New version of Proteo

Codes/MaM/MAM_RMS.c

+#define _GNU_SOURCE
+#include <stdlib.h>
+#include <stdio.h>
+#include <string.h>
+#include <unistd.h>
+#include <sched.h>
+#include <mpi.h>
+#include "MAM_RMS.h"
+#include "MAM_DataStructures.h"
+
+
+#if MAM_USE_SLURM
+#include <slurm/slurm.h>
+int MAM_I_slurm_getenv_hosts_info();
+int MAM_I_slurm_getjob_hosts_info();
+#endif
+
+int MAM_I_get_hosts_info();
+int GetCPUCount();
+
+void MAM_check_hosts() {
+  int not_filled = 1;
+  
+  #if MAM_USE_SLURM
+    not_filled = MAM_I_slurm_getjob_hosts_info();
+    if(not_filled) {
+      #if MAM_DEBUG >= 2
+        DEBUG_FUNC("WARNING - RMS info retriever failed with slurm functions. Trying with ENV variables", mall->myId, mall->numP); 
+      #endif
+      if(mall->nodelist != NULL) {
+        free(mall->nodelist);
+	mall->nodelist = NULL;
+      }
+
+      not_filled = MAM_I_slurm_getenv_hosts_info();
+    }
+  #endif
+  if(not_filled) {
+    if(mall->nodelist != NULL) {
+      free(mall->nodelist);
+      mall->nodelist = NULL;
+    }
+
+    not_filled = MAM_I_get_hosts_info();
+  }
+
+
+  if(not_filled) {
+    if(mall->myId == mall->root) printf("MAM FATAL ERROR: It has not been possible to obtain the nodelist\n");
+    fflush(stdout);
+    MPI_Abort(mall->comm, -50);
+  }
+
+  #if MAM_DEBUG >= 2
+    if(mall->myId == mall->root) {
+      DEBUG_FUNC("Obtained Nodelist", mall->myId, mall->numP); 
+      printf("NODELIST: %s\nNODE_COUNT: %d NUM_CPUS_PER_NODE: %d\n", mall->nodelist, mall->num_nodes, mall->num_cpus);
+      fflush(stdout); 
+    }
+  #endif
+}
+
+/*
+ * @brief Get if a group of processes uses an internode comunicator
+ *
+ * This function checks the physical distribution of all ranks in the
+ * original communicator passed to MaM. If all of them reside in the
+ * same host, false is returned. True is returned otherwise.
+ *
+ * @return Integer indicating if more than one node is used by the
+ * original communicator (>0) or only one (0).
+ */
+int MAM_Is_internode_group() {
+  int i, name_len, max_name_len, unique_count;
+  int myId, numP;
+  char *my_host, *all_hosts, *tested_host;
+
+  MPI_Comm_rank(mall->original_comm, &myId);
+  MPI_Comm_size(mall->original_comm, &numP);
+
+  unique_count = 0; //First node is not counted
+  if(numP == 1) return unique_count;
+
+  all_hosts = NULL;
+  my_host = (char *) malloc(MPI_MAX_PROCESSOR_NAME * sizeof(char));
+  MPI_Get_processor_name(my_host, &name_len);
+
+  MPI_Allreduce(&name_len, &max_name_len, 1, MPI_INT, MPI_MAX, mall->original_comm);
+  my_host[max_name_len] = '\0';
+  max_name_len++; // Len does not consider terminating character
+  if(myId == MAM_ROOT) {
+    all_hosts = (char *) malloc(numP * max_name_len * sizeof(char));
+  }
+  //FIXME Should be a Gatherv as each host could have unitialised chars between name_len and max_name_len
+  MPI_Gather(my_host, max_name_len, MPI_CHAR, all_hosts, max_name_len, MPI_CHAR, MAM_ROOT, mall->original_comm);
+
+  if(myId == MAM_ROOT) {
+    for (i = 1; i < numP; i++) {
+      tested_host = all_hosts + (i * max_name_len);
+      if (strcmp(my_host, tested_host) != 0) {
+        unique_count++;
+        break;
+      }
+    }
+    free(all_hosts);
+  }
+  MPI_Bcast(&unique_count, 1, MPI_INT, MAM_ROOT, mall->original_comm);
+  free(my_host);
+  return unique_count;
+}
+
+/*
+ * TODO
+ * FIXME Does not consider heterogenous machines for num_cpus
+ * FIXME Always returns 0... -- Perform error checking?
+ */
+int MAM_I_get_hosts_info() {
+  int i, j, name_len, max_name_len, unique_count, *unique_hosts;
+  char *my_host, *all_hosts, *confirmed_host, *tested_host;
+
+  all_hosts = NULL;
+  my_host = (char *) malloc(MPI_MAX_PROCESSOR_NAME * sizeof(char));
+  MPI_Get_processor_name(my_host, &name_len);
+
+  MPI_Allreduce(&name_len, &max_name_len, 1, MPI_INT, MPI_MAX, mall->comm);
+  my_host[max_name_len] = '\0';
+  max_name_len++; // Len does not consider terminating character
+  if(mall->myId == mall->root) {
+    all_hosts = (char *) malloc(mall->numP * max_name_len * sizeof(char));
+    unique_hosts = (int *) malloc(mall->numP * sizeof(int));
+    unique_hosts[0] = 0; //First host will always be unique
+    unique_count = 1;
+  }
+  //FIXME Should be a Gatherv as each host could have unitialised chars between name_len and max_name_len
+  MPI_Gather(my_host, max_name_len, MPI_CHAR, all_hosts, max_name_len, MPI_CHAR, mall->root, mall->comm);
+
+  if(mall->myId == mall->root) {
+    for (i = 1; i < mall->numP; i++) {
+      for (j = 0; j < unique_count; j++) {
+	tested_host = all_hosts + (i * max_name_len);
+	confirmed_host = all_hosts + (unique_hosts[j] * max_name_len);
+        if (strcmp(tested_host, confirmed_host) != 0) {
+	  unique_hosts[unique_count] = i;
+          unique_count++;
