Super Computing @ Home, the T-Bird Cluster

[DarkDruid]

Super Computing at home - The T-Bird Cluster
I should be updating this as I get hardware in and stuff setup.
If you want to stay on top of my progress real time, check Here[

Overr at AtlantaOC I posted a news article a while back about building a system for under $200 and they are overclockable.

Well I have been doing some research of those motherboards, and watching the price of cpu's drop down, and desided to take my own advise and buy some.

Now with all the systems I have sitting around here, I got to thinking, What am I going to do with 20 new computers?Cluster Them!! Then throw in a few TMS320C4x DSP chips and let that do some math!!

Do some math I say? and what is this business about TMS320C4x DSP chips? Well They are pushing out 1GFLOP of pure floating point. Now my T-Bird does around 2.4GFLOPS you say. But how much does that T-Bird cost? and how many can you put in a box? The TMS320C4x cost $15 each (Check the TI dist links for resalers) and you can put an unlimited amount of them on a pc and let the cpu use the DSPs as co-processors.

What does this do for us? Well lets put say 8 of the TMS320 in each node (20 nodes) and creat a CLUSTER. We would have right around 200 GFLOPS.

For a comparision it takes my TB1.33 @ 1.66 just at 220 seconds to render piece2 in povraytrace, at 2.5GFLOPs. In Theory It should take 2.75 seconds on the 20 node cluster, but in reality, I would say about 5 seconds.

Overview;

20 Nodes
4 Managers
1 Master

Each node will consist of the following;

ECS K7SEM Mobo (Built in Nic, and vid)
AMD T-Bird 1ghz 266 OCed to 1.2ghz
8 TMS320 DSP chips
256meg KingMax PC150 TinyBga

Each Manager will consist of the following;

ECS K7SEM Mobo (Built in Nic, and vid)
AMD T-Bird 1ghz 266 OCed to 1.2ghz
512meg KingMax PC150 TinyBga
46gig Maxtor Ata100 7,200rpm
2 Extra Intel 100 Pro nic

And the Master will be a

Tyan Althon Dual
2 AMD 1.4gig oced as much as I can get
1gig DDR
4 Seagate Cheatah 40gig Raid-0+1
6 Intel Pro 100 Nics

Now this is the "Long term Goal". For right now I will just be getting the nodes here and configured and use some boxes I have sitting around for the managers and master, untill I get all 20 of the nodes in. Then I will build out the managers, and save up for the master (those seagates are going to KILL me)

-------------------------

The setup will be 5 nodes in each custom built box, with just the mobo/cpu/mem, no drives, they will be diskless workstations booting off the managers that are running nfs.

All of the managers will sit in one custom case

And the master will be in a full size server case, (Water cooled of course)

Each "mini cluster" which consists of 5 nodes and 1 manager will be on a seperate network with a 100baseT switch connecting them.

Please see this diagram
Here

Each manager will be connected to its nodes, via the built in RealTec 100BaseT nic, using Gamma topology.

Then each manager will have the 2 Intel Nic Bonded together for a 200Mb pipe to the master

The Master will have 3 sets of Intel Nics Bonded with 2 connected to the Managers network and 1 connected to the outside world

I will write up more on it later, but for the time being here are some links for ya.

DSP Clusters
Beowulf Underground
Yahoo search for clusters

Peace!

 

[junthin]

Wow! This is an interesting read. 😉 Can anyone else confirm if this is acutally possible? What would you do with a setup like this? Can it perform 3D modelling well? I've been looking into Linux for a Distributed 3D Modelling program that will share the tasks with several computers thus increasing the output. Would this work for it?

 

[DarkDruid]

There are some, one mainly is POCRaytrace.

Luckly POVraytrace has a linux port and open source, so it will be semi-painless to cluster it.

Peace

 

[junthin]

Has anyone tried/verified this other than DarkDruid?

 

[Soccerman]

heh, I didn't know you could do that!

wouldn't the PCI bus be a limitation though? 133megs a second isn't exactly alot when you have massive amounts of networking and communication between cards (I'm assuming those DSPs come on a PCI card).

also, don't forget FULL DUPLEX network operation. if you have 1 Fast Ethernet connection, you get an effective 200Mbps, 100 up and 100 down. this is assuming the end link (a switch or hub) is capable of full duplex as well.
.
btw, if you actually do this, don't forget to run Seti@Home on it! hehehehe.. or at least, try to figure out a way to run Seti@Home on those DSP chips..