Good reasons to choose Ubuntu for large clusters

Carsten Aulbert carsten.aulbert at aei.mpg.de
Fri Aug 17 08:43:13 UTC 2007


Hi all,

sorry for the wrong reply to Kristian's email, but I was not subscribed
when I sent the email - but are now - and forgot to add that info to my
email.

Kristian wrote:
> I don't know why CentOS would give you any benefit over Ubuntu unless
> you were patching the kernel.  For instance, for your cluster, you may
> want support for the new Completely Fair Scheduler (CFS).  Ubuntu will
> surely include support for CFS in Ubunt 8.04 LTS, due to the kernel
> inclusion >= 2.6.23.  CentOS may take a long time to include CFS, and
> in fact, may not until RHEL6 comes along.  So, if you want cutting
> edge features for your cluster, go with Ubuntu.  Otherwise, I don't
> see how CentOS offers anything else over Ubuntu.


I fact I'm really aiming for 8.04 LTS for the new cluster, we are
currently planning here, but if CentOS would be chosen as a single
reference platform, my boss will most certainly insist on using that.
But wouldn't it be nice to have a TOP100 of TOP500 cluster running
Ubuntu[1]? But CentOS has a few god points here in our collaboration, mainly

* most of our colleagues are US based and RedHat has/had a good foot
everywhere there. And thus CentOS is a "clean" solution to continue work.

* one major pro for CentOS is that RedHat announced direct support for
Condor, a cluster batch queuing system we are using for years. It does
work on Debian/Ubuntu, but since the sources are still not open, it's
usually a big hassle.

* Also, CentOS has QFS client (of SUN fame) support, which is used
widely at Caltech and we are using that as well. My hope is that SUN is
keeping their promise to open source that package as well and it can be
"ported" to Ubuntu easily.

One big pro for Ubuntu IMHO is the FAI package. We have managed to
install a standard node in less than five minutes, almost the same time
for a file server/head node (meaning the time from first hit of the
start button to a fully installed and configured node, ready for business).

I would appreciate any more hints, if people could think of more reasons
I can add to the proposal. Right now I think the strongest points are
predictable releases, stable releases, large communities behind Ubuntu
and Debian, even interoperability of those to is great. With prevu (or
pbuilder) I can simply crate a "custom" package for my currently running
version of Ubuntu - mostly that is ;)

Thanks for any further input.

Carsten

[1] I won't promise anything in thaqt because for that we need to find a
linpack guru.




More information about the ubuntu-server mailing list