+          break;
+        }
+      }
+    }
+
+    mall->num_nodes = unique_count;
+    mall->num_cpus = GetCPUCount(); 
+    mall->nodelist_len = unique_count*max_name_len;
+    mall->nodelist = (char *) malloc(mall->nodelist_len * sizeof(char));
+
+    strcpy(mall->nodelist, ""); //FIXME Strcat can be very inneficient...
+    for (i = 0; i < unique_count; i++) {
+      confirmed_host = all_hosts + (unique_hosts[i] * max_name_len);
+      strcat(mall->nodelist, confirmed_host);
+      if (i < unique_count - 1) {
+        strcat(mall->nodelist, ",");
+      }
+    }
+
+    free(all_hosts);
+    free(unique_hosts);
+  }
+
+  free(my_host);
+  return 0;
+}
+
+/*
+ * @brief Get the total number of CPUs available to the process.
+ *
+ * This function uses sched_getaffinity to obtain the CPU affinity of the current process
+ * and counts the number of CPUs in the affinity set. It adjusts the loop based on the
+ * maximum number of CPUs allowed on the system.
+ *
+ * @return The total number of CPUs available to the process.
+ *
+ * Code obtained from: https://stackoverflow.com/questions/4586405/how-to-get-the-number-of-cpus-in-linux-using-c
+ * The code has been slightly modified.
+ */
+int GetCPUCount() {
+  cpu_set_t cs;
+  CPU_ZERO(&cs);
+  sched_getaffinity(0, sizeof(cs), &cs);
+
+  int count = 0;
+  int max_cpus = sysconf(_SC_NPROCESSORS_ONLN);
+
+  for (int i = 0; i < max_cpus; i++) {
+      if (CPU_ISSET(i, &cs)) {
+          count++;
+      } else {
+          break;
+      }
+  }
+  return count;
+}
+
+#if MAM_USE_SLURM
+/*
+ * TODO
+ */
+int MAM_I_slurm_getenv_hosts_info() {
+  char *tmp = NULL, *tmp_copy, *token;
+  int cpus, count;
+  //int i, *cpus_counts, *nodes_counts, *aux;
+  
+  tmp = getenv("SLURM_JOB_NUM_NODES");
+  if(tmp == NULL) return 1;
+  mall->num_nodes = atoi(tmp);
+  tmp = NULL;
+
+  tmp = getenv("SLURM_JOB_NODELIST");
+  if(tmp == NULL) return 1;
+  mall->nodelist_len = strlen(tmp)+1;
+  mall->nodelist = (char *) malloc(mall->nodelist_len * sizeof(char));
+  strcpy(mall->nodelist, tmp);
+  tmp = NULL;
+
+
+  //EXAMPLE - SLURM_JOB_CPUS_PER_NODE='72(x2),36'
+  //It indicates two nodes have 72 CPUs each and third has 36 cpus
+  tmp = getenv("SLURM_JOB_CPUS_PER_NODE");
+  if(tmp == NULL) return 1;
+
+  tmp_copy = (char *) malloc((strlen(tmp)+1) * sizeof(char));
+  strcpy(tmp_copy, tmp);
+  token = strtok(tmp_copy, ",");
+  //TODO When MaM considers heteregenous allocations, these will be needed instead of num_cpus.
+  //cpus_counts = (int *) malloc(mall->num_nodes * sizeof(int));
+  //nodes_counts = (int *) malloc(mall->num_nodes * sizeof(int));
+  //i = 0;
+  mall->num_cpus = 0;
+
+  while (token != NULL) {
+    // If actual token contains only one node, the second portion
+    // does not appear and sscanf does not modify "count"
+    // First portion --> "%d"
+    // Second portion -> "(x%d)"
+    count = 1;
+    if (sscanf(token, "%d(x%d)", &cpus, &count) >= 1) {
+      mall->num_cpus = cpus; // num_cpus stores the amount of cores per cpu
+      //cpus_per_node[i] = cpus;
+      //nodes_count[i] = count;
+      //i++;
+    }
+    token = strtok(NULL, ",");
+  }
+  /*
+  if(i < mall->num_nodes) {
+    aux = (int *) realloc(cpus_per_node, i * sizeof(int));
+    if(cpus_per_node != aux && cpus_per_node != NULL) free(cpus_per_node);
+    cpus_per_node = aux;
+
+    aux = (int *) realloc(nodes_counts, i * sizeof(int));
+    if(nodes_count != aux && nodes_count != NULL) free(nodes_count);
+    nodes_count = aux;
+  }
+  */
+
+  free(tmp_copy);
+  return 0;
+}
+
+/*
+ * TODO
+ * FIXME Does not consider heterogenous machines
+ */
+int MAM_I_slurm_getjob_hosts_info() {
+  int jobId, err;
+  char *tmp = NULL;
+  job_info_msg_t *j_info;
+  slurm_job_info_t last_record;
+
+  tmp = getenv("SLURM_JOB_ID");
+  if(tmp == NULL) return 1;
+  jobId = atoi(tmp);
+
+  err = slurm_load_job(&j_info, jobId, 1); // FIXME Valgrind Not freed
+  if(err) return err;
+  last_record = j_info->job_array[j_info->record_count - 1];
+
+  mall->num_nodes = last_record.num_nodes;
+  mall->num_cpus = last_record.num_cpus / last_record.num_nodes;
+
+  mall->nodelist_len = strlen(last_record.nodes)+1;
+  mall->nodelist = (char *) malloc(mall->nodelist_len * sizeof(char));
+  strcpy(mall->nodelist, last_record.nodes);
+
+  slurm_free_job_info_msg(j_info);
+  return 0;
+}
+#endif
+
+//TODO REFACTOR PARA CUANDO SE COMUNIQUE CON RMS
+    // Get Slurm job info
+    //int jobId;
+    //char *tmp;
+    //job_info_msg_t *j_info;
+    //slurm_job_info_t last_record;
+    //tmp = getenv("SLURM_JOB_ID");
+    //jobId = atoi(tmp);
+    //slurm_load_job(&j_info, jobId, 1);
+    //last_record = j_info->job_array[j_info->record_count - 1];
+    // Free JOB INFO
+    //slurm_free_job_info_msg(j_info);

