1. Pass 1（357-395行）：遍历所有 Patch，对每个 block 计算含ghost zone 的实际体积，存入 block_volumes

2. Greedy LPT（397-414行）：按体积从大到小排序，依次分配给当前负载最小的 rank
  3. Pass 2（416-555行）：原来的 block创建循环，但用 assigned_ranks[block_idx++] 替代 n_rank++，Block
  构造时直接拿到正确的 rank，内存分配在对的进程上

AMSS_NCKU_source/Parallel.C

@@ -4,6 +4,8 @@
 #include "prolongrestrict.h"
 #include "misc.h"
 #include "parameters.h"
+#include <vector>
+#include <algorithm>
 
 int Parallel::partition1(int &nx, int split_size, int min_width, int cpusize, int shape) // special for 1 diemnsion
 {
@@ -72,14 +74,14 @@ int Parallel::partition3(int *nxyz, int split_size, int *min_width, int cpusize,
   int n;
 
   block_size = shape[0] * shape[1] * shape[2];
-  n = Mymax(1, (block_size + split_size / 2) / split_size);
+  n =  Mymax(1, (block_size + split_size / 2) / split_size);
 
   maxnx = Mymax(1, shape[0] / min_width[0]);
-  maxnx = Mymin(cpusize, maxnx);
+  maxnx =  Mymin(cpusize, maxnx);
   maxny = Mymax(1, shape[1] / min_width[1]);
-  maxny = Mymin(cpusize, maxny);
+  maxny =  Mymin(cpusize, maxny);
   maxnz = Mymax(1, shape[2] / min_width[2]);
-  maxnz = Mymin(cpusize, maxnz);
+  maxnz =  Mymin(cpusize, maxnz);
   fx = (double)shape[0] / (shape[0] + shape[1] + shape[2]);
   fy = (double)shape[1] / (shape[0] + shape[1] + shape[2]);
   fz = (double)shape[2] / (shape[0] + shape[1] + shape[2]);
@@ -352,14 +354,73 @@ MyList<Block> *Parallel::distribute(MyList<Patch> *PatchLIST, int cpusize, int i
   split_size = Mymax(min_size, block_size / nodes);
   split_size = Mymax(1, split_size);
 
-  int n_rank = 0;
+  // Pass 1: compute block volumes for greedy rank assignment
+  std::vector<long> block_volumes;
   PLi = PatchLIST;
   int reacpu = 0;
   while (PLi)
   {
     Patch *PP = PLi->data;
-
     reacpu += partition3(nxyz, split_size, mmin_width, nodes, PP->shape);
+    int ibbox_here[2 * dim];
+    for (int i = 0; i < nxyz[0]; i++)
+      for (int j = 0; j < nxyz[1]; j++)
+        for (int k = 0; k < nxyz[2]; k++)
+        {
+          ibbox_here[0] = (PP->shape[0] * i) / nxyz[0];
+          ibbox_here[3] = (PP->shape[0] * (i + 1)) / nxyz[0] - 1;
+          ibbox_here[1] = (PP->shape[1] * j) / nxyz[1];
+          ibbox_here[4] = (PP->shape[1] * (j + 1)) / nxyz[1] - 1;
+          ibbox_here[2] = (PP->shape[2] * k) / nxyz[2];
+          ibbox_here[5] = (PP->shape[2] * (k + 1)) / nxyz[2] - 1;
+          if (periodic)
+          {
+            for (int d = 0; d < dim; d++) { ibbox_here[d] -= ghost_width; ibbox_here[dim + d] += ghost_width; }
+          }
+          else
+          {
+            ibbox_here[0] = Mymax(0, ibbox_here[0] - ghost_width);
+            ibbox_here[3] = Mymin(PP->shape[0] - 1, ibbox_here[3] + ghost_width);
+            ibbox_here[1] = Mymax(0, ibbox_here[1] - ghost_width);
+            ibbox_here[4] = Mymin(PP->shape[1] - 1, ibbox_here[4] + ghost_width);
+            ibbox_here[2] = Mymax(0, ibbox_here[2] - ghost_width);
+            ibbox_here[5] = Mymin(PP->shape[2] - 1, ibbox_here[5] + ghost_width);
+          }
+          long vol = 1;
+          for (int d = 0; d < dim; d++)
+            vol *= (ibbox_here[dim + d] - ibbox_here[d] + 1);
+          block_volumes.push_back(vol);
+        }
+    PLi = PLi->next;
+  }
+
+  // Greedy LPT: sort by volume descending, assign each to least-loaded rank
+  std::vector<int> assigned_ranks(block_volumes.size());
+  {
+    std::vector<int> order(block_volumes.size());
+    for (int i = 0; i < (int)order.size(); i++) order[i] = i;
+    std::sort(order.begin(), order.end(), [&](int a, int b) {
+      return block_volumes[a] > block_volumes[b];
+    });
+    std::vector<long> load(cpusize, 0);
+    for (int idx : order)
+    {
+      int min_r = 0;
+      for (int r = 1; r < cpusize; r++)
+        if (load[r] < load[min_r]) min_r = r;
+      assigned_ranks[idx] = min_r;
+      load[min_r] += block_volumes[idx];
+    }
+  }
+
+  // Pass 2: create blocks with pre-assigned ranks
+  int block_idx = 0;
+  PLi = PatchLIST;
+  while (PLi)
+  {
+    Patch *PP = PLi->data;
+
+    partition3(nxyz, split_size, mmin_width, nodes, PP->shape);
 
