Apply async Sync optimization to Z4c_class using Sync_start/finish pattern

Replaces blocking Parallel::Sync + MPI_Allreduce in Z4c_class Step() with
non-blocking MPI_Iallreduce overlapped with Sync_start/Sync_finish, matching
the pattern already used in bssn_class on cjy-oneapi-opus-hotfix. Covers both ABEtype==2
and CPBC variants (predictor + corrector = 4 call sites).

Cherry-picked optimization from afd4006, adapted to SyncCache
infrastructure instead of the separate SyncPlan API.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

AMSS_NCKU_source/MPatch.C

@@ -341,8 +341,9 @@ void Patch::Interp_Points(MyList<var> *VarList,
                           double *Shellf, int Symmetry)
 {
   // NOTE: we do not Synchnize variables here, make sure of that before calling this routine
-  int myrank;
+  int myrank, nprocs;
   MPI_Comm_rank(MPI_COMM_WORLD, &myrank);
+  MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
 
   int ordn = 2 * ghost_width;
   MyList<var> *varl;
@@ -354,24 +355,18 @@ void Patch::Interp_Points(MyList<var> *VarList,
     varl = varl->next;
   }
 
-  double *shellf;
+  memset(Shellf, 0, sizeof(double) * NN * num_var);
-  shellf = new double[NN * num_var];
-  memset(shellf, 0, sizeof(double) * NN * num_var);
 
-  // we use weight to monitor code, later some day we can move it for optimization
+  // owner_rank[j] records which MPI rank owns point j
-  int *weight;
+  // All ranks traverse the same block list so they all agree on ownership
-  weight = new int[NN];
+  int *owner_rank;
-  memset(weight, 0, sizeof(int) * NN);
+  owner_rank = new int[NN];
-
+  for (int j = 0; j < NN; j++)
-  double *DH, *llb, *uub;
+    owner_rank[j] = -1;
-  DH = new double[dim];
 
+  double DH[dim], llb[dim], uub[dim];
   for (int i = 0; i < dim; i++)
-  {
     DH[i] = getdX(i);
-  }
-  llb = new double[dim];
-  uub = new double[dim];
 
   for (int j = 0; j < NN; j++) // run along points
   {
@@ -403,12 +398,6 @@ void Patch::Interp_Points(MyList<var> *VarList,
       bool flag = true;
       for (int i = 0; i < dim; i++)
       {
-// NOTE: our dividing structure is (exclude ghost)
-// -1 0
-//       1  2
-// so (0,1) does not belong to any part for vertex structure
-// here we put (0,0.5) to left part and (0.5,1) to right part
-// BUT for cell structure the bbox is (-1.5,0.5) and (0.5,2.5), there is no missing region at all
 #ifdef Vertex
 #ifdef Cell
 #error Both Cell and Vertex are defined
@@ -433,6 +422,7 @@ void Patch::Interp_Points(MyList<var> *VarList,
       if (flag)
       {
         notfind = false;
+        owner_rank[j] = BP->rank;
         if (myrank == BP->rank)
         {
           //---> interpolation
@@ -440,14 +430,11 @@ void Patch::Interp_Points(MyList<var> *VarList,
           int k = 0;
           while (varl) // run along variables
           {
-            //              shellf[j*num_var+k] = Parallel::global_interp(dim,BP->shape,BP->X,BP->fgfs[varl->data->sgfn],
+            f_global_interp(BP->shape, BP->X[0], BP->X[1], BP->X[2], BP->fgfs[varl->data->sgfn], Shellf[j * num_var + k],
-            //	  		                                    pox,ordn,varl->data->SoA,Symmetry);
-            f_global_interp(BP->shape, BP->X[0], BP->X[1], BP->X[2], BP->fgfs[varl->data->sgfn], shellf[j * num_var + k],
                             pox[0], pox[1], pox[2], ordn, varl->data->SoA, Symmetry);
             varl = varl->next;
             k++;
           }
-          weight[j] = 1;
         }
       }
       if (Bp == ble)
@@ -456,61 +443,125 @@ void Patch::Interp_Points(MyList<var> *VarList,
     }
   }
 
-  MPI_Allreduce(shellf, Shellf, NN * num_var, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+  // Replace MPI_Allreduce with per-owner MPI_Bcast:
-  int *Weight;
+  // Group consecutive points by owner rank and broadcast each group.
-  Weight = new int[NN];
+  // Since each point's data is non-zero only on the owner rank,
-  MPI_Allreduce(weight, Weight, NN, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
+  // Bcast from owner is equivalent to Allreduce(MPI_SUM) but much cheaper.
-
-  //  misc::tillherecheck("print me");
-
-  for (int i = 0; i < NN; i++)
   {
-    if (Weight[i] > 1)
+    int j = 0;
+    while (j < NN)
+    {
+      int cur_owner = owner_rank[j];
+      if (cur_owner < 0)
       {
         if (myrank == 0)
-        cout << "WARNING: Patch::Interp_Points meets multiple weight" << endl;
-      for (int j = 0; j < num_var; j++)
-        Shellf[j + i * num_var] = Shellf[j + i * num_var] / Weight[i];
-    }
-    else if (Weight[i] == 0 && myrank == 0)
         {
           cout << "ERROR: Patch::Interp_Points fails to find point (";
-      for (int j = 0; j < dim; j++)
+          for (int d = 0; d < dim; d++)
           {
-        cout << XX[j][i];
+            cout << XX[d][j];
-        if (j < dim - 1)
+            if (d < dim - 1)
               cout << ",";
             else
               cout << ")";
           }
           cout << " on Patch (";
-      for (int j = 0; j < dim; j++)
+          for (int d = 0; d < dim; d++)
           {
-        cout << bbox[j] << "+" << lli[j] * getdX(j);
+            cout << bbox[d] << "+" << lli[d] * DH[d];
-        if (j < dim - 1)
+            if (d < dim - 1)
               cout << ",";
             else
               cout << ")--";
           }
           cout << "(";
-      for (int j = 0; j < dim; j++)
+          for (int d = 0; d < dim; d++)
           {
-        cout << bbox[dim + j] << "-" << uui[j] * getdX(j);
+            cout << bbox[dim + d] << "-" << uui[d] * DH[d];
-        if (j < dim - 1)
+            if (d < dim - 1)
               cout << ",";
             else
               cout << ")" << endl;
           }
-#if 0
+          MPI_Abort(MPI_COMM_WORLD, 1);
-       checkBlock();
+        }
-#else
+        j++;
-      cout << "splited domains:" << endl;
+        continue;
+      }
+      // Find contiguous run of points with the same owner
+      int jstart = j;
+      while (j < NN && owner_rank[j] == cur_owner)
+        j++;
+      int count = (j - jstart) * num_var;
+      MPI_Bcast(Shellf + jstart * num_var, count, MPI_DOUBLE, cur_owner, MPI_COMM_WORLD);
+    }
+  }
+
+  delete[] owner_rank;
+}
+void Patch::Interp_Points(MyList<var> *VarList,
+                          int NN, double **XX,
+                          double *Shellf, int Symmetry,
+                          int Nmin_consumer, int Nmax_consumer)
+{
+  // Targeted point-to-point overload: each owner sends each point only to
+  // the one rank that needs it for integration (consumer), reducing
+  // communication volume by ~nprocs times compared to the Bcast version.
+  int myrank, nprocs;
+  MPI_Comm_rank(MPI_COMM_WORLD, &myrank);
+  MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
+
+  int ordn = 2 * ghost_width;
+  MyList<var> *varl;
+  int num_var = 0;
+  varl = VarList;
+  while (varl)
   {
+    num_var++;
+    varl = varl->next;
+  }
+
+  memset(Shellf, 0, sizeof(double) * NN * num_var);
+
+  // owner_rank[j] records which MPI rank owns point j
+  int *owner_rank;
+  owner_rank = new int[NN];
+  for (int j = 0; j < NN; j++)
+    owner_rank[j] = -1;
+
+  double DH[dim], llb[dim], uub[dim];
+  for (int i = 0; i < dim; i++)
+    DH[i] = getdX(i);
+
+  // --- Interpolation phase (identical to original) ---
+  for (int j = 0; j < NN; j++)
+  {
+    double pox[dim];
+    for (int i = 0; i < dim; i++)
+    {
+      pox[i] = XX[i][j];
+      if (myrank == 0 && (XX[i][j] < bbox[i] + lli[i] * DH[i] || XX[i][j] > bbox[dim + i] - uui[i] * DH[i]))
+      {
+        cout << "Patch::Interp_Points: point (";
+        for (int k = 0; k < dim; k++)
+        {
+          cout << XX[k][j];
+          if (k < dim - 1)
+            cout << ",";
+          else
+            cout << ") is out of current Patch." << endl;
+        }
+        MPI_Abort(MPI_COMM_WORLD, 1);
+      }
+    }
+
     MyList<Block> *Bp = blb;
-        while (Bp)
+    bool notfind = true;
+    while (notfind && Bp)
     {
       Block *BP = Bp->data;
 
+      bool flag = true;
       for (int i = 0; i < dim; i++)
       {
 #ifdef Vertex
@@ -527,32 +578,192 @@ void Patch::Interp_Points(MyList<var> *VarList,
 #error Not define Vertex nor Cell
 #endif
 #endif
-          }
+        if (XX[i][j] - llb[i] < -DH[i] / 2 || XX[i][j] - uub[i] > DH[i] / 2)
-          cout << "(";
-          for (int j = 0; j < dim; j++)
         {
-            cout << llb[j] << ":" << uub[j];
+          flag = false;
-            if (j < dim - 1)
+          break;
-              cout << ",";
+        }
-            else
+      }
-              cout << ")" << endl;
+
+      if (flag)
+      {
+        notfind = false;
+        owner_rank[j] = BP->rank;
+        if (myrank == BP->rank)
+        {
+          varl = VarList;
+          int k = 0;
+          while (varl)
+          {
+            f_global_interp(BP->shape, BP->X[0], BP->X[1], BP->X[2], BP->fgfs[varl->data->sgfn], Shellf[j * num_var + k],
+                            pox[0], pox[1], pox[2], ordn, varl->data->SoA, Symmetry);
+            varl = varl->next;
+            k++;
+          }
+        }
       }
       if (Bp == ble)
         break;
       Bp = Bp->next;
     }
   }
-#endif
+
+  // --- Error check for unfound points ---
+  for (int j = 0; j < NN; j++)
+  {
+    if (owner_rank[j] < 0 && myrank == 0)
+    {
+      cout << "ERROR: Patch::Interp_Points fails to find point (";
+      for (int d = 0; d < dim; d++)
+      {
+        cout << XX[d][j];
+        if (d < dim - 1)
+          cout << ",";
+        else
+          cout << ")";
+      }
+      cout << " on Patch (";
+      for (int d = 0; d < dim; d++)
+      {
+        cout << bbox[d] << "+" << lli[d] * DH[d];
+        if (d < dim - 1)
+          cout << ",";
+        else
+          cout << ")--";
+      }
+      cout << "(";
+      for (int d = 0; d < dim; d++)
+      {
+        cout << bbox[dim + d] << "-" << uui[d] * DH[d];
+        if (d < dim - 1)
+          cout << ",";
+        else
+          cout << ")" << endl;
+      }
       MPI_Abort(MPI_COMM_WORLD, 1);
     }
   }
 
-  delete[] shellf;
+  // --- Targeted point-to-point communication phase ---
-  delete[] weight;
+  // Compute consumer_rank[j] using the same deterministic formula as surface_integral
-  delete[] Weight;
+  int *consumer_rank = new int[NN];
-  delete[] DH;
+  {
-  delete[] llb;
+    int mp = NN / nprocs;
-  delete[] uub;
+    int Lp = NN - nprocs * mp;
+    for (int j = 0; j < NN; j++)
+    {
+      if (j < Lp * (mp + 1))
+        consumer_rank[j] = j / (mp + 1);
+      else
+        consumer_rank[j] = Lp + (j - Lp * (mp + 1)) / mp;
+    }
+  }
+
+  // Count sends and recvs per rank
+  int *send_count = new int[nprocs];
+  int *recv_count = new int[nprocs];
+  memset(send_count, 0, sizeof(int) * nprocs);
+  memset(recv_count, 0, sizeof(int) * nprocs);
+
+  for (int j = 0; j < NN; j++)
+  {
+    int own = owner_rank[j];
+    int con = consumer_rank[j];
+    if (own == con)
+      continue; // local — no communication needed
+    if (own == myrank)
+      send_count[con]++;
+    if (con == myrank)
+      recv_count[own]++;
+  }
+
+  // Build send buffers: for each destination rank, pack (index, data) pairs
+  // Each entry: 1 int (point index j) + num_var doubles
+  int total_send = 0, total_recv = 0;
+  int *send_offset = new int[nprocs];
+  int *recv_offset = new int[nprocs];
+  for (int r = 0; r < nprocs; r++)
+  {
+    send_offset[r] = total_send;
+    total_send += send_count[r];
+    recv_offset[r] = total_recv;
+    total_recv += recv_count[r];
+  }
+
+  // Pack send buffers: each message contains (j, data[0..num_var-1]) per point
+  int stride = 1 + num_var; // 1 double for index + num_var doubles for data
+  double *sendbuf = new double[total_send * stride];
+  double *recvbuf = new double[total_recv * stride];
+
+  // Temporary counters for packing
+  int *pack_pos = new int[nprocs];
+  memset(pack_pos, 0, sizeof(int) * nprocs);
+
+  for (int j = 0; j < NN; j++)
+  {
+    int own = owner_rank[j];
+    int con = consumer_rank[j];
+    if (own != myrank || con == myrank)
+      continue;
+    int pos = (send_offset[con] + pack_pos[con]) * stride;
+    sendbuf[pos] = (double)j; // point index
+    for (int v = 0; v < num_var; v++)
+      sendbuf[pos + 1 + v] = Shellf[j * num_var + v];
+    pack_pos[con]++;
+  }
+
+  // Post non-blocking recvs and sends
+  int n_req = 0;
+  for (int r = 0; r < nprocs; r++)
+  {
+    if (recv_count[r] > 0) n_req++;
+    if (send_count[r] > 0) n_req++;
+  }
+
+  MPI_Request *reqs = new MPI_Request[n_req];
+  int req_idx = 0;
+
+  for (int r = 0; r < nprocs; r++)
+  {
+    if (recv_count[r] > 0)
+    {
+      MPI_Irecv(recvbuf + recv_offset[r] * stride,
+                recv_count[r] * stride, MPI_DOUBLE,
+                r, 0, MPI_COMM_WORLD, &reqs[req_idx++]);
+    }
+  }
+  for (int r = 0; r < nprocs; r++)
+  {
+    if (send_count[r] > 0)
+    {
+      MPI_Isend(sendbuf + send_offset[r] * stride,
+                send_count[r] * stride, MPI_DOUBLE,
+                r, 0, MPI_COMM_WORLD, &reqs[req_idx++]);
+    }
+  }
+
+  if (n_req > 0)
+    MPI_Waitall(n_req, reqs, MPI_STATUSES_IGNORE);
+
+  // Unpack recv buffers into Shellf
+  for (int i = 0; i < total_recv; i++)
+  {
+    int pos = i * stride;
+    int j = (int)recvbuf[pos];
+    for (int v = 0; v < num_var; v++)
+      Shellf[j * num_var + v] = recvbuf[pos + 1 + v];
+  }
+
+  delete[] reqs;
+  delete[] sendbuf;
+  delete[] recvbuf;
+  delete[] pack_pos;
+  delete[] send_offset;
+  delete[] recv_offset;
+  delete[] send_count;
+  delete[] recv_count;
+  delete[] consumer_rank;
+  delete[] owner_rank;
 }
 void Patch::Interp_Points(MyList<var> *VarList,
                           int NN, double **XX,
@@ -573,24 +784,22 @@ void Patch::Interp_Points(MyList<var> *VarList,
     varl = varl->next;
   }
 
-  double *shellf;
+  memset(Shellf, 0, sizeof(double) * NN * num_var);
-  shellf = new double[NN * num_var];
-  memset(shellf, 0, sizeof(double) * NN * num_var);
 
-  // we use weight to monitor code, later some day we can move it for optimization
+  // owner_rank[j] stores the global rank that owns point j
-  int *weight;
+  int *owner_rank;
-  weight = new int[NN];
+  owner_rank = new int[NN];
-  memset(weight, 0, sizeof(int) * NN);
+  for (int j = 0; j < NN; j++)
+    owner_rank[j] = -1;
 
-  double *DH, *llb, *uub;
+  // Build global-to-local rank translation for Comm_here
-  DH = new double[dim];
+  MPI_Group world_group, local_group;
+  MPI_Comm_group(MPI_COMM_WORLD, &world_group);
+  MPI_Comm_group(Comm_here, &local_group);
 
+  double DH[dim], llb[dim], uub[dim];
   for (int i = 0; i < dim; i++)
-  {
     DH[i] = getdX(i);
-  }
-  llb = new double[dim];
-  uub = new double[dim];
 
   for (int j = 0; j < NN; j++) // run along points
   {
@@ -622,12 +831,6 @@ void Patch::Interp_Points(MyList<var> *VarList,
       bool flag = true;
       for (int i = 0; i < dim; i++)
       {
-// NOTE: our dividing structure is (exclude ghost)
-// -1 0
-//       1  2
-// so (0,1) does not belong to any part for vertex structure
-// here we put (0,0.5) to left part and (0.5,1) to right part
-// BUT for cell structure the bbox is (-1.5,0.5) and (0.5,2.5), there is no missing region at all
 #ifdef Vertex
 #ifdef Cell
 #error Both Cell and Vertex are defined
@@ -652,6 +855,7 @@ void Patch::Interp_Points(MyList<var> *VarList,
       if (flag)
       {
         notfind = false;
+        owner_rank[j] = BP->rank;
         if (myrank == BP->rank)
         {
           //---> interpolation
@@ -659,14 +863,11 @@ void Patch::Interp_Points(MyList<var> *VarList,
           int k = 0;
           while (varl) // run along variables
           {
-            //              shellf[j*num_var+k] = Parallel::global_interp(dim,BP->shape,BP->X,BP->fgfs[varl->data->sgfn],
+            f_global_interp(BP->shape, BP->X[0], BP->X[1], BP->X[2], BP->fgfs[varl->data->sgfn], Shellf[j * num_var + k],
-            //	  		                                    pox,ordn,varl->data->SoA,Symmetry);
-            f_global_interp(BP->shape, BP->X[0], BP->X[1], BP->X[2], BP->fgfs[varl->data->sgfn], shellf[j * num_var + k],
                             pox[0], pox[1], pox[2], ordn, varl->data->SoA, Symmetry);
             varl = varl->next;
             k++;
           }
-          weight[j] = 1;
         }
       }
       if (Bp == ble)
@@ -675,97 +876,35 @@ void Patch::Interp_Points(MyList<var> *VarList,
     }
   }
 
-  MPI_Allreduce(shellf, Shellf, NN * num_var, MPI_DOUBLE, MPI_SUM, Comm_here);
+  // Collect unique global owner ranks and translate to local ranks in Comm_here
-  int *Weight;
+  // Then broadcast each owner's points via MPI_Bcast on Comm_here
-  Weight = new int[NN];
+  {
-  MPI_Allreduce(weight, Weight, NN, MPI_INT, MPI_SUM, Comm_here);
+    int j = 0;
+    while (j < NN)
+    {
+      int cur_owner_global = owner_rank[j];
+      if (cur_owner_global < 0)
+      {
+        // Point not found — skip (error check disabled for sub-communicator levels)
+        j++;
+        continue;
+      }
+      // Translate global rank to local rank in Comm_here
+      int cur_owner_local;
+      MPI_Group_translate_ranks(world_group, 1, &cur_owner_global, local_group, &cur_owner_local);
 
