MPI 中的 ISend 和 Recv：收到不同的值？

Question

在我的矩阵加法代码中，我使用 ISend 和标记 1 将下限传输到其他进程，但是当我编译代码时，所有其他从属进程都声称具有相同的下限。我不明白为什么？

输出：

I am process 1 and I received 1120 as lower bound
I am process 1 and my lower bound is 1120 and my upper bound is 1682
I am process 2 and I received 1120 as lower bound
I am process 2 and my lower bound is 1120 and my upper bound is 1682
Process 0 here: I am sending lower bound 0 to process 1
Process 0 here: I am sending lower bound 560 to process 2
Process 0 here: I am sending lower bound 1120 to process 3
Timings : 13.300698 Sec
I am process 3 and I received 1120 as lower bound
I am process 3 and my lower bound is 1120 and my upper bound is 1682

代码：

#define N_ROWS 1682
#define N_COLS 823
#define MASTER_TO_SLAVE_TAG 1 //tag for messages sent from master to slaves
#define SLAVE_TO_MASTER_TAG 4 //tag for messages sent from slaves to master

void readMatrix();
int rank, nproc, proc;
double matrix_A[N_ROWS][N_COLS];
double matrix_B[N_ROWS][N_COLS];
double matrix_C[N_ROWS][N_COLS];
int low_bound; //low bound of the number of rows of [A] allocated to a slave
int upper_bound; //upper bound of the number of rows of [A] allocated to a slave
int portion; //portion of the number of rows of [A] allocated to a slave
MPI_Status status; // store status of a MPI_Recv
MPI_Request request; //capture request of a MPI_Isend

int main (int argc, char *argv[]) {

    MPI_Init(&argc, &argv);
    MPI_Comm_size(MPI_COMM_WORLD, &nproc);
    MPI_Comm_rank(MPI_COMM_WORLD, &rank);

    double StartTime = MPI_Wtime();

        // -------------------> Process 0 initalizes matrices and sends work portions to other processes
        if (rank==0) {
            readMatrix();
            for (proc = 1; proc < nproc; proc++) {//for each slave other than the master
                portion = (N_ROWS / (nproc - 1)); // calculate portion without master
                low_bound = (proc - 1) * portion;
                if (((proc + 1) == nproc) && ((N_ROWS % (nproc - 1)) != 0)) {//if rows of [A] cannot be equally divided among slaves
                    upper_bound = N_ROWS; //last slave gets all the remaining rows
                } else {
                    upper_bound = low_bound + portion; //rows of [A] are equally divisable among slaves
                }
                //send the low bound first without blocking, to the intended slave
                printf("Process 0 here: I am sending lower bound %i to process %i \n",low_bound,proc);
                MPI_Isend(&low_bound, 1, MPI_INT, proc, MASTER_TO_SLAVE_TAG, MPI_COMM_WORLD, &request);
                //next send the upper bound without blocking, to the intended slave
                MPI_Isend(&upper_bound, 1, MPI_INT, proc, MASTER_TO_SLAVE_TAG + 1, MPI_COMM_WORLD, &request);
                //finally send the allocated row portion of [A] without blocking, to the intended slave
                MPI_Isend(&matrix_A[low_bound][0], (upper_bound - low_bound) * N_COLS, MPI_DOUBLE, proc, MASTER_TO_SLAVE_TAG + 2, MPI_COMM_WORLD, &request);
            }
        }

        //broadcast [B] to all the slaves
        MPI_Bcast(&matrix_B, N_ROWS*N_COLS, MPI_DOUBLE, 0, MPI_COMM_WORLD);


        // -------------------> Other processes do their work
        if (rank != 0) {
            //receive low bound from the master
            MPI_Recv(&low_bound, 1, MPI_INT, 0, MASTER_TO_SLAVE_TAG, MPI_COMM_WORLD, &status);
            printf("I am process %i and I received %i as lower bound \n",rank,low_bound);
            //next receive upper bound from the master
            MPI_Recv(&upper_bound, 1, MPI_INT, 0, MASTER_TO_SLAVE_TAG + 1, MPI_COMM_WORLD, &status);
            //finally receive row portion of [A] to be processed from the master
            MPI_Recv(&matrix_A[low_bound][0], (upper_bound - low_bound) * N_COLS, MPI_DOUBLE, 0, MASTER_TO_SLAVE_TAG + 2, MPI_COMM_WORLD, &status);
            printf("I am process %i and my lower bound is %i and my upper bound is %i \n",rank,low_bound,upper_bound);
            //do your work
            for (int i = low_bound; i < upper_bound; i++) {
                for (int j = 0; j < N_COLS; j++) {
                    matrix_C[i][j] = (matrix_A[i][j] + matrix_B[i][j]);
                }
            }
            //send back the low bound first without blocking, to the master
            MPI_Isend(&low_bound, 1, MPI_INT, 0, SLAVE_TO_MASTER_TAG, MPI_COMM_WORLD, &request);
            //send the upper bound next without blocking, to the master
            MPI_Isend(&upper_bound, 1, MPI_INT, 0, SLAVE_TO_MASTER_TAG + 1, MPI_COMM_WORLD, &request);
            //finally send the processed portion of data without blocking, to the master
            MPI_Isend(&matrix_C[low_bound][0], (upper_bound - low_bound) * N_COLS, MPI_DOUBLE, 0, SLAVE_TO_MASTER_TAG + 2, MPI_COMM_WORLD, &request);
        }

        // -------------------> Process 0 gathers the work
        ...

Answer 1

MPI_Isend() 开始非阻塞发送。因此，在不检查消息是否实际发送的情况下修改发送的缓冲区会导致发送错误的值。

这就是您提供的代码段中发生的情况，在进程for (proc = 1; proc < nproc; proc++) 的循环中

1. proc=1 : low_bound 已计算。

2. proc=1 : low_bound 被发送（非阻塞）到进程 1。

3. proc=2：low_bound 已修改。消息已损坏。

存在不同的解决方案：

• 使用阻塞发送MPI_Send()。

• 通过创建一个包含 3 个请求 MPI_Request requests[3]; MPI_Status statuses[3]; 的数组来检查消息是否完成，使用非阻塞发送并使用 MPI_Waitall() 检查请求是否完成。

   MPI_Isend(&low_bound, 1, MPI_INT, proc, MASTER_TO_SLAVE_TAG, MPI_COMM_WORLD, &requests[0]);
   MPI_Isend(..., &requests[1]);
   MPI_Isend(..., &requests[2]);
   MPI_Waitall(3, requests, statuses);
  

• 看看MPI_Scatter()和MPI_Scatterv()！

执行此操作的“通常”方法是MPI_Bcast() 矩阵的大小。然后每个进程计算其矩阵部分的大小。进程 0 计算 MPI_Scatterv() 所需的 sendcounts 和 displs。