How To: 25 duron cpu's

sirusarai

sirusarai

n00b
Joined
Dec 27, 2005
Messages
60
I'm curious how to set up a machine to run 25 duron cpu's in parellel. I have enough common sense to know that they don't go on one motherboard. So how does something like this work? Thanx,

Sirusarai
 
Yeah, you'd be talking about a performance cluster.

It all depends on what you're trying to do though. You can check out the mosix project for a nifty linux solution - it makes it way easier if you're running the same core hardware (same motherboard on each node, etc).

So what task are you trying to come up with a solution for?
 
Do a google for "Beowulf Cluster"


Hell, hit "I'm Feeling Lucky" and you're sure to find something.
 
sirusarai said:
would this be a stupid project?
Click to expand...



no, its been done, and regardless of that, it is quite interesting...however, what will you be using this for?
 
power and heat might be a problem unless you have a decent lab setup (or run it in your garage)

25 x say 200watts = a 5 kilowatt space heater, a suprising power bill, and likely a tripped circuit breaker

Still, should be a fun project for you

 
IanG said:
power and heat might be a problem unless you have a decent lab setup (or run it in your garage)

25 x say 200watts = a 5 kilowatt space heater, a suprising power bill, and likely a tripped circuit breaker

Still, should be a fun project for you
Click to expand...
I think 200w is a bit high for 25 durons. Even at 100w a piece, however, you're looking at quite a power draw.
 
It's not necessarily a bad project but it will cost a pretty penny and will be pretty involved even if you use some low end hardware. It would definately be a fun little project to pull of though, hell, if you kept enough notes and did a good write up on it, you could probably use it on a resume.
 
well i've done a little researching, and the beowulf setup looks like what we are shootin for. a few questions though, i see a lot of these setups are using gigabit switches to connect all the nodes, and i was wondering if that would in any way slow down data between the nodes or if that is more than enough to not bottleneck the dataflow. and i was wondering if there were other ways to connect your nodes.

thanks
 
I almost hate to post here...

Unless you have a specific goal in mind, you're not going to get much out of a cluster. If you just want to run F@H, don't bother setting up a cluster; just run them as individual machines. All things like Beowulf and other clustering applications (MPICH, PVM, etc.) do a lot of the time is pass messages and distribute work. Since Folding@Home is already taking care of the message-passing and work-distribution, you won't (and can't, really) get anything extra out of a "cluster."

If you're calling a "cluster" a pile of machines with no software to distribute work across the cluster, it's just a farm of machines.

The vast majority of clusters that are actually used as clusters have job-specific applications written for them that are designed from the ground up to take advantage of the work-distribution system built in to the clustering software. Most people will never have anything that can do things of the sort, nor will most people be willing to write software that actually uses that sort of cluster.

Sure, it seems cool...But you'll spend time and effort getting the thing together, then go, "okay, now what? I can't do anything with it."

As far as the interconnects goes, the faster the better. It'll run on slower interconnects, but moving data between nodes will not be as efficient. Since you're dealing with slower machines you'll most likely be all right with a slower network fabric, but if you look at all the big-nuts research clusters they all use Infiniband lately...Some still use Myrinet, but it's really falling out of favor due to a much higher cost-performance ratio than what you get with Infiniband.

I think you really need to think about what you're wanting to do before you undertake a project like this. What do you hope to gain? If it's just for the "cool" factor, it isn't worth it. If it's just for F@H, run the buzzards diskless, netbooting and running F@H individually...But setting up clustering software on them won't get you anything.
 
As to the power side of things, it will be impossible to run a 25 duron cluster on one circuit. Last time I looked a high end duron processor used about 60W. Add in power usage requirements for motherboard, ram, cooling fans, network interfaces etc, and that number will climb fast. Even using a conservative figure of 200w/system at full load you are looking at 5kw. This figure does not include any peripherals such as switches, a monitor to check individual systems etc.

