我已经实现了一个有效的MPI send/recv，但是我想更好地组织和提高性能，所以我采用了std：：vector和std：：map，并以块的形式发送数据，而不是单独发送。

// working method (Individual send/recv)


// ... here the struct fvm::F is turned into an MPI datatype STRUCT_FACE to be used ...

// main rank 0

map<int, vector<fvm::F>> faces_buf;

// ... here faces_buf is initialized and the inside vectors are properly resized using vector.resize() ...

// the collection loop in main rank 0 (collects all faces from compute ranks 1, 2, 3, ...)
for(auto& fb : faces_buf){
   int compute_rank = fb.first;
   int n = 0;
   for(auto& f : fb.second){
      MPI_Recv(&f, 1, STRUCT_FACE, compute_rank, (n++ + 100000), MPI_COMM_WORLD, MPI_STATUS_IGNORE);
   }
}

// print results of collection
for(auto& fb : faces_buf){
    for(auto& f : fb.second){
        cerr << "f.T " << f.T << endl;
        if(f.T == 0.){
            goto done;
        }
    }
}
done:;



// compute ranks 1, 2, 3, ...

vector<fvm::F> faces;

// ... here faces vector is filled with values ...

for(auto& f : faces){
   MPI_Send(&f, 1, STRUCT_FACE, 0, (f.id + 100000), MPI_COMM_WORLD);
}

// Works as intended!

。
。

// failing method (Chunk send/recv)

// ... here the struct fvm::F is turned into an MPI datatype STRUCT_FACE to be used ...

// main rank 0

map<int, vector<fvm::F>> faces_buf;

// ... here faces_buf is initialized and the inside vectors are properly resized using vector.resize() ...

// the collection loop in main rank 0 (collects all faces from compute ranks 1, 2, 3, ...)
for(auto& fb : faces_buf){
   int compute_rank = fb.first;
   MPI_Recv(fb.second.data(), fb.second.size(), STRUCT_FACE, compute_rank, (compute_rank + 100000), MPI_COMM_WORLD, MPI_STATUS_IGNORE);
}

// print results of collection
for(auto& fb : faces_buf){
    for(auto& f : fb.second){
        cerr << "f.T " << f.T << endl;
        if(f.T == 0.){
            goto done;
        }
    }
}
done:;


// compute ranks 1, 2, 3, ...

vector<fvm::F> faces;

// ... here faces vector is filled with values ...

MPI_Isend(faces.data(), faces.size(), STRUCT_FACE, 0, (rank + 100000), MPI_COMM_WORLD, &reqf);
int flagf = 0;
while(true){
   MPI_Test(&reqf, &flagf, MPI_STATUS_IGNORE);
   if(flagf){
      break;
   }
}

// Fails to recv new fvm::F values

工作方法的输出：

f.T 298.15
f.T 298.15
f.T 298.15
f.T 298.15
f.T 298.15
...
...
...

失败方法的输出：

f.T 298.15
f.T 298.15
f.T 298.15
f.T 8.48798e-315
f.T 0

为什么第二种方法会失败，而第一种方法会成功？