[Beowulf] centos5 as cluster os

[Robert G. Brown]

On Tue, 12 Feb 2008, Jon Aquilina wrote:

> whats everyones take on centos as a cluster os.

It's fine, especially right after it is released.  It is supported for
long enough to last the lifetime of a cluster, which is good -- set up
your repo mirror, set up yum, and everything just auto-updates.  Very
low maintenance, that is.

The only negative I've encountered is its advantage.  BECAUSE it is
"frozen" and has a long lifetime, it can outlive the useful life of its
libraries as things advance.  My own experience of this involved the
GSL, which was in its "infancy" in one Centos release and rapidly
surpassed it in areas of importance to me (e.g. RNGs and ODE solvers).
Yet Centos as stuck at a much earlier version of it that literally
lacked some tools I was building into things I was working on, which
then failed to back-port or run there.

This sort of problem has a number of obvious solutions, and they're not
TOO obtrusive, but one should be aware of them as they do factor into
the maintenance.  You'll save work on automated maintenance and the need
to continually reinstall relative to e.g. Fedora, but you'll likely have
to take backport some stuff FROM Fedora-current if your users discover
it and like/need it.

The one other issue that can come up is hardware related -- obviously if
Centos installs on your hardware it installs and you're done.  IF
however you add new nodes every year, and those nodes have different
motherboards or network devices or etc, it is possible that you'll hit a
motherboard that just doesn't work with the older kernel being
maintained in Centos.

This problem is most visible in the case of laptops, not so much in
clusters; if you can run Centos on a laptop you're just plain lucky.  I
use Fedora exclusively there because EVEN Fedora lags the latest
wireless devices and other laptop-oriented hardware and software tools,
but I can usually get it to work, and even if I have to struggle for a
release, the next one usually works flawlessly.  Its rapid release cycle
is actually beneficial to laptops, roughly neutral to desktops, and
probably detrimental to servers and cluster nodes.  But it can, as I
said, surface in heterogeneous clusters that rotate hardware in and out
every year.  In those clusters, to keep all the nodes running the same
thing, you'll probably want to upgrade no more slowly than Centos/RHEL
itself is released (ideally with a 3 month lag between bleeding edge
release and your update, to give the first worst bugs time to shake
out).

rgb

>
>

-- 
Robert G. Brown                            Phone(cell): 1-919-280-8443
Duke University Physics Dept, Box 90305
Durham, N.C. 27708-0305
Web: http://www.phy.duke.edu/~rgb
Book of Lilith Website: http://www.phy.duke.edu/~rgb/Lilith/Lilith.php
Lulu Bookstore: http://stores.lulu.com/store.php?fAcctID=877977