     Block *ng0, *ng;
     int shape_here[dim], ibbox_here[2 * dim];
@@ -443,10 +504,7 @@ MyList<Block> *Parallel::distribute(MyList<Patch> *PatchLIST, int cpusize, int i
             int shape_res[dim * pices];
             double bbox_res[2 * dim * pices];
             misc::dividBlock(dim, shape_here, bbox_here, pices, picef, shape_res, bbox_res, min_width);
-            ng = ng0 = new Block(dim, shape_res, bbox_res, n_rank++, ingfsi, fngfsi, PP->lev, 0); // delete through KillBlocks
+            ng = ng0 = new Block(dim, shape_res, bbox_res, assigned_ranks[block_idx++], ingfsi, fngfsi, PP->lev, 0); // delete through KillBlocks
-
-            //	       if(n_rank==cpusize) {n_rank=0; cerr<<"place one!!"<<endl;}
-
             //	       ng->checkBlock();
             if (BlL)
               BlL->insert(ng);
@@ -455,22 +513,19 @@ MyList<Block> *Parallel::distribute(MyList<Patch> *PatchLIST, int cpusize, int i
 
             for (int i = 1; i < pices; i++)
             {
-              ng = new Block(dim, shape_res + i * dim, bbox_res + i * 2 * dim, n_rank++, ingfsi, fngfsi, PP->lev, i); // delete through KillBlocks
+              ng = new Block(dim, shape_res + i * dim, bbox_res + i * 2 * dim, assigned_ranks[block_idx++], ingfsi, fngfsi, PP->lev, i); // delete through KillBlocks
-              //	        if(n_rank==cpusize) {n_rank=0; cerr<<"place two!! "<<i<<endl;}
               //	        ng->checkBlock();
               BlL->insert(ng);
             }
           }
 #else
-          ng = ng0 = new Block(dim, shape_here, bbox_here, n_rank++, ingfsi, fngfsi, PP->lev); // delete through KillBlocks
+          ng = ng0 = new Block(dim, shape_here, bbox_here, assigned_ranks[block_idx++], ingfsi, fngfsi, PP->lev); // delete through KillBlocks
           //	    ng->checkBlock();
           if (BlL)
             BlL->insert(ng);
           else
             BlL = new MyList<Block>(ng); // delete through KillBlocks
 #endif
-          if (n_rank == cpusize)
-            n_rank = 0;
 
