conf->exec_start), &displs[4]);

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

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