-  //  misc::tillherecheck("print me");
+      // Find contiguous run of points with the same owner
-  //  if(lmyrank == 0) cout<<"myrank = "<<myrank<<"print me"<<endl;
+      int jstart = j;
-
+      while (j < NN && owner_rank[j] == cur_owner_global)
-  for (int i = 0; i < NN; i++)
+        j++;
-  {
+      int count = (j - jstart) * num_var;
-    if (Weight[i] > 1)
+      MPI_Bcast(Shellf + jstart * num_var, count, MPI_DOUBLE, cur_owner_local, Comm_here);
-    {
-      if (lmyrank == 0)
-        cout << "WARNING: Patch::Interp_Points meets multiple weight" << endl;
-      for (int j = 0; j < num_var; j++)
-        Shellf[j + i * num_var] = Shellf[j + i * num_var] / Weight[i];
     }
-#if 0 // for not involved levels, this may fail     
-     else if(Weight[i] == 0 && lmyrank == 0)
-     {
-       cout<<"ERROR: Patch::Interp_Points fails to find point (";
-       for(int j=0;j<dim;j++)
-       {
-	  cout<<XX[j][i];
-	  if(j<dim-1) cout<<",";
-	  else        cout<<")";
-       }
-       cout<<" on Patch (";
-       for(int j=0;j<dim;j++)
-       {
-	  cout<<bbox[j]<<"+"<<lli[j]*getdX(j);
-	  if(j<dim-1) cout<<",";
-	  else        cout<<")--";
-       }
-       cout<<"(";
-       for(int j=0;j<dim;j++)
-       {
-	  cout<<bbox[dim+j]<<"-"<<uui[j]*getdX(j);
-	  if(j<dim-1) cout<<",";
-	  else        cout<<")"<<endl;
-       }
-#if 0
-       checkBlock();
-#else
-  cout<<"splited domains:"<<endl;
-  {
-     MyList<Block> *Bp=blb;
-     while(Bp)
-     {
-	Block *BP=Bp->data;
-
-	for(int i=0;i<dim;i++)
-	{
-#ifdef Vertex
-#ifdef Cell
-#error Both Cell and Vertex are defined
-#endif    
-          llb[i] = (feq(BP->bbox[i]    ,bbox[i]    ,DH[i]/2)) ? BP->bbox[i]+lli[i]*DH[i]     : BP->bbox[i]    +(ghost_width-0.5)*DH[i];
-          uub[i] = (feq(BP->bbox[dim+i],bbox[dim+i],DH[i]/2)) ? BP->bbox[dim+i]-uui[i]*DH[i] : BP->bbox[dim+i]-(ghost_width-0.5)*DH[i];
-#else
-#ifdef Cell
-          llb[i] = (feq(BP->bbox[i]    ,bbox[i]    ,DH[i]/2)) ? BP->bbox[i]+lli[i]*DH[i]     : BP->bbox[i]    +ghost_width*DH[i];
-          uub[i] = (feq(BP->bbox[dim+i],bbox[dim+i],DH[i]/2)) ? BP->bbox[dim+i]-uui[i]*DH[i] : BP->bbox[dim+i]-ghost_width*DH[i];
-#else
-#error Not define Vertex nor Cell
-#endif
-#endif 
-	}       
-       cout<<"(";
-       for(int j=0;j<dim;j++)
-       {
-	  cout<<llb[j]<<":"<<uub[j];
-	  if(j<dim-1) cout<<",";
-	  else        cout<<")"<<endl;
-       }
-	if(Bp == ble) break;
-	Bp=Bp->next;
-     }
-  }
-#endif       
-       MPI_Abort(MPI_COMM_WORLD,1);
-     }
-#endif
   }
 
-  delete[] shellf;
+  MPI_Group_free(&world_group);
-  delete[] weight;
+  MPI_Group_free(&local_group);
-  delete[] Weight;
+  delete[] owner_rank;
-  delete[] DH;
-  delete[] llb;
-  delete[] uub;
 }
 void Patch::checkBlock()
 {

AMSS_NCKU_source/MPatch.h

@@ -39,6 +39,10 @@ public:
 
    bool Find_Point(double *XX);
 
+   void Interp_Points(MyList<var> *VarList,
+                      int NN, double **XX,
+                      double *Shellf, int Symmetry,
+                      int Nmin_consumer, int Nmax_consumer);
    void Interp_Points(MyList<var> *VarList,
                       int NN, double **XX,
                       double *Shellf, int Symmetry, MPI_Comm Comm_here);

AMSS_NCKU_source/Parallel.C

@@ -3756,357 +3756,501 @@ void Parallel::Sync(MyList<Patch> *PatL, MyList<var> *VarList, int Symmetry)
   delete[] transfer_src;
   delete[] transfer_dst;
 }
-//
+// Merged Sync: collect all intra-patch and inter-patch grid segment lists,
-// Async Sync: split into SyncBegin (initiate MPI) and SyncEnd (wait + unpack)
+// then issue a single transfer() call instead of N+1 separate ones.
-// This allows overlapping MPI communication with computation.
+void Parallel::Sync_merged(MyList<Patch> *PatL, MyList<var> *VarList, int Symmetry)
-//
-static void transfer_begin(Parallel::TransferState *ts)
 {
-  int myrank;
+  int cpusize;
-  MPI_Comm_rank(MPI_COMM_WORLD, &myrank);
+  MPI_Comm_size(MPI_COMM_WORLD, &cpusize);
-  int cpusize = ts->cpusize;
 
-  ts->reqs = new MPI_Request[2 * cpusize];
+  MyList<Parallel::gridseg> **combined_src = new MyList<Parallel::gridseg> *[cpusize];
-  ts->stats = new MPI_Status[2 * cpusize];
+  MyList<Parallel::gridseg> **combined_dst = new MyList<Parallel::gridseg> *[cpusize];
-  ts->req_no = 0;
+  for (int node = 0; node < cpusize; node++)
-  ts->send_data = new double *[cpusize];
+    combined_src[node] = combined_dst[node] = 0;
-  ts->rec_data = new double *[cpusize];
+
-  int length;
+  // Phase A: Intra-patch ghost exchange segments
+  MyList<Patch> *Pp = PatL;
+  while (Pp)
+  {
+    Patch *Pat = Pp->data;
+    MyList<Parallel::gridseg> *dst_ghost = build_ghost_gsl(Pat);
 
     for (int node = 0; node < cpusize; node++)
     {
-    ts->send_data[node] = ts->rec_data[node] = 0;
+      MyList<Parallel::gridseg> *src_owned = build_owned_gsl0(Pat, node);
+      MyList<Parallel::gridseg> *tsrc = 0, *tdst = 0;
+      build_gstl(src_owned, dst_ghost, &tsrc, &tdst);
+
+      if (tsrc)
+      {
+        if (combined_src[node])
+          combined_src[node]->catList(tsrc);
+        else
+          combined_src[node] = tsrc;
+      }
+      if (tdst)
+      {
+        if (combined_dst[node])
+          combined_dst[node]->catList(tdst);
+        else
+          combined_dst[node] = tdst;
+      }
+
+      if (src_owned)
+        src_owned->destroyList();
+    }
+
+    if (dst_ghost)
+      dst_ghost->destroyList();
+
+    Pp = Pp->next;
+  }
+
+  // Phase B: Inter-patch buffer exchange segments
+  MyList<Parallel::gridseg> *dst_buffer = build_buffer_gsl(PatL);
+  for (int node = 0; node < cpusize; node++)
+  {
+    MyList<Parallel::gridseg> *src_owned = build_owned_gsl(PatL, node, 5, Symmetry);
+    MyList<Parallel::gridseg> *tsrc = 0, *tdst = 0;
+    build_gstl(src_owned, dst_buffer, &tsrc, &tdst);
+
+    if (tsrc)
+    {
+      if (combined_src[node])
+        combined_src[node]->catList(tsrc);
+      else
+        combined_src[node] = tsrc;
+    }
+    if (tdst)
+    {
+      if (combined_dst[node])
+        combined_dst[node]->catList(tdst);
+      else
+        combined_dst[node] = tdst;
+    }
+
+    if (src_owned)
+      src_owned->destroyList();
+  }
+  if (dst_buffer)
+    dst_buffer->destroyList();
+
+  // Phase C: Single transfer
+  transfer(combined_src, combined_dst, VarList, VarList, Symmetry);
+
+  // Phase D: Cleanup
+  for (int node = 0; node < cpusize; node++)
+  {
+    if (combined_src[node])
+      combined_src[node]->destroyList();
+    if (combined_dst[node])
+      combined_dst[node]->destroyList();
+  }
+  delete[] combined_src;
+  delete[] combined_dst;
+}
+// SyncCache constructor
+Parallel::SyncCache::SyncCache()
+    : valid(false), cpusize(0), combined_src(0), combined_dst(0),
+      send_lengths(0), recv_lengths(0), send_bufs(0), recv_bufs(0),
+      send_buf_caps(0), recv_buf_caps(0), reqs(0), stats(0), max_reqs(0),
+      lengths_valid(false)
+{
+}
+// SyncCache invalidate: free grid segment lists but keep buffers
+void Parallel::SyncCache::invalidate()
+{
+  if (!valid)
+    return;
+  for (int i = 0; i < cpusize; i++)
+  {
+    if (combined_src[i])
+      combined_src[i]->destroyList();
+    if (combined_dst[i])
+      combined_dst[i]->destroyList();
+    combined_src[i] = combined_dst[i] = 0;
+    send_lengths[i] = recv_lengths[i] = 0;
+  }
+  valid = false;
+  lengths_valid = false;
+}
+// SyncCache destroy: free everything
+void Parallel::SyncCache::destroy()
+{
+  invalidate();
+  if (combined_src) delete[] combined_src;
+  if (combined_dst) delete[] combined_dst;
+  if (send_lengths) delete[] send_lengths;
+  if (recv_lengths) delete[] recv_lengths;
+  if (send_buf_caps) delete[] send_buf_caps;
+  if (recv_buf_caps) delete[] recv_buf_caps;
+  for (int i = 0; i < cpusize; i++)
+  {
+    if (send_bufs && send_bufs[i]) delete[] send_bufs[i];
+    if (recv_bufs && recv_bufs[i]) delete[] recv_bufs[i];
+  }
+  if (send_bufs) delete[] send_bufs;
+  if (recv_bufs) delete[] recv_bufs;
+  if (reqs) delete[] reqs;
+  if (stats) delete[] stats;
+  combined_src = combined_dst = 0;
+  send_lengths = recv_lengths = 0;
+  send_buf_caps = recv_buf_caps = 0;
+  send_bufs = recv_bufs = 0;
+  reqs = 0; stats = 0;
+  cpusize = 0; max_reqs = 0;
+}
+// transfer_cached: reuse pre-allocated buffers from SyncCache
+void Parallel::transfer_cached(MyList<Parallel::gridseg> **src, MyList<Parallel::gridseg> **dst,
+                               MyList<var> *VarList1, MyList<var> *VarList2,
+                               int Symmetry, SyncCache &cache)
+{
+  int myrank;
+  MPI_Comm_size(MPI_COMM_WORLD, &cache.cpusize);
+  MPI_Comm_rank(MPI_COMM_WORLD, &myrank);
+  int cpusize = cache.cpusize;
+
+  int req_no = 0;
+  int node;
+
+  for (node = 0; node < cpusize; node++)
+  {
     if (node == myrank)
     {
-      // Local copy: pack then immediately unpack (no MPI needed)
+      int length = data_packer(0, src[myrank], dst[myrank], node, PACK, VarList1, VarList2, Symmetry);
-      if ((length = Parallel::data_packer(0, ts->transfer_src[myrank], ts->transfer_dst[myrank],
+      cache.recv_lengths[node] = length;
-                                          node, PACK, ts->VarList1, ts->VarList2, ts->Symmetry)))
+      if (length > 0)
       {
-        double *local_data = new double[length];
+        if (length > cache.recv_buf_caps[node])
-        if (!local_data)
         {
-          cout << "out of memory in transfer_begin, local copy" << endl;
+          if (cache.recv_bufs[node]) delete[] cache.recv_bufs[node];
-          MPI_Abort(MPI_COMM_WORLD, 1);
+          cache.recv_bufs[node] = new double[length];
+          cache.recv_buf_caps[node] = length;
         }
-        Parallel::data_packer(local_data, ts->transfer_src[myrank], ts->transfer_dst[myrank],
+        data_packer(cache.recv_bufs[node], src[myrank], dst[myrank], node, PACK, VarList1, VarList2, Symmetry);
-                              node, PACK, ts->VarList1, ts->VarList2, ts->Symmetry);
-        Parallel::data_packer(local_data, ts->transfer_src[node], ts->transfer_dst[node],
-                              node, UNPACK, ts->VarList1, ts->VarList2, ts->Symmetry);
-        delete[] local_data;
       }
     }
     else
     {
-      // send from this cpu to cpu#node
+      // send
-      if ((length = Parallel::data_packer(0, ts->transfer_src[myrank], ts->transfer_dst[myrank],
+      int slength = data_packer(0, src[myrank], dst[myrank], node, PACK, VarList1, VarList2, Symmetry);
-                                          node, PACK, ts->VarList1, ts->VarList2, ts->Symmetry)))
+      cache.send_lengths[node] = slength;
+      if (slength > 0)
       {
-        ts->send_data[node] = new double[length];
+        if (slength > cache.send_buf_caps[node])
-        if (!ts->send_data[node])
         {
-          cout << "out of memory in transfer_begin, send" << endl;
+          if (cache.send_bufs[node]) delete[] cache.send_bufs[node];
-          MPI_Abort(MPI_COMM_WORLD, 1);
+          cache.send_bufs[node] = new double[slength];
+          cache.send_buf_caps[node] = slength;
         }
-        Parallel::data_packer(ts->send_data[node], ts->transfer_src[myrank], ts->transfer_dst[myrank],
+        data_packer(cache.send_bufs[node], src[myrank], dst[myrank], node, PACK, VarList1, VarList2, Symmetry);
-                              node, PACK, ts->VarList1, ts->VarList2, ts->Symmetry);
+        MPI_Isend((void *)cache.send_bufs[node], slength, MPI_DOUBLE, node, 1, MPI_COMM_WORLD, cache.reqs + req_no++);
-        MPI_Isend((void *)ts->send_data[node], length, MPI_DOUBLE, node, 1, MPI_COMM_WORLD,
-                  ts->reqs + ts->req_no++);
       }
-      // receive from cpu#node to this cpu
+      // recv
-      if ((length = Parallel::data_packer(0, ts->transfer_src[node], ts->transfer_dst[node],
+      int rlength = data_packer(0, src[node], dst[node], node, UNPACK, VarList1, VarList2, Symmetry);
-                                          node, UNPACK, ts->VarList1, ts->VarList2, ts->Symmetry)))
+      cache.recv_lengths[node] = rlength;
+      if (rlength > 0)
       {
-        ts->rec_data[node] = new double[length];
+        if (rlength > cache.recv_buf_caps[node])
-        if (!ts->rec_data[node])
         {
-          cout << "out of memory in transfer_begin, recv" << endl;
+          if (cache.recv_bufs[node]) delete[] cache.recv_bufs[node];
-          MPI_Abort(MPI_COMM_WORLD, 1);
+          cache.recv_bufs[node] = new double[rlength];
+          cache.recv_buf_caps[node] = rlength;
         }
-        MPI_Irecv((void *)ts->rec_data[node], length, MPI_DOUBLE, node, 1, MPI_COMM_WORLD,
+        MPI_Irecv((void *)cache.recv_bufs[node], rlength, MPI_DOUBLE, node, 1, MPI_COMM_WORLD, cache.reqs + req_no++);
-                  ts->reqs + ts->req_no++);
       }
     }
   }
-  // NOTE: MPI_Waitall is NOT called here - that happens in transfer_end
-}
-//
-static void transfer_end(Parallel::TransferState *ts)
-{
-  // Wait for all pending MPI operations
-  MPI_Waitall(ts->req_no, ts->reqs, ts->stats);
 
-  // Unpack received data from remote ranks
+  MPI_Waitall(req_no, cache.reqs, cache.stats);
-  for (int node = 0; node < ts->cpusize; node++)
-    if (ts->rec_data[node])
-      Parallel::data_packer(ts->rec_data[node], ts->transfer_src[node], ts->transfer_dst[node],
-                            node, UNPACK, ts->VarList1, ts->VarList2, ts->Symmetry);
 
-  // Cleanup MPI buffers
+  for (node = 0; node < cpusize; node++)
-  for (int node = 0; node < ts->cpusize; node++)
+    if (cache.recv_bufs[node] && cache.recv_lengths[node] > 0)
-  {
+      data_packer(cache.recv_bufs[node], src[node], dst[node], node, UNPACK, VarList1, VarList2, Symmetry);
-    if (ts->send_data[node])
-      delete[] ts->send_data[node];
-    if (ts->rec_data[node])
-      delete[] ts->rec_data[node];
-  }
-  delete[] ts->reqs;
-  delete[] ts->stats;
-  delete[] ts->send_data;
-  delete[] ts->rec_data;
 }
-//
+// Sync_cached: build grid segment lists on first call, reuse on subsequent calls
-Parallel::SyncHandle *Parallel::SyncBegin(Patch *Pat, MyList<var> *VarList, int Symmetry)
+void Parallel::Sync_cached(MyList<Patch> *PatL, MyList<var> *VarList, int Symmetry, SyncCache &cache)
 {
+  if (!cache.valid)
+  {
     int cpusize;
     MPI_Comm_size(MPI_COMM_WORLD, &cpusize);
+    cache.cpusize = cpusize;
 
-  SyncHandle *handle = new SyncHandle;
+    // Allocate cache arrays if needed
-  handle->num_states = 1;
+    if (!cache.combined_src)
-  handle->states = new TransferState[1];
+    {
+      cache.combined_src = new MyList<Parallel::gridseg> *[cpusize];
+      cache.combined_dst = new MyList<Parallel::gridseg> *[cpusize];
+      cache.send_lengths = new int[cpusize];
+      cache.recv_lengths = new int[cpusize];
+      cache.send_bufs = new double *[cpusize];
+      cache.recv_bufs = new double *[cpusize];
+      cache.send_buf_caps = new int[cpusize];
+      cache.recv_buf_caps = new int[cpusize];
+      for (int i = 0; i < cpusize; i++)
+      {
+        cache.send_bufs[i] = cache.recv_bufs[i] = 0;
+        cache.send_buf_caps[i] = cache.recv_buf_caps[i] = 0;
+      }
+      cache.max_reqs = 2 * cpusize;
+      cache.reqs = new MPI_Request[cache.max_reqs];
+      cache.stats = new MPI_Status[cache.max_reqs];
+    }
 
