Hotfix -- No se recogian correctamente las iteraciones en Merge Shrink

Codes/Main/Main.c

@@ -71,6 +71,7 @@ int main(int argc, char *argv[]) {
     }
 
     init_group_struct(argv, argc, myId, numP);
+    //FIXME No funciona en OpenMPI
     im_child = init_malleability(myId, numP, ROOT, comm, argv[0], nodelist, num_cpus, num_nodes);
 
     if(!im_child) { //TODO REFACTOR Simplificar inicio
@@ -249,7 +250,6 @@ void iterate(double *matrix, int n, int async_comm, int iter) {
 
   operations = time / Top; //FIXME Calcular una sola vez
   
-  
   for(i=0; i < operations; i++) {
     aux += computePiSerial(n);
   }

Codes/malleability/ProcessDist.c

@@ -157,7 +157,7 @@ int check_slurm_comm(int myId, int root, int numP, MPI_Comm *child, MPI_Comm com
       int state=-10;
 
       //printf("[%d][3] Test min\n", myId); fflush(stdout);
-      //pthread_mutex_lock(&spawn_mutex);
+      //pthread_mutex_lock(&spawn_mutex); // TODO Descomentar
       MPI_Allreduce(&commState, &state, 1, MPI_INT, MPI_MIN, comm);
       //pthread_mutex_unlock(&spawn_mutex);
 
@@ -173,7 +173,7 @@ int check_slurm_comm(int myId, int root, int numP, MPI_Comm *child, MPI_Comm com
 
     } else if (slurm_data->spawn_is_single) {
 
-      //pthread_mutex_lock(&spawn_mutex);
+      //pthread_mutex_lock(&spawn_mutex); // TODO Descomentar
       MPI_Bcast(&commState, 1, MPI_INT, root, comm);
       //pthread_mutex_unlock(&spawn_mutex);
       int threads_not_spawned = pthread_equal(pthread_self(), spawn_thread);

Codes/malleability/malleabilityManager.c

@@ -170,15 +170,20 @@ int malleability_checkpoint() {
 
   } else if(state == MAL_SPAWN_PENDING || state == MAL_SPAWN_SINGLE_PENDING) { // Comprueba si el spawn ha terminado y comienza la redistribucion
     double end_real_time;
-    state = check_slurm_comm(mall->myId, mall->root, mall->numP, &(mall->intercomm), mall->comm, mall->thread_comm, &end_real_time);
-    if (state == MAL_SPAWN_COMPLETED) {  
-      mall_conf->results->spawn_time[mall_conf->grp] = MPI_Wtime() - mall_conf->results->spawn_start;
-      if(mall_conf->spawn_type == COMM_SPAWN_PTHREAD || mall_conf->spawn_type == COMM_SPAWN_MERGE_PTHREAD) {
-        mall_conf->results->spawn_real_time[mall_conf->grp] = end_real_time - mall_conf->results->spawn_start;
-      }
 
-      //TODO Si es MERGE SHRINK, metodo diferente de redistribucion de datos
-      state = start_redistribution();
+    if(mall_conf->spawn_type == COMM_SPAWN_MERGE_PTHREAD && mall->numP > mall->numC) {
+      state = shrink_redistribution(); //TODO REFACTOR
+
+    } else {
+      state = check_slurm_comm(mall->myId, mall->root, mall->numP, &(mall->intercomm), mall->comm, mall->thread_comm, &end_real_time);
+      if (state == MAL_SPAWN_COMPLETED) {  
+        mall_conf->results->spawn_time[mall_conf->grp] = MPI_Wtime() - mall_conf->results->spawn_start;
+        if(mall_conf->spawn_type == COMM_SPAWN_PTHREAD || mall_conf->spawn_type == COMM_SPAWN_MERGE_PTHREAD) {
+          mall_conf->results->spawn_real_time[mall_conf->grp] = end_real_time - mall_conf->results->spawn_start;
+        }
+        //TODO Si es MERGE SHRINK, metodo diferente de redistribucion de datos
+        state = start_redistribution();
+      }
     }
 
   } else if(state == MAL_DIST_PENDING) {
@@ -641,16 +646,71 @@ int end_redistribution() {
   return result;
 }
 
