[mpich-discuss] spawning new processes with a hostfile

Umit umitcanyilmaz at gmail.com
Mon Apr 16 08:14:36 CDT 2012


Hello all,


There are some spawn commands in my program. Now I want to specify the
nodes of my new spawned processes.  I am trying to use a hostfile for this
but  I couldn’t do it successfully.  New processes are still spawned on
next available nodes.

I added my code and outputs of my console.

*My hostfile: *

root at rck00:~> cat /shared/mpihosts

rck03

rck04

rck05



Can somebody help me? What is the problem? Can this be a bug?


*Here is my code and output of my program:*

#include "mpi.h"

#include <stdio.h>

#include <stdlib.h>



#define NUM_SPAWNS 3



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

{

    int errcodes[NUM_SPAWNS];

    MPI_Comm parentcomm, intercomm;

    int len;

    char name[MPI_MAX_PROCESSOR_NAME];

    int rank;



    MPI_Init( &argc, &argv );

    MPI_Comm_get_parent( &parentcomm );

    MPI_Comm_rank(MPI_COMM_WORLD,&rank);



    if (parentcomm == MPI_COMM_NULL)

    {

            MPI_Info info;

            MPI_Info_create( &info );

            MPI_Info_set(info, "hostfile", "/shared/mpihosts");



            MPI_Comm_spawn( "/shared/spawn/./spawn", MPI_ARGV_NULL,
NUM_SPAWNS, info, 0, MPI_COMM_WORLD, &intercomm, errcodes );



            MPI_Get_processor_name(name, &len);

            printf("I am parent process %d on %s.  \n", rank,  name);

    }

    else

    {

            MPI_Get_processor_name(name, &len);

            printf("I am a spawned process %d on %s.\n", rank,  name);

    }

    fflush(stdout);

    MPI_Finalize();

    return 0;

}



*output of my program:*

root at rck00:~> mpirun -np 1 /shared/spawn/./spawn

I am parent process 0 on rck00.

I am a spawned process 0 on rck01.

I am a spawned process 1 on rck02.

I am a spawned process 2 on rck03.





Thanks in advance,
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/mpich-discuss/attachments/20120416/0b392bdc/attachment.htm>


More information about the mpich-discuss mailing list