Cluster (Bramble...) Design Discussion (Advanced)

[Svartalf]

Quote from jamesh on September 6, 2011, 09:48
And if the GPU could be harnessed (OpenCL - not available but possible I suppose), you could really get some ooomph out of it.

Heh... You can always resort to old-school GPGPU coding by mis-using texture targets and shaders.

The only rub there is that you may have issues with doing this on a Kerrighed type cluster until someone conceives a means to express the GPGPU stuff transparently that meshes nicely with the loose SMP model there.

[jamesh]

Quote from Svartalf on September 6, 2011, 16:55
Quote from jamesh on September 6, 2011, 09:48
And if the GPU could be harnessed (OpenCL - not available but possible I suppose), you could really get some ooomph out of it.

Heh... You can always resort to old-school GPGPU coding by mis-using texture targets and shaders.

The only rub there is that you may have issues with doing this on a Kerrighed type cluster until someone conceives a means to express the GPGPU stuff transparently that meshes nicely with the loose SMP model there.

Well, since I have access to the GPU compilers etc, the easiest way would be to run stuff directly on it...although writing the vector assembler is not a trivial task.... but would be good as a proof of principle.

[Svartalf]

Quote from jamesh on September 6, 2011, 17:51
Well, since I have access to the GPU compilers etc, the easiest way would be to run stuff directly on it...although writing the vector assembler is not a trivial task.... but would be good as a proof of principle.

Well... You might not need something that sophisticated.

Pretty much every GPGPU task done "old-school" tends to be a specific complex SIMD/MIMD task that gets fed to 1 to N shaders against a region of memory that is expressed in GL/GLES as a texture or a render target. What you'd need within Kerrighed or MOSIX is a means unto which you can load the shader on a given node as the process or thread moves to a given node on the cluster- and a means to know when to unload the same when the process/thread moves elsewhere or completes.

[iAreNewb]

Quote from liz on September 6, 2011, 11:05
I have a neurophysicist buddy in Zurich who models the brains of fruit flies. Last year, he showed me a demo where he was controlling the movements of actual, honest-to-god maggots with an actual, honest-to-god WiMote. (No, I have no idea how.) Possibly the most awesome thing ever.

Perhaps we could find some volunteer arthropods and go from there.

If you've ever watched TED videos, I remember that multiple presentations had to do with neurology, one of which dealt with photoreceptors genetically engineered onto nerves, so that flashes of light could control a fruit fly's nervous signals. By specifically targeting the "critic" part of the brain (psychologically speaking, the part which judges the consequences of one's actions) with a virus that distributed this genetic modification, they could induce a sense of "bad" by flashes of light.

They had the flies crawl up and down a line, and they were flashed every time they neared one side, therefore making them feel the need to stay on the other side. This might be able to translate to Wiimote control by changing photoreceptors' sensitivity to infrared range (i.e. range of most remote controls), much like snakes have. Very experimental, but possible IMO.

Meanwhile, I'm wondering if we can combine artificial intelligence, rudimentary nervous system control, and powerful biological tools (i.e. chemoreception, self-sustenance, speed, durability, flexibility, swarm efficiency, strength, etc.) A wasp hive genetically engineered to obey an AI that attacked other wasp hives with military strategy would be amazing to watch.

[Lob0426]

Uh Oh we have moved into Bioweapons now. But if you tailor this to target africanized (so called killer bees) bees you could probably get a grant in the U.S. For research.

Another realm that would be possible is biomechanical augmentation, or armored combat suits. Small distributed control devices that can handle local control and sensor functions but coordinate as a whole would be a step in the direction. Power usage of controllers has been one of the issues in these designs. A distributed cluster design.

[traveler359]

Ive been following the Raspberry Pi for some time now, and like so many others on here am already drooling over the possibilities. While its true that one can build a more powerful cluster with pc's. Its just fitting making a Bramble out of them. The Educational possibilities would be amazing. I think I read that the GPU portion of the SoC supports open GL ES and I think OpenVG was mentioned as well. Have there been any thoughts about supporting OpenCL. The OpenCL website does mention broadcom as one of the many companies that support it. I know its a complex subject but if supported it might open up some access to the Gpu and DSP logic "I think" anyway just a thought.

[tachyon]

Well, I know I'm buying at least 4-6 to learn, and then teach clustering to my kids. Should be fun! Any thoughts on management? Is something like Oscar or Rocks looking like a fit for the noob like me. From my research tonight, I think Oscar will work.

