diff --git a/amgprec/impl/aggregator/MatchBoxPC.h b/amgprec/impl/aggregator/MatchBoxPC.h
index dfcb6f7e..0b3dcd74 100644
--- a/amgprec/impl/aggregator/MatchBoxPC.h
+++ b/amgprec/impl/aggregator/MatchBoxPC.h
@@ -234,7 +234,8 @@ extern "C"
                     staticQueue &privateQMsgType,
                     staticQueue &privateQOwner);
 
-    void clean(MilanInt myRank,
+    void clean(MilanLongInt NLVer,
+               MilanInt myRank,
                MilanLongInt MessageIndex,
                vector<MPI_Request> &SRequest,
                vector<MPI_Status> &SStatus,
@@ -245,7 +246,8 @@ extern "C"
                MilanLongInt msgInd,
                MilanLongInt *msgIndSent,
                MilanLongInt NumMessagesBundled,
-               MilanReal *msgPercent);
+               MilanReal *msgPercent,
+               omp_lock_t *MateLock);
 
     void PARALLEL_COMPUTE_CANDIDATE_MATE_B(MilanLongInt NLVer,
                                            MilanLongInt *verLocPtr,
diff --git a/amgprec/impl/aggregator/algoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP.cpp b/amgprec/impl/aggregator/algoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP.cpp
index 50930601..e45ee792 100644
--- a/amgprec/impl/aggregator/algoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP.cpp
+++ b/amgprec/impl/aggregator/algoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP.cpp
@@ -789,7 +789,8 @@ void dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP(
 #endif
     } // End of while (true)
 
-    clean(myRank,
+    clean(NLVer,
+          myRank,
           MessageIndex,
           SRequest,
           SStatus,
@@ -800,7 +801,8 @@ void dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP(
           msgInd,
           msgIndSent,
           NumMessagesBundled,
-          msgPercent);
+          msgPercent,
+          MateLock);
 
     finishTime = MPI_Wtime();
     *ph2_time = finishTime - startTime; // Time taken for Phase-2
diff --git a/amgprec/impl/aggregator/clean.cpp b/amgprec/impl/aggregator/clean.cpp
index 5204894a..d91076c9 100644
--- a/amgprec/impl/aggregator/clean.cpp
+++ b/amgprec/impl/aggregator/clean.cpp
@@ -1,79 +1,98 @@
 #include "MatchBoxPC.h"
 
-//TODO comment
-//TODO use task
-//TODO destroy the locks
+// TODO comment
+// TODO use task
+// TODO destroy the locks
 
-void clean(MilanInt myRank,
-                  MilanLongInt MessageIndex,
-                  vector<MPI_Request> &SRequest,
-                  vector<MPI_Status> &SStatus,
-                  MilanInt BufferSize,
-                  MilanLongInt *Buffer,
-                  MilanLongInt msgActual,
-                  MilanLongInt *msgActualSent,
-                  MilanLongInt msgInd,
-                  MilanLongInt *msgIndSent,
-                  MilanLongInt NumMessagesBundled,
-                  MilanReal *msgPercent)
+void clean(MilanLongInt NLVer,
+           MilanInt myRank,
+           MilanLongInt MessageIndex,
+           vector<MPI_Request> &SRequest,
+           vector<MPI_Status> &SStatus,
+           MilanInt BufferSize,
+           MilanLongInt *Buffer,
+           MilanLongInt msgActual,
+           MilanLongInt *msgActualSent,
+           MilanLongInt msgInd,
+           MilanLongInt *msgIndSent,
+           MilanLongInt NumMessagesBundled,
+           MilanReal *msgPercent,
+           omp_lock_t *MateLock)
 {
     // Cleanup Phase
 
+#pragma omp parallel
+    {
+#pragma omp master
+        {
+#pragma omp task
+            {
+
 #ifdef PRINT_DEBUG_INFO_
-    cout << "\n(" << myRank << ") Waitall= " << endl;
-    fflush(stdout);
+                cout << "\n(" << myRank << ") Waitall= " << endl;
+                fflush(stdout);
 #endif
 #ifdef DEBUG_HANG_
-    cout << "\n(" << myRank << ") Waitall " << endl;
-    fflush(stdout);
+                cout << "\n(" << myRank << ") Waitall " << endl;
+                fflush(stdout);
 #endif
-    return;
+                return;
 