Since this would most likely be classified as a continuous draw (defined as running constant for more than 4hrs at a time) electrical code (at least in my area) states that the circuit can not be placed at more than 75% load (or 1800W on a 20A circuit). That tells you that to safely power a cluster of that power draw you need 3 separate circuits (well, 2.77 to be exact)

/edit
to get actual exact numbers for your system you need to find the actual power draw under full load per system, and also account for in-efficiencies in power supplies etc.
 
defakto said:
It's not necessarily a bad project but it will cost a pretty penny and will be pretty involved even if you use some low end hardware. It would definately be a fun little project to pull of though, hell, if you kept enough notes and did a good write up on it, you could probably use it on a resume.
Click to expand...

are you talking about a college resume or a job resume? I wouldn't mind keeping tabs on crazy projects, but would something like watercooling count?
 
I'm not sure about watercooling as that, comparitivly, is an easy project. Vs getting 25 systems to work in combination well.
 
Out of curiosity, could you use something like this for a render farm?
 
QuimZ said:
Out of curiosity, could you use something like this for a render farm?
Click to expand...
Most likely...don’t big movie companies (like Pixar) use a similar scheme?
 
I use a product at work for (near) real-time simulation computers called reflective memory. The computers are connected in a ring using fiber optic cable and each computer is set up so that it has an area in memory which is shared with all the other computers. Data written to memory by any computer is broadcast to all other computers so they all can read it. Data transmission is at the speed of light and any computer can send interrupts to any or all other computers. This technology allows more exact synchronization of all computers in a network that is not possible with ethernet. The PCI-5565 cards are expensive though... $2500 per PC.
 
rayman2k2 said:
Do a google for "Beowulf Cluster"


Hell, hit "I'm Feeling Lucky" and you're sure to find something.
Click to expand...

"I have a Beowulf Cluster of atomic supermen...."
 
Frank4d said:
I use a product at work for (near) real-time simulation computers called reflective memory. The computers are connected in a ring using fiber optic cable and each computer is set up so that it has an area in memory which is shared with all the other computers. Data written to memory by any computer is broadcast to all other computers so they all can read it. Data transmission is at the speed of light and any computer can send interrupts to any or all other computers. This technology allows more exact synchronization of all computers in a network that is not possible with ethernet. The PCI-5565 cards are expensive though... $2500 per PC.
Click to expand...

:eek: I'm jealous. Where do you work? Or if you're not willing to say (understood) what sort of work do you do?
 
frank, we're going to follow you to work and gank all your machines...
that's a sweet sounding set up

I know WETA(LOTR) and Sony(FF:TSW) both used clusters, I'm not too sure about PIXAR(probably the same)

For LOTR, Weta had 3300 machines running round the clock to render the scenes. IBM opened up a production line just to create chips for Weta in August of 2004 because they were estimating it would take 347 days to render the final battle, about 250 days too long to meet the Dec. 18 release date.

http://news.bbc.co.uk/2/hi/technology/3672887.stm
 
AragornTK said:
frank, we're going to follow you to work and gank all your machines...
that's a sweet sounding set up

I know WETA(LOTR) and Sony(FF:TSW) both used clusters, I'm not too sure about PIXAR(probably the same)

For LOTR, Weta had 3300 machines running round the clock to render the scenes. IBM opened up a production line just to create chips for Weta in August of 2004 because they were estimating it would take 347 days to render the final battle, about 250 days too long to meet the Dec. 18 release date.

http://news.bbc.co.uk/2/hi/technology/3672887.stm
Click to expand...

Lucas is another one using clusters, though they can add any of their workstations into the renderfarm when needed.

"Our machines are working 24x7. When the artists go home in the evening, those workstations become part of the render farm, thanks to our own job scheduling program," Plumer said. "The jobs go into a large queue and from there are assigned to a processor - it could be any AMD processor on our network, so the render farm can grow to 4,000 machines at night."

Link

 
To the OP. I did this relatively successfully using ClusterKnoppix a while back. With that, you don't even need hard drives in the machines to get your cluster working. The problem I had was that most programs are not designed to run as a cluster (kinda like SMP support) and thus I could only start one instance of a program on each computer rather than running 25 instances from one computer on a 25 machine cluster and having the work distributed. The only benefit that I found (which was admittedly pretty neat) was that the machine that I designated as the head would offload lower importance tasks to the other machines automagically, theoretically freeing up the machine that I was working at to do more. However as we know and love Linux for not having so much background junk this didn't really help too much.

