[Beowulf] Computation on the head node

[Joe Landman]

Hi Vlad:

Vlad Manea wrote:
> Hi guys,
> 
> I have a question regarding the head node for a small cluster that I 
> would like
> to build:
> 
> which are the disadvantages of using the head node as a computation node?
> 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...

Ignoring some of the other noise on this, I would recommend you follow 
something like what Mark Hahn had suggested.

A simple "quick and dirty" guide to this can be had at 
http://www.linux-mag.com/id/3535 .  This assumes you use Rocks for 
setting the cluster up.  Others work just as well, Perceus is good kit.

As for using the head node for computation, as Mark noted, it really 
isn't a great idea, as you will be having your compute processes compete 
for resources (CPU, network, IO) with the management/file server processes.

Generally as your cluster grows, you start running into scaling issues 
with various elements.  My comments on RAID before were a caution that 
you shouldn't ignore IO (which many people do in fact ignore) when 
designing/building this system.  Likewise, if your codes are bound by 
latency (CFD/structural codes), you may wish to explore lower latency 
network fabrics for your communication layer.

Before we go down any directions, could you describe your codes, and 
what will generally be running on the cluster? This would help steer 
some of the larger design issues (Monte Carlo clusters look very 
different from CFD clusters from ...).  Many groups we have spoken to in 
the Geoscience computing arena are I/O bound (as well as FFT bound).

Joe


-- 
Joseph Landman, Ph.D
Founder and CEO
Scalable Informatics LLC,
email: landman at scalableinformatics.com
web  : http://www.scalableinformatics.com
        http://jackrabbit.scalableinformatics.com
phone: +1 734 786 8423
fax  : +1 866 888 3112
cell : +1 734 612 4615