-    MPI_Waitall(MessageIndex, &SRequest[0], &SStatus[0]);
+                MPI_Waitall(MessageIndex, &SRequest[0], &SStatus[0]);
 
-    // MPI_Buffer_attach(&Buffer, BufferSize); //Attach the Buffer
-    if (BufferSize > 0)
-    {
-        MPI_Buffer_detach(&Buffer, &BufferSize); // Detach the Buffer
-        free(Buffer);                            // Free the memory that was allocated
-    }
+                // MPI_Buffer_attach(&Buffer, BufferSize); //Attach the Buffer
+                if (BufferSize > 0)
+                {
+                    MPI_Buffer_detach(&Buffer, &BufferSize); // Detach the Buffer
+                    free(Buffer);                            // Free the memory that was allocated
+                }
+            }
 
 #ifdef PRINT_DEBUG_INFO_
-    cout << "\n(" << myRank << ")End of function to compute matching: " << endl;
-    fflush(stdout);
-    cout << "\n(" << myRank << ")myCardinality: " << myCard << endl;
-    fflush(stdout);
-    cout << "\n(" << myRank << ")Matching took " << finishTime - startTime << "seconds" << endl;
-    fflush(stdout);
-    cout << "\n(" << myRank << ")** Getting out of the matching function **" << endl;
-    fflush(stdout);
+            cout << "\n(" << myRank << ")End of function to compute matching: " << endl;
+            fflush(stdout);
+            cout << "\n(" << myRank << ")myCardinality: " << myCard << endl;
+            fflush(stdout);
+            cout << "\n(" << myRank << ")Matching took " << finishTime - startTime << "seconds" << endl;
+            fflush(stdout);
+            cout << "\n(" << myRank << ")** Getting out of the matching function **" << endl;
+            fflush(stdout);
 #endif
 #ifdef PRINT_DEBUG_INFO_
-    cout << "\n(" << myRank << ") Number of Ghost edges = " << numGhostEdges;
-    cout << "\n(" << myRank << ") Total number of potential message X 2 = " << numGhostEdges * 2;
-    cout << "\n(" << myRank << ") Number messages bundled = " << NumMessagesBundled;
-    cout << "\n(" << myRank << ") Total Individual Messages sent = " << msgInd;
-    if (msgInd > 0)
-    {
-        cout << "\n(" << myRank << ") Percentage of messages bundled = " << ((double)NumMessagesBundled / (double)(msgInd)) * 100.0 << "% \n";
-    }
-    fflush(stdout);
+            cout << "\n(" << myRank << ") Number of Ghost edges = " << numGhostEdges;
+            cout << "\n(" << myRank << ") Total number of potential message X 2 = " << numGhostEdges * 2;
+            cout << "\n(" << myRank << ") Number messages bundled = " << NumMessagesBundled;
+            cout << "\n(" << myRank << ") Total Individual Messages sent = " << msgInd;
+            if (msgInd > 0)
+            {
+                cout << "\n(" << myRank << ") Percentage of messages bundled = " << ((double)NumMessagesBundled / (double)(msgInd)) * 100.0 << "% \n";
+            }
+            fflush(stdout);
 #endif
 
-    *msgActualSent = msgActual;
-    *msgIndSent = msgInd;
-    if (msgInd > 0)
-    {
-        *msgPercent = ((double)NumMessagesBundled / (double)(msgInd)) * 100.0;
-    }
-    else
-    {
-        *msgPercent = 0;
-    }
+#pragma omp task
+            {
+                *msgActualSent = msgActual;
+                *msgIndSent = msgInd;
+                if (msgInd > 0)
+                {
+                    *msgPercent = ((double)NumMessagesBundled / (double)(msgInd)) * 100.0;
+                }
+                else
+                {
+                    *msgPercent = 0;
+                }
+            }
+            // Destroy the locks
+#pragma omp taskloop num_tasks(NUM_THREAD)
+            for (int i = 0; i < NLVer; i++)
+                omp_destroy_lock(&MateLock[i]);
 
 #ifdef DEBUG_HANG_
-    if (myRank == 0)
-        cout << "\n(" << myRank << ") Done" << endl;
-    fflush(stdout);
+            if (myRank == 0)
+                cout << "\n(" << myRank << ") Done" << endl;
+            fflush(stdout);
 #endif
+        }
+    }
 }
\ No newline at end of file