GenericSpawn.c 13.1 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
#include <stdio.h>
#include <stdlib.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <pthread.h>
#include <mpi.h>
#include <string.h>
#include "../malleabilityStates.h"
#include "ProcessDist.h"
#include "GenericSpawn.h"
#include "Baseline.h"
#include "Merge.h"
13
#include "Spawn_state.h"
14
15
16
17

// This code is a Singleton object -- Only one instance can be used at a given time and
// no multiple calls to perform diferent resizes can be performed at the same time.

18
Spawn_data *spawn_data = NULL;
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
pthread_t spawn_thread;
MPI_Comm *returned_comm;

double end_time; //FIXME REFACTOR

//--------------PRIVATE CONFIGURATION DECLARATIONS---------------//
void set_spawn_configuration(char *cmd, int num_cpus, int num_nodes, char *nodelist, int myId, int root, int initial_qty, int target_qty, int type_dist, int spawn_method, int spawn_strategies, MPI_Comm comm);
void set_basic_spawn_dtype();
void deallocate_spawn_data();

//--------------PRIVATE DECLARATIONS---------------//
void generic_spawn(MPI_Comm *child, int data_stage);

int check_single_state(MPI_Comm comm, int global_state);
int check_generic_state(MPI_Comm comm, MPI_Comm *child, int local_state, double *real_time);
int check_merge_shrink_state();

//--------------PRIVATE THREADS DECLARATIONS---------------//
int allocate_thread_spawn();
void* thread_work(void* arg);


//--------------PUBLIC FUNCTIONS---------------//

/*
 * Se solicita la creacion de un nuevo grupo de "numP" procesos con una distribucion
 * fisica "type_dist".
 *
 * Se puede solicitar en primer plano, encargandose por tanto el proceso que llama a esta funcion,
 * o en segundo plano, donde un hilo se encarga de configurar esta creacion.
 *
 * Si se pide en primer plano, al terminarla es posible llamar a "check_spawn_state()" para crear
 * los procesos.
 *
 * Si se pide en segundo plano, llamar a "check_spawn_state()" comprobara si la configuracion para
 * crearlos esta lista, y si es asi, los crea.
 *
 * Devuelve el estado de el procedimiento. Si no devuelve "MALL_SPAWN_COMPLETED", es necesario llamar a
 * "check_spawn_state()".
 */
int init_spawn(char *argv, int num_cpus, int num_nodes, char *nodelist, int myId, int initial_qty, int target_qty, int root, int type_dist, int spawn_method, int spawn_strategies, MPI_Comm comm, MPI_Comm *child) {
  int local_state;
  set_spawn_configuration(argv, num_cpus, num_nodes, nodelist, myId, root, initial_qty, target_qty, type_dist, spawn_method, spawn_strategies, comm);

  if(!spawn_data->spawn_is_async) {
64
    generic_spawn(child, MALL_NOT_STARTED);
65
66
67
68
69
70
71
72
73
74
75
76
77
    local_state = get_spawn_state(spawn_data->spawn_is_async);
    if (local_state == MALL_SPAWN_COMPLETED)
      deallocate_spawn_data();


  } else {
    local_state = spawn_data->spawn_is_single ? MALL_SPAWN_SINGLE_PENDING : MALL_SPAWN_PENDING;
    set_spawn_state(local_state, 0);
    if((spawn_data->spawn_is_single && myId == root) || !spawn_data->spawn_is_single) {
      allocate_thread_spawn();
    }
  }
    
78
  return local_state;
79
80
81
82
83
84
85
86
87
88
}

/*
 * Comprueba si una configuracion para crear un nuevo grupo de procesos esta lista,
 * y en caso de que lo este, se devuelve el communicador a estos nuevos procesos.
 */
int check_spawn_state(MPI_Comm *child, MPI_Comm comm, int data_dist_completed, double *real_time) { 
  int local_state;
  int global_state=MALL_NOT_STARTED;

89
  if(spawn_data->spawn_is_async) { // Async
90
91
    local_state = get_spawn_state(spawn_data->spawn_is_async);

92
    if(local_state == MALL_SPAWN_SINGLE_PENDING || local_state == MALL_SPAWN_SINGLE_COMPLETED) { // Single
93
94
      global_state = check_single_state(comm, local_state);

