[Beowulf] The Value Cluster

Douglas Eadline, Cluster World Magazine deadline at linux-mag.com
Sat Dec 18 12:03:52 PST 2004


On Sat, 18 Dec 2004, Jim Lux wrote:

> 
> ----- Original Message -----
> From: "Douglas Eadline, Cluster World Magazine" <deadline at linux-mag.com>
> To: "Jim Lux" <James.P.Lux at jpl.nasa.gov>
> Cc: "Beowulf Mailing List" <beowulf at beowulf.org>
> Sent: Friday, December 17, 2004 5:57 PM
> Subject: Re: [Beowulf] The Value Cluster
> 
> 
--snip--
> 
> My Microtel is somewhat noisy, as well.
> 
> In your $2500 cluster challenge you laid out some requirements, but didn't
> have something like "minimum number of processors"...  Is a cluster of two
> legal?  I know you're using 8, which seems a reasonable number to
> demonstrate lots of things (and you certainly run into scaling problems by
> that time).

Well in the magazine we mention that you can certainly use less (or more)
nodes. And, a two node/X-over cable cluster should not break the smallest
budget. We chose the 8/$2500 because it was a good balance between
nodes/cost (and as you know most small switches have 5 or 8 ports). we
also want to have a nice "pile of CPUs" to play with later.  Of course
there some other issue such as using standard household/office electrical
service, heat, and space.

BTW, we will be introducing some scripts to do "poweroff"/"Wake on LAN"
kinds of things so that if you are paying the electric bill and the
cluster is idle, you can cut down on the power usage. This is also why
homogeneous hardware is important -- "If you use what we use, then 
we can be sure it will work for you kind of thing."

> 
> Do they have to be in cases?  Think of the several clusters made of bare
> mobos stacked on a shelf, or threaded rod, or,....  Save yourself $50/case,
> and on 8 cases, you might buy another processor?
> Since the fundamental purpose of a $2500 cluster is pedagogy, it's
> instructive to go through this tradeoff for all scales, from 4 processors to
> 1024.

Well, this could get interesting. My view is that there should be no 
custom work/parts involved and there should be enough "best practice" 
information to support problems. Let me explain.

You could save $40 for the cases, but then that requires some other
kind of custom work which means you are out on a limb of sorts. 

For instance, you could be more space efficient, but the hassle of
mounting motherboards, power supplies in some custom enclosure would soon
out weigh the an easier commodity approach. Plus, if you run into
problems, there are plenty of books and web sites that explain how to
"build a basic PC" with commodity parts. 

I am often tempted to say, "Gee these Micro ATX boards are so small, I 
could just ...", but then it becomes a custom project and harder to 
reproduce. One of the goals is to provide a low cost and reproducible 
cluster on which people can play. The way I figure it, the more
"cluster play" the better.  


> 
> One suggestion I have is to assign some value or limit to "fooling around
> time", just to keep the concept of the dirt-cheap cluster sound.
> 
The idea for the magazine series is to minimize this time. We are
providing an 8 node recipe and if you follow it your chances of success
are pretty good. You can also stray a bit from the path and invest some
"fooling around time" and build a variation of the "recipe". The article
also explains much of the rational we used in selecting components. I
believe one of the biggest time and cost savings is that we identified low
cost parts that work well.  Of course, they are not "server quality", but
at least we know they work as described (which can not be guaranteed for
all low coat "value hardware"). For instance, I believe one of the
triumphs was finding an inexpensive case that was well constructed and did
not slice up you fingers. 

Doug 

----------------------------------------------------------------
Editor-in-chief                   ClusterWorld Magazine
Desk: 610.865.6061                            
Fax:  610.865.6618                www.clusterworld.com





More information about the Beowulf mailing list