Codes/MaM/MAM_RMS.h

+#ifndef MAM_RMS_H
+#define MAM_RMS_H
+
+void MAM_check_hosts();
+int MAM_Is_internode_group();
+
+#endif

Codes/MaM/MAM_Times.c

+#include "MAM_Times.h"
+#include "MAM_DataStructures.h"
+
+void def_malleability_times(MPI_Datatype *new_type);
+
+void init_malleability_times() {
+  #if MAM_DEBUG >= 4
+    DEBUG_FUNC("Initializing recording structure", mall->myId, mall->numP); fflush(stdout); MPI_Barrier(mall->comm);
+  #endif
+
+  mall_conf->times = (malleability_times_t *) malloc(sizeof(malleability_times_t));
+  if(mall_conf->times == NULL) {
+    perror("Error al crear la estructura de tiempos interna para maleabilidad\n");
+    MPI_Abort(MPI_COMM_WORLD, -5);
+  }
+
+  reset_malleability_times();
+  def_malleability_times(&mall_conf->times->times_type);
+
+  #if MAM_DEBUG >= 4
+    DEBUG_FUNC("Initialized recording structure", mall->myId, mall->numP); fflush(stdout); MPI_Barrier(mall->comm);
+  #endif
+}
+
+
+void reset_malleability_times() {
+  malleability_times_t *times = mall_conf->times;
+  times->spawn_start = 0; times->sync_start = 0; times->async_start = 0; times->user_start = 0; times->malleability_start = 0;
+  times->sync_end = 0; times->async_end = 0; times->user_end = 0; times->malleability_end = 0;
+  times->spawn_time = 0;
+}
+
+void free_malleability_times() {
+  #if MAM_DEBUG >= 4
+    DEBUG_FUNC("Freeing recording structure", mall->myId, mall->numP); fflush(stdout);
+  #endif
+  if(mall_conf->times != NULL) {
+    if(mall_conf->times->times_type != MPI_DATATYPE_NULL) {
+      MPI_Type_free(&mall_conf->times->times_type);
+      mall_conf->times->times_type = MPI_DATATYPE_NULL;
+    }
+    free(mall_conf->times);
+  }
+  #if MAM_DEBUG >= 4
+    DEBUG_FUNC("Freed recording structure", mall->myId, mall->numP); fflush(stdout);
+  #endif
+}
+
+/*
+ * @brief Returns the times used for the different steps of last reconfiguration.
+ *
+ * This function is intended to be called when a reconfiguration has ended. 
+ * It is designed to provide the necessary information for the user to perform data redistribution.
+ * 
+ * Null values can be passed if there is no interest in retreiving particular times
+ *
+ * Parameters:
+ *  - double *sp_time:   A pointer where the spawn time will be saved.
+ *  - double *sy_time:   A pointer where the sychronous data redistribution time will be saved.
+ *  - double *asy_time:  A pointer where the asychronous data redistribution time will be saved.
+ *  - double *user_time: A pointer where the user data redistribution time will be saved.
+ *  - double *mall_time: A pointer where the malleability time will be saved.
+ */
+void MAM_Retrieve_times(double *sp_time, double *sy_time, double *asy_time, double *user_time, double *mall_time) {
+  malleability_times_t *times = mall_conf->times;
+  if(sp_time != NULL)   *sp_time = times->spawn_time;
+  if(sy_time != NULL)   *sy_time = times->sync_end - times->sync_start;
+  if(asy_time != NULL)  *asy_time = times->async_end - times->async_start;
+  if(user_time != NULL) *user_time = times->user_end - times->user_start;
+  if(mall_time != NULL) *mall_time = times->malleability_end - times->malleability_start;
+}
+
+void malleability_times_broadcast(int root) {
+  MPI_Bcast(mall_conf->times, 1, mall_conf->times->times_type, root, mall->intercomm);
+}
+
+void def_malleability_times(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;
+
+  // Se pasa el vector a traves de la direccion de "mall_conf"
+  // Rellenar vector displs
+  MPI_Get_address(mall_conf->times, &dir);
+
+  // Obtener direccion base
+  MPI_Get_address(&(mall_conf->times->spawn_time), &displs[0]);
+  MPI_Get_address(&(mall_conf->times->sync_start), &displs[1]);
+  MPI_Get_address(&(mall_conf->times->async_start), &displs[2]);
+  MPI_Get_address(&(mall_conf->times->user_start), &displs[3]);
+  MPI_Get_address(&(mall_conf->times->malleability_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);
+}