95
    } else if(local_state == MALL_SPAWN_ADAPT_POSTPONE && data_dist_completed) { // Start Merge Shrink Async
96
97
      global_state = check_merge_shrink_state();

98
    } else if(local_state == MALL_SPAWN_PENDING || local_state == MALL_SPAWN_COMPLETED || local_state == MALL_DIST_ADAPTED) { // Baseline
99
100
101
102
103
104
105
      global_state = check_generic_state(comm, child, local_state, real_time);

    } else {
      printf("Error Check spawn: Configuracion invalida\n");
      MPI_Abort(MPI_COMM_WORLD, -1);
      return -10;
    }
106
  } else if(spawn_data->spawn_method == MALL_SPAWN_MERGE){ // Start Merge shrink Sync
107
108
109
    generic_spawn(child, MALL_DIST_COMPLETED);
    global_state = get_spawn_state(spawn_data->spawn_is_async);
  }
110
111
  if(global_state == MALL_SPAWN_COMPLETED || global_state == MALL_DIST_ADAPTED)
    deallocate_spawn_data();
112
113
114
115
116
117
118
119
120
121
122
  return global_state;
}

/*
 * Funcion bloqueante de los hijos para asegurar que todas las tareas del paso
 * de creacion de los hijos se terminan correctamente.
 *
 * Ademas los hijos obtienen informacion basica de los padres
 * para el paso de redistribucion de datos (Numeros de procesos y Id del Root).
 *
 */
123
void malleability_connect_children(int myId, int numP, int root, MPI_Comm comm, int *numP_parents, int *root_parents, MPI_Comm *parents) {
124
125
126
127
128
129
130
131
132
133
134
135
  spawn_data = (Spawn_data *) malloc(sizeof(Spawn_data));
  spawn_data->root = root;
  spawn_data->myId = myId;
  spawn_data->spawn_qty = numP;
  spawn_data->target_qty = numP;
  spawn_data->comm = comm;

  set_basic_spawn_dtype();
  MPI_Bcast(spawn_data, 1, spawn_data->dtype, MALLEABILITY_ROOT, *parents);

  switch(spawn_data->spawn_method) {
    case MALL_SPAWN_BASELINE:
136
      baseline(*spawn_data, parents);
137
138
      break;
    case MALL_SPAWN_MERGE:
139
140
      spawn_data->target_qty += *numP_parents;
      merge(*spawn_data, parents, MALL_NOT_STARTED);
141
142
143
      break;
  }

144
145
146
  *root_parents = spawn_data->root_parents;
  *numP_parents = spawn_data->initial_qty;

147
148
149
150
  MPI_Type_free(&(spawn_data->dtype));
  free(spawn_data);
}

151
152
153
154
155
156
157
158
159
160
161
162
163
/*
 * Función para obtener si entre las estrategias elegidas, se utiliza
 * la estrategia pasada como segundo argumento.
 *
 * Devuelve en "result" 1(Verdadero) si utiliza la estrategia, 0(Falso) en caso
 * contrario.
 */
int malleability_spawn_contains_strat(int spawn_strategies, int strategy, int *result) {
  int value = spawn_strategies % strategy ? 0 : 1;
  if(result != NULL) *result = value;
  return value;
}

164
165
166
167
168
169
170
171
172
173
174
175
176
177
//--------------PRIVATE CONFIGURATION FUNCTIONS---------------//
/*
 * Agrupa en una sola estructura todos los datos de configuración necesarios
 * e inicializa las estructuras necesarias.
 */
void set_spawn_configuration(char *cmd, int num_cpus, int num_nodes, char *nodelist, int myId, int root, int initial_qty, int target_qty, int type_dist, int spawn_method, int spawn_strategies, MPI_Comm comm) {
  spawn_data = (Spawn_data *) malloc(sizeof(Spawn_data));

  spawn_data->myId = myId;
  spawn_data->root = root;
  spawn_data->root_parents = root;
  spawn_data->initial_qty = initial_qty;
  spawn_data->target_qty = target_qty;
  spawn_data->spawn_method = spawn_method;
178
179
  malleability_spawn_contains_strat(spawn_strategies, MALL_SPAWN_SINGLE, &(spawn_data->spawn_is_single));
  malleability_spawn_contains_strat(spawn_strategies, MALL_SPAWN_PTHREAD, &(spawn_data->spawn_is_async));
180
181
182
183
184
185
186
187
  spawn_data->comm = comm;

