[Beowulf] anyone using SALT on your clusters?

Mark Hahn hahn at mcmaster.ca
Thu Jun 27 18:02:10 PDT 2013


>> why the big deal about cf-engine/salt/puppet/chef/etc?
>
> Well with xCAT (which we use) the big thing is node auto-discovery and
> naming. If we physically replace a node we remove its MAC address from
> xCAT and when we turn it on it boots into a discovery image, gets
> named based on its switch port (via a system you define to it),
> programs the BMC/IMM appropriately and then lights up its attention light.
>
> When we brought up our new Intel system recently we just plugged it
> in, turned it on and waited for all the little blue lights to come on
> to say it was appropriately configured. :-)

sure, I like auto-configuring clusters too, but do you need salt-like
tools for this?  the way I do it is that the pxe default boots into 
the discovery kernel, which in rc.local just probes MAC/IPMI/port
and hands it to the admin node, which updates the DB and switches pxe
to a normal node boot.

> So we can have a library of osimage definitions and switch groups
> between them.

sure - we currently use onesis, which has groups/roles too, though
it doesn't put the customization all on one spot, (still very obvious.)
we don't actually try to configure all our stuff from one place,
though (partly for historic and geographic reasons, I guess.)  I'm not
really sure we'd want *everthing* intertwingled, though - for instance 
the admin/compute/login part of a cluster is basically independent from
the luster server nodes.  maybe we're just not ambitious enough in 
our automation.


More information about the Beowulf mailing list