Codes/MaM/MAM_Times.h

+#ifndef MAM_TIMES_H
+#define MAM_TIMES_H
+
+#include <mpi.h>
+
+void init_malleability_times();
+void reset_malleability_times();
+void free_malleability_times();
+
+void malleability_times_broadcast(int root);
+
+void MAM_I_retrieve_times(double *sp_time, double *sy_time, double *asy_time, double *user_time, double *mall_time);
+
+#endif

Codes/MaM/MAM_Times_retrieve.h

+
+#ifndef MAM_TIMES_RETRIEVE_H
+#define MAM_TIMES_RETRIEVE_H
+
+void MAM_Retrieve_times(double *sp_time, double *sy_time, double *asy_time, double *user_time, double *mall_time);
+#endif

Codes/malleability/malleabilityTypes.c → Codes/MaM/MAM_Types.c

-#include "malleabilityTypes.h"
+#include "MAM_Types.h"
+#include "MAM_DataStructures.h"
+#include "MAM_Configuration.h"
 
 
-void init_malleability_data_struct(malleability_data_t *data_struct, int size);
-void realloc_malleability_data_struct(malleability_data_t *data_struct, int qty_to_add);
+void init_malleability_data_struct(malleability_data_t *data_struct, size_t size);
+void realloc_malleability_data_struct(malleability_data_t *data_struct, size_t qty_to_add);
 
 void def_malleability_entries(malleability_data_t *data_struct_rep, malleability_data_t *data_struct_dist, MPI_Datatype *new_type);
 void def_malleability_qty_type(malleability_data_t *data_struct_rep, malleability_data_t *data_struct_dist, MPI_Datatype *new_type);
@@ -20,26 +22,58 @@ void def_malleability_qty_type(malleability_data_t *data_struct_rep, malleabilit
  * todos los padres. La nueva serie "data" solo representa los datos
  * que tiene este padre.
  */
-void add_data(void *data, int total_qty, int type, int request_qty, malleability_data_t *data_struct) {
-  int i;
+void add_data(void *data, size_t total_qty, MPI_Datatype type, size_t request_qty, malleability_data_t *data_struct) {
+  size_t i;
   
   if(data_struct->entries == 0) {
-    init_malleability_data_struct(data_struct, MALLEABILITY_INIT_DATA_QTY);
+    init_malleability_data_struct(data_struct, MAM_TYPES_INIT_DATA_QTY);
   } else if(data_struct->entries == data_struct->max_entries) {
-    realloc_malleability_data_struct(data_struct, MALLEABILITY_INIT_DATA_QTY);
+    realloc_malleability_data_struct(data_struct, MAM_TYPES_INIT_DATA_QTY);
   }
   
   data_struct->qty[data_struct->entries] = total_qty;
   data_struct->types[data_struct->entries] = type;
   data_struct->arrays[data_struct->entries] = data;
+  data_struct->request_qty[data_struct->entries] = request_qty;
 
-  data_struct->requests[data_struct->entries] = (MPI_Request *) malloc(request_qty * sizeof(MPI_Request));
-  for(i=0; i < request_qty; i++) {
-    data_struct->requests[data_struct->entries][i] = MPI_REQUEST_NULL;
+  if(request_qty) {
+    data_struct->requests[data_struct->entries] = (MPI_Request *) malloc(request_qty * sizeof(MPI_Request));
+    for(i=0; i < request_qty; i++) {
+      data_struct->requests[data_struct->entries][i] = MPI_REQUEST_NULL;
+    }
   }
   data_struct->entries+=1;
 }
 
+/*
+ * Modifica en la estructura de datos a comunicar con los hijos
+ * un set de datos de un total "total_qty" distribuido entre
+ * todos los padres. La nueva serie "data" solo representa los datos
+ * que tiene este padre.
+ */
+void modify_data(void *data, size_t index, size_t total_qty, MPI_Datatype type, size_t request_qty, malleability_data_t *data_struct) {
+  size_t i;
+  
+  if(data_struct->entries < index) { // Index does not exist
+    return;
+  }
+  if(data_struct->requests[index] != NULL) {
+    //free(data_struct->requests[index]); TODO Error when trying to free
+    data_struct->requests[index] = NULL;
+  }
+
+  data_struct->qty[index] = total_qty;
+  data_struct->types[index] = type;
+  data_struct->arrays[index] = data;
+  data_struct->request_qty[index] = request_qty;
+
+  if(request_qty) {
+    data_struct->requests[index] = (MPI_Request *) malloc(request_qty * sizeof(MPI_Request));
+    for(i=0; i < request_qty; i++) {
+      data_struct->requests[index][i] = MPI_REQUEST_NULL;
+    }
+  }
+}
 
 /*
  * Comunicar desde los padres a los hijos las estructuras de datos sincronas o asincronas
@@ -48,37 +82,42 @@ void add_data(void *data, int total_qty, int type, int request_qty, malleability
  * En el argumento "root" todos tienen que indicar quien es el proceso raiz de los padres
  * unicamente.
  */
-void comm_data_info(malleability_data_t *data_struct_rep, malleability_data_t *data_struct_dist, int is_children_group, int myId, int root, MPI_Comm intercomm) {
-  int i, rootBcast = MPI_PROC_NULL;
+void comm_data_info(malleability_data_t *data_struct_rep, malleability_data_t *data_struct_dist, int is_children_group) {
+  int type_size;
+  size_t i, j;
   MPI_Datatype entries_type, struct_type;
 