-  TransferState *ts = &handle->states[0];
-  ts->cpusize = cpusize;
-  ts->VarList1 = VarList;
-  ts->VarList2 = VarList;
-  ts->Symmetry = Symmetry;
-  ts->owns_gsl = true;
-
-  ts->dst = build_ghost_gsl(Pat);
-  ts->src = new MyList<Parallel::gridseg> *[cpusize];
-  ts->transfer_src = new MyList<Parallel::gridseg> *[cpusize];
-  ts->transfer_dst = new MyList<Parallel::gridseg> *[cpusize];
     for (int node = 0; node < cpusize; node++)
     {
-    ts->src[node] = build_owned_gsl0(Pat, node);
+      cache.combined_src[node] = cache.combined_dst[node] = 0;
-    build_gstl(ts->src[node], ts->dst, &ts->transfer_src[node], &ts->transfer_dst[node]);
+      cache.send_lengths[node] = cache.recv_lengths[node] = 0;
     }
 
-  transfer_begin(ts);
+    // Build intra-patch segments (same as Sync_merged Phase A)
-
-  return handle;
-}
-//
-Parallel::SyncHandle *Parallel::SyncBegin(MyList<Patch> *PatL, MyList<var> *VarList, int Symmetry)
-{
-  int cpusize;
-  MPI_Comm_size(MPI_COMM_WORLD, &cpusize);
-
-  // Count patches
-  int num_patches = 0;
     MyList<Patch> *Pp = PatL;
-  while (Pp) { num_patches++; Pp = Pp->next; }
-
-  SyncHandle *handle = new SyncHandle;
-  handle->num_states = num_patches + 1; // intra-patch transfers + 1 inter-patch transfer
-  handle->states = new TransferState[handle->num_states];
-
-  // Intra-patch sync: for each patch, build ghost lists and initiate transfer
-  int idx = 0;
-  Pp = PatL;
     while (Pp)
     {
-    TransferState *ts = &handle->states[idx];
+      Patch *Pat = Pp->data;
-    ts->cpusize = cpusize;
+      MyList<Parallel::gridseg> *dst_ghost = build_ghost_gsl(Pat);
-    ts->VarList1 = VarList;
-    ts->VarList2 = VarList;
-    ts->Symmetry = Symmetry;
-    ts->owns_gsl = true;
-
-    ts->dst = build_ghost_gsl(Pp->data);
-    ts->src = new MyList<Parallel::gridseg> *[cpusize];
-    ts->transfer_src = new MyList<Parallel::gridseg> *[cpusize];
-    ts->transfer_dst = new MyList<Parallel::gridseg> *[cpusize];
       for (int node = 0; node < cpusize; node++)
       {
-      ts->src[node] = build_owned_gsl0(Pp->data, node);
+        MyList<Parallel::gridseg> *src_owned = build_owned_gsl0(Pat, node);
-      build_gstl(ts->src[node], ts->dst, &ts->transfer_src[node], &ts->transfer_dst[node]);
+        MyList<Parallel::gridseg> *tsrc = 0, *tdst = 0;
+        build_gstl(src_owned, dst_ghost, &tsrc, &tdst);
+        if (tsrc)
+        {
+          if (cache.combined_src[node])
+            cache.combined_src[node]->catList(tsrc);
+          else
+            cache.combined_src[node] = tsrc;
         }
-
+        if (tdst)
-    transfer_begin(ts);
+        {
-
+          if (cache.combined_dst[node])
-    idx++;
+            cache.combined_dst[node]->catList(tdst);
+          else
+            cache.combined_dst[node] = tdst;
+        }
+        if (src_owned) src_owned->destroyList();
+      }
+      if (dst_ghost) dst_ghost->destroyList();
       Pp = Pp->next;
     }
 
-  // Inter-patch sync: buffer zone exchange between patches
+    // Build inter-patch segments (same as Sync_merged Phase B)
-  {
+    MyList<Parallel::gridseg> *dst_buffer = build_buffer_gsl(PatL);
-    TransferState *ts = &handle->states[idx];
-    ts->cpusize = cpusize;
-    ts->VarList1 = VarList;
-    ts->VarList2 = VarList;
-    ts->Symmetry = Symmetry;
-    ts->owns_gsl = true;
-
-    ts->dst = build_buffer_gsl(PatL);
-    ts->src = new MyList<Parallel::gridseg> *[cpusize];
-    ts->transfer_src = new MyList<Parallel::gridseg> *[cpusize];
-    ts->transfer_dst = new MyList<Parallel::gridseg> *[cpusize];
     for (int node = 0; node < cpusize; node++)
     {
-      ts->src[node] = build_owned_gsl(PatL, node, 5, Symmetry);
+      MyList<Parallel::gridseg> *src_owned = build_owned_gsl(PatL, node, 5, Symmetry);
-      build_gstl(ts->src[node], ts->dst, &ts->transfer_src[node], &ts->transfer_dst[node]);
+      MyList<Parallel::gridseg> *tsrc = 0, *tdst = 0;
+      build_gstl(src_owned, dst_buffer, &tsrc, &tdst);
+      if (tsrc)
+      {
+        if (cache.combined_src[node])
+          cache.combined_src[node]->catList(tsrc);
+        else
+          cache.combined_src[node] = tsrc;
+      }
+      if (tdst)
+      {
+        if (cache.combined_dst[node])
+          cache.combined_dst[node]->catList(tdst);
+        else
+          cache.combined_dst[node] = tdst;
+      }
+      if (src_owned) src_owned->destroyList();
+    }
+    if (dst_buffer) dst_buffer->destroyList();
+
+    cache.valid = true;
   }
 
-    transfer_begin(ts);
+  // Use cached lists with buffer-reusing transfer
-  }
+  transfer_cached(cache.combined_src, cache.combined_dst, VarList, VarList, Symmetry, cache);
-
-  return handle;
 }
-//
+// Sync_start: pack and post MPI_Isend/Irecv, return immediately
-void Parallel::SyncEnd(SyncHandle *handle)
+void Parallel::Sync_start(MyList<Patch> *PatL, MyList<var> *VarList, int Symmetry,
+                          SyncCache &cache, AsyncSyncState &state)
 {
-  if (!handle)
+  // Ensure cache is built
-    return;
+  if (!cache.valid)
-
-  // Wait for all pending transfers and unpack
-  for (int i = 0; i < handle->num_states; i++)
   {
-    TransferState *ts = &handle->states[i];
+    // Build cache (same logic as Sync_cached)
-    transfer_end(ts);
-
-    // Cleanup grid segment lists only if this state owns them
-    if (ts->owns_gsl)
-    {
-      if (ts->dst)
-        ts->dst->destroyList();
-      for (int node = 0; node < ts->cpusize; node++)
-      {
-        if (ts->src[node])
-          ts->src[node]->destroyList();
-        if (ts->transfer_src[node])
-          ts->transfer_src[node]->destroyList();
-        if (ts->transfer_dst[node])
-          ts->transfer_dst[node]->destroyList();
-      }
-      delete[] ts->src;
-      delete[] ts->transfer_src;
-      delete[] ts->transfer_dst;
-    }
-  }
-
-  delete[] handle->states;
-  delete handle;
-}
-//
-// SyncPreparePlan: Pre-build grid segment lists for a patch list.
-// The plan can be reused across multiple SyncBeginWithPlan calls
-// as long as the mesh topology does not change (no regridding).
-//
-Parallel::SyncPlan *Parallel::SyncPreparePlan(MyList<Patch> *PatL, int Symmetry)
-{
     int cpusize;
     MPI_Comm_size(MPI_COMM_WORLD, &cpusize);
+    cache.cpusize = cpusize;
 
-  // Count patches
+    if (!cache.combined_src)
-  int num_patches = 0;
-  MyList<Patch> *Pp = PatL;
-  while (Pp) { num_patches++; Pp = Pp->next; }
-
-  SyncPlan *plan = new SyncPlan;
-  plan->num_entries = num_patches + 1; // intra-patch + 1 inter-patch
-  plan->Symmetry = Symmetry;
-  plan->entries = new SyncPlanEntry[plan->num_entries];
-
-  // Intra-patch entries: ghost zone exchange within each patch
-  int idx = 0;
-  Pp = PatL;
-  while (Pp)
     {
-    SyncPlanEntry *pe = &plan->entries[idx];
+      cache.combined_src = new MyList<Parallel::gridseg> *[cpusize];
-    pe->cpusize = cpusize;
+      cache.combined_dst = new MyList<Parallel::gridseg> *[cpusize];
-    pe->dst = build_ghost_gsl(Pp->data);
+      cache.send_lengths = new int[cpusize];
-    pe->src = new MyList<Parallel::gridseg> *[cpusize];
+      cache.recv_lengths = new int[cpusize];
-    pe->transfer_src = new MyList<Parallel::gridseg> *[cpusize];
+      cache.send_bufs = new double *[cpusize];
-    pe->transfer_dst = new MyList<Parallel::gridseg> *[cpusize];
+      cache.recv_bufs = new double *[cpusize];
+      cache.send_buf_caps = new int[cpusize];
+      cache.recv_buf_caps = new int[cpusize];
+      for (int i = 0; i < cpusize; i++)
+      {
+        cache.send_bufs[i] = cache.recv_bufs[i] = 0;
+        cache.send_buf_caps[i] = cache.recv_buf_caps[i] = 0;
+      }
+      cache.max_reqs = 2 * cpusize;
+      cache.reqs = new MPI_Request[cache.max_reqs];
+      cache.stats = new MPI_Status[cache.max_reqs];
+    }
+
     for (int node = 0; node < cpusize; node++)
     {
-      pe->src[node] = build_owned_gsl0(Pp->data, node);
+      cache.combined_src[node] = cache.combined_dst[node] = 0;
-      build_gstl(pe->src[node], pe->dst, &pe->transfer_src[node], &pe->transfer_dst[node]);
+      cache.send_lengths[node] = cache.recv_lengths[node] = 0;
     }
-    idx++;
+
+    MyList<Patch> *Pp = PatL;
+    while (Pp)
+    {
+      Patch *Pat = Pp->data;
+      MyList<Parallel::gridseg> *dst_ghost = build_ghost_gsl(Pat);
+      for (int node = 0; node < cpusize; node++)
+      {
+        MyList<Parallel::gridseg> *src_owned = build_owned_gsl0(Pat, node);
+        MyList<Parallel::gridseg> *tsrc = 0, *tdst = 0;
+        build_gstl(src_owned, dst_ghost, &tsrc, &tdst);
+        if (tsrc)
+        {
+          if (cache.combined_src[node])
+            cache.combined_src[node]->catList(tsrc);
+          else
+            cache.combined_src[node] = tsrc;
+        }
+        if (tdst)
+        {
+          if (cache.combined_dst[node])
+            cache.combined_dst[node]->catList(tdst);
+          else
+            cache.combined_dst[node] = tdst;
+        }
+        if (src_owned) src_owned->destroyList();
+      }
+      if (dst_ghost) dst_ghost->destroyList();
       Pp = Pp->next;
     }
 
-  // Inter-patch entry: buffer zone exchange between patches
+    MyList<Parallel::gridseg> *dst_buffer = build_buffer_gsl(PatL);
-  {
-    SyncPlanEntry *pe = &plan->entries[idx];
-    pe->cpusize = cpusize;
-    pe->dst = build_buffer_gsl(PatL);
-    pe->src = new MyList<Parallel::gridseg> *[cpusize];
-    pe->transfer_src = new MyList<Parallel::gridseg> *[cpusize];
-    pe->transfer_dst = new MyList<Parallel::gridseg> *[cpusize];
     for (int node = 0; node < cpusize; node++)
     {
-      pe->src[node] = build_owned_gsl(PatL, node, 5, Symmetry);
+      MyList<Parallel::gridseg> *src_owned = build_owned_gsl(PatL, node, 5, Symmetry);
-      build_gstl(pe->src[node], pe->dst, &pe->transfer_src[node], &pe->transfer_dst[node]);
+      MyList<Parallel::gridseg> *tsrc = 0, *tdst = 0;
+      build_gstl(src_owned, dst_buffer, &tsrc, &tdst);
+      if (tsrc)
+      {
+        if (cache.combined_src[node])
+          cache.combined_src[node]->catList(tsrc);
+        else
+          cache.combined_src[node] = tsrc;
       }
+      if (tdst)
+      {
+        if (cache.combined_dst[node])
+          cache.combined_dst[node]->catList(tdst);
+        else
+          cache.combined_dst[node] = tdst;
+      }
+      if (src_owned) src_owned->destroyList();
+    }
+    if (dst_buffer) dst_buffer->destroyList();
+    cache.valid = true;
   }
 
-  return plan;
+  // Now pack and post async MPI operations
+  int myrank;
+  MPI_Comm_rank(MPI_COMM_WORLD, &myrank);
+  int cpusize = cache.cpusize;
+  state.req_no = 0;
+  state.active = true;
+
+  MyList<Parallel::gridseg> **src = cache.combined_src;
+  MyList<Parallel::gridseg> **dst = cache.combined_dst;
+
+  for (int node = 0; node < cpusize; node++)
+  {
+    if (node == myrank)
+    {
+      int length;
+      if (!cache.lengths_valid) {
+        length = data_packer(0, src[myrank], dst[myrank], node, PACK, VarList, VarList, Symmetry);
+        cache.recv_lengths[node] = length;
+      } else {
+        length = cache.recv_lengths[node];
+      }
+      if (length > 0)
+      {
+        if (length > cache.recv_buf_caps[node])
+        {
+          if (cache.recv_bufs[node]) delete[] cache.recv_bufs[node];
+          cache.recv_bufs[node] = new double[length];
+          cache.recv_buf_caps[node] = length;
+        }
+        data_packer(cache.recv_bufs[node], src[myrank], dst[myrank], node, PACK, VarList, VarList, Symmetry);
+      }
+    }
+    else
+    {
+      int slength;
+      if (!cache.lengths_valid) {
+        slength = data_packer(0, src[myrank], dst[myrank], node, PACK, VarList, VarList, Symmetry);
+        cache.send_lengths[node] = slength;
+      } else {
+        slength = cache.send_lengths[node];
+      }
+      if (slength > 0)
+      {
+        if (slength > cache.send_buf_caps[node])
+        {
+          if (cache.send_bufs[node]) delete[] cache.send_bufs[node];
+          cache.send_bufs[node] = new double[slength];
+          cache.send_buf_caps[node] = slength;
+        }
+        data_packer(cache.send_bufs[node], src[myrank], dst[myrank], node, PACK, VarList, VarList, Symmetry);
+        MPI_Isend((void *)cache.send_bufs[node], slength, MPI_DOUBLE, node, 2, MPI_COMM_WORLD, cache.reqs + state.req_no++);
+      }
+      int rlength;
+      if (!cache.lengths_valid) {
+        rlength = data_packer(0, src[node], dst[node], node, UNPACK, VarList, VarList, Symmetry);
+        cache.recv_lengths[node] = rlength;
+      } else {
+        rlength = cache.recv_lengths[node];
+      }
+      if (rlength > 0)
+      {
+        if (rlength > cache.recv_buf_caps[node])
+        {
+          if (cache.recv_bufs[node]) delete[] cache.recv_bufs[node];
+          cache.recv_bufs[node] = new double[rlength];
+          cache.recv_buf_caps[node] = rlength;
+        }
+        MPI_Irecv((void *)cache.recv_bufs[node], rlength, MPI_DOUBLE, node, 2, MPI_COMM_WORLD, cache.reqs + state.req_no++);
+      }
+    }
+  }
+  cache.lengths_valid = true;
 }
-//
+// Sync_finish: wait for async MPI operations and unpack
-void Parallel::SyncFreePlan(SyncPlan *plan)
+void Parallel::Sync_finish(SyncCache &cache, AsyncSyncState &state,
+                           MyList<var> *VarList, int Symmetry)
 {
-  if (!plan)
+  if (!state.active)
     return;
 
-  for (int i = 0; i < plan->num_entries; i++)
+  MPI_Waitall(state.req_no, cache.reqs, cache.stats);
-  {
-    SyncPlanEntry *pe = &plan->entries[i];
-    if (pe->dst)
-      pe->dst->destroyList();
-    for (int node = 0; node < pe->cpusize; node++)
-    {
-      if (pe->src[node])
-        pe->src[node]->destroyList();
-      if (pe->transfer_src[node])
-        pe->transfer_src[node]->destroyList();
-      if (pe->transfer_dst[node])
-        pe->transfer_dst[node]->destroyList();
-    }
-    delete[] pe->src;
-    delete[] pe->transfer_src;
-    delete[] pe->transfer_dst;
-  }
-  delete[] plan->entries;
-  delete plan;
-}
-//
-// SyncBeginWithPlan: Use pre-built GSLs from a SyncPlan to initiate async transfer.
-// This avoids the O(cpusize * blocks^2) cost of rebuilding GSLs on every call.
-//
-Parallel::SyncHandle *Parallel::SyncBeginWithPlan(SyncPlan *plan, MyList<var> *VarList)
-{
-  return SyncBeginWithPlan(plan, VarList, VarList);
-}
-//
-Parallel::SyncHandle *Parallel::SyncBeginWithPlan(SyncPlan *plan, MyList<var> *VarList1, MyList<var> *VarList2)
-{
-  SyncHandle *handle = new SyncHandle;
-  handle->num_states = plan->num_entries;
-  handle->states = new TransferState[handle->num_states];
 
-  for (int i = 0; i < plan->num_entries; i++)
+  int cpusize = cache.cpusize;
-  {
+  MyList<Parallel::gridseg> **src = cache.combined_src;
-    SyncPlanEntry *pe = &plan->entries[i];
+  MyList<Parallel::gridseg> **dst = cache.combined_dst;
-    TransferState *ts = &handle->states[i];
 
-    ts->cpusize = pe->cpusize;
+  for (int node = 0; node < cpusize; node++)
-    ts->VarList1 = VarList1;
+    if (cache.recv_bufs[node] && cache.recv_lengths[node] > 0)
-    ts->VarList2 = VarList2;
+      data_packer(cache.recv_bufs[node], src[node], dst[node], node, UNPACK, VarList, VarList, Symmetry);
-    ts->Symmetry = plan->Symmetry;
-    ts->owns_gsl = false; // GSLs are owned by the plan, not this handle
 
