Simple Analyzing program

[Robert G. Brown]

On Wed, 25 Oct 2000 mogema at stonemedia.com wrote:

>           Recently, I decided to make a small 5 node cluster for my
> Science Fair project, and I am in need of alot of guidance. First off, I
> need some software that would need to do two simple things. 1) Be able
> to run across a Beowulf and on a single computer, and 2) Show the
> immense processing capabilities of Beowulf versus a single computer. If
> anyone is familiar with Seti at Home, or DNETC, I was thinking of a
> modified version of something along those lines, that, on the Beowulf,
> would run across the network and send its statistics to a proxy node,
> that would display the overall results of all the nodes combined. Then,
> for the "antagonist" part of the science fair project, run the same
> program on a single computer, to show how a Beowulf can improve
> performance, at a semi affordable rate, to applications as simple as web
> servers, to as big as database processing.

One of the "funnest" things you can run in parallel as a demo is the
parallel mandelbrot set generator that comes in source with PVM (I think
there is a similar toy available for MPI as well).  You can rubberband
your way down into the set and actually see bands of the new image being
returned by the various nodes, which makes it easy and fun for folks to
play with at the fair itself and also lends itself to the presentation.

Another common demo is povray (the "Persistence Of Vision" Raytracer).
There is a PVM version of it.  At Linux Expo a year ago we distributed
this across some 48 PII processors and it could render the image in
something like one second.  Again it works by chopping the image up into
chunks and farming the chunks out to nodes, and then gluing them back
together again in the final assembly.

Both PVM and MPI come with some limited demo/benchmark codes included
for testing purposes, and if you do a websearch you can probably find
some other relatively interesting applets (like the Mandelbrot set toy)
that use them and are good for demos.  Seti or RC5 are good
demonstrations of the power of distributed computation or clustering,
but they are SO embarrassingly parallel that they really aren't that
great a demonstration of beowulfery.  Not that the mandelbrot set toy is
really any better, but at least it uses PVM in a "real" master/slave
sort of calculation rather than a script based distribution of
truly independent chores.

>           As of right now, I am still working on getting enough
> equipment for the project. I am soon going to be contacting companies
> such as Dell, Cisco, and 3Com, to see if they may have any student
> programs I could become involved with to somehow get the equipment
> through donations or through a lease / borrow of some sort. If anyone
> knows of any organization / company / person that may be able to
> financially help would be greatly appreciated, as well as any other
> comments or suggestions.

Wish I could help.  My only suggestion is to try to get moderately
balanced equipment.  Demonstrating parallel speedup works best when (for
example) the mandelbrot set generator pops through all the strips in
about the same time.  If you have a much slower processor in the bunch,
the strip it does will typically hang out long after the rest of the
figure is complete which does illustrate an important feature of
efficient parallel programming, but not exactly the way you want to
illustrate it...;-)

Note that if your school has a computer classroom with (say) a bunch of
more or less identical pentium-class systems on a switched network (or
even a hub) you can convert it into your beowulf WITHOUT ANY MONEY AND
WITHOUT TOUCHING THEIR DISKS by using the new Scyld beowulf
distribution.  Well, you'll have to buy a set of the Scyld CD's, but at
$2 each this is about the right level of personal investment for a
science fair project.  You'll probably also want/need the use of a
single machines WITH a disk you can write to (which could be a laptop or
your own personal system) to bring in preconfigured as a Scyld "master"
node.  That way you can set the demos up on it and test them, and can
then literally assemble the beowulf in real time the day of the fair --
assuming that the fair is held at your school and that you can set up
signs directing the judges and guests to the computer room.

The most you might need to do in this case is beg, borrow or buy an
ethernet switch for the classroom systems, which are unfortunately most
likely on a hub.  Buy will cost you around $100 (for 5-8 ports), and you
can keep it and use it to set up your HOME beowulf (just like mine:-).
Beg involves groveling in front of the store manager of a store that
sells them, convincing him or her that perhaps you or your school will
end up buying one after they see how it enables some real "computer
science" to be done in their computer lab.  Borrow involves finding
somebody on this list "nearby" with a switch they can loan you for a day
or two.

Good luck.

>           The applications I mentioned earlier, Seti and Dnetc, I just
> realized, in the context of how I envisioned it being set up, would not
> be true parallel processing, would it? Not all nodes are working
> together as one "supercomputer", instead they each operate on their own
> and tell a "master node" how they are coming along and that master node
> just takes all the statistics, adds them up, and makes it seems like one
> large multi-node computer. ... Or should I go get another Pepsi for more
> caffeine so I can actually understand what I just wrote..
> 
> 
> 
>           Well, If anyone understands what I am trying to tackle at this
> young of an age, please speak up. I will need all the help I can get.
> Every teensy tiny bit of help will be appreciated.

I don't know if the above helps at all, but I >>do<< know that Don
Becker, Erik Hendriks, and the rest of the Scyld gang turned a bunch of
systems (the public "email cluster") they had never seen before into a
beowulf in about 30 minutes of work at the Atlanta Linux Showcase and
Conference using the Scyld beowulf CD's, and that when they were done
they simply removed the CD's and rebooted the systems and poof! they
were an email cluster again.  Didn't touch their hard drives.

Heck, Scyld might even be willing to DONATE a handful of CD's to your
project if $2/CD is too much for you -- I can think of no better
demonstration of the power of their tool than a student converting a
school cluster -- nondestructively -- into a beowulf for a science fair
project.  What a coup for them, for beowulfery, and for Linux in
general!  It might even convince your school's computer person to
consider making linux a part of the school's regular computer
curriculum!  Try doing THAT with WinXX!

   rgb

-- 
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