Xbox Linux Cluster

aeiz writes "Adam Cecchetti put together a Linux cluster using 3 modded Xboxes and was quite surprised with the results. He used the Mandrake Linux distro." illumin8 summarizes Cecchetti's conclusion after investigating "the cost-effectiveness (or lack thereof) and trouble involved in setting up a 3 node Linux cluster based on the Microsoft hardware. The end result: A cheaper Walmart PC would perform better at the same task."

Comments

Re:Quick Summary... and a Why?

[Kentamanos]

I'm a bit confused as well...

Why would you want to buy an XBox for this in general? It's somewhat difficult to determine the cost factors here, but I would think you could buy just as powerful of a MB/CPU/RAM combo (actually, you could get a better CPU easily) cheaper. Heck, it's hard to even find people still selling 733 Celerons.

The only thing that might blow the budget is the video card you would need (it's a specialized G2, right?), and if you're clustering them, what's the point of having a good video card?

Seems like a fun little project, but the article seemed pretty light on actual performance data, etc.

Well if you...

[Solidblu]

Well if you happen to have 4 Xbox's for 16 player halo you could always cluster 3 of them together to do setiathome because you probably aren't gonna use all 4 for most other things..

Bad Writeup

[hakker]

Not to be a troll, but this guy seems to be claiming he knows what he's doing. From his writeup, I very much doubt that. Performance modelling a cluster is FAR more complicated that "this is how long it takes to compile a kernel on one box, and this is how long in parallel." There are all kinds of issues such as MPICH's underlying communication structure (Tree, Linear, Cube, whatever). The fact that he could not get his MPI test program results consistant tells me that he has no idea how to configure it. Furthermore, when is the last time you saw a Linux kernel take 48 minutes to compile, on any Pentium 3 machine? The XBOX is capable of building a kernel far faster than that.. As another reader pointed out, read the specs.

Oh yeah, one final note. I *hate* how everyone always thinks that they were the first one to think of building a Xbox or PS2 cluster. Kudos to this guy for actually coming through, but his scientific reporting skills leave much to be desired.

Trolling... perhaps

[josh crawley]

I see people talking how this is not cost effective, but there's something everyone's missing. Where's all that money invested into? The graphics card. Wasnt there a slashdot story that pumped data into the gfx card as vector equasions and then read the output of the data? And for some reason, I remember a GF4200 putting out a theoritical 1 Tflop.

Perhaps this isnt soo cost uneffective?

Re:Trolling... perhaps

[RealAlaskan]

Is the graphics card doing GOOD floating point, or is it doing fast floating point? It is possible to take short cuts which won't screw up many applications (probably including graphics), but which can be absolutely unacceptable if your answer has to be right. Here is a link to a paper which should expand on that idea.

I seem to recall that when compiling BLAS for an AMD chip, one of the ./configure options was to use the 3DNow! extensions. There was a warning against doing that, since the speedup came at the expense of accuracy, in some situations. See here .

Re:Quick Summary... and a Why?

[Ghengis]

if you're clustering them, what's the point of having a good video card?

Not that you would use X-boxes for this, but on a similar note, some visualization clusters are specifically meant to harness the power of the video cards. The MB, CPU's and ram are just there because they have to be there (you can't plug 1024 video cards into one x86 box!)

Why not OpenMosix?

[ReelOddeeo]

If you're trying to cluster XBox'es, why not use OpenMosix? I think everything about it is cool. The load is distributed automatically. Applications don't have to be specially written. All of the boxes can see each other's filesystem? Maybe there is some downside to using OpenMosix for clustering?

I guess Saddam

[asv108]

Will not be upgrading his ps2 cluster for quite some time.

Re:Xbox mods should focus on a PURPOSE

[Rolman]

Well, yes and no. There is a reason why the bandwidth requirements for a video chip are so high. The GPU needs really fast access to Geometry, Texture and Shader data.

For a rendering farm node, the Xbox has some harsh limitations on memory size (64MB in UMA for both CPU and GPU), also the 100Megabits/s network interface (actually like 12MegaBytes/s) is not that fast when comparing to the 6 or so GigaBytes/s the chip needs.

So, the 64MB system memory will impose a big limit on how much data can be pumped to the GPU (remember you need to store the results in memory too, uncompressed), and the network interface will not help to alleviate the memory load fast enough. This makes it useless for something more complex than home-made graphics, and then a regular PC would make a better job for a home user.

If the Xbox had a lot more memory, it would be easier to make a valuable rendering farm node out of it.

Re:Looks as if MS has succeded.

[aminorex]

On the contrary!

While an XBox is not competetive with a PC as a compute node
or as a desktop system, it is much superior to a PC as a
dedicated appliance, providing a network service, because
it occupies less space and consumes less power. That's what
I use mine for. And I pay $150.00 each for refurbs, not the
$328 fullkit that the author of the referenced article pays.

By having different network services on different dedicated
boxes, I gain uptime. When I need to frob the web server,
the file server remains stable. When I need to upgrade the
fileserver, my wife can still read her mail.