-    // Borrow GSL pointers from the plan (do NOT free them in SyncEnd)
+  state.active = false;
-    ts->transfer_src = pe->transfer_src;
-    ts->transfer_dst = pe->transfer_dst;
-    ts->src = pe->src;
-    ts->dst = pe->dst;
-
-    transfer_begin(ts);
-  }
-
-  return handle;
 }
 // collect buffer grid segments or blocks for the periodic boundary condition of given patch
 // ---------------------------------------------------
@@ -5142,6 +5286,203 @@ void Parallel::OutBdLow2Himix(MyList<Patch> *PatcL, MyList<Patch> *PatfL,
   delete[] transfer_src;
   delete[] transfer_dst;
 }
+
+// Restrict_cached: cache grid segment lists, reuse buffers via transfer_cached
+void Parallel::Restrict_cached(MyList<Patch> *PatcL, MyList<Patch> *PatfL,
+                               MyList<var> *VarList1, MyList<var> *VarList2,
+                               int Symmetry, SyncCache &cache)
+{
+  if (!cache.valid)
+  {
+    int cpusize;
+    MPI_Comm_size(MPI_COMM_WORLD, &cpusize);
+    cache.cpusize = cpusize;
+
+    if (!cache.combined_src)
+    {
+      cache.combined_src = new MyList<Parallel::gridseg> *[cpusize];
+      cache.combined_dst = new MyList<Parallel::gridseg> *[cpusize];
+      cache.send_lengths = new int[cpusize];
+      cache.recv_lengths = new int[cpusize];
+      cache.send_bufs = new double *[cpusize];
+      cache.recv_bufs = new double *[cpusize];
+      cache.send_buf_caps = new int[cpusize];
+      cache.recv_buf_caps = new int[cpusize];
+      for (int i = 0; i < cpusize; i++)
+      {
+        cache.send_bufs[i] = cache.recv_bufs[i] = 0;
+        cache.send_buf_caps[i] = cache.recv_buf_caps[i] = 0;
+      }
+      cache.max_reqs = 2 * cpusize;
+      cache.reqs = new MPI_Request[cache.max_reqs];
+      cache.stats = new MPI_Status[cache.max_reqs];
+    }
+
+    MyList<Parallel::gridseg> *dst = build_complete_gsl(PatcL);
+    for (int node = 0; node < cpusize; node++)
+    {
+      MyList<Parallel::gridseg> *src_owned = build_owned_gsl(PatfL, node, 2, Symmetry);
+      build_gstl(src_owned, dst, &cache.combined_src[node], &cache.combined_dst[node]);
+      if (src_owned) src_owned->destroyList();
+    }
+    if (dst) dst->destroyList();
+
+    cache.valid = true;
+  }
+
+  transfer_cached(cache.combined_src, cache.combined_dst, VarList1, VarList2, Symmetry, cache);
+}
+
+// OutBdLow2Hi_cached: cache grid segment lists, reuse buffers via transfer_cached
+void Parallel::OutBdLow2Hi_cached(MyList<Patch> *PatcL, MyList<Patch> *PatfL,
+                                  MyList<var> *VarList1, MyList<var> *VarList2,
+                                  int Symmetry, SyncCache &cache)
+{
+  if (!cache.valid)
+  {
+    int cpusize;
+    MPI_Comm_size(MPI_COMM_WORLD, &cpusize);
+    cache.cpusize = cpusize;
+
+    if (!cache.combined_src)
+    {
+      cache.combined_src = new MyList<Parallel::gridseg> *[cpusize];
+      cache.combined_dst = new MyList<Parallel::gridseg> *[cpusize];
+      cache.send_lengths = new int[cpusize];
+      cache.recv_lengths = new int[cpusize];
+      cache.send_bufs = new double *[cpusize];
+      cache.recv_bufs = new double *[cpusize];
+      cache.send_buf_caps = new int[cpusize];
+      cache.recv_buf_caps = new int[cpusize];
+      for (int i = 0; i < cpusize; i++)
+      {
+        cache.send_bufs[i] = cache.recv_bufs[i] = 0;
+        cache.send_buf_caps[i] = cache.recv_buf_caps[i] = 0;
+      }
+      cache.max_reqs = 2 * cpusize;
+      cache.reqs = new MPI_Request[cache.max_reqs];
+      cache.stats = new MPI_Status[cache.max_reqs];
+    }
+
+    MyList<Parallel::gridseg> *dst = build_buffer_gsl(PatfL);
+    for (int node = 0; node < cpusize; node++)
+    {
+      MyList<Parallel::gridseg> *src_owned = build_owned_gsl(PatcL, node, 4, Symmetry);
+      build_gstl(src_owned, dst, &cache.combined_src[node], &cache.combined_dst[node]);
+      if (src_owned) src_owned->destroyList();
+    }
+    if (dst) dst->destroyList();
+
+    cache.valid = true;
+  }
+
+  transfer_cached(cache.combined_src, cache.combined_dst, VarList1, VarList2, Symmetry, cache);
+}
+
+// OutBdLow2Himix_cached: same as OutBdLow2Hi_cached but uses transfermix for unpacking
+void Parallel::OutBdLow2Himix_cached(MyList<Patch> *PatcL, MyList<Patch> *PatfL,
+                                     MyList<var> *VarList1, MyList<var> *VarList2,
+                                     int Symmetry, SyncCache &cache)
+{
+  if (!cache.valid)
+  {
+    int cpusize;
+    MPI_Comm_size(MPI_COMM_WORLD, &cpusize);
+    cache.cpusize = cpusize;
+
+    if (!cache.combined_src)
+    {
+      cache.combined_src = new MyList<Parallel::gridseg> *[cpusize];
+      cache.combined_dst = new MyList<Parallel::gridseg> *[cpusize];
+      cache.send_lengths = new int[cpusize];
+      cache.recv_lengths = new int[cpusize];
+      cache.send_bufs = new double *[cpusize];
+      cache.recv_bufs = new double *[cpusize];
+      cache.send_buf_caps = new int[cpusize];
+      cache.recv_buf_caps = new int[cpusize];
+      for (int i = 0; i < cpusize; i++)
+      {
+        cache.send_bufs[i] = cache.recv_bufs[i] = 0;
+        cache.send_buf_caps[i] = cache.recv_buf_caps[i] = 0;
+      }
+      cache.max_reqs = 2 * cpusize;
+      cache.reqs = new MPI_Request[cache.max_reqs];
+      cache.stats = new MPI_Status[cache.max_reqs];
+    }
+
+    MyList<Parallel::gridseg> *dst = build_buffer_gsl(PatfL);
+    for (int node = 0; node < cpusize; node++)
+    {
+      MyList<Parallel::gridseg> *src_owned = build_owned_gsl(PatcL, node, 4, Symmetry);
+      build_gstl(src_owned, dst, &cache.combined_src[node], &cache.combined_dst[node]);
+      if (src_owned) src_owned->destroyList();
+    }
+    if (dst) dst->destroyList();
+
+    cache.valid = true;
+  }
+
+  // Use transfermix instead of transfer for mix-mode interpolation
+  int myrank;
+  MPI_Comm_size(MPI_COMM_WORLD, &cache.cpusize);
+  MPI_Comm_rank(MPI_COMM_WORLD, &myrank);
+  int cpusize = cache.cpusize;
+
+  int req_no = 0;
+  for (int node = 0; node < cpusize; node++)
+  {
+    if (node == myrank)
+    {
+      int length = data_packermix(0, cache.combined_src[myrank], cache.combined_dst[myrank], node, PACK, VarList1, VarList2, Symmetry);
+      cache.recv_lengths[node] = length;
+      if (length > 0)
+      {
+        if (length > cache.recv_buf_caps[node])
+        {
+          if (cache.recv_bufs[node]) delete[] cache.recv_bufs[node];
+          cache.recv_bufs[node] = new double[length];
+          cache.recv_buf_caps[node] = length;
+        }
+        data_packermix(cache.recv_bufs[node], cache.combined_src[myrank], cache.combined_dst[myrank], node, PACK, VarList1, VarList2, Symmetry);
+      }
+    }
+    else
+    {
+      int slength = data_packermix(0, cache.combined_src[myrank], cache.combined_dst[myrank], node, PACK, VarList1, VarList2, Symmetry);
+      cache.send_lengths[node] = slength;
+      if (slength > 0)
+      {
+        if (slength > cache.send_buf_caps[node])
+        {
+          if (cache.send_bufs[node]) delete[] cache.send_bufs[node];
+          cache.send_bufs[node] = new double[slength];
+          cache.send_buf_caps[node] = slength;
+        }
+        data_packermix(cache.send_bufs[node], cache.combined_src[myrank], cache.combined_dst[myrank], node, PACK, VarList1, VarList2, Symmetry);
+        MPI_Isend((void *)cache.send_bufs[node], slength, MPI_DOUBLE, node, 1, MPI_COMM_WORLD, cache.reqs + req_no++);
+      }
+      int rlength = data_packermix(0, cache.combined_src[node], cache.combined_dst[node], node, UNPACK, VarList1, VarList2, Symmetry);
+      cache.recv_lengths[node] = rlength;
+      if (rlength > 0)
+      {
+        if (rlength > cache.recv_buf_caps[node])
+        {
+          if (cache.recv_bufs[node]) delete[] cache.recv_bufs[node];
+          cache.recv_bufs[node] = new double[rlength];
+          cache.recv_buf_caps[node] = rlength;
+        }
+        MPI_Irecv((void *)cache.recv_bufs[node], rlength, MPI_DOUBLE, node, 1, MPI_COMM_WORLD, cache.reqs + req_no++);
+      }
+    }
+  }
+
+  MPI_Waitall(req_no, cache.reqs, cache.stats);
+
+  for (int node = 0; node < cpusize; node++)
+    if (cache.recv_bufs[node] && cache.recv_lengths[node] > 0)
+      data_packermix(cache.recv_bufs[node], cache.combined_src[node], cache.combined_dst[node], node, UNPACK, VarList1, VarList2, Symmetry);
+}
+
 // collect all buffer grid segments or blocks for given patch
 MyList<Parallel::gridseg> *Parallel::build_buffer_gsl(Patch *Pat)
 {

AMSS_NCKU_source/Parallel.h

@@ -81,53 +81,43 @@ namespace Parallel
                    int Symmetry);
   void Sync(Patch *Pat, MyList<var> *VarList, int Symmetry);
   void Sync(MyList<Patch> *PatL, MyList<var> *VarList, int Symmetry);
+  void Sync_merged(MyList<Patch> *PatL, MyList<var> *VarList, int Symmetry);
 
-  // Async Sync: overlap MPI communication with computation
+  struct SyncCache {
-  struct TransferState
+    bool valid;
-  {
+    int cpusize;
+    MyList<gridseg> **combined_src;
+    MyList<gridseg> **combined_dst;
+    int *send_lengths;
+    int *recv_lengths;
+    double **send_bufs;
+    double **recv_bufs;
+    int *send_buf_caps;
+    int *recv_buf_caps;
     MPI_Request *reqs;
     MPI_Status *stats;
-    int req_no;
+    int max_reqs;
-    double **send_data;
+    bool lengths_valid;
-    double **rec_data;
+    SyncCache();
-    int cpusize;
+    void invalidate();
-    MyList<gridseg> **transfer_src;
+    void destroy();
-    MyList<gridseg> **transfer_dst;
-    MyList<gridseg> **src;
-    MyList<gridseg> *dst;
-    MyList<var> *VarList1;
-    MyList<var> *VarList2;
-    int Symmetry;
-    bool owns_gsl; // true if this state owns and should free the GSLs
   };
-  struct SyncHandle
-  {
-    TransferState *states;
-    int num_states;
-  };
-  SyncHandle *SyncBegin(Patch *Pat, MyList<var> *VarList, int Symmetry);
-  SyncHandle *SyncBegin(MyList<Patch> *PatL, MyList<var> *VarList, int Symmetry);
-  void SyncEnd(SyncHandle *handle);
 
-  // Cached GSL plan: pre-build grid segment lists once, reuse across multiple Sync calls
+  void Sync_cached(MyList<Patch> *PatL, MyList<var> *VarList, int Symmetry, SyncCache &cache);
-  struct SyncPlanEntry
+  void transfer_cached(MyList<gridseg> **src, MyList<gridseg> **dst,
-  {
+                       MyList<var> *VarList1, MyList<var> *VarList2,
-    int cpusize;
+                       int Symmetry, SyncCache &cache);
-    MyList<gridseg> **transfer_src;
+
-    MyList<gridseg> **transfer_dst;
+  struct AsyncSyncState {
-    MyList<gridseg> **src;
+    int req_no;
-    MyList<gridseg> *dst;
+    bool active;
+    AsyncSyncState() : req_no(0), active(false) {}
   };
-  struct SyncPlan
+
-  {
+  void Sync_start(MyList<Patch> *PatL, MyList<var> *VarList, int Symmetry,
-    SyncPlanEntry *entries;
+                  SyncCache &cache, AsyncSyncState &state);
-    int num_entries;
+  void Sync_finish(SyncCache &cache, AsyncSyncState &state,
-    int Symmetry;
+                   MyList<var> *VarList, int Symmetry);
-  };
-  SyncPlan *SyncPreparePlan(MyList<Patch> *PatL, int Symmetry);
-  void SyncFreePlan(SyncPlan *plan);
-  SyncHandle *SyncBeginWithPlan(SyncPlan *plan, MyList<var> *VarList);
-  SyncHandle *SyncBeginWithPlan(SyncPlan *plan, MyList<var> *VarList1, MyList<var> *VarList2);
   void OutBdLow2Hi(Patch *Patc, Patch *Patf,
                    MyList<var> *VarList1 /* source */, MyList<var> *VarList2 /* target */,
                    int Symmetry);
@@ -140,6 +130,15 @@ namespace Parallel
   void OutBdLow2Himix(MyList<Patch> *PatcL, MyList<Patch> *PatfL,
                       MyList<var> *VarList1 /* source */, MyList<var> *VarList2 /* target */,
                       int Symmetry);
+  void Restrict_cached(MyList<Patch> *PatcL, MyList<Patch> *PatfL,
+                       MyList<var> *VarList1, MyList<var> *VarList2,
+                       int Symmetry, SyncCache &cache);
+  void OutBdLow2Hi_cached(MyList<Patch> *PatcL, MyList<Patch> *PatfL,
+                          MyList<var> *VarList1, MyList<var> *VarList2,
+                          int Symmetry, SyncCache &cache);
+  void OutBdLow2Himix_cached(MyList<Patch> *PatcL, MyList<Patch> *PatfL,
+                             MyList<var> *VarList1, MyList<var> *VarList2,
+                             int Symmetry, SyncCache &cache);
   void Prolong(Patch *Patc, Patch *Patf,
                MyList<var> *VarList1 /* source */, MyList<var> *VarList2 /* target */,
                int Symmetry);

AMSS_NCKU_source/Z4c_class.C

@@ -186,12 +186,6 @@ void Z4c_class::Step(int lev, int YN)
   int ERROR = 0;
 
   MyList<ss_patch> *sPp;
-
-  // Pre-build grid segment lists once for this level's patches.
-  // These are reused across predictor + 3 corrector SyncBegin calls,
-  // avoiding O(cpusize * blocks^2) rebuild each time.
-  Parallel::SyncPlan *sync_plan = Parallel::SyncPreparePlan(GH->PatL[lev], Symmetry);
-
   // Predictor
   MyList<Patch> *Pp = GH->PatL[lev];
   while (Pp)
@@ -327,26 +321,7 @@ void Z4c_class::Step(int lev, int YN)
     }
     Pp = Pp->next;
   }
-  // Start async ghost zone exchange - overlaps with error check and Shell computation
+  // NOTE: error check deferred to after Shell Patch computation to reduce MPI_Allreduce calls
-  Parallel::SyncHandle *sync_pre = Parallel::SyncBeginWithPlan(sync_plan, SynchList_pre);
-
-  // check error information (overlaps with MPI transfer)
-  {
-    int erh = ERROR;
-    MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
-  }
-  if (ERROR)
-  {
-    Parallel::SyncEnd(sync_pre); sync_pre = 0;
-    Parallel::Dump_Data(GH->PatL[lev], StateList, 0, PhysTime, dT_lev);
-    if (myrank == 0)
-    {
-      if (ErrorMonitor->outfile)
-        ErrorMonitor->outfile << "find NaN in state variables at t = " << PhysTime 
-                              << ", lev = " << lev << endl;
-      MPI_Abort(MPI_COMM_WORLD, 1);
-    }
-  }
 
 #ifdef WithShell
   // evolve Shell Patches
@@ -478,26 +453,16 @@ void Z4c_class::Step(int lev, int YN)
       sPp = sPp->next;
     }
   }
-  // check error information
+  // Non-blocking error reduction overlapped with Sync to hide Allreduce latency
+  MPI_Request err_req_pre;
   {
     int erh = ERROR;
-    MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
+    MPI_Iallreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD, &err_req_pre);
-  }
-  if (ERROR)
-  {
-    Parallel::SyncEnd(sync_pre); sync_pre = 0;
-    SH->Dump_Data(StateList, 0, PhysTime, dT_lev);
-    if (myrank == 0)
-    {
-      if (ErrorMonitor->outfile)
-        ErrorMonitor->outfile << "find NaN in state variables on Shell Patches at t = " << PhysTime << endl;
-      MPI_Abort(MPI_COMM_WORLD, 1);
-    }
   }
 #endif
 
-  // Complete async ghost zone exchange
+  Parallel::AsyncSyncState async_pre;
-  if (sync_pre) Parallel::SyncEnd(sync_pre);
+  Parallel::Sync_start(GH->PatL[lev], SynchList_pre, Symmetry, sync_cache_pre[lev], async_pre);
 
 #ifdef WithShell
   if (lev == 0)
@@ -516,6 +481,24 @@ void Z4c_class::Step(int lev, int YN)
     }
   }
 #endif
+  Parallel::Sync_finish(sync_cache_pre[lev], async_pre, SynchList_pre, Symmetry);
+
+#ifdef WithShell
+  // Complete non-blocking error reduction and check
+  MPI_Wait(&err_req_pre, MPI_STATUS_IGNORE);
+  if (ERROR)
+  {
+    Parallel::Dump_Data(GH->PatL[lev], StateList, 0, PhysTime, dT_lev);
+    SH->Dump_Data(StateList, 0, PhysTime, dT_lev);
+    if (myrank == 0)
+    {
+      if (ErrorMonitor->outfile)
+        ErrorMonitor->outfile << "find NaN in state variables at t = " << PhysTime
+                              << ", lev = " << lev << endl;
+      MPI_Abort(MPI_COMM_WORLD, 1);
+    }
+  }
+#endif
 
   // for black hole position
   if (BH_num > 0 && lev == GH->levels - 1)
@@ -705,27 +688,7 @@ void Z4c_class::Step(int lev, int YN)
       Pp = Pp->next;
     }
 
-    // Start async ghost zone exchange - overlaps with error check and Shell computation
+    // NOTE: error check deferred to after Shell Patch computation to reduce MPI_Allreduce calls
-    Parallel::SyncHandle *sync_cor = Parallel::SyncBeginWithPlan(sync_plan, SynchList_cor);
-
-    // check error information (overlaps with MPI transfer)
-    {
-      int erh = ERROR;
-      MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
-    }
-    if (ERROR)
-    {
-      Parallel::SyncEnd(sync_cor); sync_cor = 0;
-      Parallel::Dump_Data(GH->PatL[lev], SynchList_pre, 0, PhysTime, dT_lev);
-      if (myrank == 0)
-      {
-        if (ErrorMonitor->outfile)
-          ErrorMonitor->outfile << "find NaN in RK4 substep#" << iter_count 
-                                << " variables at t = " << PhysTime 
-                                << ", lev = " << lev << endl;
-        MPI_Abort(MPI_COMM_WORLD, 1);
-      }
-    }
 
 #ifdef WithShell
     // evolve Shell Patches
@@ -866,27 +829,16 @@ void Z4c_class::Step(int lev, int YN)
         sPp = sPp->next;
       }
     }
-    // check error information
+    // Non-blocking error reduction overlapped with Sync to hide Allreduce latency
+    MPI_Request err_req_cor;
     {
       int erh = ERROR;
-      MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
+      MPI_Iallreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD, &err_req_cor);
-    }
-    if (ERROR)
-    {
-      Parallel::SyncEnd(sync_cor); sync_cor = 0;
-      SH->Dump_Data(SynchList_pre, 0, PhysTime, dT_lev);
-      if (myrank == 0)
-      {
-        if (ErrorMonitor->outfile)
-          ErrorMonitor->outfile << "find NaN on Shell Patches in RK4 substep#" << iter_count 
-                                << " variables at t = " << PhysTime << endl;
-        MPI_Abort(MPI_COMM_WORLD, 1);
-      }
     }
 #endif
 
