[Beowulf] Clusters and Distro Lifespans

[hahn at physics.mcmaster.ca]

>> Fedora isn't the only kid on the block, and given the short support
>> lifetime I would recommend against using it in a cluster.

that's an odd statement - do you specifically try to keep the software
on a cluster unchanged for years?

>> Fedora is clearly aimed and billed as a cutting edge distro, which will
>> have a release every six months.

a 2/year release cycle is not the same as being unsupported in 6 months.

> i)  Is the practice of buying clusters with pre-installed distros popular?

it's a tradeoff, mainly in the human parts of TCO.  few places have such an
excess of talented and non-saturated people that they are eager to take on
new or extra workloads.  I find this quite insidious - the organization may
feel it needs to minimize risk if a guru leaves, for instance.  does this 
mean that all decisions should be aimed at being maintainable by any muggle
who's got MSCE?

> ii) Would it be better to develop our own installation process for
> clusters so that upgrades, in terms of distros, can be rolled out
> easily?  I feel like i'm tied in some way to the supplier of our
> cluster for upgrades.

how many distros do you need?

> iii) Do people regularly upgrade their clusters in relation to
> distros?  I guess this is like asking how long is a piece of string
> because everyone's needs are different.

I think it's more a matter of taste/comfort.  I hate running a kernel
that's more than 6 months old, since I'm often aware of problems fixed
during that time, etc.  if your distro has only a 2.4 kernel, for instance,
IMO, you are losing quite a lot.  if your distro has a 2.6.9 kernel,
you're still missing, for instance, sata smart support.

mostly, I think people fall into two camps: conservative, where the hardware
is chosen to be at least 8 months behind the cutting edge, and which will
have a frozen-in-time distro for ~3 years service life.  or else the somewhat
less risk-averse, happy-go-lucky, guru-advantage-taking approach of jumping 
onto good hardware when it's obviously right, and keeping software
up-to-date.

all that said, distros are mostly eye-candy.  to a HPC cluster, what matters
the most is hardware, then clustering/management tools, then glibc, then
other user-level packages (GSL, lapack/blas, compilers).  most of these 
things are either easy/ductile (cluster management), or merely repackaged
by the distro.