*quick edit*

I'm not an electrical engineer, but there's got to be a way to combine power supplies, to get a one wall wart to many r-pi's powered? That my next research path. It can't be as easy as getting a small pc power supply and splicing out the 12vs

*quick edit 2*

The power supply thing was bugging me before bed, I think the medical device industry has us covered. Could someone smarter than I look at these: http://www.trcelectronics.com/.....ulti.shtml and see if they got the juice to power our berries? Goodnight!

[Svartalf]

Quote from traveler359 on September 15, 2011, 03:37
Ive been following the Raspberry Pi for some time now, and like so many others on here am already drooling over the possibilities. While its true that one can build a more powerful cluster with pc's. Its just fitting making a Bramble out of them. The Educational possibilities would be amazing. I think I read that the GPU portion of the SoC supports open GL ES and I think OpenVG was mentioned as well. Have there been any thoughts about supporting OpenCL. The OpenCL website does mention broadcom as one of the many companies that support it. I know its a complex subject but if supported it might open up some access to the Gpu and DSP logic "I think" anyway just a thought.

FYI, you don't need OpenCL to do GPGPU coding- it just makes it easier.

Here's a few links to info on how one would do some of this without OpenCL:

http://www.mathematik.uni-dort.....orial.html
http://gpgpu.org/developer/leg.....phics-apis

There's a reason I started this thread...a Bramble's not likely to burn the doors off of a top 500 supercomputer, but the educational aspects are really, really good without blowing a bunch of cash on trying some of these ideas out along with coming up with clustering solutions in that space.

[Svartalf]

Quote from tachyon on September 15, 2011, 03:58
Well, I know I'm buying at least 4-6 to learn, and then teach clustering to my kids. Should be fun! Any thoughts on management? Is something like Oscar or Rocks looking like a fit for the noob like me. From my research tonight, I think Oscar will work.

Not sure. Management's going to be dependent on which cluster type you assemble.


I'm not an electrical engineer, but there's got to be a way to combine power supplies, to get a one wall wart to many r-pi's powered? That my next research path. It can't be as easy as getting a small pc power supply and splicing out the 12vs


It can be that simple- the only problem with doing that is that you're going to have fun keeping the voltage stable on the rail coming out of the switcher you'd use.


The power supply thing was bugging me before bed, I think the medical device industry has us covered. Could someone smarter than I look at these: http://www.trcelectronics.com/.....ulti.shtml and see if they got the juice to power our berries? Goodnight!

They're FDA rateable, heavily isolated switchers. They'll work- and well. They're also expensive and vast overkill for a Bramble.

[srarthur]

Would something like a Guitar effects power chain / snake not do the job?
Similar to:
http://www.amazon.co.uk/Power-.....038;sr=8-4 - would support an 8 branch bramble plus hub plus HDD
or
http://www.amazon.co.uk/Stompb.....038;sr=8-3 - would support an 4 branch bramble plus 4 port hub plus HDD
or even just something like:
http://www.amazon.co.uk/PSU-FX.....38;sr=8-43 and daisy chain them together to support a larger bramble?

I'm not up on my electronics personnally

[tachyon]

Quote from Svartalf on September 15, 2011, 17:44

The power supply thing was bugging me before bed, I think the medical device industry has us covered. Could someone smarter than I look at these: http://www.trcelectronics.com/.....ulti.shtml and see if they got the juice to power our berries? Goodnight!

They're FDA rateable, heavily isolated switchers. They'll work- and well. They're also expensive and vast overkill for a Bramble.

Good! If those are overkill, do you have an example of a middlekill (as opposed to underkill ) solution? The search continues.

[tachyon]

Quote from srarthur on September 15, 2011, 19:14
Would something like a Guitar effects power chain / snake not do the job?
Similar to:
http://www.amazon.co.uk/Power-.....038;sr=8-4 - would support an 8 branch bramble plus hub plus HDD
or
http://www.amazon.co.uk/Stompb.....038;sr=8-3 - would support an 4 branch bramble plus 4 port hub plus HDD
or even just something like:
http://www.amazon.co.uk/PSU-FX.....38;sr=8-43 and daisy chain them together to support a larger bramble?

I'm not up on my electronics personnally

Nice! Is the hard part matching up the voltages to ensure we don't blow our berries and ensuring enough power for all of them under load?

Going to give something like this a try. Thanks for the reference links

[rickyjames]

I fully understand that the bramble proposal discussed here is a hobbyist / educational effort and not some race to the bleeding edge of technology, but if you're interested in clustering at all you'll be interested in this:

http://arstechnica.com/busines.....nt%29

If nothing else, this shows that educating oneself and especially kids about clustering is indeed a worthwhile effort with much upside in the future.