-  if(is_children_group) {
-    rootBcast = root;
-  } else {
-    if(myId == root) rootBcast = MPI_ROOT;
-  }
-
   // Mandar primero numero de entradas
   def_malleability_entries(data_struct_dist, data_struct_rep, &entries_type);
-  MPI_Bcast(MPI_BOTTOM, 1, entries_type, rootBcast, intercomm);
+  MPI_Bcast(MPI_BOTTOM, 1, entries_type, mall->root_collectives, mall->intercomm);
 
   if(is_children_group) {
-    if(data_struct_rep->entries != 0) init_malleability_data_struct(data_struct_rep, data_struct_rep->entries);
-    if(data_struct_dist->entries != 0) init_malleability_data_struct(data_struct_dist, data_struct_dist->entries);
+    if(data_struct_rep->entries != 0) { init_malleability_data_struct(data_struct_rep, data_struct_rep->entries); }
+    if(data_struct_dist->entries != 0) { init_malleability_data_struct(data_struct_dist, data_struct_dist->entries); } //FIXME Valgrind not freed
   }
 
   def_malleability_qty_type(data_struct_dist, data_struct_rep, &struct_type);
-  MPI_Bcast(MPI_BOTTOM, 1, struct_type, rootBcast, intercomm); //FIXME Doy error
+  MPI_Bcast(MPI_BOTTOM, 1, struct_type, mall->root_collectives, mall->intercomm);
 
   if(is_children_group) {
-    //data_struct->requests[data_struct->entries] = (MPI_Request *) malloc(request_qty * sizeof(MPI_Request)); FIXME Crear los requests?
-    //data_struct->requests[data_struct->entries][i] = MPI_REQUEST_NULL;
-    
     for(i=0; i < data_struct_rep->entries; i++) {
-      data_struct_rep->arrays[i] = (void *) malloc(data_struct_rep->qty[i] * sizeof(int)); //TODO Tener en cuenta que no siempre es int
+      MPI_Type_size(data_struct_rep->types[i], &type_size);
+      data_struct_rep->arrays[i] = (void *) malloc(data_struct_rep->qty[i] * (size_t) type_size); //FIXME This memory is not freed -- How should be done?
+      if(data_struct_rep->request_qty[i]) {
+        data_struct_rep->requests[i] = (MPI_Request *) malloc(data_struct_rep->request_qty[i] * sizeof(MPI_Request));
+        for(j=0; j < data_struct_rep->request_qty[i]; j++) {
+          data_struct_rep->requests[i][j] = MPI_REQUEST_NULL;
+        }
+      }
     }
     for(i=0; i < data_struct_dist->entries; i++) {
-      data_struct_dist->arrays[i] = (void *) malloc(data_struct_dist->qty[i] * sizeof(int)); //TODO Tener en cuenta que no siempre es int
+      data_struct_dist->arrays[i] = (void *) NULL; // TODO Se podria inicializar aqui?
+      if(data_struct_dist->request_qty[i]) {
+        data_struct_dist->requests[i] = (MPI_Request *) malloc(data_struct_dist->request_qty[i] * sizeof(MPI_Request));
+        for(j=0; j < data_struct_dist->request_qty[i]; j++) {
+          data_struct_dist->requests[i][j] = MPI_REQUEST_NULL;
+        }
+      }
     }
   }
 
@@ -97,14 +136,22 @@ void comm_data_info(malleability_data_t *data_struct_rep, malleability_data_t *d
  * caracteristicas de localización y uso. Se inicializa para utilizar hasta
  * "size" elementos.
  */
-void init_malleability_data_struct(malleability_data_t *data_struct, int size) {
+void init_malleability_data_struct(malleability_data_t *data_struct, size_t size) {
+  size_t i;
+
   data_struct->max_entries = size;
-  data_struct->qty = (int *) malloc(size * sizeof(int));
-  data_struct->types = (int *) malloc(size * sizeof(int));
+  data_struct->qty = (size_t *) malloc(size * sizeof(size_t));
+  data_struct->types = (MPI_Datatype *) malloc(size * sizeof(MPI_Datatype));
+  data_struct->request_qty = (size_t *) malloc(size * sizeof(size_t));
   data_struct->requests = (MPI_Request **) malloc(size * sizeof(MPI_Request *));
+  data_struct->windows = (MPI_Win *) malloc(size * sizeof(MPI_Win));
   data_struct->arrays = (void **) malloc(size * sizeof(void *));
 
-  data_struct->request_ibarrier = MPI_REQUEST_NULL;
+  for(i=0; i<size; i++) { //calloc and memset does not ensure a NULL value
+    data_struct->requests[i] = NULL;
+    data_struct->windows[i] = MPI_WIN_NULL;
+    data_struct->arrays[i] = NULL;
+  }
 }
 