-    // Complete async ghost zone exchange
+    Parallel::AsyncSyncState async_cor;
-    if (sync_cor) Parallel::SyncEnd(sync_cor);
+    Parallel::Sync_start(GH->PatL[lev], SynchList_cor, Symmetry, sync_cache_cor[lev], async_cor);
 
 #ifdef WithShell
     if (lev == 0)
@@ -904,6 +856,25 @@ void Z4c_class::Step(int lev, int YN)
              << " seconds! " << endl;
       }
     }
+#endif
+    Parallel::Sync_finish(sync_cache_cor[lev], async_cor, SynchList_cor, Symmetry);
+
+#ifdef WithShell
+    // Complete non-blocking error reduction and check
+    MPI_Wait(&err_req_cor, MPI_STATUS_IGNORE);
+    if (ERROR)
+    {
+      Parallel::Dump_Data(GH->PatL[lev], SynchList_pre, 0, PhysTime, dT_lev);
+      SH->Dump_Data(SynchList_pre, 0, PhysTime, dT_lev);
+      if (myrank == 0)
+      {
+        if (ErrorMonitor->outfile)
+          ErrorMonitor->outfile << "find NaN in RK4 substep#" << iter_count
+                                << " variables at t = " << PhysTime
+                                << ", lev = " << lev << endl;
+        MPI_Abort(MPI_COMM_WORLD, 1);
+      }
+    }
 #endif
     // for black hole position
     if (BH_num > 0 && lev == GH->levels - 1)
@@ -1060,8 +1031,6 @@ void Z4c_class::Step(int lev, int YN)
       Porg0[ithBH][2] = Porg1[ithBH][2];
     }
   }
-
-  Parallel::SyncFreePlan(sync_plan);
 }
 #else
 // for constraint preserving boundary (CPBC)
@@ -1095,10 +1064,6 @@ void Z4c_class::Step(int lev, int YN)
   int ERROR = 0;
 
   MyList<ss_patch> *sPp;
-
-  // Pre-build grid segment lists once for this level's patches.
-  Parallel::SyncPlan *sync_plan = Parallel::SyncPreparePlan(GH->PatL[lev], Symmetry);
-
   // Predictor
   MyList<Patch> *Pp = GH->PatL[lev];
   while (Pp)
@@ -1276,22 +1241,7 @@ void Z4c_class::Step(int lev, int YN)
 	 }
   }
 #endif
-  // check error information
+  // NOTE: error check deferred to after Shell Patch computation to reduce MPI_Allreduce calls
-  {
-    int erh = ERROR;
-    MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
-  }
-  if (ERROR)
-  {
-    Parallel::Dump_Data(GH->PatL[lev], StateList, 0, PhysTime, dT_lev);
-    if (myrank == 0)
-    {
-      if (ErrorMonitor->outfile)
-        ErrorMonitor->outfile << "find NaN in state variables at t = " << PhysTime 
-                              << ", lev = " << lev << endl;
-      MPI_Abort(MPI_COMM_WORLD, 1);
-    }
-  }
 
   // evolve Shell Patches
   if (lev == 0)
@@ -1566,28 +1516,15 @@ void Z4c_class::Step(int lev, int YN)
   }
 #endif
   }
-  // Start async ghost zone exchange - overlaps with error check
+  // Non-blocking error reduction overlapped with Sync to hide Allreduce latency
-  Parallel::SyncHandle *sync_pre = Parallel::SyncBeginWithPlan(sync_plan, SynchList_pre);
+  MPI_Request err_req_pre;
-
-  // check error information (overlaps with MPI transfer)
   {
     int erh = ERROR;
-    MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
+    MPI_Iallreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD, &err_req_pre);
-  }
-  if (ERROR)
-  {
-    Parallel::SyncEnd(sync_pre); sync_pre = 0;
-    SH->Dump_Data(StateList, 0, PhysTime, dT_lev);
-    if (myrank == 0)
-    {
-      if (ErrorMonitor->outfile)
-        ErrorMonitor->outfile << "find NaN in state variables on Shell Patches at t = " << PhysTime << endl;
-      MPI_Abort(MPI_COMM_WORLD, 1);
-    }
   }
 
-  // Complete async ghost zone exchange
+  Parallel::AsyncSyncState async_pre;
-  if (sync_pre) Parallel::SyncEnd(sync_pre);
+  Parallel::Sync_start(GH->PatL[lev], SynchList_pre, Symmetry, sync_cache_pre[lev], async_pre);
 
   if (lev == 0)
   {
@@ -1649,6 +1586,22 @@ void Z4c_class::Step(int lev, int YN)
   }
 #endif
   }
+  Parallel::Sync_finish(sync_cache_pre[lev], async_pre, SynchList_pre, Symmetry);
+
+  // Complete non-blocking error reduction and check
+  MPI_Wait(&err_req_pre, MPI_STATUS_IGNORE);
+  if (ERROR)
+  {
+    Parallel::Dump_Data(GH->PatL[lev], StateList, 0, PhysTime, dT_lev);
+    SH->Dump_Data(StateList, 0, PhysTime, dT_lev);
+    if (myrank == 0)
+    {
+      if (ErrorMonitor->outfile)
+        ErrorMonitor->outfile << "find NaN in state variables at t = " << PhysTime
+                              << ", lev = " << lev << endl;
+      MPI_Abort(MPI_COMM_WORLD, 1);
+    }
+  }
 
   // for black hole position
   if (BH_num > 0 && lev == GH->levels - 1)
@@ -1870,23 +1823,7 @@ void Z4c_class::Step(int lev, int YN)
       Pp = Pp->next;
     }
 
-    // check error information
+    // NOTE: error check deferred to after Shell Patch computation to reduce MPI_Allreduce calls
-    {
-      int erh = ERROR;
-      MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
-    }
-    if (ERROR)
-    {
-      Parallel::Dump_Data(GH->PatL[lev], SynchList_pre, 0, PhysTime, dT_lev);
-      if (myrank == 0)
-      {
-        if (ErrorMonitor->outfile)
-          ErrorMonitor->outfile << "find NaN in RK4 substep#" << iter_count 
-                                << " variables at t = " << PhysTime 
-                                << ", lev = " << lev << endl;
-        MPI_Abort(MPI_COMM_WORLD, 1);
-      }
-    }
 
     // evolve Shell Patches
     if (lev == 0)
@@ -2132,29 +2069,15 @@ void Z4c_class::Step(int lev, int YN)
         sPp = sPp->next;
       }
     }
-    // Start async ghost zone exchange - overlaps with error check
+    // Non-blocking error reduction overlapped with Sync to hide Allreduce latency
-    Parallel::SyncHandle *sync_cor = Parallel::SyncBeginWithPlan(sync_plan, SynchList_cor);
+    MPI_Request err_req_cor;
-
-    // check error information (overlaps with MPI transfer)
     {
       int erh = ERROR;
-      MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
+      MPI_Iallreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD, &err_req_cor);
-    }
-    if (ERROR)
-    {
-      Parallel::SyncEnd(sync_cor); sync_cor = 0;
-      SH->Dump_Data(SynchList_pre, 0, PhysTime, dT_lev);
-      if (myrank == 0)
-      {
-        if (ErrorMonitor->outfile)
-          ErrorMonitor->outfile << "find NaN on Shell Patches in RK4 substep#" << iter_count 
-                                << " variables at t = " << PhysTime << endl;
-        MPI_Abort(MPI_COMM_WORLD, 1);
-      }
     }
 
-    // Complete async ghost zone exchange
+    Parallel::AsyncSyncState async_cor;
-    if (sync_cor) Parallel::SyncEnd(sync_cor);
+    Parallel::Sync_start(GH->PatL[lev], SynchList_cor, Symmetry, sync_cache_cor[lev], async_cor);
 
     if (lev == 0)
     {
@@ -2204,6 +2127,23 @@ void Z4c_class::Step(int lev, int YN)
     }
 // end smooth
 #endif
+    Parallel::Sync_finish(sync_cache_cor[lev], async_cor, SynchList_cor, Symmetry);
+
+    // Complete non-blocking error reduction and check
+    MPI_Wait(&err_req_cor, MPI_STATUS_IGNORE);
+    if (ERROR)
+    {
+      Parallel::Dump_Data(GH->PatL[lev], SynchList_pre, 0, PhysTime, dT_lev);
+      SH->Dump_Data(SynchList_pre, 0, PhysTime, dT_lev);
+      if (myrank == 0)
+      {
+        if (ErrorMonitor->outfile)
+          ErrorMonitor->outfile << "find NaN in RK4 substep#" << iter_count
+                                << " variables at t = " << PhysTime
+                                << ", lev = " << lev << endl;
+        MPI_Abort(MPI_COMM_WORLD, 1);
+      }
+    }
 
     // for black hole position
     if (BH_num > 0 && lev == GH->levels - 1)
@@ -2380,8 +2320,6 @@ void Z4c_class::Step(int lev, int YN)
 	  DG_List->clearList();
 	}
 #endif
-
-  Parallel::SyncFreePlan(sync_plan);
 }
 #endif
 #undef MRBD

AMSS_NCKU_source/bssn_class.C

@@ -730,6 +730,12 @@ void bssn_class::Initialize()
     PhysTime = StartTime;
     Setup_Black_Hole_position();
   }
+
+  // Initialize sync caches (per-level, for predictor and corrector)
+  sync_cache_pre = new Parallel::SyncCache[GH->levels];
+  sync_cache_cor = new Parallel::SyncCache[GH->levels];
+  sync_cache_rp_coarse = new Parallel::SyncCache[GH->levels];
+  sync_cache_rp_fine = new Parallel::SyncCache[GH->levels];
 }
 
 //================================================================================================
@@ -981,6 +987,32 @@ bssn_class::~bssn_class()
   delete Azzz;
 #endif
 
+  // Destroy sync caches before GH
+  if (sync_cache_pre)
+  {
+    for (int i = 0; i < GH->levels; i++)
+      sync_cache_pre[i].destroy();
+    delete[] sync_cache_pre;
+  }
+  if (sync_cache_cor)
+  {
+    for (int i = 0; i < GH->levels; i++)
+      sync_cache_cor[i].destroy();
+    delete[] sync_cache_cor;
+  }
+  if (sync_cache_rp_coarse)
+  {
+    for (int i = 0; i < GH->levels; i++)
+      sync_cache_rp_coarse[i].destroy();
+    delete[] sync_cache_rp_coarse;
+  }
+  if (sync_cache_rp_fine)
+  {
+    for (int i = 0; i < GH->levels; i++)
+      sync_cache_rp_fine[i].destroy();
+    delete[] sync_cache_rp_fine;
+  }
+
   delete GH;
 #ifdef WithShell
   delete SH;
@@ -2181,6 +2213,7 @@ void bssn_class::Evolve(int Steps)
     GH->Regrid(Symmetry, BH_num, Porgbr, Porg0,
                SynchList_cor, OldStateList, StateList, SynchList_pre,
                fgt(PhysTime - dT_mon, StartTime, dT_mon / 2), ErrorMonitor);
+    for (int il = 0; il < GH->levels; il++) { sync_cache_pre[il].invalidate(); sync_cache_cor[il].invalidate(); sync_cache_rp_coarse[il].invalidate(); sync_cache_rp_fine[il].invalidate(); }
 #endif
 
 #if (REGLEV == 0 && (PSTR == 1 || PSTR == 2))
@@ -2396,6 +2429,7 @@ void bssn_class::RecursiveStep(int lev)
   GH->Regrid_Onelevel(lev, Symmetry, BH_num, Porgbr, Porg0,
                       SynchList_cor, OldStateList, StateList, SynchList_pre,
                       fgt(PhysTime - dT_lev, StartTime, dT_lev / 2), ErrorMonitor);
+  for (int il = 0; il < GH->levels; il++) { sync_cache_pre[il].invalidate(); sync_cache_cor[il].invalidate(); sync_cache_rp_coarse[il].invalidate(); sync_cache_rp_fine[il].invalidate(); }
 #endif
 }
 
@@ -2574,6 +2608,7 @@ void bssn_class::ParallelStep()
   GH->Regrid_Onelevel(GH->mylev, Symmetry, BH_num, Porgbr, Porg0,
                       SynchList_cor, OldStateList, StateList, SynchList_pre,
                       fgt(PhysTime - dT_lev, StartTime, dT_lev / 2), ErrorMonitor);
+  for (int il = 0; il < GH->levels; il++) { sync_cache_pre[il].invalidate(); sync_cache_cor[il].invalidate(); sync_cache_rp_coarse[il].invalidate(); sync_cache_rp_fine[il].invalidate(); }
 #endif
 }
 
@@ -2740,6 +2775,7 @@ void bssn_class::ParallelStep()
         GH->Regrid_Onelevel(lev + 1, Symmetry, BH_num, Porgbr, Porg0,
                             SynchList_cor, OldStateList, StateList, SynchList_pre,
                             fgt(PhysTime - dT_levp1, StartTime, dT_levp1 / 2), ErrorMonitor);
+        for (int il = 0; il < GH->levels; il++) { sync_cache_pre[il].invalidate(); sync_cache_cor[il].invalidate(); sync_cache_rp_coarse[il].invalidate(); sync_cache_rp_fine[il].invalidate(); }
 
         //               a_stream.clear();
         //               a_stream.str("");
@@ -2754,6 +2790,7 @@ void bssn_class::ParallelStep()
       GH->Regrid_Onelevel(lev, Symmetry, BH_num, Porgbr, Porg0,
                           SynchList_cor, OldStateList, StateList, SynchList_pre,
                           fgt(PhysTime - dT_lev, StartTime, dT_lev / 2), ErrorMonitor);
+      for (int il = 0; il < GH->levels; il++) { sync_cache_pre[il].invalidate(); sync_cache_cor[il].invalidate(); sync_cache_rp_coarse[il].invalidate(); sync_cache_rp_fine[il].invalidate(); }
 
       //               a_stream.clear();
       //               a_stream.str("");
@@ -2772,6 +2809,7 @@ void bssn_class::ParallelStep()
           GH->Regrid_Onelevel(lev - 1, Symmetry, BH_num, Porgbr, Porg0,
                               SynchList_cor, OldStateList, StateList, SynchList_pre,
                               fgt(PhysTime - dT_lev, StartTime, dT_levm1 / 2), ErrorMonitor);
+          for (int il = 0; il < GH->levels; il++) { sync_cache_pre[il].invalidate(); sync_cache_cor[il].invalidate(); sync_cache_rp_coarse[il].invalidate(); sync_cache_rp_fine[il].invalidate(); }
 
           //               a_stream.clear();
           //               a_stream.str("");
@@ -2787,6 +2825,7 @@ void bssn_class::ParallelStep()
           GH->Regrid_Onelevel(lev - 1, Symmetry, BH_num, Porgbr, Porg0,
                               SynchList_cor, OldStateList, StateList, SynchList_pre,
                               fgt(PhysTime - dT_lev, StartTime, dT_levm1 / 2), ErrorMonitor);
+          for (int il = 0; il < GH->levels; il++) { sync_cache_pre[il].invalidate(); sync_cache_cor[il].invalidate(); sync_cache_rp_coarse[il].invalidate(); sync_cache_rp_fine[il].invalidate(); }
 
           //               a_stream.clear();
           //               a_stream.str("");
@@ -3035,12 +3074,6 @@ void bssn_class::Step(int lev, int YN)
   int ERROR = 0;
 
   MyList<ss_patch> *sPp;
-
-  // Pre-build grid segment lists once for this level's patches.
-  // These are reused across predictor + 3 corrector SyncBegin calls,
-  // avoiding O(cpusize * blocks^2) rebuild each time.
-  Parallel::SyncPlan *sync_plan = Parallel::SyncPreparePlan(GH->PatL[lev], Symmetry);
-
   // Predictor
   MyList<Patch> *Pp = GH->PatL[lev];
   while (Pp)
@@ -3164,26 +3197,7 @@ void bssn_class::Step(int lev, int YN)
     }
     Pp = Pp->next;
   }
-
+  // NOTE: error check deferred to after Shell Patch computation to reduce MPI_Allreduce calls
-  // Start async ghost zone exchange - overlaps with error check and Shell computation
-  Parallel::SyncHandle *sync_pre = Parallel::SyncBeginWithPlan(sync_plan, SynchList_pre);
-
-  // check error information (overlaps with MPI transfer)
-  {
-    int erh = ERROR;
-    MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
-  }
-  if (ERROR)
-  {
-    Parallel::SyncEnd(sync_pre); sync_pre = 0;
-    Parallel::Dump_Data(GH->PatL[lev], StateList, 0, PhysTime, dT_lev);
-    if (myrank == 0)
-    {
-      if (ErrorMonitor->outfile)
-        ErrorMonitor->outfile << "find NaN in state variables at t = " << PhysTime << ", lev = " << lev << endl;
-      MPI_Abort(MPI_COMM_WORLD, 1);
-    }
-  }
 
 #ifdef WithShell
   // evolve Shell Patches
@@ -3327,27 +3341,16 @@ void bssn_class::Step(int lev, int YN)
 #endif
   }
 
-  // check error information
+  // Non-blocking error reduction overlapped with Sync to hide Allreduce latency
+  MPI_Request err_req;
   {
     int erh = ERROR;
-    MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
+    MPI_Iallreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD, &err_req);
-  }
-
-  if (ERROR)
-  {
-    Parallel::SyncEnd(sync_pre); sync_pre = 0;
-    SH->Dump_Data(StateList, 0, PhysTime, dT_lev);
-    if (myrank == 0)
-    {
-      if (ErrorMonitor->outfile)
-        ErrorMonitor->outfile << "find NaN in state variables on Shell Patches at t = " << PhysTime << endl;
-      MPI_Abort(MPI_COMM_WORLD, 1);
-    }
   }
 #endif
 
-  // Complete async ghost zone exchange
+  Parallel::AsyncSyncState async_pre;
-  if (sync_pre) Parallel::SyncEnd(sync_pre);
+  Parallel::Sync_start(GH->PatL[lev], SynchList_pre, Symmetry, sync_cache_pre[lev], async_pre);
 
 #ifdef WithShell
   if (lev == 0)
@@ -3366,6 +3369,23 @@ void bssn_class::Step(int lev, int YN)
     }
   }
 #endif
+  Parallel::Sync_finish(sync_cache_pre[lev], async_pre, SynchList_pre, Symmetry);
+
+#ifdef WithShell
+  // Complete non-blocking error reduction and check
+  MPI_Wait(&err_req, MPI_STATUS_IGNORE);
+  if (ERROR)
+  {
+    Parallel::Dump_Data(GH->PatL[lev], StateList, 0, PhysTime, dT_lev);
+    SH->Dump_Data(StateList, 0, PhysTime, dT_lev);
+    if (myrank == 0)
+    {
+      if (ErrorMonitor->outfile)
+        ErrorMonitor->outfile << "find NaN in state variables at t = " << PhysTime << ", lev = " << lev << endl;
+      MPI_Abort(MPI_COMM_WORLD, 1);
+    }
+  }
+#endif
 
 #if (MAPBH == 0)
   // for black hole position
@@ -3541,28 +3561,7 @@ void bssn_class::Step(int lev, int YN)
       Pp = Pp->next;
     }
 
