Well , with the compitition picking up round here i was thinking of ways to OPTIMIZE F@Holding. This weekend , my Son **in -law ** / business partner came up and we got to messing around in my shop. He asked me about the couple rigs i have in the corner , so i told him about Folding@Home. He then told me about a way to F@H and produce INSANE amounts of WU's. Mind you this is gonna take some time, like a few months, and more than likely cost a bit of cash, but its for a good cause AND it's also something , more and more , i REALLY wanna build............

So without further babblin' , i've desided my next pc build will be a BEOWULF cluster !!!!!! If your wondering what one is , basiclly a bunch of pc's all linked together, and say one instance of F@H spread out amoung all the CPU's to complete the task faster. Now for the COOL part, here is a random pic of one BEOWULF cluster.


and heres a less slick lookin one

Whoa dude! That's gonna be some HEAVY artillery for the @rmy!

You gonna make 'em camo??

ooooo that is nice. So how many wu's do you think the cluster can do in a week?

THE ARMY THANKS YOU!!!!!!

Beowulf (computing)
http://en.wikipedia.org/wiki/Beowulf_(computing)

Quote: Originally Posted by Pharaoh ooooo that is nice. So how many wu's do you think the cluster can do in a week?

Well, my plan is to use around 12 to 15 1.6GHz or higher cpu's with 512mb each , and i'm hoping it will trun out around 17 to 25 WU's a day. Thats just my math and thinking that a 1.6GHZ puts out roughly 1 WU a day, but my ultimate goal is to make the cluster put out 1 WU every hour

Yea alot of guys do that with Seti@home. Good luck setting it all up.

Sweet, imagine someone walking into the next lan with one!

Is F@H MPI aware?

As fas as I know it will not actually cluster so you really will just be able to install it on each node. Why cluster then?

Not knocking the idea and it's a great way to use a cluster in it's idle time but you wont see any benefit from it. I plan to run F@H on my cluster when I finish the planning and build it, probably close to the end of the year.

Of course if you want to build the cluster just for linux geekery I'll be happy to provide advice and info if you want it.

Quote: Originally Posted by thedemonxanth Is F@H MPI aware? As fas as I know it will not actually cluster so you really will just be able to install it on each node. Why cluster then? Not knocking the idea and it's a great way to use a cluster in it's idle time but you wont see any benefit from it. I plan to run F@H on my cluster when I finish the planning and build it, probably close to the end of the year. Of course if you want to build the cluster just for linux geekery I'll be happy to provide advice and info if you want it.

Although i myself am a Linux NOOB to the extreme, my son has been useing Linux for 10+ years and also doing IT work with all the major OS's. What he told me was it should be possible , but it gonna take some serious programming and research, which is his dept, im just the hardware Redneck . BTW Thanks for your offer and will definatly be takin you up on it

This is interesting. That is like doing Distributed Distributed computing.....

Gromacs, is an MPI threaded app but the F@H version is not. There are talks of them developing a new code that will work on SMP's and HPC's. You're on the other end of this battle from me but I'll help you out a bit.

What you are wanting to do is called parallel jobs. This means that all of the nodes work as one. The other type of job is serial. This means that each node is working on it's own. When and if the MPI code comes out, you'll want to use a low latency interconnect. Gigabit is NOT low latency. It may work for that slow of a system but gigabit won't be your best bet. Cost wise, you'll probably want to look into Dolphin or Myrinet interconnects. For larger clusters running parallel jobs, I recommend using Infiniband. Latencies with IB are down into the ~2ms rance. Bandwidth is up to 60GBps. The average is ~4ms latency and 10GBps. However the DDR versions are out now and costs are reasonable for ~3ms latency and 20GBps of bandwidth.

Many people think that 100/1000 is fine until they check their node usage and find that 50% of the nodes are idle waiting for info from another node at any given moment. Thats a lot of wasted time and cycles. A high-speed interconnect will save this time and precious cycles for a good cause.
Swing by some of the places that specialize in clusters for more info. My favorite is http://www.advancedclustering.com.

Wow , thanks for the info , and the link i got my bedtime reading for the next few days now , also thanks for "supplying weapons " as it were for the enemy dont worry i wont tell .

Oh and some BLING your way for the info, help, link and for F@Holdin' **even if it aint for team 596 **

Theres a linux os especially for folding..Its called overfoldix

http://sourceforge.net/projects/ofdx/

Quote: Originally Posted by HigHTecHReDNecK Wow , thanks for the info , and the link i got my bedtime reading for the next few days now , also thanks for "supplying weapons " as it were for the enemy dont worry i wont tell . Oh and some BLING your way for the info, help, link and for F@Holdin' **even if it aint for team 596 **

No problem. I work there and design clusters all day. Suse linux Pro 9.3 has the most compatibility with the interconnects right now. V10 is out but isnt fully supported yet.

But look out, because our demos are about to become free again and I have a feeling that 4 nodes of dualcore opterons are going to be folding for 35275.

YOU WORK THERE !!!!!!!!! and i though my shop was cool , DaYuM i wanna play room like the one pictured on your **work** web site .

You just became my NEW BESTEST BUDDY !!!! If ya wouldnt mind , keep me updated on when the the Demo is released , i'd like to take it out on a test drive , i have five P3 1.0GHz cpu's so far just about ready to go

Hows about this for a F@H Juggernaught

Attached Thumbnails

 

HOLY CRAP!

Its funny there are no Keyboards or Mouses connected

That A LOT of MoBo's !!!!!!!!!!!!!!!!!!!!!!!!!!!!

** Thats actually kinda what i'm planning BTW, im gonna put all the boards in cabinets so as to save space by not having alot of cases around.

Damn, HigHTecHReDNecK , when you do something, it's gotta be BIG, MEAN & NASTy!!!
Looking forward in seeing this "cluster" 'crush some numbers!

TDR