MPI 点对点通信到集体通信

Question

我正在学习 MPI，我正在尝试将我的 MPI 程序从点对点通信转换为 MPI Collectives ..

下面是我使用 MPI 点对点通信进行矩阵乘法的代码片段...

int i;
    if(rank == 0) {
        for(i = 1; i < size; i++){
            MPI_Send(&rows, 1, MPI_INT, i, 0, MPI_COMM_WORLD);
            MPI_Send(&columns, 1, MPI_INT, i, 0, MPI_COMM_WORLD);
        }
    } else {
        MPI_Recv(&rows, 1, MPI_INT, 0, 0, MPI_COMM_WORLD, &status);
        MPI_Recv(&columns, 1, MPI_INT, 0, 0, MPI_COMM_WORLD, &status);
    }   

    int local_block_size = rows / size;
    int process, column_pivot;

    if(rank == 0) {
        for(i = 1; i < size; i++){
            MPI_Send((matrix_1D_mapped + (i * (local_block_size * rows))), (local_block_size * rows), MPI_DOUBLE, i, 0, MPI_COMM_WORLD);
            MPI_Send((rhs + (i * local_block_size)), local_block_size, MPI_DOUBLE, i, 0, MPI_COMM_WORLD);
        }
        for(i = 0; i < local_block_size * rows; i++){
            matrix_local_block[i] = matrix_1D_mapped[i];
        }
        for(i = 0; i < local_block_size; i++){
            rhs_local_block[i] = rhs[i];
        }
    } else {
        MPI_Recv(matrix_local_block, local_block_size * rows, MPI_DOUBLE, 0, 0, MPI_COMM_WORLD, &status);
        MPI_Recv(rhs_local_block, local_block_size, MPI_DOUBLE, 0, 0, MPI_COMM_WORLD, &status);
    }

我正在考虑将 MPI_Send 替换为 MPI_Bcast ......这是正确的方法吗？

Answer 1

对于第一次通信，发送给所有接收者的数据实际上是相同的，因此MPI_Bcast是正确的做法。第二次通信将更大数组的不同块分发给接收者，这是作为一个集体与 MPI_Scatter 一起完成的。注意scatter在通信中包含了root rank，所以可以省略手动本地copy

MPI 点对点通信到集体通信

MPI Point to Point Communication to Collective Communication

c++

mpi