has anyone here set up a beowulf(cluster) successfully?

if so, what programs did you use to manage the cpus and what compilers are best to parralize programs? i read some of the beowulf howto and heard that there are compilers that can automatically parralize a program, given the number of cpus, memory, etc. it would be cool to set up a cluster because it would be cheap, and i have a bunch of 200 mb harddrives laying around... i can mount the clients filesystems on an scsi harddrive hooked to the server...

heheh thats some cool sh*t

CLUSTERS RULE...
BUT LINUX DOESNT LIKE ME YET...

I think you have a bit of a overrated idea of a beowulf cluster. Four pentium 100mhz in a beowulf cluster won't necessarily be faster then one 350mhz pentium.
I personally never heard of any compilers which can automatically adapt a program for a beowulf cluster. But they could be pretty well possible, although I think they would be very limited. Getting real speed out of a beowulf cluster is done by making programs which have been designed for a system like that. You can't just take any program and use some compiler switch to make it optimal for parallel processing.

Anyways I've never set up a beowulf system personally so I wouldn't know about all the details. But from what I've read about beowulf clusters is that they are only effecient if the programs are designed to run on a cluster which most programs aren't.

But it's a nice experiment so you might still want to try and build one since it will teach you a lot about linux.

id just like to set up a cluster for the experience. yeah, youre right about a cluster not necessarily being faster than a single cpu box, but it all depends on the programs you run. if you compile a program so that it does number crunching on two or three boxes, giving that the program is large enough for you to want to use more than one cpu (so you arent overkilling with more cpus than you need) you can get a program to run fast as hell! and since many linux games are open source (and large) im hoping to be able to get some 3d games "parralized" with the help of automated compilers and my meager skills as a programmer.

You would have a bandwidth pipe unless you ran gigabit, and even then you'd run into latency issues with memory.

When a game is running, it's reading and writing from CPU memory to Video Memory(or even System Memory to Video Memory). This access time is generally extremely fast <1ns).
Putting all of this onto a network would add rougly 10ns to the access time(not to mention the time to crunch on a P100) and then another 10ns return(this is taking into effect the CPU->Memory->HD->Network->HD->Memory->CPU->Repeat cycle). Games are also not able to predict movement. Imagine if you were in a driving game on a beowulf system. How can the CPU know your going to say...turn right at a given time? IT can't until you actually do it. Hope you see what I'm getting at here.

I'm by far not the best judge of this, but my system architecture course tells me that it would be painfully slow.

A beowulf system is much much better for number crunching large amounts of data(think of SETI @ Home). They work best when you can segment your data into large chunks that each processor can work on individually and then reconstruct on one or two other processors.

It would be interesting to build a beowulf, but without the necessary software it would basically just sit there eating up space. My future goal is to write a video render program that could run in a beowulf cluster, and make it opensource of course.

J