Bugfix around communicators. Most of them must be disconnected to ensure zombies can terminate.

Codes/malleability/malleabilityDataStructures.c

@@ -84,8 +84,8 @@ void MAM_print_comms_state() {
  * Función para modificar los comunicadores principales de MaM
  */
 void MAM_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->thread_comm != MPI_COMM_WORLD) MPI_Comm_disconnect(&(mall->thread_comm));
+  if(mall->comm != MPI_COMM_WORLD) MPI_Comm_disconnect(&(mall->comm));
 
   MPI_Comm_dup(comm, &(mall->thread_comm));
   MPI_Comm_dup(comm, &(mall->comm));

Codes/malleability/malleabilityDataStructures.h

@@ -18,6 +18,7 @@ typedef struct {
   double spawn_start, spawn_time;
   double sync_start,  sync_end;
   double async_start, async_end;
+  double user_start, user_end;
   double malleability_start, malleability_end;
 
   MPI_Datatype times_type;

Codes/malleability/spawn_methods/Baseline.c

@@ -116,7 +116,7 @@ void multiple_strat_parents(Spawn_data spawn_data, MPI_Comm comm, MPI_Comm *inte
 
   MPI_Comm_connect(port_name, MPI_INFO_NULL, mall->root, comm, child);
   for(i=0; i<spawn_data.total_spawns; i++) {
-    MPI_Comm_free(&intercomms[i]);
+    MPI_Comm_disconnect(&intercomms[i]);
   }
   free(port_name);
 }
@@ -153,7 +153,7 @@ void multiple_strat_children(MPI_Comm *parents) {
     MPI_Comm_connect(port_name, MPI_INFO_NULL, mall->root, mall->comm, &intercomm);
     MPI_Bcast(&total_spawns, 1, MPI_INT, mall->root, intercomm); // FIXME Seems inneficient - Should be performed by parent root?
     MPI_Intercomm_merge(intercomm, 1, &newintracomm); // Get last ranks
-    MPI_Comm_free(&intercomm);
+    MPI_Comm_disconnect(&intercomm);
   } else { 
     start = 1; 
     MPI_Comm_dup(mall->comm, &newintracomm);
@@ -163,9 +163,9 @@ void multiple_strat_children(MPI_Comm *parents) {
   for(i=start; i<total_spawns; i++) {
     MPI_Comm_accept(port_name, MPI_INFO_NULL, mall->root, newintracomm, &intercomm);
     MPI_Bcast(&total_spawns, 1, MPI_INT, rootBcast, intercomm); // FIXME Seems inneficient - Should be performed by parent root?
-    if(newintracomm != MPI_COMM_WORLD) MPI_Comm_free(&newintracomm);
+    if(newintracomm != MPI_COMM_WORLD) MPI_Comm_disconnect(&newintracomm);
     MPI_Intercomm_merge(intercomm, 0, &newintracomm); // Get first ranks
-    MPI_Comm_free(&intercomm);
+    MPI_Comm_disconnect(&intercomm);
 
     if(new_root) {
       MPI_Send(&aux, 1, MPI_CHAR, stat.MPI_SOURCE, stat.MPI_TAG, parents_comm); // Ensures order in the created intracommunicator
@@ -181,8 +181,8 @@ void multiple_strat_children(MPI_Comm *parents) {
 
   if(new_root) MPI_Close_port(port_name);
   free(port_name);
-  MPI_Comm_free(&newintracomm);
-  MPI_Comm_free(parents);
+  MPI_Comm_disconnect(&newintracomm);
+  MPI_Comm_disconnect(parents);
   *parents = intercomm;
 }
 
@@ -207,7 +207,7 @@ void single_strat_parents(Spawn_data spawn_data, MPI_Comm *child) {
   MPI_Comm_connect(port_name, MPI_INFO_NULL, mall->root, spawn_data.comm, &newintercomm);
 
   if(mall->myId == mall->root)
-    MPI_Comm_free(child);
+    MPI_Comm_disconnect(child);
   free(port_name);
   *child = newintercomm;
 }
@@ -237,6 +237,6 @@ void single_strat_children(MPI_Comm *parents) {
     MPI_Close_port(port_name);
   }
   free(port_name);
-  MPI_Comm_free(parents);
+  MPI_Comm_disconnect(parents);
   *parents = newintercomm;
 }

Codes/malleability/spawn_methods/Merge.c

@@ -57,7 +57,7 @@ void merge_adapt_expand(MPI_Comm *child, int is_children_group) {
 
   MPI_Intercomm_merge(*child, is_children_group, &new_comm); //El que pone 0 va primero
 
-  MPI_Comm_free(child);
+  MPI_Comm_disconnect(child);
   *child = new_comm;
 }
 
@@ -72,6 +72,7 @@ void merge_adapt_expand(MPI_Comm *child, int is_children_group) {
 void merge_adapt_shrink(int numC, MPI_Comm *child, MPI_Comm comm, int myId) {
   int color = MPI_UNDEFINED;
 
+  if(*child != MPI_COMM_NULL && *child != MPI_COMM_WORLD) MPI_Comm_disconnect(child);
   if(myId < numC) {
       color = 1;  
   }