diff --git a/CMakeLists-NonLibraryDefs.txt b/CMakeLists-NonLibraryDefs.txt
index 8128c6b..88f5387 100644
--- a/CMakeLists-NonLibraryDefs.txt
+++ b/CMakeLists-NonLibraryDefs.txt
@@ -70,6 +70,7 @@ option(WITH_CUDA " build dendro with cuda " OFF)
 option(BUILD_WITH_PETSC " build dendro with PETSC " OFF)
 option(USE_FD_INTERP_FOR_UNZIP "use FD style interpolation for unzip" OFF)
 option(MPIX_CUDA_AWARE_SUPPORT "use FD style interpolation for unzip" OFF)
+option(DVEC_ZERO_ALLOC "Make the memory allocation of dvec on CPU initialize as zeros" OFF)
 
 # option(KWAY "K parameter for alltoallv_kway" 128)
 set(KWAY
@@ -207,4 +208,8 @@ if(USE_FD_INTERP_FOR_UNZIP)
   add_definitions(-DUSE_FD_INTERP_FOR_UNZIP)
 endif()
 
+if(DVEC_ZERO_ALLOC)
+  add_definitions(-DDVEC_ZERO_ALLOC)
+endif()
+
 add_definitions(-DMATVEC_PROFILE)
diff --git a/ODE/include/enuts.h b/ODE/include/enuts.h
index 3a29384..e69835f 100644
--- a/ODE/include/enuts.h
+++ b/ODE/include/enuts.h
@@ -766,6 +766,13 @@ namespace ts
         
         m_uiBVec.clear();
 
+        // deallocate m_uiStVec, which is based on the number of stages
+        for (unsigned int i = 0; i < m_uiNumStages; i++) {
+            m_uiStVec[i].destroy_vector();
+        }
+        m_uiStVec.clear(); 
+        
+
         m_uiBlkTimeLevMinMax.clear();
         
         m_uiActiveBlkIDs.clear();
@@ -2883,4 +2890,4 @@ namespace ts
      */
     template<typename T>
     void sync_blk_padding(const ot::Mesh* pMesh, const T* const dgWVec,std::vector<ts::BlockTimeStep<T>>& bVec,  unsigned int blk, unsigned int bVecIndex,unsigned int dof=1);
-}
\ No newline at end of file
+}
diff --git a/include/dvec.h b/include/dvec.h
index d33bedf..694e303 100644
--- a/include/dvec.h
+++ b/include/dvec.h
@@ -186,8 +186,15 @@ namespace ot
             #ifdef __CUDACC__
                 m_data_ptr = GPUDevice::host_malloc<T>(m_size);
             #else 
+
+            #ifdef DVEC_ZERO_ALLOC
+                m_data_ptr = (T*) calloc(m_size, sizeof(T));
+            #else
                 m_data_ptr = (T*) malloc(sizeof(T)*m_size);
             #endif
+            // end DVEC_ZERO_ALLOC
+            #endif
+            // end __CUDACC__ if
             
         }else if(m_vec_loc == DVEC_LOC::DEVICE)
         {   
@@ -207,9 +214,6 @@ namespace ot
     template<typename T,typename I>
     void DVector<T,I>::set_vec_ptr(T*& ptr, const ot::Mesh* pMesh, DVEC_TYPE type, DVEC_LOC loc, unsigned int dof, bool allocate_ghost)
     {
-        if(!(pMesh->isActive()))
-            return;
-        
         m_dof       = dof;
         m_comm      = pMesh->getMPICommunicator();
         m_vec_type  = type;
@@ -231,6 +235,9 @@ namespace ot
             MPI_Abort(m_comm,0);
         }
 
+        if(!(pMesh->isActive()))
+            return;
+
         m_data_ptr = ptr;
         return;
     }