[Beowulf] Rocks Cluster and 2 Ethernet networks

[mechti01]

Hi Reuti,

Thanks for your help.  I have not installed Rocks just yet.  Can you
explain to me what the ch_p4 device of MPI_CH is?  Nochmals, vielen Dank!

-Timo


> Hi,
>
>> I'm considering installing the Rocks cluster distro on a cluster that
>> uses
>> only ethernet.  As I understand it, eth0 (or first network interface) is
>> used for administration and also message passing if no other high speed
>> interface is present (e.g. myrinet).  My question is, if each of my
>> compute node have two ethernet interfaces, say eth0 and eth1, can the
>> cluster be configured that message passing takes place only over eth1?
>> It
>> would be nice to have an interface devoted to just message passing.  If
>> it
>> is possible, how would I go about setting it up?  If it's not possible,
>> is
>> there are a lot of performance loss due to the fact that other tasks
>> (such
>> as administration, etc.) are also taking place over eth0?  Thanks in
>> advance for your help.
>
> do you want to use the ch_p4 device of MPICH for communication? Then you
> simply
> have to set the machinefile for mpirun to include only the names of the
> second
> interface in all nodes. Maybe your queuingsystem can do this already for
> you.
> Furthermore, you have to change the setting in mpirun.args that way, that
> instead:
>
> MPI_HOST=`hostname`
>
> will be substituded with the name of the second interface. E.g.
>
> MPI_HOST=`hostname | sed "s/^node/internal/"`
>
> to change the name from node001 to internal001 or whatever names you use.
> Otherwise your machinefile will be scanned in a wrong way (wrong
> distribution
> of the processes to the nodes in the end), and the communication back from
> the
> slaves to the head node of the job will still use the wrong interface. You
> can
> simply include this at the beginning of the mpirun.arg file. If it's
> already
> set, it will no be set later in the script.
>
> Cheers - Reuti
>


--