[BBLISA] RAM for cluster net boot?

[Daniel Hagerty]

Daniel Feenberg <feenberg at nber.org> writes:

> Its the convinience of having only one image to keep up to date, of being 
> able to make universal changes with a single edit, of knowing that none of 
> the systems was out of sync, the ability to swap OS with a single edit 
> to /etc/dhcpd.conf etc. Of course, it is nice to have local swap 
> and /tmp, the convinience comes from effective statelessness which a local 
> disk used only for swap and /tmp does not vitiate.

    This was a reasonable answer ten years ago.  These days, it's a
rare shop that can get by on one server that everything else NFS
mounts root from.  I realize you probably don't do that, I'm just
trying to point out that the places where these properties are
critical are much broader than places where diskless root is the
correct tool to achieve them.

> I understand that cfengine is an alternative, but not always a superior 
> alternative. In our case the amount of ethernet traffic from program 
> loading is trivial compared to data I/O, so there is little benefit to 
> us in keeping /usr local. Scott may be in a similar position.

    I am the wrong person to offer unbiased opinions of cfengine.
I've found it easier to roll something from scratch that did what I
needed, rather than deal with cfengine's enormous brain damage.

    lcfg, bcfg2, and puppet all seem much more reasonable.  I know
that CERN is managing an ungodly number of machines with lcfg; the
scaling problems they've report are computation in compiling profiles
for individual hosts, not bandwidth.

    If you can actually beat the I/O of the config management system
by a substantial margin, going diskless seems reasonable.  I'm
doubtful that it's common in the face of more modern configuration
management.