There's also the issue of networking the cluster. Either you go gboe or fiber or ethernet. Two are costly, one severely hampers the speed of a cluster. Not a really good solution there either.

In summary, it's a neat project and you should do it if you're doing it as a learning experiment, but if you want to fold like a master origami professional on meth I'd sell the 25 Duron machines on ebay and grab as many opteron 165's or Pentium D's as budget allows.
 
this is going to be a a project done just for the sake of doing it. we have too much time on our hands.

and in my first post when i said "25 durons" what i really meant was "cluster of really cheap old computers" :)

but anyways, is there a way to measure the speed of such a machine? i know if i am using 25x 1.8ghz durons i dont have a 45ghz beast..

and all of you who have experience using clusters, tell me more about the setups you were using..(hardware used, what were they used for..)

thanx
 
sirusarai said:
I'm curious how to set up a machine to run 25 duron cpu's in parellel. I have enough common sense to know that they don't go on one motherboard. So how does something like this work? Thanx,

Sirusarai
Click to expand...

In this context, stop thinking cpus but rather think computers. You physically cluster computers and virtually cluster processors.
From what you wrote, I don't think you intend to cluster computers.

Now, with that said, I think you meant to talk of linking multiple motherboards together via direct hardware connections to form a cluster.
Well, this is not yet a consumer level type of thing. A few institutions are experimenting with it. Google and you will find some info on it.
 
Opteration said:
they need to open it up for a F@H type deal, that would be such a cool cause to contribute too.
Click to expand...

But if they did that, it would take even longer to come out with another Star Wars movie.
 
superkdogg said:
But if they did that, it would take even longer to come out with another Star Wars movie.
Click to expand...

After these last three, I think I can deal with that.

Going back to the cluster... have you tried looking in to ITX boards? A bit slower (and *possibly* more expensive, depending on what you go for), but definately less power draw and definately less space.
 
those ITX boards look like a good idea, and were you thinking of just clustering those systems, or actually connecting those boards together in some fashion i am not aware of :p
 
sirusarai said:
but anyways, is there a way to measure the speed of such a machine? i know if i am using 25x 1.8ghz durons i dont have a 45ghz beast..
Click to expand...

That's good; you're ahead of many people around here with only that realization.

Most clusters use a total count of teraflops to measure their capacity to do work. They'll use a distributed version of the LINKPACK benchmark. If you follow that lead, you'll be able to compare your cluster to http://www.top500.org/ pretty directly.

That you're doing this without an intended application seems a little strange to me. You're going to spend hundreds of dollars on cabling and infrastructure alone, not to mention the machines, but you don't know what such a configuration is used for or what you will use it for yourself?
 
mikeblas said:
That's good; you're ahead of many people around here with only that realization.

Most clusters use a total count of teraflops to measure their capacity to do work. They'll use a distributed version of the LINKPACK benchmark. If you follow that lead, you'll be able to compare your cluster to http://www.top500.org/ pretty directly.

That you're doing this without an intended application seems a little strange to me. You're going to spend hundreds of dollars on cabling and infrastructure alone, not to mention the machines, but you don't know what such a configuration is used for or what you will use it for yourself?
Click to expand...

He's learning.

Whether it's that it costs a lot to build and maintain or how to run a program using a cluster, is unknown, but he's learning.
 
ScHpAnKy said:
Whether it's that it costs a lot to build and maintain or how to run a program using a cluster, is unknown, but he's learning.
Click to expand...

Sure, and I'm all for learning -- in a smart way.

For the prior fact, there's easier ways to learn than to do the exercise.

For the latter fact, I'd still figure someone would plan ahead to the intended application before embarking on the hardware build... that would influence the choice of hardware and the software development strategy.
 
You must log in or register to reply here.
Back
Top