Building a Teraflop Donated Beowulf Cluster

A number of people have written in about the new Teraflop Project aka Project Übermensch. It's an interesting idea-these folks want to get essentially the equivalent of 10757 AMD KII 350s, and turn it into a monster Beowulf cluster. In exchange for donating a machine to the project, you get a month of full bore processing power from your old machine, as well as a for-life e-mail address. They've got an address on the site to send machines to-but how often do you think one of these things is gonna break? I'd hate to sys-admin thousands of old boxen.

Comments

Where are they going to put this thing?

[ElJefe]

Let's have a show of hands: how many of you have actually seen a Beowulf cluster? Where I work (Caltech's CACR), we have a 114-node system, and it's pretty damn big. You can't just have all of the nodes packed densly: you need to be able to access the backs for networking, power, etc. It takes up a pretty sizeable amount of floor, and reaches up to the (rather high) ceiling.

Here's a couple of pictures. The one up top is just one side of it.

These guys want to make one that's over 100 times as big! Can you imagine the network cable nightmare? Not to mention the power requirements. Makes you feel sorry for the technician that has to set it up.

The other big problem with a large cluster is network latency. You can reduce the effects of this by passing larger packets of info, but there's still a limit that you reach. Just because you make something 100 times bigger doesn't mean it'll be 100 times better.

I also think that the software configuration would play a major role in the efficiency. I'd rather trust trained scientist (not me; I'm just a student), who's been working with large-scale parallel machines for years to set this up, not some tech guys who thought it'd be a neat idea. But maybe I'm just pessimistic.

Still perfectly happy with my 1-node PII... -ElJefe

Re:Where are they going to put this thing?

[drudd]

To be more efficient they'd probably have to have multiple layers of clustering, i.e. have a gigantic Beowolf cluster which is made up of smaller Beowolf clusters.

Assuming they can get the networking and power resources and have the warehouse necessary to house all of these computers, there's very little you can actually do in such a massively parallel system. The hardest part would be simply feeding it enough data to keep all the nodes processing concurrently.

A much more efficient use of such massive resources would be to split them into smaller clusters of 50-100 computers in size, and allocate time individually. Then when a really large project comes along you can merge a few of the smaller clusters together.

IMHO, if this project is for real, it's still a pretty shabby deal. Granted that old 386 box you have is pretty near worthless in today's sub $500 market, but email addresses are cheap too. Donate that box to Goodwill and give some poor child the chance to familiarize himself with the technology necessary to succeed in the next millennium.

Doug