[Steven Pigeon] got his hands on ten iPaq computers that a friend acquired through an eBay auction. The older machines were in good condition but the march of technology had left them behind as casualties. He’s given them new life by assembling a cluster. The first order of business was testing the hardware to make sure it’s working. [Steven] used memtest86+ that comes along with the Ubuntu distribution of Linux to find one bad memory chip in the bunch (a revelation that took 10 hours to discover on the slow hardware). He assembled the unit above with MDF as a support structure and threaded rod to hang the boards. He ended up with a beautiful module and his next step is to choose the operating system that will pull the whole thing together.
We find this build every bit as beautiful as the file cabinet cluster. It’ll be interesting to check back with him and see what kind of performance he can get out of it.
what would be the point of a thing like this? what can it do?
@sam Folding@home comes to mind http://folding.stanford.edu/ maybe home automation, you could do a lot with old machine :)
From the mention of Memtest86, I take it these are the iPAQ desktops, not the iPAQ PDA’s (which are ARM-based)?
@sam
It looks like he wants to build a single image cluster.. that is a bunch of machines that operate as if they were a single machine.
I’m not sure how good that is going to work out though, the machines aren’t fast or low power, he only has a 100Mbit interconnect (I doubt the switch can manage wire speed for everything connected either),.. now that you can get machines with multiple “cores” there’s even less of a point to this.
More interesting would be a high availability cluster of some sort. With that setup he has a good setup to play with stuff like DRDB and the clustering FS that are springing up all over the place.
I have a feeling this is a creation that lacks a real application though.
Heh, I had one of these in the past, if I recall the specs…
Pentium 3 1Ghz, 1Gb max SDRAM (2 slots).
@Desssstruction: Right. They came with either a P3 or a Celeron, anywhere between 500 mhz and 1 ghz.
I fail to see the point in this. You’d get much more performance out of a Wolfdale Pentium or Q6600, which are both dirt cheap nowadays. Just the power consumption is going to add up over time, never mind their initial cost.
I actually did something like this my first year in college, bought 25 k6-233’s from the high school they were gonna get rid of for 2 bucks a piece. Using openmosix (now LinuxPMI) you can get a pretty nifty, fairly easy to setup SSI.
I really was interested in the idea for something like the LTSP, so you could PXE boot say 30 old machines from a decent server or two, and still get performance for a school lab like environment. When i did the test, 12 233’s were something like 40% faster than my dual 800 i tested with doing blender rendering. It was an interesting excerise, I actually wrote a multi-process search algo for one of my data structures classes and presented that.
What ever happened to the spirit of doing something just to do it? We all know he can go out and buy a machine for $100 that will have more power. But he won’t learn anything from doing that. The practical application of this project and many like it is what is learned from doing it and tucking that information away knowing that it’s there if you ever need it, regardless how slim the chance.
What happened to you people? I can only assume that at one point you were 10 years old and got excited when you saw something on the curb come trash day that you could drag home on your bike and take apart. Where did that spirit go? What happened to the journey being much more fruitful than the outcome?
I have wanted to do something like this for a long time, unfortunately I was unable to find software (OS and applications) to run on one of these.
For software, you probably want MPI. That’s what professional HPC people use, and it’s free.
beowulf cluster!
I am on a P3 733 coppermine as my primary machine/homeserver for 11 years now. They are surprisingly more resolute and capable than some folks give them credit for.
As others have also said, it would be much more efficient to use something else. As a proof of concept and follow through though, good on him :)
When I read the title, I pictured the PDAs which would have been really funny but also neat.
Very cool :) very impressed
@blue carbuncle P3 733 you will be in for a treat when you upgrade with hardware that old.
@blue carbuncle : I have that same processor in one of my machines, and it is still surprisingly powerful. For most day-to day activities, it is plenty powerful. I even got some newer games running on it (Far Cry, HL2, even Quake 4).
Hmm…overkill?
My main workstation with a 650 watt power supply has more muscle and uses a fraction of electricity (looks like this cluster has 8 power supplies, each probably about 350 watts or less?)
Personally, just about anything that says “Compaq” on it ends up in my parts bin or disappears on CraigsList.
hm, yeah, well…
Nice as a “because I can” project…
But 1800$ you can get a fully packed IBM Blade Center with 14 blades, which have two dual core opterons @2.2GHz, 4 Gigs of RAM and a 76GB 15.000k SAS drive each. All that nicely wrapped with dual power supplys remote management and gigabit ethernet connection..
that comes down to approx. 135$ per machine…
@ Marvin
Where?
@ Marvin
Yea, where? That could come in handy.
“My main workstation with a 650 watt power supply has more muscle and uses a fraction of electricity ”
I absolutely agree with this and I would say the main problem is the power consumption. With a new active fsp supply you save up to 80% of the energy. This is why I throw out all my old machines.
Also why would you have cluster at home when you can get access for more and more like Amazon E2C. In case you would use it for distributed password cracking those are also exist, see freerainbowtables, wpacracker, plaintext info etc.
