MPI_Reduce w user function and non contiguous data

Question

I'm trying to do some simple MPI Projects (w MPICH) but while doing so I experienced a problem I neither understand nor am able to solve (propably bc I misunderstand the doc). So what I basically wanted to do was to pass a struct to MPI_Reduce in order to make some operations on it and return the result back to the root process.

To do so I tried two different ways. First way was to use MPI_Pack to successively pack the struct's elements in a buffer and unpack them in my user function.

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

#define MAX_PACKING_BUF_SIZE 100

typedef struct {
    double x;
    int n;
} BE_Type;

int rank = -1;

void operation(void *invec, void *inoutvec, int *length, MPI_Datatype *type)
{
    uint8_t *buf  = (uint8_t*) invec;
    double *res     = (double*) inoutvec;
    int pos;
    BE_Type value;

    printf("[%d] len: %d\n", rank, *length);

    pos = 0;
    MPI_Unpack(buf, *length, &pos, &value.x, 1, MPI_DOUBLE, MPI_COMM_WORLD);
    MPI_Unpack(buf, *length, &pos, &value.n, 1, MPI_INT,    MPI_COMM_WORLD);

    printf("[%d] x: %lf, n: %d\n", rank, value.x, value.n);

    *res    += value.x;
}

int main (int argc, char *argv[])
{
    int rc, pos;
    int root    = 0;
    double res  = 0.0;
    MPI_Op my_op;
    BE_Type value;
    uint8_t buf[MAX_PACKING_BUF_SIZE];

    rc = MPI_Init(&argc,&argv);
    if (rc != MPI_SUCCESS) printf("ERROR\n");
    rc = MPI_Comm_rank(MPI_COMM_WORLD, &rank);
    if (rc != MPI_SUCCESS) printf("ERROR\n");
    rc = MPI_Op_create( (MPI_User_function*) operation, 1, &my_op);
    if (rc != MPI_SUCCESS) printf("ERROR\n");

    value.x = 1.0;
    value.n = rank;

    pos = 0;
    MPI_Pack(&value.x, 1, MPI_DOUBLE, buf, MAX_PACKING_BUF_SIZE, &pos, MPI_COMM_WORLD);
    MPI_Pack(&value.n, 1, MPI_INT,    buf, MAX_PACKING_BUF_SIZE, &pos, MPI_COMM_WORLD);

    rc = MPI_Reduce(buf, &res, pos, MPI_PACKED, my_op, root, MPI_COMM_WORLD);
    if (rc != MPI_SUCCESS) printf("ERROR\n");

    if (rank == root) {
        printf("res: %lf\n", res);
    }
}

This code, however, leads to the following results (started 4 processes):

[0] len: 12
[2] len: 12
[2] x: 1.000000, n: 3
[0] x: 1.000000, n: 1
[0] len: 12
[0] x: 2.000000, n: 2
res: 4.000000

So first of all I'm wondering why my function is only called three times instead of four? And second of all (and that's my main question): Why is that x value altered at one point?

Interestingly that x value is altered as well using the second way I tested which was defining a new datatype:

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

typedef struct {
    double x;
    int n;
} BE_Type;

int rank = -1;

void operation(void *invec, void *inoutvec, int *length, MPI_Datatype *type)
{
    BE_Type *value  = (BE_Type*) invec;
    double *res     = (double*) inoutvec;

    printf("[%d] x: %lf, n: %d\n", rank, value->x, value->n);

    *res    += value->x;
}

int main (int argc, char *argv[])
{
    int rc, pos;
    int root    = 0;
    double res  = 0.0;
    MPI_Op my_op;
    BE_Type value;

    MPI_Datatype MPI_BE_Type;
    int blocklens[2];
    MPI_Aint indices[2];
    MPI_Datatype old_types[2];

    rc = MPI_Init(&argc,&argv);
    if (rc != MPI_SUCCESS) printf("ERROR\n");
    rc = MPI_Comm_rank(MPI_COMM_WORLD, &rank);
    if (rc != MPI_SUCCESS) printf("ERROR\n");
    rc = MPI_Op_create( (MPI_User_function*) operation, 1, &my_op);
    if (rc != MPI_SUCCESS) printf("ERROR\n");

    blocklens[0] = 1;
    blocklens[1] = 1;
    old_types[0] = MPI_DOUBLE;
    old_types[1] = MPI_INT;
    MPI_Address(&value.x, &indices[0]);
    MPI_Address(&value.n, &indices[1]);
    indices[1] = indices[1] - indices[0];
    indices[0] = 0;
    MPI_Type_struct(2, blocklens, indices, old_types, &MPI_BE_Type);
    MPI_Type_commit(&MPI_BE_Type);

    value.x = 1.0;
    value.n = rank;

    rc = MPI_Reduce(&value, &res, 1, MPI_BE_Type, my_op, root, MPI_COMM_WORLD);
    if (rc != MPI_SUCCESS) printf("ERROR\n");

    if (rank == root) {
        printf("res: %lf\n", res);
    }
}

Results to (4 processes started):

[2] x: 1.000000, n: 3
[0] x: 1.000000, n: 1
[0] x: 2.000000, n: 2
res: 4.000000

So I guess I'm just misunderstanding something or using it wrongly. Every helped is appreciated. Thanks!

Answer 1

So first of all I'm wondering why my function is only called three times instead of four?

Adding up N values, takes N-1 additions. The same holds for any operation. No matter how you rearrange them.

And second of all (and that's my main question): Why is that x value altered at one point?

The operation is performed as a tree (typically a binomial tree). In your case it looks something like this:

Ranks   0 1 2 3

        1 1 1 1
        |/  | /
        +   +
        2   2
        |  /
        | /
        +
        4

Operations must always be associative, so this is a valid way to compute the result.

I wouldn't really recommend the first way of packing/unpacking in the custom reduction operation. Anyway: your reduction function must use both in , and inout as BE_Type ! The function also must work for any length . So it could look like:

void operation(void *invec, void *inoutvec, int *length, MPI_Datatype *type)
{
    BE_Type *value = (BE_Type*) invec;
    BE_Type *res   = (BE_Type*) inoutvec;

    for (int i = 0; i < *length; i++) {
        res[i].x += value[i].x;
        // should probably do something with n
    }
}