           // set PP->blb
           if (i == 0 && j == 0 && k == 0)
@@ -3504,7 +3559,7 @@ int Parallel::data_packermix(double *data, MyList<Parallel::gridseg> *src, MyLis
 
   return size_out;
 }
-//
+
 void Parallel::transfer(MyList<Parallel::gridseg> **src, MyList<Parallel::gridseg> **dst,
                         MyList<var> *VarList1 /* source */, MyList<var> *VarList2 /*target */,
                         int Symmetry)
@@ -3512,13 +3567,20 @@ void Parallel::transfer(MyList<Parallel::gridseg> **src, MyList<Parallel::gridse
   int myrank, cpusize;
   MPI_Comm_size(MPI_COMM_WORLD, &cpusize);
   MPI_Comm_rank(MPI_COMM_WORLD, &myrank);
-
+/*
+  // Early exit: if no gridseg pairs exist for any node, skip all work
+  {
+    bool has_segs = false;
+    for (int n = 0; n < cpusize; n++) {
+      if (src[n] && dst[n]) { has_segs = true; break; }
+    }
+    if (!has_segs) return;
+  }
+*/
   int node;
 
-  MPI_Request *reqs;
+  MPI_Request *reqs = new MPI_Request[2 * cpusize];
-  MPI_Status *stats;
+  MPI_Status *stats = new MPI_Status[2 * cpusize];
-  reqs = new MPI_Request[2 * cpusize];
-  stats = new MPI_Status[2 * cpusize];
   int req_no = 0;
 
   double **send_data, **rec_data;
@@ -3527,49 +3589,41 @@ void Parallel::transfer(MyList<Parallel::gridseg> **src, MyList<Parallel::gridse
   int length;
 
   for (node = 0; node < cpusize; node++)
-  {
     send_data[node] = rec_data[node] = 0;
+
+  // 第1步: 本地拷贝 + 所有 Irecv
+  for (node = 0; node < cpusize; node++)
+  {
     if (node == myrank)
     {
       if (length = data_packer(0, src[myrank], dst[myrank], node, PACK, VarList1, VarList2, Symmetry))
       {
         rec_data[node] = new double[length];
-        if (!rec_data[node])
-        {
-          cout << "out of memory when new in short transfer, place 1" << endl;
-          MPI_Abort(MPI_COMM_WORLD, 1);
-        }
         data_packer(rec_data[node], src[myrank], dst[myrank], node, PACK, VarList1, VarList2, Symmetry);
       }
     }
     else
     {
-      // send from this cpu to cpu#node
-      if (length = data_packer(0, src[myrank], dst[myrank], node, PACK, VarList1, VarList2, Symmetry))
-      {
-        send_data[node] = new double[length];
-        if (!send_data[node])
-        {
-          cout << "out of memory when new in short transfer, place 2" << endl;
-          MPI_Abort(MPI_COMM_WORLD, 1);
-        }
-        data_packer(send_data[node], src[myrank], dst[myrank], node, PACK, VarList1, VarList2, Symmetry);
-        MPI_Isend((void *)send_data[node], length, MPI_DOUBLE, node, 1, MPI_COMM_WORLD, reqs + req_no++);
-      }
-      // receive from cpu#node to this cpu
       if (length = data_packer(0, src[node], dst[node], node, UNPACK, VarList1, VarList2, Symmetry))
       {
         rec_data[node] = new double[length];
-        if (!rec_data[node])
-        {
-          cout << "out of memory when new in short transfer, place 3" << endl;
-          MPI_Abort(MPI_COMM_WORLD, 1);
-        }
         MPI_Irecv((void *)rec_data[node], length, MPI_DOUBLE, node, 1, MPI_COMM_WORLD, reqs + req_no++);
       }
     }
   }
-  // wait for all requests to complete
+
+  // 第2步: pack + Isend
+  for (node = 0; node < cpusize; node++)
+  {
+    if (node == myrank) continue;
+    if (length = data_packer(0, src[myrank], dst[myrank], node, PACK, VarList1, VarList2, Symmetry))
+    {
+      send_data[node] = new double[length];
+      data_packer(send_data[node], src[myrank], dst[myrank], node, PACK, VarList1, VarList2, Symmetry);
+      MPI_Isend((void *)send_data[node], length, MPI_DOUBLE, node, 1, MPI_COMM_WORLD, reqs + req_no++);
+    }
+  }
+
   MPI_Waitall(req_no, reqs, stats);
 
   for (node = 0; node < cpusize; node++)