[Beowulf] Rockport 441 Tech Design: CANPUTER

[Krugger]

So you have 4 RedHat 8.0 and 2 switches.

Phase 1
Think what is the problems you want to solve. This determines what the
use for the machines is and what is important for you.
Do you want something to crunch numbers using MPI? Or is it more like
HA cluster that provides redundancy and therefor high availability of
services?

Phase 2:
Define the arch.
- Are you going to bundle interfaces together so that it doubles the
bandwidth or do you prefer to split the traffic, one switch manages
filesystem IO and the other MPI.
- If it is a fileserver should you use a distributed filesystem
- If it is HA, how do you detect failures and keep the fallback
servers in sync with the main servers
- How are you going to manage the systems? Are you doing images so
that all the systems stay the same as they are upgraded or are you
using a NFS root.

Phase 3:
Once it is all setup you want to tune and experiment to check that the
problem you defined is actually being tackled in the most effective
way.
- For example, tcp or udp NFS connections. IPv4 or IPv6, proper
compiler optimizations, changes to the sysctl of tcp, hdparm, and
anything you can think of that can be tuned. Because you can only see
if you have real world good performance with actual testing, although
it doesn't hurt to think it through first.

It is not just a question of building one, the maintance of a cluster
is quite importante too. A Beowulf cluster usually will have some
scripts for tasks as adding users across nodes, license managment(when
there is non-free software in the mix) and other custom stuff(renicing
or PBS).

On 6/1/06, Bjorn Nilsson <bnil89 at yahoo.com> wrote:
>
> 670 BNIL                   ATD: Adv. Technology & Design            5/24/06
>
>
>   To whomever this may concern it is in the Beowulf project of the Rockport
> High School Advanced Technology and Design Class, where a special team of
> computer enthusiasts have encountered a steadfast blockade that prohibits
> the intuitively gifted group from continuing the important experience of an
> educational project. At the moment the team is in despair as to have only 4
> boxes (computers/nodes) running on Linux Red Hat 8.0. Although these boxes
> are a triumphant when compared to the unfortunate failure that our group has
> encountered. The setup for IP addresses has been completed and the boxes
> themselves are identical in all respects. The four boxes that I speak of are
> DELL Precision 220s, which contain the required specifications for building
> a super computer. Nine boxes is the correct amount of computers we had
> established as our goal. Other misc. hardware that we have installed into
> our project is 2 Belkin 4-ports, which enable us to efficiently switch from
> computer to computer, 9 CAT 5 Ethernet cables, and one Bay Networks hub. We
> have also had miniscule success in the area of communication; the command:
> ifconfig has allowed us to generate a positive establishment in the
> speculation of sanctioning "ping tests".  What can you assess from our
> current situation? Is there a series of required steps that are in a complex
> and complicated form that need to be instituted within the production of the
> super computer? Does Linux need to be propagated or properly formatted, with
> individual elements to fit the infrastructure and purpose of our cluster of
> computers? Please help us; it is imperative that we make a small amount of
> success before the end of school (June 16).
>
>  ________________________________
> Yahoo! Messenger with Voice. PC-to-Phone calls for ridiculously low rates.
>
>
>
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit
> http://www.beowulf.org/mailman/listinfo/beowulf
>
>
>