Fix MPI communicator leak

CHANGELOG.md

@@ -27,6 +27,7 @@
 - [[PR 663]](https://github.com/lanl/parthenon/pull/663) Change bvals_in_one to use sparse boundary buffers and add flux_correction in one.
 
 ### Fixed (not changing behavior/API/variables/...)
+- [[PR 841]](https://github.com/parthenon-hpc-lab/parthenon/pull/841) De-allocate MPI communicators when freeing `Reduction` objects
 - [[PR 843]](https://github.com/parthenon-hpc-lab/parthenon/pull/843) Add guard rails to prolongation/restriction infrastructure
 - [[PR 832]](https://github.com/parthenon-hpc-lab/parthenon/pull/833) Fix movie2d script after it broke due to change in HDF5 format
 - [[PR 820]](https://github.com/parthenon-hpc-lab/parthenon/pull/820) Fix XDMF spec to conform to standard and handle scalar and vector variables

src/utils/reductions.hpp

@@ -13,6 +13,7 @@
 #ifndef UTILS_REDUCTIONS_HPP_
 #define UTILS_REDUCTIONS_HPP_
 
+#include <memory>
 #include <vector>
 
 #include <config.hpp>
@@ -59,20 +60,24 @@ struct ReductionBase {
   T val;
 #ifdef MPI_PARALLEL
   MPI_Request req;
-  MPI_Comm comm;
+  std::shared_ptr<MPI_Comm> pcomm;
 #endif
   bool active = false;
   ReductionBase() {
 #ifdef MPI_PARALLEL
-    MPI_Comm_dup(MPI_COMM_WORLD, &comm);
+    // Store the communicator in a shared_ptr, so that
+    // MPI_Comm_free is called, but only when the last
+    // copy of this ReductionBase is destroyed.
+    pcomm = std::shared_ptr<MPI_Comm>(new MPI_Comm, MPI_Comm_free);
+    PARTHENON_MPI_CHECK(MPI_Comm_dup(MPI_COMM_WORLD, pcomm.get()));
 #endif
   }
 
   TaskStatus CheckReduce() {
     if (!active) return TaskStatus::complete;
     int check = 1;
 #ifdef MPI_PARALLEL
-    MPI_Test(&req, &check, MPI_STATUS_IGNORE);
+    PARTHENON_MPI_CHECK(MPI_Test(&req, &check, MPI_STATUS_IGNORE));
 #endif
     if (check) {
       active = false;
@@ -87,9 +92,10 @@ struct AllReduce : public ReductionBase<T> {
   TaskStatus StartReduce(MPI_Op op) {
     if (this->active) return TaskStatus::complete;
 #ifdef MPI_PARALLEL
-    MPI_Iallreduce(MPI_IN_PLACE, contiguous_container::data(this->val),
-                   contiguous_container::size(this->val), GetContainerMPIType(this->val),
-                   op, this->comm, &(this->req));
+    PARTHENON_MPI_CHECK(
+        MPI_Iallreduce(MPI_IN_PLACE, contiguous_container::data(this->val),
+                       contiguous_container::size(this->val),
+                       GetContainerMPIType(this->val), op, *(this->pcomm), &(this->req)));
 #endif
     this->active = true;
     return TaskStatus::complete;
@@ -102,14 +108,16 @@ struct Reduce : public ReductionBase<T> {
     if (this->active) return TaskStatus::complete;
 #ifdef MPI_PARALLEL
     if (Globals::my_rank == n) {
-      MPI_Ireduce(MPI_IN_PLACE, contiguous_container::data(this->val),
-                  contiguous_container::size(this->val), GetContainerMPIType(this->val),
-                  op, n, this->comm, &(this->req));
+      PARTHENON_MPI_CHECK(MPI_Ireduce(MPI_IN_PLACE, contiguous_container::data(this->val),
+                                      contiguous_container::size(this->val),
+                                      GetContainerMPIType(this->val), op, n,
+                                      *(this->pcomm), &(this->req)));
 
     } else {
-      MPI_Ireduce(contiguous_container::data(this->val), nullptr,
-                  contiguous_container::size(this->val), GetContainerMPIType(this->val),
-                  op, n, this->comm, &(this->req));
+      PARTHENON_MPI_CHECK(MPI_Ireduce(contiguous_container::data(this->val), nullptr,
+                                      contiguous_container::size(this->val),
+                                      GetContainerMPIType(this->val), op, n,
+                                      *(this->pcomm), &(this->req)));
     }
 #endif
     this->active = true;