-    // Start async ghost zone exchange - overlaps with error check and Shell computation
+    // NOTE: error check deferred to after Shell Patch computation to reduce MPI_Allreduce calls
-    Parallel::SyncHandle *sync_cor = Parallel::SyncBeginWithPlan(sync_plan, SynchList_cor);
-
-    // check error information (overlaps with MPI transfer)
-    {
-      int erh = ERROR;
-      MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
-    }
-
-    if (ERROR)
-    {
-      Parallel::SyncEnd(sync_cor); sync_cor = 0;
-      Parallel::Dump_Data(GH->PatL[lev], SynchList_pre, 0, PhysTime, dT_lev);
-      if (myrank == 0)
-      {
-        if (ErrorMonitor->outfile)
-          ErrorMonitor->outfile << "find NaN in RK4 substep#" << iter_count 
-                                << " variables at t = " << PhysTime 
-                                << ", lev = " << lev << endl;
-        MPI_Abort(MPI_COMM_WORLD, 1);
-      }
-    }
 
 #ifdef WithShell
     // evolve Shell Patches
@@ -3702,28 +3701,16 @@ void bssn_class::Step(int lev, int YN)
         sPp = sPp->next;
       }
     }
-    // check error information
+    // Non-blocking error reduction overlapped with Sync to hide Allreduce latency
+    MPI_Request err_req_cor;
     {
       int erh = ERROR;
-      MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
+      MPI_Iallreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD, &err_req_cor);
-    }
-    if (ERROR)
-    {
-      Parallel::SyncEnd(sync_cor); sync_cor = 0;
-      SH->Dump_Data(SynchList_pre, 0, PhysTime, dT_lev);
-      if (myrank == 0)
-      {
-        if (ErrorMonitor->outfile)
-          ErrorMonitor->outfile << "find NaN on Shell Patches in RK4 substep#" 
-                                << iter_count << " variables at t = " 
-                                << PhysTime << endl;
-        MPI_Abort(MPI_COMM_WORLD, 1);
-      }
     }
 #endif
 
-    // Complete async ghost zone exchange
+    Parallel::AsyncSyncState async_cor;
-    if (sync_cor) Parallel::SyncEnd(sync_cor);
+    Parallel::Sync_start(GH->PatL[lev], SynchList_cor, Symmetry, sync_cache_cor[lev], async_cor);
 
 #ifdef WithShell
     if (lev == 0)
@@ -3742,6 +3729,25 @@ void bssn_class::Step(int lev, int YN)
       }
     }
 #endif
+    Parallel::Sync_finish(sync_cache_cor[lev], async_cor, SynchList_cor, Symmetry);
+
+#ifdef WithShell
+    // Complete non-blocking error reduction and check
+    MPI_Wait(&err_req_cor, MPI_STATUS_IGNORE);
+    if (ERROR)
+    {
+      Parallel::Dump_Data(GH->PatL[lev], SynchList_pre, 0, PhysTime, dT_lev);
+      SH->Dump_Data(SynchList_pre, 0, PhysTime, dT_lev);
+      if (myrank == 0)
+      {
+        if (ErrorMonitor->outfile)
+          ErrorMonitor->outfile << "find NaN in RK4 substep#" << iter_count
+                                << " variables at t = " << PhysTime
+                                << ", lev = " << lev << endl;
+        MPI_Abort(MPI_COMM_WORLD, 1);
+      }
+    }
+#endif
 
 #if (MAPBH == 0)
     // for black hole position
@@ -3914,8 +3920,6 @@ void bssn_class::Step(int lev, int YN)
       Porg0[ithBH][2] = Porg1[ithBH][2];
     }
   }
-
-  Parallel::SyncFreePlan(sync_plan);
 }
 
 //================================================================================================
@@ -4055,22 +4059,7 @@ void bssn_class::Step(int lev, int YN)
     }
     Pp = Pp->next;
   }
-  // check error information
+  // NOTE: error check deferred to after Shell Patch computation to reduce MPI_Allreduce calls
-  {
-    int erh = ERROR;
-    MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
-  }
-  if (ERROR)
-  {
-    Parallel::Dump_Data(GH->PatL[lev], StateList, 0, PhysTime, dT_lev);
-    if (myrank == 0)
-    {
-      if (ErrorMonitor->outfile)
-        ErrorMonitor->outfile << "find NaN in state variables at t = " << PhysTime 
-                              << ", lev = " << lev << endl;
-      MPI_Abort(MPI_COMM_WORLD, 1);
-    }
-  }
 
 #ifdef WithShell
   // evolve Shell Patches
@@ -4211,25 +4200,16 @@ void bssn_class::Step(int lev, int YN)
   }
 #endif
   }
-  // check error information
+  // Non-blocking error reduction overlapped with Sync to hide Allreduce latency
+  MPI_Request err_req;
   {
     int erh = ERROR;
-    MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
+    MPI_Iallreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD, &err_req);
-  }
-  if (ERROR)
-  {
-    SH->Dump_Data(StateList, 0, PhysTime, dT_lev);
-    if (myrank == 0)
-    {
-      if (ErrorMonitor->outfile)
-        ErrorMonitor->outfile << "find NaN in state variables on Shell Patches at t = " 
-                              << PhysTime << endl;
-      MPI_Abort(MPI_COMM_WORLD, 1);
-    }
   }
 #endif
 
-  Parallel::Sync(GH->PatL[lev], SynchList_pre, Symmetry);
+  Parallel::AsyncSyncState async_pre;
+  Parallel::Sync_start(GH->PatL[lev], SynchList_pre, Symmetry, sync_cache_pre[lev], async_pre);
 
 #ifdef WithShell
   if (lev == 0)
@@ -4248,6 +4228,24 @@ void bssn_class::Step(int lev, int YN)
     }
   }
 #endif
+  Parallel::Sync_finish(sync_cache_pre[lev], async_pre, SynchList_pre, Symmetry);
+
+#ifdef WithShell
+  // Complete non-blocking error reduction and check
+  MPI_Wait(&err_req, MPI_STATUS_IGNORE);
+  if (ERROR)
+  {
+    Parallel::Dump_Data(GH->PatL[lev], StateList, 0, PhysTime, dT_lev);
+    SH->Dump_Data(StateList, 0, PhysTime, dT_lev);
+    if (myrank == 0)
+    {
+      if (ErrorMonitor->outfile)
+        ErrorMonitor->outfile << "find NaN in state variables at t = " << PhysTime
+                              << ", lev = " << lev << endl;
+      MPI_Abort(MPI_COMM_WORLD, 1);
+    }
+  }
+#endif
 
   // for black hole position
   if (BH_num > 0 && lev == GH->levels - 1)
@@ -4407,23 +4405,7 @@ void bssn_class::Step(int lev, int YN)
       Pp = Pp->next;
     }
 
-    // check error information
+    // NOTE: error check deferred to after Shell Patch computation to reduce MPI_Allreduce calls
-    {
-      int erh = ERROR;
-      MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
-    }
-    if (ERROR)
-    {
-      Parallel::Dump_Data(GH->PatL[lev], SynchList_pre, 0, PhysTime, dT_lev);
-      if (myrank == 0)
-      {
-        if (ErrorMonitor->outfile)
-          ErrorMonitor->outfile << "find NaN in RK4 substep#" << iter_count 
-                                << " variables at t = " << PhysTime 
-                                << ", lev = " << lev << endl;
-        MPI_Abort(MPI_COMM_WORLD, 1);
-      }
-    }
 
 #ifdef WithShell
     // evolve Shell Patches
@@ -4563,25 +4545,16 @@ void bssn_class::Step(int lev, int YN)
         sPp = sPp->next;
       }
     }
-    // check error information
+    // Non-blocking error reduction overlapped with Sync to hide Allreduce latency
+    MPI_Request err_req_cor;
     {
       int erh = ERROR;
-      MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
+      MPI_Iallreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD, &err_req_cor);
-    }
-    if (ERROR)
-    {
-      SH->Dump_Data(SynchList_pre, 0, PhysTime, dT_lev);
-      if (myrank == 0)
-      {
-        if (ErrorMonitor->outfile)
-          ErrorMonitor->outfile << "find NaN on Shell Patches in RK4 substep#" << iter_count 
-                                << " variables at t = " << PhysTime << endl;
-        MPI_Abort(MPI_COMM_WORLD, 1);
-      }
     }
 #endif
 
-    Parallel::Sync(GH->PatL[lev], SynchList_cor, Symmetry);
+    Parallel::AsyncSyncState async_cor;
+    Parallel::Sync_start(GH->PatL[lev], SynchList_cor, Symmetry, sync_cache_cor[lev], async_cor);
 
 #ifdef WithShell
     if (lev == 0)
@@ -4599,6 +4572,25 @@ void bssn_class::Step(int lev, int YN)
              << " seconds! " << endl;
       }
     }
+#endif
+    Parallel::Sync_finish(sync_cache_cor[lev], async_cor, SynchList_cor, Symmetry);
+
+#ifdef WithShell
+    // Complete non-blocking error reduction and check
+    MPI_Wait(&err_req_cor, MPI_STATUS_IGNORE);
+    if (ERROR)
+    {
+      Parallel::Dump_Data(GH->PatL[lev], SynchList_pre, 0, PhysTime, dT_lev);
+      SH->Dump_Data(SynchList_pre, 0, PhysTime, dT_lev);
+      if (myrank == 0)
+      {
+        if (ErrorMonitor->outfile)
+          ErrorMonitor->outfile << "find NaN in RK4 substep#" << iter_count
+                                << " variables at t = " << PhysTime
+                                << ", lev = " << lev << endl;
+        MPI_Abort(MPI_COMM_WORLD, 1);
+      }
+    }
 #endif
     // for black hole position
     if (BH_num > 0 && lev == GH->levels - 1)
@@ -4838,12 +4830,6 @@ void bssn_class::Step(int lev, int YN)
   int ERROR = 0;
 
   MyList<ss_patch> *sPp;
-
-  // Pre-build grid segment lists once for this level's patches.
-  // These are reused across predictor + 3 corrector SyncBegin calls,
-  // avoiding O(cpusize * blocks^2) rebuild each time.
-  Parallel::SyncPlan *sync_plan = Parallel::SyncPreparePlan(GH->PatL[lev], Symmetry);
-
   // Predictor
   MyList<Patch> *Pp = GH->PatL[lev];
   while (Pp)
@@ -4970,17 +4956,21 @@ void bssn_class::Step(int lev, int YN)
 
   //   misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"after Predictor rhs calculation");
 
-  // Start async ghost zone exchange - overlaps with error check and BH position
+  // Non-blocking error reduction overlapped with Sync to hide Allreduce latency
-  Parallel::SyncHandle *sync_pre = Parallel::SyncBeginWithPlan(sync_plan, SynchList_pre);
+  MPI_Request err_req;
-
-  // check error information (overlaps with MPI transfer)
   {
     int erh = ERROR;
-    MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, GH->Commlev[lev]);
+    MPI_Iallreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, GH->Commlev[lev], &err_req);
   }
+
+  //   misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"before Predictor sync");
+
+  Parallel::Sync_cached(GH->PatL[lev], SynchList_pre, Symmetry, sync_cache_pre[lev]);
+
+  // Complete non-blocking error reduction and check
+  MPI_Wait(&err_req, MPI_STATUS_IGNORE);
   if (ERROR)
   {
-    Parallel::SyncEnd(sync_pre); sync_pre = 0;
     Parallel::Dump_Data(GH->PatL[lev], StateList, 0, PhysTime, dT_lev);
     if (myrank == 0)
     {
@@ -4990,11 +4980,6 @@ void bssn_class::Step(int lev, int YN)
     }
   }
 
-  //   misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"before Predictor sync");
-
-  // Complete async ghost zone exchange
-  if (sync_pre) Parallel::SyncEnd(sync_pre);
-
 #if (MAPBH == 0)
   // for black hole position
   if (BH_num > 0 && lev == GH->levels - 1)
@@ -5172,17 +5157,23 @@ void bssn_class::Step(int lev, int YN)
 
     //   misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"before Corrector error check");
 
-    // Start async ghost zone exchange - overlaps with error check and BH position
+    // Non-blocking error reduction overlapped with Sync to hide Allreduce latency
-    Parallel::SyncHandle *sync_cor = Parallel::SyncBeginWithPlan(sync_plan, SynchList_cor);
+    MPI_Request err_req_cor;
-
-    // check error information (overlaps with MPI transfer)
     {
       int erh = ERROR;
-      MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, GH->Commlev[lev]);
+      MPI_Iallreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, GH->Commlev[lev], &err_req_cor);
     }
+
+    //    misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"before Corrector sync");
+
+    Parallel::Sync_cached(GH->PatL[lev], SynchList_cor, Symmetry, sync_cache_cor[lev]);
+
+    //    misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"after Corrector sync");
+
+    // Complete non-blocking error reduction and check
+    MPI_Wait(&err_req_cor, MPI_STATUS_IGNORE);
     if (ERROR)
     {
-      Parallel::SyncEnd(sync_cor); sync_cor = 0;
       Parallel::Dump_Data(GH->PatL[lev], SynchList_pre, 0, PhysTime, dT_lev);
       if (myrank == 0)
       {
@@ -5194,13 +5185,6 @@ void bssn_class::Step(int lev, int YN)
       }
     }
 
-    //    misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"before Corrector sync");
-
-    // Complete async ghost zone exchange
-    if (sync_cor) Parallel::SyncEnd(sync_cor);
-
-    //    misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"after Corrector sync");
-
 #if (MAPBH == 0)
     // for black hole position
     if (BH_num > 0 && lev == GH->levels - 1)
@@ -5313,8 +5297,6 @@ void bssn_class::Step(int lev, int YN)
 
   //     if(myrank==GH->start_rank[lev]) cout<<GH->mylev<<endl;
   //     misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"complet GH Step");
-
-  Parallel::SyncFreePlan(sync_plan);
 }
 
 //================================================================================================
@@ -5486,21 +5468,11 @@ void bssn_class::SHStep()
 #if (PSTR == 1 || PSTR == 2)
 //   misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"before Predictor's error check");
 #endif
-  // check error information
+  // Non-blocking error reduction overlapped with Synch to hide Allreduce latency
+  MPI_Request err_req;
   {
     int erh = ERROR;
-    MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
+    MPI_Iallreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD, &err_req);
-  }
-
-  if (ERROR)
-  {
-    SH->Dump_Data(StateList, 0, PhysTime, dT_lev);
-    if (myrank == 0)
-    {
-      if (ErrorMonitor->outfile)
-        ErrorMonitor->outfile << "find NaN in state variables on Shell Patches at t = " << PhysTime << endl;
-      MPI_Abort(MPI_COMM_WORLD, 1);
-    }
   }
 
   {
@@ -5518,6 +5490,19 @@ void bssn_class::SHStep()
     }
   }
 
+  // Complete non-blocking error reduction and check
+  MPI_Wait(&err_req, MPI_STATUS_IGNORE);
+  if (ERROR)
+  {
+    SH->Dump_Data(StateList, 0, PhysTime, dT_lev);
+    if (myrank == 0)
+    {
+      if (ErrorMonitor->outfile)
+        ErrorMonitor->outfile << "find NaN in state variables on Shell Patches at t = " << PhysTime << endl;
+      MPI_Abort(MPI_COMM_WORLD, 1);
+    }
+  }
+
   // corrector
   for (iter_count = 1; iter_count < 4; iter_count++)
   {
@@ -5660,21 +5645,11 @@ void bssn_class::SHStep()
         sPp = sPp->next;
       }
     }
-    // check error information
+    // Non-blocking error reduction overlapped with Synch to hide Allreduce latency
+    MPI_Request err_req_cor;
     {
       int erh = ERROR;
-      MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
+      MPI_Iallreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD, &err_req_cor);
-    }
-    if (ERROR)
-    {
-      SH->Dump_Data(SynchList_pre, 0, PhysTime, dT_lev);
-      if (myrank == 0)
-      {
-        if (ErrorMonitor->outfile)
-          ErrorMonitor->outfile << "find NaN on Shell Patches in RK4 substep#" << iter_count 
-                                << " variables at t = " << PhysTime << endl;
-        MPI_Abort(MPI_COMM_WORLD, 1);
-      }
     }
 
     {
@@ -5692,6 +5667,20 @@ void bssn_class::SHStep()
       }
     }
 
+    // Complete non-blocking error reduction and check
+    MPI_Wait(&err_req_cor, MPI_STATUS_IGNORE);
+    if (ERROR)
+    {
+      SH->Dump_Data(SynchList_pre, 0, PhysTime, dT_lev);
+      if (myrank == 0)
+      {
+        if (ErrorMonitor->outfile)
+          ErrorMonitor->outfile << "find NaN on Shell Patches in RK4 substep#" << iter_count
+                                << " variables at t = " << PhysTime << endl;
+        MPI_Abort(MPI_COMM_WORLD, 1);
+      }
+    }
+
     sPp = SH->PatL;
     while (sPp)
     {
@@ -5820,7 +5809,7 @@ void bssn_class::RestrictProlong(int lev, int YN, bool BB,
 //       misc::tillherecheck(GH->Commlev[GH->mylev],GH->start_rank[GH->mylev],a_stream.str());
 #endif
 
-      Parallel::Sync(GH->PatL[lev - 1], SynchList_pre, Symmetry);
+      Parallel::Sync_cached(GH->PatL[lev - 1], SynchList_pre, Symmetry, sync_cache_rp_coarse[lev]);
 
 #if (PSTR == 1 || PSTR == 2)
 //       a_stream.clear();
@@ -5830,21 +5819,11 @@ void bssn_class::RestrictProlong(int lev, int YN, bool BB,
 #endif
 
 #if (RPB == 0)
-      Ppc = GH->PatL[lev - 1];
-      while (Ppc)
-      {
-        Pp = GH->PatL[lev];
-        while (Pp)
-        {
 #if (MIXOUTB == 0)
-          Parallel::OutBdLow2Hi(Ppc->data, Pp->data, SynchList_pre, SL, Symmetry);
+      Parallel::OutBdLow2Hi(GH->PatL[lev - 1], GH->PatL[lev], SynchList_pre, SL, Symmetry);
 #elif (MIXOUTB == 1)
-          Parallel::OutBdLow2Himix(Ppc->data, Pp->data, SynchList_pre, SL, Symmetry);
+      Parallel::OutBdLow2Himix(GH->PatL[lev - 1], GH->PatL[lev], SynchList_pre, SL, Symmetry);
 #endif
-          Pp = Pp->next;
-        }
-        Ppc = Ppc->next;
-      }
 #elif (RPB == 1)
       //       Parallel::OutBdLow2Hi_bam(GH->PatL[lev-1],GH->PatL[lev],SynchList_pre,SL,Symmetry);
       Parallel::OutBdLow2Hi_bam(GH->PatL[lev - 1], GH->PatL[lev], SynchList_pre, SL, GH->bdsul[lev], Symmetry);
@@ -5881,7 +5860,7 @@ void bssn_class::RestrictProlong(int lev, int YN, bool BB,
 //       misc::tillherecheck(GH->Commlev[GH->mylev],GH->start_rank[GH->mylev],a_stream.str());
 #endif
 
-      Parallel::Sync(GH->PatL[lev - 1], SL, Symmetry);
+      Parallel::Sync_cached(GH->PatL[lev - 1], SL, Symmetry, sync_cache_rp_coarse[lev]);
 
 #if (PSTR == 1 || PSTR == 2)
 //       a_stream.clear();
@@ -5891,21 +5870,11 @@ void bssn_class::RestrictProlong(int lev, int YN, bool BB,
 #endif
 