 /*
@@ -112,43 +159,83 @@ void init_malleability_data_struct(malleability_data_t *data_struct, int size) {
  * caracteristicas de localización y uso. Se anyaden "size" entradas nuevas
  * a las ya existentes.
  */
-void realloc_malleability_data_struct(malleability_data_t *data_struct, int qty_to_add) {
-  int *qty_aux, *types_aux, needed;
+void realloc_malleability_data_struct(malleability_data_t *data_struct, size_t qty_to_add) {
+  size_t i, needed, *qty_aux, *request_qty_aux;
+  MPI_Datatype *types_aux;
+  MPI_Win *windows_aux;
   MPI_Request **requests_aux;
   void **arrays_aux;
 
   needed = data_struct->max_entries + qty_to_add;
-  qty_aux = (int *) realloc(data_struct->qty, needed * sizeof(int));
-  types_aux = (int *) realloc(data_struct->types, needed * sizeof(int));
+  qty_aux = (size_t *) realloc(data_struct->qty, needed * sizeof(int));
+  types_aux = (MPI_Datatype *) realloc(data_struct->types, needed * sizeof(MPI_Datatype));
+  request_qty_aux = (size_t *) realloc(data_struct->request_qty, needed * sizeof(int));
   requests_aux = (MPI_Request **) realloc(data_struct->requests, needed * sizeof(MPI_Request *));
+  windows_aux = (MPI_Win *) realloc(data_struct->windows, needed * sizeof(MPI_Win));
   arrays_aux = (void **) realloc(data_struct->arrays, needed * sizeof(void *));
 
-  if(qty_aux == NULL || arrays_aux == NULL || requests_aux == NULL || types_aux == NULL) {
+  if(qty_aux == NULL || arrays_aux == NULL || requests_aux == NULL || types_aux == NULL || request_qty_aux == NULL || windows_aux == NULL) {
     fprintf(stderr, "Fatal error - No se ha podido realojar la memoria constante de datos a redistribuir/comunicar\n");
     MPI_Abort(MPI_COMM_WORLD, 1);
   }
 
+  for(i=data_struct->max_entries; i<needed; i++) { //realloc does not ensure a NULL value
+    requests_aux[i] = NULL;
+    windows_aux[i] = MPI_WIN_NULL;
+    arrays_aux[i] = NULL;
+  }
+
+  // Check if old array can be freed
+  if(data_struct->qty != qty_aux && data_struct->qty != NULL) free(data_struct->qty);
+  if(data_struct->types != types_aux && data_struct->types != NULL) free(data_struct->types);
+  if(data_struct->request_qty != request_qty_aux && data_struct->request_qty != NULL) free(data_struct->request_qty);
+  if(data_struct->requests != requests_aux && data_struct->requests != NULL) free(data_struct->requests);
+  if(data_struct->windows != windows_aux && data_struct->windows != NULL) free(data_struct->windows);
+  if(data_struct->arrays != arrays_aux && data_struct->arrays != NULL) free(data_struct->arrays);
+
   data_struct->qty = qty_aux;
   data_struct->types = types_aux;
+  data_struct->request_qty = request_qty_aux;
   data_struct->requests = requests_aux;
+  data_struct->windows = windows_aux;
   data_struct->arrays = arrays_aux;
   data_struct->max_entries = needed;
 }
 
 void free_malleability_data_struct(malleability_data_t *data_struct) {
-  int i, max;
+  size_t i, j, max;
 
   max = data_struct->entries;
   if(max != 0) {
-    for(i=0; i<max; i++) {
-      //free(data_struct->arrays[i]); //FIXME Valores alojados con 1 elemento no se liberan?
-      //free(data_struct->requests[i]); //TODO Plantear como crearlos
+    if(data_struct->qty != NULL) {
+      free(data_struct->qty);
+    }
+    if(data_struct->types != NULL) {
+      free(data_struct->types);
+    }
+    if(data_struct->requests != NULL && data_struct->request_qty != NULL) {
+      for(i=0; i<max; i++) {
+        if(data_struct->requests[i] != NULL) {
+	  for(j=0; j<data_struct->request_qty[i]; j++) {
+	    if(data_struct->requests[i][j] != MPI_REQUEST_NULL) {
+              MPI_Request_free(&(data_struct->requests[i][j]));
+	      data_struct->requests[i][j] = MPI_REQUEST_NULL;
+	    }
+	  }
+          free(data_struct->requests[i]);
+	}
+      }
+      free(data_struct->request_qty);
+      free(data_struct->requests);  
+    }
+
+    if(data_struct->windows != NULL) {
+      free(data_struct->windows);
     }
 
-    free(data_struct->qty);
-    free(data_struct->types);
-    free(data_struct->requests);
-    free(data_struct->arrays);
+    if(data_struct->arrays != NULL) {
+      free(data_struct->arrays);
+    }
   }
 }
 
@@ -166,10 +253,11 @@ void def_malleability_entries(malleability_data_t *data_struct_rep, malleability
   int counts = 2;
   int blocklengths[counts];
   MPI_Aint displs[counts];
-  MPI_Datatype types[counts];
+  MPI_Datatype types[counts], type_size_t;
+  MPI_Type_match_size(MPI_TYPECLASS_INTEGER, sizeof(size_t), &type_size_t);
 
   blocklengths[0] = blocklengths[1] = 1;
-  types[0] = types[1] = MPI_INT;
+  types[0] = types[1] = type_size_t;
 
   // Obtener direccion base
   MPI_Get_address(&(data_struct_rep->entries), &displs[0]);
@@ -187,20 +275,27 @@ void def_malleability_entries(malleability_data_t *data_struct_rep, malleability
  * TODO Refactor?
  */
 void def_malleability_qty_type(malleability_data_t *data_struct_rep, malleability_data_t *data_struct_dist, MPI_Datatype *new_type) {
-  int counts = 4;
+  int counts = 6;
   int blocklengths[counts];
   MPI_Aint displs[counts];
-  MPI_Datatype types[counts];
+  MPI_Datatype types[counts], type_size_t;
+  MPI_Type_match_size(MPI_TYPECLASS_INTEGER, sizeof(size_t), &type_size_t);
 