[rickyjames]

To get this thread back on track after my last post, let me just note that we are lucky that the Rpi is baselining Debian Linux for the primary OS / that Debian supports the ARM processor in Rpi. There is a vast array of Debian clustering software and work that has been done previously and should be able to be ported fairly readily to the Rpi. If you're going to build a Bramble, a great place to start are these links:

http://debianclusters.org/inde.....rs_Project

http://debianclusters.org/inde...../Main_Page

http://wiki.debian.org/DebianBeowulf

http://idea.uab.es/mcreel/PelicanHPC/

[Bacan]

I had thoughts of a school room of kids out on the playground, each with a portable R-Pi. Each R-Pi running some form of Mesh / "Smart Dust" network OS. Then having them experience the network reconfigure itself as the kids moved about the playground. Graphically have it show who is near and far, and what packets route through their R-Pi to reach the other machines.

[Svartalf]

Interesting experiment...but if parents saw it, it might be viewed as "creepy".

[jshriver]

Created an account to show my interest. I've been wanting some of these since they were first announced and one of the projects is to create a 4 or 8 node mini cluster. So if others are interested in the same I'd like to help or at least share my experience once these are in the wild.

[reiuyi]

Besides the choices in design, I am interested in what could be done with a cluster of raspberry pis

Due to their low power consumption and relative high speed, I can imagine using a small stack for archiving large amounts of documents and other kinds of data. I've seen those http://www.archive.org/web/petabox.php and they're still not very efficient at storing data as for kW/tb. What I am trying to say is that for those interested in archival (like me), a Raspberry pi is the perfect mix of low power and high speed. Imagine being able to store 10tb at your home using only 50W of electricity!

[obarthelemy]

"Imagine being able to store 10tb at your home using only 50W of electricity!"
That's very close to possible. 1 E-350 MB, 4x3TB HD... 30+4x10 = 70W under load. You can probably go lower underclocking the E-350 and making sure you use 1.35V RAM on a compatible MB. Maybe the newer Atoms will consume even less, but the E-350 is a good underclocker/undervolter already.

I don't quite understand why a storage solution would need a cluster... You'll be constrained by I/O, then RAM, rather than CPU ?

[Svartalf]

Quote from obarthelemy on October 10, 2011, 02:47
I don't quite understand why a storage solution would need a cluster... You'll be constrained by I/O, then RAM, rather than CPU ?


Depends on your desires. Fault tolerance would be one of the reasons you'd want that. RAID1 isn't the same thing as a clustered filesystem with atomic transactions on the system- and I'd rather have the latter rather than the former, even if there's a slight expense to the speed as a result. The other one would be that your storage could scale out easily and transparently with the right clustered system. Not so easily done with other solutions. A hint would be what Red Hat just bought recently.

[Svartalf]

Quote from reiuyi on October 9, 2011, 19:33
Due to their low power consumption and relative high speed, I can imagine using a small stack for archiving large amounts of documents and other kinds of data. I've seen those http://www.archive.org/web/petabox.php and they're still not very efficient at storing data as for kW/tb. What I am trying to say is that for those interested in archival (like me), a Raspberry pi is the perfect mix of low power and high speed. Imagine being able to store 10tb at your home using only 50W of electricity!

Biggest problem there is the latency injected by the USB leg that everything, including Ethernet, is hooked to. Now, making the 10TB be replicatable elsewhere and to be able to stack another 2-10 TB on top of the 10...that's a bit interesting. Drawbacks using R-Pi's for it for anything other than a learning tool for these techniques and technologies, would be that the USB is the only "bus" for the device that can manage what you'd need out of it. This puts a throughput bottleneck on the cluster.

[Josh]

Never knew what a cluster was (I always though it was some kind of networking thing for servers and such). But, now I do (thanks to this page: http://www.webstreet.com/super.....mputer.htm) and I'm definitely going to build my very own Bramble cluster!

Since I'm a little new to this, how much power would you get out of... let's say, 4 RPi's?

[MorgUK]

I am very interested in Bramble. We need someone to take charge of this project and get as much done before release. I have little to offer but I can create and host a website and graphics. If anyone has the knowledge to start this project please step forward.

[Josh]

We really don't need a fancy website or out. It's just a cluster.
What I'm hoping for though, is for the RPi website to have a Projects page at the Wiki. Although, it is easy enough to create Wiki page, so maybe the Bramble project can go there...

[FiftyOne]

Do a google search for SeaMicro atom server. You might get a few ideas how to go about this clustering