[mpich-discuss] mpd and login node

[Rajeev Thakur]

Steve, 
      If you use -1 combined with the -machinefile option to mpiexec, it
should create processes only on the hosts in the machine file.

% mpiexec -1 -machinefile ~/mpd.hosts -n 20 hostname

Rajeev  

PS: This is the first time I have seen car companies as host names :-).



> -----Original Message-----
> From: mpich-discuss-bounces at mcs.anl.gov 
> [mailto:mpich-discuss-bounces at mcs.anl.gov] On Behalf Of Stephen Siegel
> Sent: Saturday, February 14, 2009 9:33 AM
> To: mpich-discuss at mcs.anl.gov
> Subject: Re: [mpich-discuss] mpd and login node
> 
> Rajeev: if I use the -1 option with mpiexec, and my cluster 
> has 8 nodes, that when I run mpiexec with 9 processes, the 
> 9th process I THINK gets run on the local node, which I don't 
> want to happen.
> 
> In any case, I found another way to handle the situation 
> which also takes into account that each node has two 
> processors, so I want ranks  
> 0 and 1 to map to node1, ranks 1 and 2 to map to node2 etc.   What I  
> did was this....(porsche is the local node, the others have 
> names node1=honda, ...):
> 
> siegel at porsche ~ $ cat ~/mpd.hosts   /* I will also use this file as  
> machinefile for mpiexec...*/
> node1:2
> node2:2
> node3:2
> node4:2
> node5:2
> node6:2
> node7:2
> node8:2
> siegel at porsche ~ $ mpdboot -n 9    /* NOTE: 9=8+1 because the local  
> host counts as 1 in mpd ring */
> siegel at porsche ~ $ mpdtrace     /* all 9 machines appear...*/
> porsche
> nissan
> honda
> subaru
> toyota
> suzuki
> acura
> mazda
> hyundai
> siegel at porsche ~ $ cd code/mpi/hello/
> siegel at porsche ~/code/mpi/hello $ cat hello2.c 
> #include<stdio.h> #include<stdlib.h> #include<unistd.h> 
> #include<mpi.h>
> 
> int main(int argc,char *argv[]) {
>    int rank;
>    char *name = (char*)malloc(200*sizeof(char));
> 
>    gethostname(name, 200);
>    MPI_Init(&argc, &argv);
>    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
>    printf("Hello from process %d running on %s.\n", rank, name);
>    fflush(stdout);
>    MPI_Finalize();
>    return 0;
> }
> 
> siegel at porsche ~/code/mpi/hello $ mpiexec -machinefile ~/mpd.hosts -n
> 16 ./hello2
> Hello from process 0 running on honda.cis.udel.edu.
> Hello from process 4 running on subaru.cis.udel.edu.
> Hello from process 2 running on toyota.cis.udel.edu.
> Hello from process 1 running on honda.cis.udel.edu.
> Hello from process 5 running on subaru.cis.udel.edu.
> Hello from process 14 running on suzuki.cis.udel.edu.
> Hello from process 3 running on toyota.cis.udel.edu.
> Hello from process 10 running on acura.cis.udel.edu.
> Hello from process 15 running on suzuki.cis.udel.edu.
> Hello from process 12 running on mazda.cis.udel.edu.
> Hello from process 13 running on mazda.cis.udel.edu.
> Hello from process 11 running on acura.cis.udel.edu.
> Hello from process 8 running on hyundai.cis.udel.edu.
> Hello from process 6 running on nissan.cis.udel.edu.
> Hello from process 9 running on hyundai.cis.udel.edu.
> Hello from process 7 running on nissan.cis.udel.edu.
> 
> siegel at porsche ~/code/mpi/hello $
> 
> Looks like exactly what I want: 2 procs on each node, and no 
> proc running on local node.
> 
> Does this seem like a reasonable way to set things up?
> 
> -Steve
> 
> On Feb 13, 2009, at 10:24 PM, Rajeev Thakur wrote:
> 
> > Use the "-1" option to mpiexec to prevent the first process 
> from being 
> > run on the local node.
> >
> > Rajeev
> >
> >> -----Original Message-----
> >> From: mpich-discuss-bounces at mcs.anl.gov 
> >> [mailto:mpich-discuss-bounces at mcs.anl.gov] On Behalf Of Stephen 
> >> Siegel
> >> Sent: Friday, February 13, 2009 7:21 PM
> >> To: mpich-discuss at mcs.anl.gov
> >> Subject: Re: [mpich-discuss] mpd and login node
> >>
> >> No queuing system.   Each user just runs mpd, and has a 
> machine file
> >> that lists the nodes in their home directory.   They are 
> supposed to
> >> terminate the mpd ring when they are done.   We are not too
> >> concerned
> >> with overloading---this is just being used by an introductory 
> >> parallel programming class.  A queuing system would probably be a 
> >> good idea though.  All the machines are running linux.
> >>
> >> On Feb 13, 2009, at 8:14 PM, Reuti wrote:
> >>
> >>> Am 14.02.2009 um 02:06 schrieb Stephen Siegel:
> >>>
> >>>> I am running mpich2 with mpd on a small cluster.   We have a
> >>>> separate logon node which users will log on to, compile, and
> >>>> execute programs using mpiexec.   We want to configure mpd
> >> so that
> >>>> it will distribute the processes over the nodes of the cluster,
> >>>> excluding the logon node.   For some reason we could not
> >> figure out
> >>>> how to do this----the logon node itself ends up getting
> >> some of the
> >>>> processes.   Does anyone know how to do this?  Thanks.
> >>>
> >>> How are you distributing other jobs right now and avoid
> >> overloading of
> >>> certain machines - you have a queuing system like SUN Grid Engine?
> >>>
> >>> -- Reuti
>