-  types[0] = types[1] = types[2] = types[3] = MPI_INT;
-  blocklengths[0] = blocklengths[1] = data_struct_rep->entries;
-  blocklengths[2] = blocklengths[3] = data_struct_dist->entries;
+  types[0] = types[1] = types[3] = types[4] = type_size_t;
+  types[2] = types[5] = MPI_INT;
+  blocklengths[0] = blocklengths[1] = blocklengths[2] = data_struct_rep->entries;
+  blocklengths[3] = blocklengths[4] = blocklengths[5] = data_struct_dist->entries;
 
   MPI_Get_address((data_struct_rep->qty), &displs[0]);
-  MPI_Get_address((data_struct_rep->types), &displs[1]);
-  MPI_Get_address((data_struct_dist->qty), &displs[2]);
-  MPI_Get_address((data_struct_dist->types), &displs[3]);
+  MPI_Get_address((data_struct_rep->request_qty), &displs[1]);
+  MPI_Get_address((data_struct_rep->types), &displs[2]); // MPI_Datatype uses typedef int to be declared
+  MPI_Get_address((data_struct_dist->qty), &displs[3]);
+  MPI_Get_address((data_struct_dist->request_qty), &displs[4]);
+  MPI_Get_address((data_struct_dist->types), &displs[5]); // MPI_Datatype uses typedef int to be declared
 
   MPI_Type_create_struct(counts, blocklengths, displs, types, new_type);
   MPI_Type_commit(new_type);
 }
+
+
+

Codes/malleability/malleabilityTypes.h → Codes/MaM/MAM_Types.h

+#ifndef MAM_TYPES_H
+#define MAM_TYPES_H
+
 #include <stdio.h>
 #include <stdlib.h>
 #include <mpi.h>
 #include <fcntl.h>
 #include <sys/stat.h>
-#include "malleabilityStates.h"
+#include "MAM_Constants.h"
 
-#define MALLEABILITY_INIT_DATA_QTY 100
+#define MAM_TYPES_INIT_DATA_QTY 100
 
 typedef struct {
-  int entries; // Indica numero de vectores a comunicar (replicated data)
-  int max_entries;
-  MPI_Request request_ibarrier; // Request para indicar que los padres esperan a que los hijos terminen de recibir
-  int *qty; // Indica numero de elementos en cada subvector de sync_array
-  int *types;
+  size_t entries; // Indica numero de vectores a comunicar (replicated data)
+  size_t max_entries;
+  size_t *qty; // Indica numero de elementos en cada subvector de sync_array
+  MPI_Datatype *types;
 
   // Vector de vectores de request. En cada elemento superior se indican los requests a comprobar para dar por finalizada 
   // la comunicacion de ese dato
+  size_t *request_qty;
   MPI_Request **requests; 
+  MPI_Win *windows;
   void **arrays; // Cada subvector es una serie de datos a comunicar
 
 } malleability_data_t;
 
-void add_data(void *data, int total_qty, int type, int request_qty, malleability_data_t *data_struct);
-void comm_data_info(malleability_data_t *data_struct_rep, malleability_data_t *data_struct_dist, int is_children_group, int myId, int root, MPI_Comm intercomm);
+void add_data(void *data, size_t total_qty, MPI_Datatype type, size_t request_qty, malleability_data_t *data_struct);
+void modify_data(void *data, size_t index, size_t total_qty, MPI_Datatype type, size_t request_qty, malleability_data_t *data_struct);
+void comm_data_info(malleability_data_t *data_struct_rep, malleability_data_t *data_struct_dist, int is_children_group);
 void free_malleability_data_struct(malleability_data_t *data_struct);
+
+#endif

