[lug] NFS cross-mounts

[Paul E Condon]

On 20100618_083441, Vince Dean wrote:
> I'm working with a small cluster of Linux servers and desktop machines.
> Most of the machines export some of their local file systems to all the
> others via NFS.  The intent is that the disks on each machine are
> resources which should be shared with all the users.
> 
> This results in what is known as NFS cross-mounts, where two machines
> each depend on an NFS mount from the other.  Our administrators are
> concerned that there is a considerable risk of deadlocks which can only
> be resolved by rebooting one or more machines.  However, my copy of the
> O'Reilly book "Managing NFS and NIS" claims that "This deadlock problem
> goes away when your NFS clients use the automounter in place of
> hard-mounts.", page 415.
> 
> Are NFS cross-mounts evil even if you use the automounter?
> 
> There is a larger issue here: whether it makes sense to have such cyclic
> dependencies, especially involving desktop machines, at all.  That
> arrangement puts us in a situation where shutting down one machine will
> inconvenience an unknown number of users on other machines. You could
> make a strong case that machines should only depend on services which
> are provided by carefully-maintained servers in machine rooms, and that
> the dependencies should should be simple and should form an acyclic graph.
> 
> We're working on that larger design question and I'd appreciate any
> comments.

My experience is strictly as an intelligent user of poorly maintained
corporate networks: 

Don't expect to have people obey rules. People will turn off their
office computer when they go home, at least some will some of the
time. It is part of the new ethic of green.

If you are willing to entrust business data to reside on desktop class
hardware, you would be better off to have a few desktop class boxes
dedicated to being 'servers'. Have more than one, because they will 
crash from time to time. But at least when something wierd happens you
have a limited number of boxes to check for a pulled power cord.

Basically the whole idea seems wierd to me as a user. 

Maybe, as an academic exercise, it would be interesting to try as a
faculty/staff/student project in a department of computer science,
maybe. Not, impossible. But maybe there are more interesting projects
to spend time and effort on. A good candidate for students to get
real world experience with failed project syndrome.

-- 
Paul E Condon           
pecondon at mesanetworks.net