MPI Reduce 和 Broadcast 工作，但导致未来 return 值失败

Question

我正在使用 MPI 为 class 实施 Dijkstras 算法。我的老师也不知道为什么这会被破坏，并允许我在这里 post。

我的问题出在 chooseVertex 函数中。该程序适用于 1 个处理器，但是当我运行使用 2 个处理器时，处理器 0 无法 return leastPostition，即使我能够在 [= 之前的行上打印 leastPosition 的内容21=]。我的代码：

#include "mpi.h"
#include <stdlib.h>
#include <stdio.h>

#define min(x,y) ((x) > (y) ? (y) : (x))
#define MASTER 0
#define INFINTY 100000

void dijkstra(int, int, int **, int *, int, int);
int chooseVertex(int *, int, int *, int, int);

int main(int argc, char* argv[])
{
int rank, size, i, j;
//Initialize MPI
MPI_Status status;
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &size);

//Initialize graph
int src = 0;
int n = 12;
int **edge = (int**) malloc(n * sizeof(int *));
for (i = 0; i < n; i++)
    edge[i] = (int *)malloc(n * sizeof(int));
int dist[12];

//Set all graph lengths to infinity
for (i = 0; i < n; i++)
{
    for (j = 0; j < n; j++)
    {
        if (i == j) { edge[i][j] = 0; }
        else { edge[i][j] = INFINTY; }
    }
}
//set graph edge lengths
edge[0][3] = 5;
edge[0][6] = 13;
edge[1][5] = 12;
edge[2][1] = 7;
edge[3][2] = 9;
edge[3][4] = 2;
edge[4][7] = 3;
edge[5][10] = 1;
edge[5][11] = 4;
edge[6][9] = 9;
edge[7][8] = 4;
edge[8][9] = 10;
edge[8][10] = 7;
edge[9][10] = 6;
edge[10][11] = 1;

dijkstra(src, n, edge, dist, rank, size);

if(rank == MASTER){ printf("The distance is %d", dist[n - 1]); }

MPI_Finalize();
return 0;
}

//called by dijkstras function below
int chooseVertex(int *dist, int n, int *found, int rank, int size) {
int i, tmp, partition, lower, upper, leastPosition;
int least = INFINTY;

//set the number of nodes wach processor will work with
partition = n / size;
lower = rank * partition;
upper = lower + partition;

//used for MPI_Reduce
struct {
    int pos;
    int val;
} sendBuffr, recvBuffr;

//calculate least position
for (i = lower; i < upper; i++) {
    tmp = dist[i];
    if ((!found[i]) && (tmp < least)) {
        least = tmp;
        leastPosition = i;
    }
}
//if all nodes checked are INFINITY, go with last node checked
if (least == INFINTY) leastPosition = i;

//set the send buffer for MPI_Reduce
sendBuffr.val = least;
sendBuffr.pos = leastPosition;

//Rank 0 processor has correct least position and value
MPI_Reduce(&sendBuffr, &recvBuffr, 1, MPI_DOUBLE_INT, MPI_MINLOC, MASTER, MPI_COMM_WORLD);


if (rank == MASTER) leastPosition = recvBuffr.pos;

//Update all processors to have correct position
MPI_Bcast(&leastPosition, 1, MPI_INT, MASTER, MPI_COMM_WORLD);

//Print the contents of leastPosition on rank 0 for debugging
if(rank == MASTER) printf("LeastPosition for rank %d is: %d\n", rank, leastPosition);
fflush(stdout);

return leastPosition;
}

void dijkstra(int SOURCE, int n, int **edge, int *dist, int rank, int size)          
{
int i, j, count, partition, lower, upper, *found, *sendBuffer;

j = INFINTY;
sendBuffer = (int *)malloc(n * sizeof(int));
found = (int *)calloc(n, sizeof(int));

partition = n / size;
lower = rank * partition;
upper = lower + partition;

//set the distance array
for (i = 0; i < n; i++) {
    found[i] = 0;
    dist[i] = edge[SOURCE][i];
    sendBuffer[i] = dist[i];
}

found[SOURCE] = 1;
count = 1;

//Dijkstra loop
while (count < n) {
    printf("before ChooseVertex:  rank %d reporting\n", rank);
    fflush(stdout);
    j = chooseVertex(dist, n, found, rank, size);
    printf("after ChooseVertex:   rank %d reporting\n", rank);
    fflush(stdout);
    count++;

    found[j] = 1;
    for (i = lower; i < upper; i++) {
        if (!found[i])
        {
            dist[i] = min(dist[i], dist[j] + edge[j][i]);
            sendBuffer[i] = dist[i];
        }
    }
    MPI_Reduce(sendBuffer, dist, n, MPI_INT, MPI_MIN, MASTER, MPI_COMM_WORLD);
    MPI_Bcast(dist, n, MPI_INT, MASTER, MPI_COMM_WORLD);
}
}

示例错误消息：

before ChooseVertex:  rank 1 reporting
before ChooseVertex:  rank 0 reporting
LeastPosition for rank 0 is: 3
after ChooseVertex:   rank 1 reporting
after ChooseVertex:   rank 0 reporting
before ChooseVertex:  rank 1 reporting
before ChooseVertex:  rank 0 reporting
after ChooseVertex:   rank 1 reporting
LeastPosition for rank 0 is: 4
after ChooseVertex:   rank 0 reporting
before ChooseVertex:  rank 0 reporting
before ChooseVertex:  rank 1 reporting
LeastPosition for rank 0 is: 7
after ChooseVertex:   rank 1 reporting

job aborted:
[ranks] message

[0] process exited without calling finalize

[1] terminated

---- error analysis -----

[0] on My-ComputerName
Assignmet3PP ended prematurely and may have crashed. exit code 3

---- error analysis -----

Answer 1

你的减少命令是：

MPI_Reduce(&sendBuffr, &recvBuffr, 1, MPI_DOUBLE_INT, MPI_MINLOC, MASTER, MPI_COMM_WORLD);

通过使用 MPI_DOUBLE_INT，您表示您正在发送一个包含两个变量的结构：一个 double 后跟一个 int。然而，这不是您的结构：您只有 2 个整数。因此你应该使用 MPI_2INT。这些类型源自 this source。或者，您可以使用向量创建自己的类型。

示例修复是：

MPI_Reduce(&sendBuffr, &recvBuffr, 1, MPI_2INT, MPI_MINLOC, MASTER, MPI_COMM_WORLD);

此外，可以通过 MPI_Allreduce().

轻松地将缩减和广播合并为一个步骤

MPI Reduce 和 Broadcast 工作，但导致未来 return 值失败

MPI Reduce and Broadcast work, but cause a future return value to fail

c

mpi

ms-mpi