在 c 中使用 MPI_Reduce

Question

我正在尝试使用 MPI 库在 c 中编写程序。 在我的程序中，我正在解决 TSP（但不是使用任何特殊算法......）。 我的输入参数是int nCites、int xCoord[] 和int yCoord[]。 我将它们捆绑到Coordinates 并使用MPI_Bcast 使它们可用于所有线程。

我的问题是：在计算完每个线程中所有路径的权重后，我想将它们减少为一个结果，最好的一个。我试过使用MPI_Reduce，但是有些东西，这就是我感到困惑的地方，会导致分段错误（仅在一个线程中，通常是root）。

这是主要代码和结构：

typedef struct Coordinates_t {
    int* x;
    int* y;
    int n;
} Coordinates;

typedef struct PathAndLength_t {
    int* path;
    int pathSize;
    int length;
} PathAndLength;

void comparePaths(void* a, void* b, int* len, MPI_Datatype* dataType) {
    ...
}

int tsp_main(int nCites, int xCoord[], int yCoord[], int P[]){
    int numOfProcs, rank;
    if (MPI_Comm_size(MPI_COMM_WORLD, &numOfProcs))
        throw "Error: MPI_Comm_size failed";
    if (MPI_Comm_rank(MPI_COMM_WORLD, &rank))
        throw "Error: MPI_Comm_rank failed";

    Coordinates crds;
    crds.x = xCoord;
    crds.y = yCoord;
    crds.n = nCites;

    MPI_Datatype data;
    createInDataType(&crds, &data);

    if (MPI_Bcast(&crds, 1, data, 0, MPI_COMM_WORLD))
        throw "Error: MPI_Comm_size failed";

        ...

    PathAndLength* pal = (PathAndLength*)malloc(sizeof(PathAndLength));
    pal->path = (int*)malloc(sizeof(int)*crds.n);

    pal->length = min_length;
    for (int i = 0; i < crds.n; ++i) {
        (pal->path)[i] = min_path[i];
    }
    pal->pathSize = crds.n;
    MPI_Datatype outDatatype;
    MPI_Op op;

    createOutDataType(pal, &outDatatype);

    MPI_Op_create(&comparePaths, 1, &op);

    PathAndLength* result = (PathAndLength*)malloc(sizeof(PathAndLength));
    result->path = (int*)malloc(sizeof(int)*crds.n);
    MPI_Reduce(pal, result, crds.n, outDatatype, op, 0, MPI_COMM_WORLD);

    ...

    return result->length;
}

这些是我在代码中使用的createOutDataType 和createInDataType：

void createInDataType(Coordinates* indata, MPI_Datatype* message_type_ptr) {
    // Build a derived datatype
    int block_lengths[3];
    MPI_Aint displacements[3];
    MPI_Aint addresses[4];
    MPI_Datatype typelist[3];

    // First specify the types
    typelist[0] = typelist[1] = typelist[2] = MPI_INT;

    // Specify the number of elements of each type
    block_lengths[0] = block_lengths[1] = indata->n;
    block_lengths[2] = 1;

    // Calculate the displacements of the members relative to indata
    MPI_Address(indata, &addresses[0]);
    MPI_Address(indata->x, &addresses[1]);
    MPI_Address(indata->y, &addresses[2]);
    MPI_Address(&indata->n, &addresses[3]);
    displacements[0] = addresses[1] - addresses[0];
    displacements[1] = addresses[2] - addresses[0];
    displacements[2] = addresses[3] - addresses[0];

    // Create the derived type
    MPI_Type_struct(3, block_lengths, displacements, typelist, message_type_ptr);

    // Commit it so that it can be used
    MPI_Type_commit(message_type_ptr);
}

void createOutDataType(PathAndLength* outdata, MPI_Datatype* message_type_ptr) {
    // Build a derived datatype
    int block_lengths[2];
    MPI_Aint displacements[2];
    MPI_Aint addresses[3];
    MPI_Datatype typelist[2];

    // First specify the types
    typelist[0] = MPI_INT;
    typelist[1] = MPI_INT;

    // Specify the number of elements of each type
    block_lengths[0] = outdata->pathSize;
    block_lengths[1] = 1;

    // Calculate the displacements of the members relative to outdata
    MPI_Address(outdata, &addresses[0]);
    MPI_Address(outdata->path, &addresses[1]);
    MPI_Address(&outdata->length, &addresses[2]);
    displacements[0] = addresses[1] - addresses[0];
    displacements[1] = addresses[2] - addresses[0];

    // Create the derived type
    MPI_Type_struct(2, block_lengths, displacements, typelist, message_type_ptr);

    // Commit it so that it can be used
    MPI_Type_commit(message_type_ptr);
}

很抱歉包含这么多代码，但我无法确定哪些无关紧要（如果有的话）...
谢谢。

Answer 1

PathAndLength* result = (PathAndLength*)malloc(sizeof(PathAndLength));
result->path = (int*)malloc(sizeof(int)*crds.n);
MPI_Reduce(pal, result, crds.n, outDatatype, op, 0, MPI_COMM_WORLD);

您将crds.n*outDatatypes 接收到大小为sizeof(PathAndLength) 的结果缓冲区中。你似乎有设计缺陷。