On Thu, 15 Jul 1999 [EMAIL PROTECTED] wrote:
> Howdy. I'm in the midst of cobbling together a beowulf cluster. I want
> to be able to demo some kind of application on the beast that will
> easily show speedup vs. scalability with 1 node, 2 nodes, 3
> nodes...etc..16 nodes. Also, the demo should be as flashy as possible(to
> impress some rather calcified managers). Any ideas/suggestions? Thanks.
Hmm... Sounds like a question better asked on a more Beowulf speciffic
mailing list... This is an SMP list.
Anyway, it all depends on what sort of thing it is going to use it for. If
you are after something visually impressive, then you might want to
demonstrate the speed of rendering a complex 3D image, with a program such
as PovRay.
There is already a walkthrough about settign up an Extreme
Linux / Beowulf cluster and then testing the rendering performance
increase over a standalone machine on it. The PVM (or was it MPI?) patches
exist for PovRay, and they will allow you to efficiently distribute the
workload across the whole cluster. You should be able to find this
information fairly easily. I haven't got my bookmarks handy, so I cannot
paste the URLs straight in here...
It would be most beneficcial of all the
machines in the cluster are the same (or if making them identical is not
possible, then as similar in performance, especially on the CPU front, as
possible), because from what I read of it, it looks like distributing the
workload works better with machines of similar spec.
Then you can demonsrate to them how long it takes to render a picture on
the screen on a single machine, and how long it takes on a cluster.
If you have done it properly, the performance should scale pretty well.
On the other hand, that sort of performance demonstration may not be
adequate. It all depends on what they want to see. However, the rendering
performance does demonstrate the over-all performance increase of an
application that is written to support the advantage of having a cluster,
and which requires lots of processing power.
Of course, in your case, an application more related to it's final purpose
might be a better demonstration.
If you simply want to run one single-threaded application on the cluster,
you will get no performance increase whatsoever. It all depends.
If you want to run lots of programs independently on the cluster, for
example, for a web server farm, or like me, you want to be able to
train lots of neural networks in parallel, then what you may want to
consider is a software such as Mosix (do a search on the web, and you will
find it). This will automatically try to distribute the processes onto a
cluster.
Without more speciffic information on the intended use of the cluster it
is difficult to be more specific with the answer...
Hope that helps.
Gordan
-
Linux SMP list: FIRST see FAQ at http://www.irisa.fr/prive/mentre/smp-faq/
To Unsubscribe: send "unsubscribe linux-smp" to [EMAIL PROTECTED]
