BEOWULF multi CPU computing project [cpu clusters w/ linux]

[DPA]

Hey gang, i joined into a project that is currently underway at campus, and the project is to link about 15 P2 400MHz computers together via network to make them "think" as one mind. In other words the objective is to create a server from these towers and make them all execute on one singular task.

I have a couple of questions for those of you who know what to do.

A. They are using Redhat 7.3 which has absolutely no GUI in it at all, and everything is run via command line entries. Is it a smarter/easier move to install something like mandrake 9 or another version of linux that is free that uses a GUI like KDE or Gnome?

B. There appears to be a "server" computer for this cluster of computers that has more ram and more hard drive in it than the rest of the other boxes, will it be capable of working with the cluster computers on a singular task? or does it manage the process of dividing the work load amongst the slave computers?

any input is appreciated on this, especially if any of you have done something like this. I'd like to know what you did and if you'd be willing to help point me in the right direction.

The professor who is somewhat in charge/lead here--more of a democracy in here--is using a software package called "beowulf" so if that rings any bells for any of you, please respond.

thanks

Kip

May all your questions be answered....

http://www.canonical.org/~kragen/beowulf-faq.txt

Robert Richmond

[Redwolf]

OpenMosix is your *best friend* in clustering. It gives a GUI for clustering stuff. Very, very, VERY nice and good.

http://openmosix.sourceforge.net/

You set it up on a 'server' machine, which can be any machine as long as it has the package installed. The other machines don't even need hard drives as long as the system is set up right.

Unfortuantely they don't have a 2.6 kernel out yet, but it is certainly being worked on. It also isn't Beowulf, but it's the same basic concept. Beowulf is alot harder to pull off.

Good resources for Beowulf itself:

http://www.beowulf.org/ <Duh
http://www.linuxhpc.org/
http://newton.gsfc.nasa.gov/thehive/

I have seen OpenMosix in action. Nice stuff, especially for first-time grid computing users.

The kernel extension technique makes everything transparent to the end-user. Unless you show them the hardware, they wouldn't have a clue that they are not setting in front of a single Linux workstation.

BTW, Xgrid is one of the better grid engines I have seen outside of Sun's proprietary system, but it is only for the Apple Mac.

[Redwolf]

Whle I'm thinking about it, what kind of applications are you going to be running on the cluster?

[DPA]

Im not sure that the project has a definite goal for what is to be computed as an end result.. i think that they are more focused upon getting to square 54 (actually getting the cluster to work)

One professor did mention to me that if i would like that perhaps i could try using it for a school based rendering project--the total power of the cluster should be around 17,000-21,000 MIPS

*on another note, i did not know that you could run a cluster with computers that lacked harddrives--this is very intriguing--the noise ratio of the rack would drop incredibly.

thank you very much for the links and the suggestions, they will most definately be perused to a great extent.

if you have any suggestions on what to do with the cpu power (except DC efforts) please give the suggestion, im thinking that they are looking along the lines of math and reasoning projects.

[Redwolf]

Suggestions for a Linux Cluster:

POV_Ray rendering (Photo-realistic images basied on 3D raytracing):
www.povray.org

Blender (3D Modeling) rendering (you could go almost real time here!):
www.blender.org

Mathematica 5 (equation solver-commercial software):
http://www.wolfram.com/

Just a few ideas.

I would have this professor put out a staff bulletin and ask other departments if they have uses for a high-performance machine like this. Basically, anything that involves really complicated mathematics that can be broken down into parts (like machine 1 renders pixeel 1, machine 2 pixel 2, or machine 1 solves for square 1, machine 2 square 2, etc).