[MPICH] Multi-homed and default configuration

[Ralph Butler]

Hi Steve:

On May 24, 2006, at 4:21 PM, Steve Kargl wrote:

> I've set up mpich2 on a small cluster.  The master node,
> where users will connect via the internet, is multi-homed
> with a FQDN and a 192.168 subnet.  I want mpich2 to run
> on the 192.168 subnet.  In reading the docs, this can be
> accomplished by
>
> mpdboot --ifhn=192.168.0.10 -n 3 -f mpd.hosts
>
> where 192.168.0.10 is in the master node and mpd.hosts
> contains 192.168.0.{11,12,13,14,15}.
>


The --ifhn option to mpdboot ONLY affects the local machine where
the first mpd is always started.  Then, you can specify other ifhn
values for other hosts in the mpd.hosts file, e.g.:
     node2 ifhn=192.168.1.2


> I've scanned the installation and user's guides and the
> mpich-discuss archive, but can't find the answers to
> the following questions.
>
> Is it possible to have a /etc/mpdboot.conf file and
> /etc/mpd.hosts file so that a ring of mpd servers can
> be started by root such that the few users on the
> system will use this ring?
>

No, there is no mpdboot.config file.  But, you can start mpds as
root and have users use them to run their own jobs.  The users
have to set the variable named MPD_USE_ROOT_MPD to 1 either in their
env or in their ~/mdp.conf files.  I believe the name of this var
is MIS-spelled in the manual with the current release.


> It appears that --ncpus= option will assume that all
> nodes have the same numbers of cpus, which may not be
> the case.  Is it possible to set the number of cpus in
> a node via the mpd.hosts file?
>
>


Like ifhn, --ncpus is for the local mpd only.  You can specify it in
the mpd.hosts file this way:
     node1:4
which says that there are 4 cpus on node1.  Yes, I know it should
have been handled the same way as ifhn (i.e. with a ncpus keyword),
but it was left this way for historical reasons.  :-)

> -- 
> Steve
>

--ralph