I built an 8 core server with 16gb ram for compiling my apps because it can do the job with multiple threads in seconds but the machine mostly idling. Running tools like setihome are kinda pointless…
The problem of power consumption/low efficiency that has been mentioned is the main reason I haven’t done this with my old machines.
If they haven’t already, someone needs to design an open-source, scalable (modular?), switching-mode power supply that uses inexpensive components (perhaps recycled from old AT/ATX supplies) and an easy to make single-sided board, and puts out 3.3, +/-5, and +/-12 volts. Maybe it could also provide (optional) remote diagnostic/management over USB or the Internet.
Actually, I’d be happy with a basic, open-source ATX-class SMPS with complete schematic and a write-up about the design process…
If he ran this off of one high quality supply that was loaded to achieve high efficiency it would be a good idea.
I love the threaded rod supports, looks like a pain to swap boards though.
Cool hack.
The Xbox 1 just got removed from XBLive, now is the time to pick up a boatload of Xboxes to do this with.
Solder 2 points and flash them with a corrupt memory overflow. You only need one working drive to flash an entire stack of them.
@Marvin
Where? The cheapest blade server I could find on IBM’s site was $1.3k per blade, and a 14 blade chassis is $3.7k and up. 1.3*14+3.7=$21.9k
On the SMPS front I found these sites by Lazar Rozenblat with lots of info:
http://www.smps.us
smpspowersupply.com
Also, since nobody’s mentioned it yet, what about EMI and RFI shielding? You gotta keep what you make in and what others (and the earth itself) make out.
It would suck to get near the end of crunching some huge calculation and get your results corrupted by interference affecting the RAM bus. Or even worse, to not even know that the results are wrong…
@R
Here, for example: http://www.itsco.de/Server/IBM_Bladecenter_E_Chassis_8677_inkl_14_x_Server_IBM_Blade_LS20_8850_i39_3105.htm
used of course, and in germany…
I’m sure you’ll find something like that in every country…
thats kind of cool! and good price on those blade servers. hack on
I know this isn’t the best hardware to use to make a cluster, but what would be?
I render 3D animation on a quadcore that often takes days for a few seconds worth of video. I’d like to render longer projects but a recent 30 second render would have taken over a week but i had to cancel it. Even boosting the speed to a six core it would still be a 4-5 day render and I want to work on much longer renders. A top end AMD 6 core chip is the most i can upgrade to while keeping the same motherboard.
If i were to get into NET render in Cinema 4D, I could use a bunch of machines at once. I’d only need each one to have lan and the best speed/power consumption at the best price.
Any recomendations on hardware?
I’m condidering real cheap AM2 motherboards with DDR2 Memory and those cheap Athlon X4’S like the 630’s.
In the long run it may be cheaper to pay more for 6 core setups with the power saving but I’m really not sure.
Any advice appreciated.
pilot: lol. That would be pretty bad. Nah I actually have a dual core laptop and the wife has a quad core laptop for heavier stuff and portability. The main box is just the desktop workhorse for copying the library and netflix’s entire collection and any other digital crap the wife requests outside of my own data hoarding lol. It is a good little machine though. I will probably freeze it in carbonite when it finally dies ;) It has saved the lives of hundreds of other machines and can absorb more rum than I can :)
@techpops: what software are you using for 3d animations, just this week I’ve played with drqueue which is designed to have computers render a frame at a time, and to render them all. Also been trying out luxrender which can use many computers to render an image as it throws a lot of power towards computations. Also be on the lookout for rendering engines that support OpenCL or CUDA.
I use cinema 4D which can render out to other computers as long as theyre on the network. Cuda looks exciting if you’re not a cinema 4d user since it doesnt support gpu rendering. I think there are render engines you can use with it but ive got a good workflow going now all my cinema 4d textures made to use the built in advanced renderer. Switching now would really be tough.
I don’t really see why you guys are commenting on the supposedly “huge” power consumption of this cluster. These mobos are all from PDA handhelds. The entire load is probably under 120W.
@biggie smalls: Look at the picture. Do those look like iPaqs to you? ;)
Better to learn clustering on something you won’t care if you break something. Once he has this in the bag, building expensive clusters won’t be as shocking. Building something for a first time, especially clusters, price tags can get out of hand. The experience of building a cluster for the prices of power supply, and the back bone, and I am guessing his friend gave him the boards or it was a low, low, one time only price.
I personally salute him, there are things other than rendering that clusters are good for.
PROTIP: Power factor correction, active or otherwise, has negligible impact on efficiency. It does, however, help to avoid incurring the wrath of your friendly neighborhood ham or audiophile.
Very similar to a cluster I built a while ago.
http://www.flickr.com/photos/markgosling/sets/72157594343066943/
Agreed, you can learn a lot from building this, and having separate machines while doing cluster programming is IMO more fun than having a virutal cluster.
ECC is pretty nice to have if you do these kinds of things..
for those of you who are worried about speed, systems, etc…. one of the biggest issues are the inter connects.
you need to get good hardware that runs @ wire speed in order to see erious preformance gains.
I work for a cisco eseller (www.knowlagecomputers.com) and we get in all kinds of equipment. but since we’re mostly looking for the Cisco stuff out of a shipment, we usally let things like blade centers go for cheap!!!