MPI_Win_fence（）有什么用

Question

以下代码为我提供了正确答案

MPI_Win win;
MPI_Win_create(xyz, nsd*nnc, sizeof(double), MPI_INFO_NULL, MPI_COMM_WORLD, &win);
MPI_Win_fence(0, win);

for(i=0; i<nnl; i++)
{
    int gb_node_num = nodeLToG[i];
    int rank_to_get_from = gb_node_num/mnc;
    int buffer_position = gb_node_num%mnc;

    MPI_Get(&x_cord[i], 1, MPI_DOUBLE, rank_to_get_from, nsd*buffer_position, 1, MPI_DOUBLE, win);
    MPI_Get(&y_cord[i], 1, MPI_DOUBLE, rank_to_get_from, nsd*buffer_position+1, 1, MPI_DOUBLE, win);    
}   


MPI_Win_fence(0, win);

for(i=0;i<nnl;i++)
{
    lNode[i].setX(x_cord[i]);
    lNode[i].setY(y_cord[i]);   
}

如果我将MPI_Win_fence（）进一步向下移动，如下所示，则不会得到结果。为什么会有所作为？

for(i=0; i<nnl; i++)
{
    int gb_node_num = nodeLToG[i];
    int rank_to_get_from = gb_node_num/mnc;
    int buffer_position = gb_node_num%mnc;

    MPI_Get(&x_cord[i], 1, MPI_DOUBLE, rank_to_get_from, nsd*buffer_position, 1, MPI_DOUBLE, win);
    MPI_Get(&y_cord[i], 1, MPI_DOUBLE, rank_to_get_from, nsd*buffer_position+1, 1, MPI_DOUBLE, win);    
}   


for(i=0;i<nnl;i++)
{
    lNode[i].setX(x_cord[i]);
    lNode[i].setY(y_cord[i]);   
}

MPI_Win_fence(0, win);

我们应该如何使用MPI_Win_fence（）？你能指出一些参考吗？