[Beowulf] quick and dirty cluster demo

Robert G. Brown rgb at phy.duke.edu
Fri Jun 23 11:51:41 EDT 2006


On Thu, 22 Jun 2006, Jim Lux wrote:

> I was recently asked about how one would go about doing a quick cluster demo 
> with the following configuration:
>
> a laptop running a TBD version of Linux
> a bunch (4-16) of diskless nodes  (implying that they need to netboot with 
> PXE, etc.)
> all connected with ethernet..
>
> Seems like something like warewulf would do this nicely, but are there other 
> alternatives? ClusterKnoppix?
>
> And then, what's a nice demo showing speedup?  POVray rendering?  glitz and 
> flashy is important, raw speed is not.  It would be nice if you had something 
> that took, say, 30 seconds to run on just the laptop, and then, as you add 
> nodes, it gets faster and faster.

POVray is "traditional", but I've also used xep (the mandelbrot set
program that comes with PVM) to good effect.  You can "see" the cluster
nodes returning mandelbrot strips in differential time with xep (some
strips require a lot more work than others).  It isn't so easy to find
an application that can automatically add nodes as they become
available -- a lot of computations configure the virtual cluster when
the program is initiated and then don't reconfigure in real time as it
executes.  xep is one of the few demos that does this -- you can add
nodes in real time and it increases its worker count accordingly and
automatically.  xpvm (running at the same time) becomes a fairly nifty
way of "watching the nodes work"...

Unfortunately, the demos have grown less impressive over time as Moore's
Law has eaten into the time required to do the computations on a single
box.  Time was that the mandelbrot set required minutes to compute a
given rubberbanding to a large depth (or even hours).  pvm and a nice
little cluster reduced it to seconds.  Now, however, my 32 bit
speed-shifting laptop CPU can do most updates in a few seconds to start
with, and it is half the speed of a reaonably fast/current i386 node,
maybe 1/3 to 1/4 the speed of an AMD64 or opteron.  I haven't run POVray
in demo mode for a long time, but I'm guessing that the standard vase
demo renders fairly quickly with one CPU and like lightning with a small
cluster at this point, as it was maybe 30 seconds with one CPU vs a
second or two with a small cluster six or seven years ago (scaled by
Moore's law).

So you may have to find some way to beef up the "work" a bit to get
visibly better scaling and drive the point home.

   rgb

>
> James Lux, P.E.
> Spacecraft Radio Frequency Subsystems Group
> Flight Communications Systems Section
> Jet Propulsion Laboratory, Mail Stop 161-213
> 4800 Oak Grove Drive
> Pasadena CA 91109
> tel: (818)354-2075
> fax: (818)393-6875
>
>
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit 
> http://www.beowulf.org/mailman/listinfo/beowulf
>

-- 
Robert G. Brown	                       http://www.phy.duke.edu/~rgb/
Duke University Dept. of Physics, Box 90305
Durham, N.C. 27708-0305
Phone: 1-919-660-2567  Fax: 919-660-2525     email:rgb at phy.duke.edu


_______________________________________________
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf



More information about the Beowulf mailing list