  set_basic_spawn_dtype();

  switch(spawn_data->spawn_method) {
    case MALL_SPAWN_BASELINE:
      spawn_data->spawn_qty = spawn_data->target_qty;
      spawn_data->already_created = 0;
188
      break;
189
190
191
    case MALL_SPAWN_MERGE:
      spawn_data->spawn_qty = spawn_data->target_qty - spawn_data->initial_qty;
      spawn_data->already_created = spawn_data->initial_qty;
192
      break;
193
194
195
  }

  if(spawn_data->spawn_is_async) {
196
    init_spawn_state();
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
  }

  if(spawn_data->myId == spawn_data->root) {
    physical_struct_create(target_qty, spawn_data->already_created, num_cpus, num_nodes, nodelist, type_dist, MALL_DIST_STRING, &(spawn_data->dist));

    //COPY PROGRAM NAME
    spawn_data->cmd = malloc(strlen(cmd) * sizeof(char));
    strcpy(spawn_data->cmd, cmd);

  } else {
    spawn_data->cmd = malloc(1 * sizeof(char));
    spawn_data->mapping = MPI_INFO_NULL; //It is only needed for the root process
  }
}

/*
 * Crea un tipo derivado para mandar 4 enteros con informacion
 * basica a los hijos. Son datos necesarios para que terminen
 * la creacion de procesos.
 */
void set_basic_spawn_dtype() {
  int i, counts = 4;
  int blocklengths[] = {1, 1, 1, 1};
  MPI_Aint displs[counts], dir;
  MPI_Datatype types[counts];

  // Rellenar vector types
  types[0] = types[1] = types[2] = types[3] = MPI_INT;

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

  MPI_Get_address(&(spawn_data->root_parents), &displs[0]);
  MPI_Get_address(&(spawn_data->initial_qty), &displs[1]);
  MPI_Get_address(&(spawn_data->spawn_is_single), &displs[2]);
  MPI_Get_address(&(spawn_data->spawn_method), &displs[3]);

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

  MPI_Type_create_struct(counts, blocklengths, displs, types, &(spawn_data->dtype));
  MPI_Type_commit(&(spawn_data->dtype));
}

/*
 * Libera una estructura de datos spawn_data
 * junto a la destrucion de aquellas estructuras que utiliza.
 */
void deallocate_spawn_data() {
245
246
247
248
249
250
251
252
  if(spawn_data == NULL) return;

  if(spawn_data->cmd != NULL) {
    free(spawn_data->cmd);
  }
  if(spawn_data->dtype != MPI_DATATYPE_NULL) {
    MPI_Type_free(&(spawn_data->dtype));
  }
253
254
255
256
  if(spawn_data->mapping != MPI_INFO_NULL) {
    MPI_Info_free(&(spawn_data->mapping));
  }
  if(spawn_data->spawn_is_async) {
257
    free_spawn_state();
258
259
  }
  free(spawn_data); 
260
  spawn_data = NULL;
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
}


//--------------PRIVATE SPAWN CREATION FUNCTIONS---------------//

/*
 * Funcion generica para la creacion de procesos. Obtiene la configuracion
 * y segun esta, elige como deberian crearse los procesos.
 *
 * Cuando termina, modifica la variable global para indicar este cambio
 */
void generic_spawn(MPI_Comm *child, int data_stage) {
  int local_state;

  // WORK
  if(spawn_data->myId == spawn_data->root) { //SET MAPPING
    processes_dist(spawn_data->dist, &(spawn_data->mapping));
  }
  switch(spawn_data->spawn_method) {
    case MALL_SPAWN_BASELINE:
      local_state = baseline(*spawn_data, child);
      break;
    case MALL_SPAWN_MERGE:
      local_state = merge(*spawn_data, child, data_stage);
      break;
  }
  // END WORK

289
  set_spawn_state(local_state, spawn_data->spawn_is_async);
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
  end_time = MPI_Wtime();
}