+
+///=============================================
+///=============================================
+///=============================================
+double time_adapt;
+int state_shrink=0; //TODO Refactor
+pthread_t thread_shrink;
+MPI_Comm comm_shrink;
+
+int thread_shrink_creation();
+void *thread_shrink_work();
+/*
+ * Crea una hebra para ejecutar una comunicación en segundo plano.
+ */
+int thread_shrink_creation() {
+  if(pthread_create(&thread_shrink, NULL, thread_shrink_work, NULL)) {
+    printf("Error al crear el hilo\n");
+    MPI_Abort(MPI_COMM_WORLD, -1);
+    return -1;
+  }
+  return MAL_SPAWN_PENDING;
+}
+void* thread_shrink_work() {
+  proc_adapt_shrink(mall->numC, &comm_shrink, mall->myId);
+  state_shrink=2;
+  pthread_exit(NULL);
+}
+///=============================================
+///=============================================
+///=============================================
 int shrink_redistribution() {
-    double time_adapt = MPI_Wtime();
+    int global_state;
     MPI_Comm aux_comm;
-    MPI_Comm_dup(mall->comm, &aux_comm);
 
-    proc_adapt_shrink( mall->numC, &(mall->comm), mall->myId);
+    if(mall_conf->spawn_type == COMM_SPAWN_MERGE_PTHREAD) {
+      if(state_shrink == 0) {
+        time_adapt = MPI_Wtime();
+	state_shrink = 1;
+        MPI_Comm_dup(mall->comm, &comm_shrink);
+        thread_shrink_creation();
+	return MAL_SPAWN_PENDING;
+      } else if(state_shrink>0) {
+        MPI_Allreduce(&state_shrink, &global_state, 1, MPI_INT, MPI_MIN, mall->comm);
+
+	if(global_state < 2) return MAL_SPAWN_PENDING;
+        if(pthread_join(thread_shrink, NULL)) { 
+          printf("Error al esperar al hilo\n");
+          MPI_Abort(MPI_COMM_WORLD, -1);
+          return -10;
+        }
+        MPI_Comm_dup(mall->comm, &aux_comm);
+	mall->comm = comm_shrink;
+      }
+
+    } else {
+      time_adapt = MPI_Wtime();
+      MPI_Comm_dup(mall->comm, &aux_comm);
+      proc_adapt_shrink( mall->numC, &(mall->comm), mall->myId);
+    }
+
+    //TODO REFACTOR -- Que solo la llamada de collect iters este fuera de los hilos
     zombies_collect_suspended(aux_comm, mall->myId, mall->numP, mall->numC, mall->root, (void *) mall_conf->results, mall->user_comm);
-    MPI_Comm_free(&aux_comm);
     
     if(mall->myId < mall->numC) {
+      MPI_Comm_free(&aux_comm);
       MPI_Comm_dup(mall->comm, &aux_comm);
       mall->thread_comm = aux_comm;
       MPI_Comm_dup(mall->comm, &aux_comm);

Codes/malleability/malleabilityStates.h

@@ -19,14 +19,18 @@
 #define COMM_PHY_NODES 1
 #define COMM_PHY_CPU 2
 
-// TODO Separar PTHREAD
+// SPAWN METHODS
 #define COMM_SPAWN_SERIAL 0
 #define COMM_SPAWN_PTHREAD 1
 #define COMM_SPAWN_MERGE 2
 #define COMM_SPAWN_MERGE_PTHREAD 3
-
+//#define COMM_SPAWN_BASELINE 0
+//#define COMM_SPAWN_MERGE 1
+//SPAWN STRATEGIES
 #define COMM_SPAWN_MULTIPLE 0
 #define COMM_SPAWN_SINGLE 1
+//#define COMM_SPAWN_SERIAL 0
+//#define COMM_SPAWN_PTHREAD 1
 
 #define MAL_USE_NORMAL 0
 #define MAL_USE_IBARRIER 1

Codes/recordMachinefile.sh

@@ -15,3 +15,4 @@ elif [ $dist == "cpu" ]; then
 fi
 
 $dir/Recordnodelist.o $numP $dist
+echo $numP

Codes/runBase.sh

 #!/bin/bash
 
+#SBATCH -p P1
 #SBATCH -N 1
 #SBATCH --exclude=c01,c00,c02
 
@@ -17,6 +18,7 @@ module load mpich-3.4.1-noucx
 numP=$(bash recordMachinefile.sh $1)
 
 mpirun -print-all-exitcodes -f hostfile.o$SLURM_JOB_ID $dir$codeDir/a.out $1 $2 $nodelist $nodes
+#mpirun -np $numP $dir$codeDir/a.out $1 $2 $nodelist $nodes
 rm hostfile.o$SLURM_JOB_ID
 
 echo "END RUN"