[Beowulf] Computation on the head node

[Mark Hahn]

> which are the disadvantages of using the head node as a computation node?

just that other headnode activities will interfere with compute jobs
and vice versa.  this may not matter, depending on your workload,
habits of users (compilation, etc).  mainly you can just pick out 
certain cases where interference would be very bad - memory intensive 
jobs, or memory-intensive login activity (compilation?), or IO-intensive.
if you run collective-intensive MPI jobs, you always want to avoid 
introducing jitter (extraneous activity on the job's nodes).

> Since I plan to have NFS on it, I guess there will be problems with 
> comunication
> which probably will slow down the whole cluster, but I am not sure since this
> is my first attempt to built one...

best practice is to separate administrative services from both compute nodes
and login nodes.  consider, for instance, that an NFS server doesn't have
much need for a lot of compute power, and under many workloads, needs only
modest amounts of memory.  separating NFS service from login activity is 
appealing from a security standpoint, since it means the login node can be 
less trusted as its more exposed.  from a failure-mode point of view,
bundling admin services onto the NFS server makes some sense, since if one 
fails, the other is likley to be irrelevant (but naturally, both should be 
separate from the depredations of users on login node(s).)

partitioning by function also lets you decouple nodes: they can differ in 
hardware config, software config, update cycles, etc.  (it's mainly login 
and compute nodes that should, by principle of least surprise, have nearly
the same architecture and software config.)