//--------------PRIVATE THREAD FUNCTIONS---------------//

/*
 * Aloja la memoria para un hilo auxiliar dedicado a la creacion de procesos.
 * No se puede realizar un "join" sobre el hilo y el mismo libera su memoria
 * asociado al terminar.
 */
int allocate_thread_spawn() {
  if(pthread_create(&spawn_thread, NULL, thread_work, NULL)) {
    printf("Error al crear el hilo de SPAWN\n");
    MPI_Abort(MPI_COMM_WORLD, -1);
    return -1;
  }
  if(pthread_detach(spawn_thread)) {
    printf("Error when detaching spawning thread\n");
    MPI_Abort(MPI_COMM_WORLD, -1);
    return -1;
  }
  return 0;
}

/*
 * Funcion llamada por un hilo para que este se encarge
 * de configurar la creacion de un nuevo grupo de procesos.
 *
 * Una vez esta lista la configuracion y es posible crear los procesos
 * se avisa al hilo maestro.
 */
void* thread_work(void* arg) {
323
  int local_state, repeat = 0;
324
325
  returned_comm = (MPI_Comm *) malloc(sizeof(MPI_Comm));
 
326
327
328
329
  generic_spawn(returned_comm, MALL_NOT_STARTED);

  local_state = get_spawn_state(MALL_SPAWN_PTHREAD);
  if(local_state == MALL_SPAWN_ADAPT_POSTPONE) {
330
    // El grupo de procesos se terminara de juntar tras la redistribucion de datos
331
332
    repeat = 1;
    local_state = wait_wakeup();
333
  }
334
  if (repeat) generic_spawn(returned_comm, MALL_DIST_COMPLETED);
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355

  pthread_exit(NULL);
}

/*
 * Comprueba si una creacion de procesos asincrona en el
 * paso "single" ha terminado. 
 * Si no ha terminado se mantiene el estado 
 * "MALL_SPAWN_SINGLE_PENDING".
 *
 * Si ha terminado se crean los hilos auxiliares para 
 * los procesos no root y se devuelve el estado
 * "MALL_SPAWN_PENDING".
 */
int check_single_state(MPI_Comm comm, int global_state) {
  MPI_Bcast(&global_state, 1, MPI_INT, spawn_data->root, comm);

  // Non-root processes join root to finalize the spawn
  // They also must join if the application has ended its work
  if(global_state == MALL_SPAWN_SINGLE_COMPLETED) { 
    global_state = MALL_SPAWN_PENDING;
356
    set_spawn_state(global_state, MALL_SPAWN_PTHREAD);
357

358
359
    //int threads_not_spawned = pthread_equal(pthread_self(), spawn_thread);
    if(spawn_data->myId != spawn_data->root) { //&& threads_not_spawned) {
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
      allocate_thread_spawn(spawn_data);
    }
  }
  return global_state;
}

/*
 * Comprueba si una creación de procesos asincrona en el
 * paso "generic" ha terminado.
 * Si no ha terminado devuelve el estado 
 * "MALL_SPAWN_PENDING".
 *
 * Si ha terminado libera la memoria asociada a spawn_data
 * y devuelve el estado "MALL_SPAWN_COMPLETED".
 */
int check_generic_state(MPI_Comm comm, MPI_Comm *child, int local_state, double *real_time) {
  int global_state;

  MPI_Allreduce(&local_state, &global_state, 1, MPI_INT, MPI_MIN, comm);
379
380
  if(global_state == MALL_SPAWN_COMPLETED || global_state == MALL_DIST_ADAPTED) {
    set_spawn_state(global_state, MALL_SPAWN_PTHREAD);
381
382
383
384
385
386
387
388
389
390
391
392
393
394
    *child = *returned_comm;
    deallocate_spawn_data(spawn_data);
    *real_time=end_time;
  }
  return global_state;
}

/*
 * Permite a una reduccion merge asincrona
 * de procesos que estaba a la espera de que la
 * distribucion de los datos se completase continue.
 */
int check_merge_shrink_state() {
  // FIXME Pasar como caso especial para evitar iteracion no necesaria
395
396
397
  int global_state = MALL_SPAWN_PENDING;
  set_spawn_state(global_state, MALL_SPAWN_PTHREAD);
  return global_state;
398
}