Codes/MaM/MAM_Zombies.c

+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <fcntl.h>
+#include <unistd.h>
+#include <mpi.h>
+#include <signal.h>
+#include "MAM_Zombies.h"
+#include "MAM_DataStructures.h"
+
+#define PIDS_QTY 320
+//TODO Add option to allow the usage of signal USR2 or not.
+//This code asumes ROOT of each group will be the last to be zombified
+//
+
+void MAM_I_zombies_collect(int new_zombies);
+void MAM_I_zombies_split();
+void MAM_I_zombies_suspend();
+int MAM_I_zombies_awake();
+void zombies_handler_usr2() {}
+
+int *pids = NULL;
+int zombies_qty = 0;
+
+
+void MAM_Zombies_service_init() {
+  zombies_qty = 0;
+  pids = malloc(PIDS_QTY * sizeof(int));
+
+  for(int i=0; i<PIDS_QTY; i++) {
+    pids[i] = 0;
+  }
+}
+
+int MAM_Zombies_service_free() {
+  int request_abort = MAM_I_zombies_awake();
+  free(pids);
+  return request_abort;
+}
+
+
+void MAM_Zombies_update() {
+  int myId, numP, new_zombies;
+
+  MPI_Comm_rank(mall->original_comm, &myId);
+  MPI_Comm_size(mall->original_comm, &numP);
+
+  MPI_Allreduce(&mall->zombie, &new_zombies, 1, MPI_INT, MPI_SUM, mall->original_comm);
+  if(new_zombies && new_zombies < numP) {
+    MAM_I_zombies_collect(new_zombies);
+    MAM_I_zombies_split();
+    MAM_I_zombies_suspend();
+    if(myId == MAM_ROOT) zombies_qty += new_zombies;
+  }
+}
+
+void MAM_I_zombies_collect(int new_zombies) {
+  int pid = getpid();
+  int *pids_counts, *pids_displs;
+  int i, count, active;
+  int myId, numP;
+
+  MPI_Comm_rank(mall->original_comm, &myId);
+  MPI_Comm_size(mall->original_comm, &numP);
+  pids_counts = (int *) malloc(numP * sizeof(int));
+  pids_displs = (int *) malloc(numP * sizeof(int));
+
+  #if MAM_DEBUG > 2
+    if(myId == MAM_ROOT){ DEBUG_FUNC("Collecting zombies", mall->myId, mall->numP); } fflush(stdout);
+  #endif
+
+  count = mall->zombie;
+  if(myId == MAM_ROOT) {
+    active = numP - new_zombies;
+    for(i=0; i < active; i++) {
+      pids_counts[i] = 0;
+    }
+    pids_displs[i-1] = -1;
+    for(; i< active+new_zombies; i++) {
+      pids_counts[i] = 1;
+      pids_displs[i] = (pids_displs[i-1] + 1) + zombies_qty;
+    }
+  }
+  MPI_Gatherv(&pid, count, MPI_INT, pids, pids_counts, pids_displs, MPI_INT, MAM_ROOT, mall->original_comm);
+  free(pids_counts);
+  free(pids_displs);
+}
+
+void MAM_I_zombies_split() {
+  int myId, color;
+  MPI_Comm new_original_comm;
+
+  MPI_Comm_rank(mall->original_comm, &myId);
+  color = mall->zombie ? MPI_UNDEFINED : 1;
+  MPI_Comm_split(mall->original_comm, color, myId, &new_original_comm);
+
+  if(mall->original_comm != MPI_COMM_WORLD) MPI_Comm_free(&mall->original_comm);
+  if(new_original_comm != MPI_COMM_NULL) MPI_Comm_set_name(new_original_comm, "MAM_ORIGINAL");
+  mall->original_comm = new_original_comm;
+}
+
+void MAM_I_zombies_suspend() {
+  struct sigaction act;
+  if(!mall->zombie) return;
+
+  sigemptyset(&act.sa_mask);
+  act.sa_flags=0;
+  act.sa_handler=zombies_handler_usr2;
+
+  sigaction(SIGUSR2, &act, NULL);
+
+  sigset_t set;
+  sigprocmask(SIG_SETMASK,NULL,&set);
+
+  sigsuspend(&set);
+}
+
+int MAM_I_zombies_awake() {
+  if(mall->internode_group && zombies_qty) return 1; //Request Abort
+  for(int i=0; i < zombies_qty; i++) { // Despertar a los zombies
+    kill(pids[i], SIGUSR2);
+  }
+  zombies_qty = 0;
+  return 0; //Normal termination
+}

Codes/MaM/MAM_Zombies.h

+#ifndef MAM_ZOMBIES_H
+#define MAM_ZOMBIES_H
+
+void MAM_Zombies_service_init();
+int MAM_Zombies_service_free();
+void MAM_Zombies_update();
+
+#endif

Codes/MaM/Makefile

+CC = gcc
+MCC = mpicc
+#C_FLAGS_ALL = -Wconversion -Wpedantic
+C_FLAGS = -Wall -Wextra -Wshadow -Wfatal-errors
+LD_FLAGS = -lm -pthread
+
+MAM_USE_SLURM ?= 0
+MAM_USE_BARRIERS ?= 0
+MAM_DEBUG ?= 0
+DEF = -DMAM_USE_SLURM=$(MAM_USE_SLURM) -DMAM_USE_BARRIERS=$(MAM_USE_BARRIERS) -DMAM_DEBUG=$(MAM_DEBUG)
+
+ifeq ($(MAM_USE_SLURM),1)
+    LD_FLAGS += -lslurm
+endif
+
+ifeq ($(shell test $(MAM_DEBUG) -gt 0; echo $$?),0)
+    C_FLAGS += -g
+endif
+
+# Final library
+LIB = libmam.so
+# Put all auto generated stuff to this build dir.
+BUILD_DIR = ./build
+
+# List of all directories where source files are located
+SRCDIRS = . spawn_methods distribution_methods
+
+# List of all .c source files.
+C_FILES = $(foreach dire, $(SRCDIRS), $(wildcard $(dire)/*.c))
+
+# All .o files go to build dir.
+OBJ = $(C_FILES:%.c=$(BUILD_DIR)/%.o)
+# Gcc will create these .d files containing dependencies.
+DEP = $(OBJ:%.o=%.d)
+
+# BASIC RULES
+.PHONY : clean clear install
+
+all: install
+
+clean:
+	-rm $(BUILD_DIR)/$(LIB) $(OBJ) $(DEP)
+clear:
+	-rm -rf $(BUILD_DIR)
+
+install: $(LIB)
+	echo "Done"
+
+# SPECIFIC RULES
+# Default target named after the binary.
+$(LIB) : $(BUILD_DIR)/$(LIB)
+
+# Actual target of the binary - depends on all .o files.
+$(BUILD_DIR)/$(LIB) : $(OBJ)
+	$(MCC) $(C_FLAGS) $^ -shared -o $@ $(LD_FLAGS)
+
+# Include all .d files
+# .d files are used for knowing the dependencies of each source file
+-include $(DEP)
+
+# Build target for every single object file.
+# The potential dependency on header files is covered
+# by calling `-include $(DEP)`.
+# The -MMD flags additionaly creates a .d file with
+# the same name as the .o file.
+$(BUILD_DIR)/%.o : %.c
+	@mkdir -p $(@D)
+	$(MCC) $(C_FLAGS) $(DEF) -fpic -MMD -c $< -o $@