 #if (RPB == 0)
-      Ppc = GH->PatL[lev - 1];
-      while (Ppc)
-      {
-        Pp = GH->PatL[lev];
-        while (Pp)
-        {
 #if (MIXOUTB == 0)
-          Parallel::OutBdLow2Hi(Ppc->data, Pp->data, SL, SL, Symmetry);
+      Parallel::OutBdLow2Hi(GH->PatL[lev - 1], GH->PatL[lev], SL, SL, Symmetry);
 #elif (MIXOUTB == 1)
-          Parallel::OutBdLow2Himix(Ppc->data, Pp->data, SL, SL, Symmetry);
+      Parallel::OutBdLow2Himix(GH->PatL[lev - 1], GH->PatL[lev], SL, SL, Symmetry);
 #endif
-          Pp = Pp->next;
-        }
-        Ppc = Ppc->next;
-      }
 #elif (RPB == 1)
       //       Parallel::OutBdLow2Hi_bam(GH->PatL[lev-1],GH->PatL[lev],SL,SL,Symmetry);
       Parallel::OutBdLow2Hi_bam(GH->PatL[lev - 1], GH->PatL[lev], SL, SL, GH->bdsul[lev], Symmetry);
@@ -5919,7 +5888,7 @@ void bssn_class::RestrictProlong(int lev, int YN, bool BB,
 #endif
     }
 
-    Parallel::Sync(GH->PatL[lev], SL, Symmetry);
+    Parallel::Sync_cached(GH->PatL[lev], SL, Symmetry, sync_cache_rp_fine[lev]);
 
 #if (PSTR == 1 || PSTR == 2)
 //    a_stream.clear();
@@ -5977,24 +5946,14 @@ void bssn_class::RestrictProlong_aux(int lev, int YN, bool BB,
       Parallel::Restrict_bam(GH->PatL[lev - 1], GH->PatL[lev], SL, SynchList_pre, GH->rsul[lev], Symmetry);
 #endif
 
-      Parallel::Sync(GH->PatL[lev - 1], SynchList_pre, Symmetry);
+      Parallel::Sync_cached(GH->PatL[lev - 1], SynchList_pre, Symmetry, sync_cache_rp_coarse[lev]);
 
 #if (RPB == 0)
-      Ppc = GH->PatL[lev - 1];
-      while (Ppc)
-      {
-        Pp = GH->PatL[lev];
-        while (Pp)
-        {
 #if (MIXOUTB == 0)
-          Parallel::OutBdLow2Hi(Ppc->data, Pp->data, SynchList_pre, SL, Symmetry);
+      Parallel::OutBdLow2Hi(GH->PatL[lev - 1], GH->PatL[lev], SynchList_pre, SL, Symmetry);
 #elif (MIXOUTB == 1)
-          Parallel::OutBdLow2Himix(Ppc->data, Pp->data, SynchList_pre, SL, Symmetry);
+      Parallel::OutBdLow2Himix(GH->PatL[lev - 1], GH->PatL[lev], SynchList_pre, SL, Symmetry);
 #endif
-          Pp = Pp->next;
-        }
-        Ppc = Ppc->next;
-      }
 #elif (RPB == 1)
       //       Parallel::OutBdLow2Hi_bam(GH->PatL[lev-1],GH->PatL[lev],SynchList_pre,SL,Symmetry);
       Parallel::OutBdLow2Hi_bam(GH->PatL[lev - 1], GH->PatL[lev], SynchList_pre, SL, GH->bdsul[lev], Symmetry);
@@ -6009,31 +5968,21 @@ void bssn_class::RestrictProlong_aux(int lev, int YN, bool BB,
       Parallel::Restrict_bam(GH->PatL[lev - 1], GH->PatL[lev], SL, SL, GH->rsul[lev], Symmetry);
 #endif
 
-      Parallel::Sync(GH->PatL[lev - 1], SL, Symmetry);
+      Parallel::Sync_cached(GH->PatL[lev - 1], SL, Symmetry, sync_cache_rp_coarse[lev]);
 
 #if (RPB == 0)
-      Ppc = GH->PatL[lev - 1];
-      while (Ppc)
-      {
-        Pp = GH->PatL[lev];
-        while (Pp)
-        {
 #if (MIXOUTB == 0)
-          Parallel::OutBdLow2Hi(Ppc->data, Pp->data, SL, SL, Symmetry);
+      Parallel::OutBdLow2Hi(GH->PatL[lev - 1], GH->PatL[lev], SL, SL, Symmetry);
 #elif (MIXOUTB == 1)
-          Parallel::OutBdLow2Himix(Ppc->data, Pp->data, SL, SL, Symmetry);
+      Parallel::OutBdLow2Himix(GH->PatL[lev - 1], GH->PatL[lev], SL, SL, Symmetry);
 #endif
-          Pp = Pp->next;
-        }
-        Ppc = Ppc->next;
-      }
 #elif (RPB == 1)
       //       Parallel::OutBdLow2Hi_bam(GH->PatL[lev-1],GH->PatL[lev],SL,SL,Symmetry);
       Parallel::OutBdLow2Hi_bam(GH->PatL[lev - 1], GH->PatL[lev], SL, SL, GH->bdsul[lev], Symmetry);
 #endif
     }
 
-    Parallel::Sync(GH->PatL[lev], SL, Symmetry);
+    Parallel::Sync_cached(GH->PatL[lev], SL, Symmetry, sync_cache_rp_fine[lev]);
   }
 }
 
@@ -6084,24 +6033,14 @@ void bssn_class::RestrictProlong(int lev, int YN, bool BB)
       Parallel::Restrict_bam(GH->PatL[lev - 1], GH->PatL[lev], SynchList_cor, SynchList_pre, GH->rsul[lev], Symmetry);
 #endif
 
-      Parallel::Sync(GH->PatL[lev - 1], SynchList_pre, Symmetry);
+      Parallel::Sync_cached(GH->PatL[lev - 1], SynchList_pre, Symmetry, sync_cache_rp_coarse[lev]);
 
 #if (RPB == 0)
-      Ppc = GH->PatL[lev - 1];
-      while (Ppc)
-      {
-        Pp = GH->PatL[lev];
-        while (Pp)
-        {
 #if (MIXOUTB == 0)
-          Parallel::OutBdLow2Hi(Ppc->data, Pp->data, SynchList_pre, SynchList_cor, Symmetry);
+      Parallel::OutBdLow2Hi(GH->PatL[lev - 1], GH->PatL[lev], SynchList_pre, SynchList_cor, Symmetry);
 #elif (MIXOUTB == 1)
-          Parallel::OutBdLow2Himix(Ppc->data, Pp->data, SynchList_pre, SynchList_cor, Symmetry);
+      Parallel::OutBdLow2Himix(GH->PatL[lev - 1], GH->PatL[lev], SynchList_pre, SynchList_cor, Symmetry);
 #endif
-          Pp = Pp->next;
-        }
-        Ppc = Ppc->next;
-      }
 #elif (RPB == 1)
       //       Parallel::OutBdLow2Hi_bam(GH->PatL[lev-1],GH->PatL[lev],SynchList_pre,SynchList_cor,Symmetry);
       Parallel::OutBdLow2Hi_bam(GH->PatL[lev - 1], GH->PatL[lev], SynchList_pre, SynchList_cor, GH->bdsul[lev], Symmetry);
@@ -6118,31 +6057,21 @@ void bssn_class::RestrictProlong(int lev, int YN, bool BB)
       Parallel::Restrict_bam(GH->PatL[lev - 1], GH->PatL[lev], SynchList_cor, StateList, GH->rsul[lev], Symmetry);
 #endif
 
-      Parallel::Sync(GH->PatL[lev - 1], StateList, Symmetry);
+      Parallel::Sync_cached(GH->PatL[lev - 1], StateList, Symmetry, sync_cache_rp_coarse[lev]);
 
 #if (RPB == 0)
-      Ppc = GH->PatL[lev - 1];
-      while (Ppc)
-      {
-        Pp = GH->PatL[lev];
-        while (Pp)
-        {
 #if (MIXOUTB == 0)
-          Parallel::OutBdLow2Hi(Ppc->data, Pp->data, StateList, SynchList_cor, Symmetry);
+      Parallel::OutBdLow2Hi(GH->PatL[lev - 1], GH->PatL[lev], StateList, SynchList_cor, Symmetry);
 #elif (MIXOUTB == 1)
-          Parallel::OutBdLow2Himix(Ppc->data, Pp->data, StateList, SynchList_cor, Symmetry);
+      Parallel::OutBdLow2Himix(GH->PatL[lev - 1], GH->PatL[lev], StateList, SynchList_cor, Symmetry);
 #endif
-          Pp = Pp->next;
-        }
-        Ppc = Ppc->next;
-      }
 #elif (RPB == 1)
       //       Parallel::OutBdLow2Hi_bam(GH->PatL[lev-1],GH->PatL[lev],StateList,SynchList_cor,Symmetry);
       Parallel::OutBdLow2Hi_bam(GH->PatL[lev - 1], GH->PatL[lev], StateList, SynchList_cor, GH->bdsul[lev], Symmetry);
 #endif
     }
 
-    Parallel::Sync(GH->PatL[lev], SynchList_cor, Symmetry);
+    Parallel::Sync_cached(GH->PatL[lev], SynchList_cor, Symmetry, sync_cache_rp_fine[lev]);
   }
 }
 
@@ -6172,21 +6101,11 @@ void bssn_class::ProlongRestrict(int lev, int YN, bool BB)
       }
 
 #if (RPB == 0)
-      Ppc = GH->PatL[lev - 1];
-      while (Ppc)
-      {
-        Pp = GH->PatL[lev];
-        while (Pp)
-        {
 #if (MIXOUTB == 0)
-          Parallel::OutBdLow2Hi(Ppc->data, Pp->data, SynchList_pre, SynchList_cor, Symmetry);
+      Parallel::OutBdLow2Hi(GH->PatL[lev - 1], GH->PatL[lev], SynchList_pre, SynchList_cor, Symmetry);
 #elif (MIXOUTB == 1)
-          Parallel::OutBdLow2Himix(Ppc->data, Pp->data, SynchList_pre, SynchList_cor, Symmetry);
+      Parallel::OutBdLow2Himix(GH->PatL[lev - 1], GH->PatL[lev], SynchList_pre, SynchList_cor, Symmetry);
 #endif
-          Pp = Pp->next;
-        }
-        Ppc = Ppc->next;
-      }
 #elif (RPB == 1)
       //       Parallel::OutBdLow2Hi_bam(GH->PatL[lev-1],GH->PatL[lev],SynchList_pre,SynchList_cor,Symmetry);
       Parallel::OutBdLow2Hi_bam(GH->PatL[lev - 1], GH->PatL[lev], SynchList_pre, SynchList_cor, GH->bdsul[lev], Symmetry);
@@ -6195,21 +6114,11 @@ void bssn_class::ProlongRestrict(int lev, int YN, bool BB)
     else // no time refinement levels and for all same time levels
     {
 #if (RPB == 0)
-      Ppc = GH->PatL[lev - 1];
-      while (Ppc)
-      {
-        Pp = GH->PatL[lev];
-        while (Pp)
-        {
 #if (MIXOUTB == 0)
-          Parallel::OutBdLow2Hi(Ppc->data, Pp->data, StateList, SynchList_cor, Symmetry);
+      Parallel::OutBdLow2Hi(GH->PatL[lev - 1], GH->PatL[lev], StateList, SynchList_cor, Symmetry);
 #elif (MIXOUTB == 1)
-          Parallel::OutBdLow2Himix(Ppc->data, Pp->data, StateList, SynchList_cor, Symmetry);
+      Parallel::OutBdLow2Himix(GH->PatL[lev - 1], GH->PatL[lev], StateList, SynchList_cor, Symmetry);
 #endif
-          Pp = Pp->next;
-        }
-        Ppc = Ppc->next;
-      }
 #elif (RPB == 1)
       //       Parallel::OutBdLow2Hi_bam(GH->PatL[lev-1],GH->PatL[lev],StateList,SynchList_cor,Symmetry);
       Parallel::OutBdLow2Hi_bam(GH->PatL[lev - 1], GH->PatL[lev], StateList, SynchList_cor, GH->bdsul[lev], Symmetry);
@@ -6225,10 +6134,10 @@ void bssn_class::ProlongRestrict(int lev, int YN, bool BB)
 #else
       Parallel::Restrict_after(GH->PatL[lev - 1], GH->PatL[lev], SynchList_cor, StateList, Symmetry);
 #endif
-      Parallel::Sync(GH->PatL[lev - 1], StateList, Symmetry);
+      Parallel::Sync_cached(GH->PatL[lev - 1], StateList, Symmetry, sync_cache_rp_coarse[lev]);
     }
 
-    Parallel::Sync(GH->PatL[lev], SynchList_cor, Symmetry);
+    Parallel::Sync_cached(GH->PatL[lev], SynchList_cor, Symmetry, sync_cache_rp_fine[lev]);
   }
 }
 #undef MIXOUTB

AMSS_NCKU_source/bssn_class.h

@@ -126,6 +126,11 @@ public:
        MyList<var> *OldStateList, *DumpList;
        MyList<var> *ConstraintList;
 
+       Parallel::SyncCache *sync_cache_pre;  // per-level cache for predictor sync
+       Parallel::SyncCache *sync_cache_cor;  // per-level cache for corrector sync
+       Parallel::SyncCache *sync_cache_rp_coarse;  // RestrictProlong sync on PatL[lev-1]
+       Parallel::SyncCache *sync_cache_rp_fine;    // RestrictProlong sync on PatL[lev]
+
        monitor *ErrorMonitor, *Psi4Monitor, *BHMonitor, *MAPMonitor;
        monitor *ConVMonitor;
        surface_integral *Waveshell;

AMSS_NCKU_source/makefile.inc

@@ -10,14 +10,15 @@ filein  = -I/usr/include/ -I${MKLROOT}/include
 ## Added -lifcore for Intel Fortran runtime and -limf for Intel math library
 LDLIBS  = -L${MKLROOT}/lib -lmkl_intel_lp64 -lmkl_sequential -lmkl_core -lifcore -limf -lpthread -lm -ldl
 
-## Aggressive optimization flags:
+## Aggressive optimization flags + PGO Phase 2 (profile-guided optimization)
-## -O3: Maximum optimization
+## -fprofile-instr-use: use collected profile data to guide optimization decisions
-## -xHost: Optimize for the host CPU architecture (Intel/AMD compatible)
+##   (branch prediction, basic block layout, inlining, loop unrolling)
-## -fp-model fast=2: Aggressive floating-point optimizations
+PROFDATA     = ../../pgo_profile/default.profdata
-## -fma: Enable fused multiply-add instructions
 CXXAPPFLAGS  = -O3 -xHost -fp-model fast=2 -fma -ipo \
+               -fprofile-instr-use=$(PROFDATA) \
                -Dfortran3 -Dnewc -I${MKLROOT}/include
 f90appflags  = -O3 -xHost -fp-model fast=2 -fma -ipo \
+               -fprofile-instr-use=$(PROFDATA) \
                -align array64byte -fpp -I${MKLROOT}/include
 f90          = ifx
 f77          = ifx

AMSS_NCKU_source/surface_integral.C

@@ -220,16 +220,9 @@ void surface_integral::surf_Wave(double rex, int lev, cgh *GH, var *Rpsi4, var *
     pox[2][n] = rex * nz_g[n];
   }
 
-  double *shellf;
-  shellf = new double[n_tot * InList];
-
-  GH->PatL[lev]->data->Interp_Points(DG_List, n_tot, pox, shellf, Symmetry);
-
   int mp, Lp, Nmin, Nmax;
-
   mp = n_tot / cpusize;
   Lp = n_tot - cpusize * mp;
-
   if (Lp > myrank)
   {
     Nmin = myrank * mp + myrank;
@@ -241,6 +234,11 @@ void surface_integral::surf_Wave(double rex, int lev, cgh *GH, var *Rpsi4, var *
     Nmax = Nmin + mp - 1;
   }
 
+  double *shellf;
+  shellf = new double[n_tot * InList];
+
+  GH->PatL[lev]->data->Interp_Points(DG_List, n_tot, pox, shellf, Symmetry, Nmin, Nmax);
+
   //|~~~~~> Integrate the dot product of Dphi with the surface normal.
 
   double *RP_out, *IP_out;
@@ -363,8 +361,17 @@ void surface_integral::surf_Wave(double rex, int lev, cgh *GH, var *Rpsi4, var *
   }
   //|------+  Communicate and sum the results from each processor.
 
-  MPI_Allreduce(RP_out, RP, NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+  {
-  MPI_Allreduce(IP_out, IP, NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+    double *RPIP_out = new double[2 * NN];
+    double *RPIP = new double[2 * NN];
+    memcpy(RPIP_out, RP_out, NN * sizeof(double));
+    memcpy(RPIP_out + NN, IP_out, NN * sizeof(double));
+    MPI_Allreduce(RPIP_out, RPIP, 2 * NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+    memcpy(RP, RPIP, NN * sizeof(double));
+    memcpy(IP, RPIP + NN, NN * sizeof(double));
+    delete[] RPIP_out;
+    delete[] RPIP;
+  }
 
   //|------= Free memory.
 
@@ -556,8 +563,17 @@ void surface_integral::surf_Wave(double rex, int lev, cgh *GH, var *Rpsi4, var *
   }
   //|------+  Communicate and sum the results from each processor.
 
-  MPI_Allreduce(RP_out, RP, NN, MPI_DOUBLE, MPI_SUM, Comm_here);
+  {
-  MPI_Allreduce(IP_out, IP, NN, MPI_DOUBLE, MPI_SUM, Comm_here);
+    double *RPIP_out = new double[2 * NN];
+    double *RPIP = new double[2 * NN];
+    memcpy(RPIP_out, RP_out, NN * sizeof(double));
+    memcpy(RPIP_out + NN, IP_out, NN * sizeof(double));
+    MPI_Allreduce(RPIP_out, RPIP, 2 * NN, MPI_DOUBLE, MPI_SUM, Comm_here);
+    memcpy(RP, RPIP, NN * sizeof(double));
+    memcpy(IP, RPIP + NN, NN * sizeof(double));
+    delete[] RPIP_out;
+    delete[] RPIP;
+  }
 
   //|------= Free memory.
 
@@ -735,8 +751,17 @@ void surface_integral::surf_Wave(double rex, int lev, ShellPatch *GH, var *Rpsi4
   }
   //|------+  Communicate and sum the results from each processor.
 
-  MPI_Allreduce(RP_out, RP, NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+  {
-  MPI_Allreduce(IP_out, IP, NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+    double *RPIP_out = new double[2 * NN];
+    double *RPIP = new double[2 * NN];
+    memcpy(RPIP_out, RP_out, NN * sizeof(double));
+    memcpy(RPIP_out + NN, IP_out, NN * sizeof(double));
+    MPI_Allreduce(RPIP_out, RPIP, 2 * NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+    memcpy(RP, RPIP, NN * sizeof(double));
+    memcpy(IP, RPIP + NN, NN * sizeof(double));
+    delete[] RPIP_out;
+    delete[] RPIP;
+  }
 
   //|------= Free memory.
 
@@ -984,8 +1009,17 @@ void surface_integral::surf_Wave(double rex, int lev, ShellPatch *GH,
   }
   //|------+  Communicate and sum the results from each processor.
 
