queueTransfer optimization in processMatchedVertices

amgprec/impl/aggregator/MatchBoxPC.h

@@ -65,8 +65,8 @@
 
 using namespace std;
 
-#define NUM_THREAD 4
+const int NUM_THREAD = 2;
-#define UCHUNK 5
+const int UCHUNK = 50;
 
 const MilanLongInt REQUEST = 1;
 const MilanLongInt SUCCESS = 2;

amgprec/impl/aggregator/processMatchedVertices.cpp

@@ -199,10 +199,11 @@ void processMatchedVertices(
                                 Message[1] = w;       // GHOST
                                 Message[2] = REQUEST; // TYPE
                                                       // Send a Request (Asynchronous)
-                                                      //#pragma omp master
+
-                                                      //                                {
+                                printf("Send case 2: (%ld, %ld, %ld)\n", Message[0], Message[1], Message[2]);
+                                fflush(stdout);
                                 MPI_Bsend(&Message[0], 3, TypeMap<MilanLongInt>(), ghostOwner, ComputeTag, comm);
-//                                }
+
 #pragma omp atomic
                                 (*msgActual)++;
                             }
@@ -254,10 +255,10 @@ void processMatchedVertices(
                                         Message[1] = w;       // GHOST
                                         Message[2] = FAILURE; // TYPE
                                                               // Send a Request (Asynchronous)
-                                                              //#pragma omp master
+
-                                                              //                                        {
+                                        printf("Send case 4: (%ld, %ld, %ld)\n", Message[0], Message[1], Message[2]);
+                                        fflush(stdout);
                                         MPI_Bsend(&Message[0], 3, TypeMap<MilanLongInt>(), ghostOwner, ComputeTag, comm);
-//                                        }
 #pragma omp atomic
                                         (*msgActual)++;
                                     }
@@ -300,10 +301,9 @@ void processMatchedVertices(
                                 Message[2] = SUCCESS; // TYPE
 
                                 // Send a Request (Asynchronous)
-                                //#pragma omp master
+                                // printf("Send case 5: (%ld, %ld, %ld)\n", Message[0], Message[1], Message[2]);
-                                //                                {
+                                fflush(stdout);
                                 MPI_Bsend(&Message[0], 3, TypeMap<MilanLongInt>(), ghostOwner, ComputeTag, comm);
-//                                }
 #pragma omp atomic
                                 (*msgActual)++;
                             }
@@ -327,39 +327,17 @@ void processMatchedVertices(
                         } // End of switch
 
                     } // End of inner for
-
-                    // TODO privateU.size() < UCHUNK could be commented but it generate errors, why?
-                    if (privateU.size() > UCHUNK || U.empty())
-                    {
-#pragma omp critical(U)
-                        {
-                            while (!privateU.empty())
-                                U.push_back(privateU.pop_back());
-                        }
-
-#ifndef error
-#pragma omp critical(privateMsg)
-                        {
-                            while (!privateQLocalVtx.empty())
-                            {
-                                QLocalVtx.push_back(privateQLocalVtx.pop_back());
-                                QGhostVtx.push_back(privateQGhostVtx.pop_back());
-                                QMsgType.push_back(privateQMsgType.pop_back());
-                                QOwner.push_back(privateQOwner.pop_back());
-                            }
-                        }
-
-#endif
-                    } // End of private.size()
                 }
             } // End of outer for
-        }     // End of while ( !U.empty() )
+
-        queuesTransfer(U, privateU, QLocalVtx,
+            queuesTransfer(U, privateU, QLocalVtx,
-                       QGhostVtx,
+                           QGhostVtx,
-                       QMsgType, QOwner, privateQLocalVtx,
+                           QMsgType, QOwner, privateQLocalVtx,
-                       privateQGhostVtx,
+                           privateQGhostVtx,
-                       privateQMsgType,
+                           privateQMsgType,
-                       privateQOwner);
+                           privateQOwner);
+
+        } // End of while ( !U.empty() )
 
 #ifdef COUNT_LOCAL_VERTEX
         printf("Count local vertexes: %ld for thread %d of processor %d\n",

amgprec/impl/aggregator/processMessages.cpp

@@ -28,6 +28,8 @@ void processMessages(
     staticQueue &U)
 {
 
+//#define PRINT_DEBUG_INFO_
+
     MilanInt Sender;
     MPI_Status computeStatus;
     MilanLongInt bundleSize, w;
@@ -92,6 +94,8 @@ void processMessages(
 
     if (Message[2] == SIZEINFO)
     {
+        //printf("Inizio sizeinfo\n");
+        fflush(stdout);
 
 #ifdef PRINT_DEBUG_INFO_
         cout << "\n(" << myRank << ")Received bundled message from Process " << Sender << " Size= " << Message[0] << endl;
@@ -124,6 +128,9 @@ void processMessages(
         cout << endl;
         fflush(stdout);
 #endif
+
+        //printf("Fine sizeinfo\n");
+        fflush(stdout);
     }
     else
     { // Just a single message: