Wanted: Turn-Key 10-Node Beowulf Cluster

forgotten password writes: "I'd just started working on my morning M&Ms, when I was asked where my group can buy a good turn-key ~2CPUx10-node Beowulf cluster in two hours. I suspect the time frame is longer than that, although the window-of-opportunity for the money is apparently on the order of days, and a quote before the procurement meeting would help. Any ideas? Who's good? What it should cost? Thanks!" If you're quick, maybe you can become the world's newest manufacturer of custom beowulf clusters.

Comments

Re:What do you run on the darn thing ?

[cr0sh]

Most of the apps for Beowulf clusters tend to be scientifically oriented apps. It really comes down to there being only certain kinds of problems that can be broken down into parallel processes that can be executed concurently, then the answers "recombined" at the end to get the result.

IIRC, Mosix allows for using the machine as one large machine, essentially allowing each process (or groups of processes) a single CPU, but funneling the result back to the main controller, so to the user, it looks like one large machine. This is different from a standard parallel processing Beowulf cluster, which behave like classical parallel processing machines.

For that domain (parallel processing), aside from coding your own stuff (hard to due, even if you are a master coder, from what I understand - due to having to understand what classes of problems can be broken down into parallel tasks, and then actually applying that knowledge to real tasks), there is one application that would prove to be "fun" - Ray Tracing.

Fortunately, POVRay has a paralleled version available, for Beowulf clusters. I don't have a link, but I know it exists (heck, you can probably get it at the POVRay site).

One other disclaimer: Everything I have said should be interpreted as "coming out my ass", simply because I have no experience at building or using Beowulf clusters or any other parallel processing architectures. Most of what I know about them have come from /. discussions, various web sites of Beowulf cluster projects, books, and magazine articles (as well as one funky book I picked up at the library detailing programming in Fortran for a Connection Machine - eeepp!). The POVRay stuff is true, though...

Been There Done That

[elinenbe]

This summer I was employed at a finite element analysis company in Philadelphia, and I designed, budgeted, and built a 10 processor AMD System (Octavian) based on gigabit ethernet, 1.2 Gig MP AMD chips on Tyan mobo's with half a gig of RAM per node, and it did not cost more then $13000 plus some setup time. (This was with a gigabit switch, etc AND the current cost with dropping prices would be less then $10000) The computer was designed to run LS-DYNA (a Livermore software finite analysis program) and it has not let them down.

Here are some benchmarks:

Octavian Benchmark runs
as of 8/17/01

Problem description:
Acetabular cup with a spherical metal ball compressing the liner into the
shell.
The effect of holes used for screw fixation to the bone is included
Total Mesh Statistics
45814 Nodes
37696 8-noded solid elements
2 contact pairs
Dynamic Relaxation Solution:

Execution Statistics
2 Processors : 62 Minutes
6 Processors: 34 minutes
10 Processors: 24 minutes

Analysis of the execution throughput indicicates a linearly increasing speed
(1/wall clock time)

And here is a review of the results:

the cluster is performing BEAUTIFULLY, and we have been crunching problems on it pretty much nonstop since it was brought online. It has really saved our butts, as we could never have met some key project deadlines without the speed. I've included some bechmarking stats below FYI for a contact problem that took only 24 minutes to run on octavian using 10 cpus. The comparable time on our $50K dual CPU octane workstation is 3 hours and 26 minutes, which translates into a speed up of 8.6 times for about 1/5 the cost.

So, What I am trying to say is build the thing yourself. You will know much more about the system, and you will be able to install any software you want to without having to deal with "customer service". Also, it will save you a bundle as a turnkey solution is nearly 3 times the price. (Even if this cluster was built with Myrinet it would still be far less then any of the pre-built solutions) Lastly, design the cluster for what you need. If your problems involves lots of RAM, then spen money there. If CPU is the bottleneck spent money there. If communication is the bottleneck....

Best of Luck,
Eric