-  MPI_Allreduce(RP_out, RP, NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+  {
-  MPI_Allreduce(IP_out, IP, NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+    double *RPIP_out = new double[2 * NN];
+    double *RPIP = new double[2 * NN];
+    memcpy(RPIP_out, RP_out, NN * sizeof(double));
+    memcpy(RPIP_out + NN, IP_out, NN * sizeof(double));
+    MPI_Allreduce(RPIP_out, RPIP, 2 * NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+    memcpy(RP, RPIP, NN * sizeof(double));
+    memcpy(IP, RPIP + NN, NN * sizeof(double));
+    delete[] RPIP_out;
+    delete[] RPIP;
+  }
 
   //|------= Free memory.
 
@@ -1419,8 +1453,17 @@ void surface_integral::surf_Wave(double rex, int lev, ShellPatch *GH,
   }
   //|------+  Communicate and sum the results from each processor.
 
-  MPI_Allreduce(RP_out, RP, NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+  {
-  MPI_Allreduce(IP_out, IP, NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+    double *RPIP_out = new double[2 * NN];
+    double *RPIP = new double[2 * NN];
+    memcpy(RPIP_out, RP_out, NN * sizeof(double));
+    memcpy(RPIP_out + NN, IP_out, NN * sizeof(double));
+    MPI_Allreduce(RPIP_out, RPIP, 2 * NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+    memcpy(RP, RPIP, NN * sizeof(double));
+    memcpy(IP, RPIP + NN, NN * sizeof(double));
+    delete[] RPIP_out;
+    delete[] RPIP;
+  }
 
   //|------= Free memory.
 
@@ -1854,8 +1897,17 @@ void surface_integral::surf_Wave(double rex, int lev, cgh *GH,
   }
   //|------+  Communicate and sum the results from each processor.
 
-  MPI_Allreduce(RP_out, RP, NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+  {
-  MPI_Allreduce(IP_out, IP, NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+    double *RPIP_out = new double[2 * NN];
+    double *RPIP = new double[2 * NN];
+    memcpy(RPIP_out, RP_out, NN * sizeof(double));
+    memcpy(RPIP_out + NN, IP_out, NN * sizeof(double));
+    MPI_Allreduce(RPIP_out, RPIP, 2 * NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+    memcpy(RP, RPIP, NN * sizeof(double));
+    memcpy(IP, RPIP + NN, NN * sizeof(double));
+    delete[] RPIP_out;
+    delete[] RPIP;
+  }
 
   //|------= Free memory.
 
@@ -2040,8 +2092,17 @@ void surface_integral::surf_Wave(double rex, int lev, NullShellPatch2 *GH, var *
   }
   //|------+  Communicate and sum the results from each processor.
 
-  MPI_Allreduce(RP_out, RP, NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+  {
-  MPI_Allreduce(IP_out, IP, NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+    double *RPIP_out = new double[2 * NN];
+    double *RPIP = new double[2 * NN];
+    memcpy(RPIP_out, RP_out, NN * sizeof(double));
+    memcpy(RPIP_out + NN, IP_out, NN * sizeof(double));
+    MPI_Allreduce(RPIP_out, RPIP, 2 * NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+    memcpy(RP, RPIP, NN * sizeof(double));
+    memcpy(IP, RPIP + NN, NN * sizeof(double));
+    delete[] RPIP_out;
+    delete[] RPIP;
+  }
 
   //|------= Free memory.
 
@@ -2226,8 +2287,17 @@ void surface_integral::surf_Wave(double rex, int lev, NullShellPatch *GH, var *R
   }
   //|------+  Communicate and sum the results from each processor.
 
-  MPI_Allreduce(RP_out, RP, NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+  {
-  MPI_Allreduce(IP_out, IP, NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+    double *RPIP_out = new double[2 * NN];
+    double *RPIP = new double[2 * NN];
+    memcpy(RPIP_out, RP_out, NN * sizeof(double));
+    memcpy(RPIP_out + NN, IP_out, NN * sizeof(double));
+    MPI_Allreduce(RPIP_out, RPIP, 2 * NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+    memcpy(RP, RPIP, NN * sizeof(double));
+    memcpy(IP, RPIP + NN, NN * sizeof(double));
+    delete[] RPIP_out;
+    delete[] RPIP;
+  }
 
   //|------= Free memory.
 
@@ -2314,25 +2384,9 @@ void surface_integral::surf_MassPAng(double rex, int lev, cgh *GH, var *chi, var
     pox[2][n] = rex * nz_g[n];
   }
 
-  double *shellf;
-  shellf = new double[n_tot * InList];
-
-  // we have assumed there is only one box on this level,
-  // so we do not need loop boxes
-  GH->PatL[lev]->data->Interp_Points(DG_List, n_tot, pox, shellf, Symmetry);
-
-  double Mass_out = 0;
-  double ang_outx, ang_outy, ang_outz;
-  double p_outx, p_outy, p_outz;
-  ang_outx = ang_outy = ang_outz = 0.0;
-  p_outx = p_outy = p_outz = 0.0;
-  const double f1o8 = 0.125;
-
   int mp, Lp, Nmin, Nmax;
-
   mp = n_tot / cpusize;
   Lp = n_tot - cpusize * mp;
-
   if (Lp > myrank)
   {
     Nmin = myrank * mp + myrank;
@@ -2344,6 +2398,20 @@ void surface_integral::surf_MassPAng(double rex, int lev, cgh *GH, var *chi, var
     Nmax = Nmin + mp - 1;
   }
 
+  double *shellf;
+  shellf = new double[n_tot * InList];
+
+  // we have assumed there is only one box on this level,
+  // so we do not need loop boxes
+  GH->PatL[lev]->data->Interp_Points(DG_List, n_tot, pox, shellf, Symmetry, Nmin, Nmax);
+
+  double Mass_out = 0;
+  double ang_outx, ang_outy, ang_outz;
+  double p_outx, p_outy, p_outz;
+  ang_outx = ang_outy = ang_outz = 0.0;
+  p_outx = p_outy = p_outz = 0.0;
+  const double f1o8 = 0.125;
+
   double Chi, Psi;
   double Gxx, Gxy, Gxz, Gyy, Gyz, Gzz;
   double gupxx, gupxy, gupxz, gupyy, gupyz, gupzz;
@@ -2464,15 +2532,13 @@ void surface_integral::surf_MassPAng(double rex, int lev, cgh *GH, var *chi, var
     }
   }
 
-  MPI_Allreduce(&Mass_out, &mass, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+  {
-
+    double scalar_out[7] = {Mass_out, ang_outx, ang_outy, ang_outz, p_outx, p_outy, p_outz};
-  MPI_Allreduce(&ang_outx, &sx, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+    double scalar_in[7];
-  MPI_Allreduce(&ang_outy, &sy, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+    MPI_Allreduce(scalar_out, scalar_in, 7, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-  MPI_Allreduce(&ang_outz, &sz, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+    mass = scalar_in[0]; sx = scalar_in[1]; sy = scalar_in[2]; sz = scalar_in[3];
-
+    px = scalar_in[4]; py = scalar_in[5]; pz = scalar_in[6];
-  MPI_Allreduce(&p_outx, &px, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+  }
-  MPI_Allreduce(&p_outy, &py, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-  MPI_Allreduce(&p_outz, &pz, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
 
 #ifdef GaussInt
   mass = mass * rex * rex * dphi * factor;
@@ -2735,15 +2801,13 @@ void surface_integral::surf_MassPAng(double rex, int lev, cgh *GH, var *chi, var
     }
   }
 
-  MPI_Allreduce(&Mass_out, &mass, 1, MPI_DOUBLE, MPI_SUM, Comm_here);
+  {
-
+    double scalar_out[7] = {Mass_out, ang_outx, ang_outy, ang_outz, p_outx, p_outy, p_outz};
-  MPI_Allreduce(&ang_outx, &sx, 1, MPI_DOUBLE, MPI_SUM, Comm_here);
+    double scalar_in[7];
-  MPI_Allreduce(&ang_outy, &sy, 1, MPI_DOUBLE, MPI_SUM, Comm_here);
+    MPI_Allreduce(scalar_out, scalar_in, 7, MPI_DOUBLE, MPI_SUM, Comm_here);
-  MPI_Allreduce(&ang_outz, &sz, 1, MPI_DOUBLE, MPI_SUM, Comm_here);
+    mass = scalar_in[0]; sx = scalar_in[1]; sy = scalar_in[2]; sz = scalar_in[3];
-
+    px = scalar_in[4]; py = scalar_in[5]; pz = scalar_in[6];
-  MPI_Allreduce(&p_outx, &px, 1, MPI_DOUBLE, MPI_SUM, Comm_here);
+  }
-  MPI_Allreduce(&p_outy, &py, 1, MPI_DOUBLE, MPI_SUM, Comm_here);
-  MPI_Allreduce(&p_outz, &pz, 1, MPI_DOUBLE, MPI_SUM, Comm_here);
 
 #ifdef GaussInt
   mass = mass * rex * rex * dphi * factor;
@@ -3020,15 +3084,13 @@ void surface_integral::surf_MassPAng(double rex, int lev, ShellPatch *GH, var *c
     }
   }
 
-  MPI_Allreduce(&Mass_out, &mass, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+  {
-
+    double scalar_out[7] = {Mass_out, ang_outx, ang_outy, ang_outz, p_outx, p_outy, p_outz};
-  MPI_Allreduce(&ang_outx, &sx, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+    double scalar_in[7];
-  MPI_Allreduce(&ang_outy, &sy, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+    MPI_Allreduce(scalar_out, scalar_in, 7, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-  MPI_Allreduce(&ang_outz, &sz, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+    mass = scalar_in[0]; sx = scalar_in[1]; sy = scalar_in[2]; sz = scalar_in[3];
-
+    px = scalar_in[4]; py = scalar_in[5]; pz = scalar_in[6];
-  MPI_Allreduce(&p_outx, &px, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+  }
-  MPI_Allreduce(&p_outy, &py, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-  MPI_Allreduce(&p_outz, &pz, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
 
 #ifdef GaussInt
   mass = mass * rex * rex * dphi * factor;
@@ -3607,8 +3669,17 @@ void surface_integral::surf_Wave(double rex, cgh *GH, ShellPatch *SH,
   }
   //|------+  Communicate and sum the results from each processor.
 
-  MPI_Allreduce(RP_out, RP, NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+  {
-  MPI_Allreduce(IP_out, IP, NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+    double *RPIP_out = new double[2 * NN];
+    double *RPIP = new double[2 * NN];
+    memcpy(RPIP_out, RP_out, NN * sizeof(double));
+    memcpy(RPIP_out + NN, IP_out, NN * sizeof(double));
+    MPI_Allreduce(RPIP_out, RPIP, 2 * NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+    memcpy(RP, RPIP, NN * sizeof(double));
+    memcpy(IP, RPIP + NN, NN * sizeof(double));
+    delete[] RPIP_out;
+    delete[] RPIP;
+  }
 
   //|------= Free memory.
 

makefile_and_run.py

@@ -15,12 +15,13 @@ import time
 ## taskset ensures all child processes inherit the CPU affinity mask
 ## This forces make and all compiler processes to use only nohz_full cores (4-55, 60-111)
 ## Format: taskset -c 4-55,60-111 ensures processes only run on these cores
-NUMACTL_CPU_BIND = "taskset -c 0-111"
+#NUMACTL_CPU_BIND = "taskset -c 0-111"
+NUMACTL_CPU_BIND = "taskset -c 16-47,64-95"
 
 ## Build parallelism configuration
 ## Use nohz_full cores (4-55, 60-111) for compilation: 52 + 52 = 104 cores
 ## Set make -j to utilize available cores for faster builds
-BUILD_JOBS = 104
+BUILD_JOBS = 96
 
 
 ##################################################################
@@ -117,6 +118,7 @@ def run_ABE():
     
     if (input_data.GPU_Calculation == "no"):
         mpi_command         = NUMACTL_CPU_BIND + " mpirun -np " + str(input_data.MPI_processes) + " ./ABE"
+        #mpi_command         = " mpirun -np " + str(input_data.MPI_processes) + " ./ABE"
         mpi_command_outfile = "ABE_out.log"
     elif (input_data.GPU_Calculation == "yes"):
         mpi_command         = NUMACTL_CPU_BIND + " mpirun -np " + str(input_data.MPI_processes) + " ./ABEGPU"
@@ -158,7 +160,8 @@ def run_TwoPunctureABE():
     print(                                                          )
     
     ## Define the command to run
-    TwoPuncture_command         = NUMACTL_CPU_BIND + " ./TwoPunctureABE"
+    #TwoPuncture_command         = NUMACTL_CPU_BIND + " ./TwoPunctureABE"
+    TwoPuncture_command         = " ./TwoPunctureABE"
     TwoPuncture_command_outfile = "TwoPunctureABE_out.log"
 
     ## Execute the command with subprocess.Popen and stream output

pgo_profile/PGO_Profile_Analysis.md

@@ -0,0 +1,97 @@
+# AMSS-NCKU PGO Profile Analysis Report
+
+## 1. Profiling Environment
+
+| Item | Value |
+|------|-------|
+| Compiler | Intel oneAPI DPC++/C++ 2025.3.0 (icpx/ifx) |
+| Instrumentation Flag | `-fprofile-instr-generate` |
+| Optimization Level (instrumented) | `-O2 -xHost -fma` |
+| MPI Processes | 1 (single process to avoid MPI+instrumentation deadlock) |
+| Profile File | `default_9725750769337483397_0.profraw` (327 KB) |
+| Merged Profile | `default.profdata` (394 KB) |
+| llvm-profdata | `/home/intel/oneapi/compiler/2025.3/bin/compiler/llvm-profdata` |
+
+## 2. Reduced Simulation Parameters (for profiling run)
+
+| Parameter | Production Value | Profiling Value |
+|-----------|-----------------|-----------------|
+| MPI_processes | 64 | 1 |
+| grid_level | 9 | 4 |
+| static_grid_level | 5 | 3 |
+| static_grid_number | 96 | 24 |
+| moving_grid_number | 48 | 16 |
+| largest_box_xyz_max | 320^3 | 160^3 |
+| Final_Evolution_Time | 1000.0 | 10.0 |
+| Evolution_Step_Number | 10,000,000 | 1,000 |
+| Detector_Number | 12 | 2 |
+
+## 3. Profile Summary
+
+| Metric | Value |
+|--------|-------|
+| Total instrumented functions | 1,392 |
+| Functions with non-zero counts | 117 (8.4%) |
+| Functions with zero counts | 1,275 (91.6%) |
+| Maximum function entry count | 386,459,248 |
+| Maximum internal block count | 370,477,680 |
+| Total block count | 4,198,023,118 |
+
+## 4. Top 20 Hotspot Functions
+
+| Rank | Total Count | Max Block Count | Function | Category |
+|------|------------|-----------------|----------|----------|
+| 1 | 1,241,601,732 | 370,477,680 | `polint_` | Interpolation |
+| 2 | 755,994,435 | 230,156,640 | `prolong3_` | Grid prolongation |
+| 3 | 667,964,095 | 3,697,792 | `compute_rhs_bssn_` | BSSN RHS evolution |
+| 4 | 539,736,051 | 386,459,248 | `symmetry_bd_` | Symmetry boundary |
+| 5 | 277,310,808 | 53,170,728 | `lopsided_` | Lopsided FD stencil |
+| 6 | 155,534,488 | 94,535,040 | `decide3d_` | 3D grid decision |
+| 7 | 119,267,712 | 19,266,048 | `rungekutta4_rout_` | RK4 time integrator |
+| 8 | 91,574,616 | 48,824,160 | `kodis_` | Kreiss-Oliger dissipation |
+| 9 | 67,555,389 | 43,243,680 | `fderivs_` | Finite differences |
+| 10 | 55,296,000 | 42,246,144 | `misc::fact(int)` | Factorial utility |
+| 11 | 43,191,071 | 27,663,328 | `fdderivs_` | 2nd-order FD derivatives |
+| 12 | 36,233,965 | 22,429,440 | `restrict3_` | Grid restriction |
+| 13 | 24,698,512 | 17,231,520 | `polin3_` | Polynomial interpolation |
+| 14 | 22,962,942 | 20,968,768 | `copy_` | Data copy |
+| 15 | 20,135,696 | 17,259,168 | `Ansorg::barycentric(...)` | Spectral interpolation |
+| 16 | 14,650,224 | 7,224,768 | `Ansorg::barycentric_omega(...)` | Spectral weights |
+| 17 | 13,242,296 | 2,871,920 | `global_interp_` | Global interpolation |
+| 18 | 12,672,000 | 7,734,528 | `sommerfeld_rout_` | Sommerfeld boundary |
+| 19 | 6,872,832 | 1,880,064 | `sommerfeld_routbam_` | Sommerfeld boundary (BAM) |
+| 20 | 5,709,900 | 2,809,632 | `l2normhelper_` | L2 norm computation |
+
+## 5. Hotspot Category Breakdown
+
+Top 20 functions account for ~98% of total execution counts:
+
+| Category | Functions | Combined Count | Share |
+|----------|-----------|---------------|-------|
+| Interpolation / Prolongation / Restriction | polint_, prolong3_, restrict3_, polin3_, global_interp_, Ansorg::* | ~2,093M | ~50% |
+| BSSN RHS + FD stencils | compute_rhs_bssn_, lopsided_, fderivs_, fdderivs_ | ~1,056M | ~25% |
+| Boundary conditions | symmetry_bd_, sommerfeld_rout_, sommerfeld_routbam_ | ~559M | ~13% |
+| Time integration | rungekutta4_rout_ | ~119M | ~3% |
+| Dissipation | kodis_ | ~92M | ~2% |
+| Utilities | misc::fact, decide3d_, copy_, l2normhelper_ | ~256M | ~6% |
+
+## 6. Conclusions
+
+1. **Profile data is valid**: 1,392 functions instrumented, 117 exercised with ~4.2 billion total counts.
+2. **Hotspot concentration is high**: Top 5 functions alone account for ~76% of all counts, which is ideal for PGO — the compiler has strong branch/layout optimization targets.
+3. **Fortran numerical kernels dominate**: `polint_`, `prolong3_`, `compute_rhs_bssn_`, `symmetry_bd_`, `lopsided_` are all Fortran routines in the inner evolution loop. PGO will optimize their branch prediction and basic block layout.
+4. **91.6% of functions have zero counts**: These are code paths for unused features (GPU, BSSN-EScalar, BSSN-EM, Z4C, etc.). PGO will deprioritize them, improving instruction cache utilization.
+5. **Profile is representative**: Despite the reduced grid size, the code path coverage matches production — the same kernels (RHS, prolongation, restriction, boundary) are exercised. PGO branch probabilities from this profile will transfer well to full-scale runs.
+
+## 7. PGO Phase 2 Usage
+
+To apply the profile, use the following flags in `makefile.inc`:
+
+```makefile
+CXXAPPFLAGS = -O3 -xHost -fp-model fast=2 -fma -ipo \
+              -fprofile-instr-use=/home/amss/AMSS-NCKU/pgo_profile/default.profdata \
+              -Dfortran3 -Dnewc -I${MKLROOT}/include
+f90appflags = -O3 -xHost -fp-model fast=2 -fma -ipo \
+              -fprofile-instr-use=/home/amss/AMSS-NCKU/pgo_profile/default.profdata \
+              -align array64byte -fpp -I${MKLROOT}/include
+```