MPI_Scatterv example

scatterv.c

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

#define SIZE 4

int main(int argc, char *argv[])
{
    int rank, size; // for storing this process' rank, and the number of processes

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

    int *sendcounts;            // array describing how many elements to send to each process
    int *displs;                // array describing the displacements where each segment begins
    int rem = (SIZE*SIZE)%size; // elements remaining after division among processes
    int sum = 0;                // Sum of counts. Used to calculate displacements
    char rec_buf[100];          // buffer where the received data should be stored

    // the data to be distributed
    char data[SIZE][SIZE] = {
        {'a', 'b', 'c', 'd'},
        {'e', 'f', 'g', 'h'},
        {'i', 'j', 'k', 'l'},
        {'m', 'n', 'o', 'p'}
    };

    sendcounts = malloc(sizeof(int)*size);
    displs = malloc(sizeof(int)*size);

    // calculate send counts and displacements
    for (int i = 0; i < size; i++) {
        sendcounts[i] = (SIZE*SIZE)/size;
        if (rem > 0) {
            sendcounts[i]++;
            rem--;
        }

        displs[i] = sum;
        sum += sendcounts[i];
    }

    // print calculated send counts and displacements for each process
    if (0 == rank) {
        for (int i = 0; i < size; i++) {
            printf("sendcounts[%d] = %d\tdispls[%d] = %d\n", i, sendcounts[i], i, displs[i]);
        }
    }

    // divide the data among processes as described by sendcounts and displs
    MPI_Scatterv(&data, sendcounts, displs, MPI_CHAR, &rec_buf, 100, MPI_CHAR, 0, MPI_COMM_WORLD);

    // print what each process received
    printf("%d: ", rank);
    for (int i = 0; i < sendcounts[rank]; i++) {
        printf("%c\t", rec_buf[i]);
    }
    printf("\n");

    MPI_Finalize();

    free(sendcounts);
    free(displs);

    return 0;
}

Thanks for you code :) It helps me a lot.

Cool. Glad to hear you found it useful. :)

Thanks a lot. This a easy to understand tutorial compared to other sources.

Thanks for your sample code!

The variable "rem" is initialized when size is uninitialized.
I think that "rem" should be initialized after calling MPI_Comm_size().

I get "Floating point exception (signal 8)"...

Thx for this sample!

Thank you for this sample 👍 . I guess rem should be calculated after MPI_Comm_size(MPI_COMM_WORLD, &size);

Hello! thank you so much for the sample.

I had one question. Shouldn't lines 30-43 be also enclosed in "if (rank==0)" conditional? I'm not sure why every processor would have to allocate memory & calculate values for sendcounts and displs.

Thank you and I have the same question as anjennakwon06

Thanks for this example!

Is there a reason why you reinitialise these variables?

int *sendcounts;    // array describing how many elements to send to each process // line 11
int *displs;        // array describing the displacements where each segment begins // line 12

int *sendcounts = malloc(sizeof(int)*size); // line 30
int *displs = malloc(sizeof(int)*size); // line 31

Can I Call MPI_Scatterv only in few processes ?
for example only for process which rank<=20 ?

mpirun noticed that process rank 0 with PID 28921 on node sateesh-OptiPlex-9020 exited on signal 8 (Floating point exception)....??
can anyone help me

mpirun noticed that process rank 0 with PID 28921 on node sateesh-OptiPlex-9020 exited on signal 8 (Floating point exception)....??
can anyone help me

Me too. It has been 3 years and still no answer.

Uninitialized variable size is used at line 14 to calculate the remainder. But, size gets the number of processes later, at line 28. It can fixed moving lines 26-28 at the beginning of main(), between lines 10 and 11.

For those who want to know what it prints out without compiling:

1: e    f       g       h
sendcounts[0] = 4       displs[0] = 0
sendcounts[1] = 4       displs[1] = 4
sendcounts[2] = 4       displs[2] = 8
sendcounts[3] = 4       displs[3] = 12
0: a    b       c       d
3: m    n       o       p
2: i    j       k       l

Great, thanks

LO DE ABAJO ES EL CODIGO CORRECTO:
.
.
.
.
.
.
.
.
.

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

#define N 4

int main(int argc, char *argv[])
{
int rank, numProcs; // para almacenar rango número de procesos

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

int *sendcounts;    // matriz que describe cuántos elementos enviar a cada proceso
int *displs;        // array que describe los desplazamientos donde comienza cada segmento

int rem = (N*N)%numProcs; // Elementos que quedan después de la división entre procesos
int sum = 0;                // Suma de cuentas. Se utiliza para calcular desplazamientos
char rec_buf[100];          // buffer donde se deben almacenar los datos recibidos

// los datos a distribuir
char data[N][N] = {
    {'a', 'b', 'c', 'd'},
    {'e', 'f', 'g', 'h'},
    {'i', 'j', 'k', 'l'},
    {'m', 'n', 'o', 'p'}
};


sendcounts = malloc(sizeof(int)*numProcs);
displs = malloc(sizeof(int)*numProcs);

// calcular conteos y desplazamientos de envío
for (int i = 0; i < numProcs; i++) {
    sendcounts[i] = (N*N)/numProcs;

    if (rem > 0) {
        sendcounts[i]++;
        rem--;
    }

    displs[i] = sum;
    sum += sendcounts[i];
}

// imprimir los conteos y desplazamientos de envío calculados para cada proceso
if (0 == rank) {
    for (int i = 0; i < numProcs; i++) {
        printf("sendcounts[%d] = %d\tdispls[%d] = %d\n", i, sendcounts[i], i, displs[i]);
    }
}

// dividir los datos entre procesos según lo descrito por sendcounts y displs
MPI_Scatterv(&data, sendcounts, displs, MPI_CHAR, &rec_buf, 100, MPI_CHAR, 0, MPI_COMM_WORLD);

// imprime lo que recibió cada proceso
printf("%d: ", rank);
for (int i = 0; i < sendcounts[rank]; i++) {
    printf("%c\t", rec_buf[i]);
}
printf("\n");

MPI_Finalize();

free(sendcounts);
free(displs);

return 0;

}

Thanks a lot, very helpful ❤️

Thank you for the code, very helpfull