Those new Shuttle cube sized PC's are pretty cool, and I was thinking that if you hooked several together, you can set up a cluster, which I know someone has done.

But then you could get a dual processor Hypterthreaded board.

So my question is, what would be faster: 4-5 Shuttles, or a dual hyper threaded machine?

Basically I am going to be doing mathematical/graphical computations if that helps.

I don't know much in the way of technical details on this subject, but one advantage of a cluster is easy expansion: you can increase its processing power just by adding more machines. With a dual-CPU machine, it's a little harder to upgrade. Don't know which would be faster initially, but the cluster has the potential to be faster.

If it were optimized right I would say the cluster. Each processor would not have to share memory for one thing... Of course a lot of it depends on how good you are at parralellization of your tasks. The better you are at splitting up the tasks the better the cluster will be.

Edit:
Just felt the need to throw in the hole example...
It takes a man 1 minute to dig a hole.
If you have 30 men and need to dig 30 holes how long would it take you?
If you had to dig 1 hole and had 30 men how long would it take you?

"So my question is, what would be faster: 4-5 Shuttles, or a dual hyper threaded machine?"

Part of the answer is dependent on how your application is set up. How many daemons does your application spawn? If it is just a single daemon then hyperthreading will not increase your speed because you will only use one CPU at a time anyway.

Clustering has an application program that breaks your application program into chunks and submits chunks to all the CPUs in the cluster so that all of the CPUs are kept busy simultaneously. Here is a good, short explanation of calculating pi on a cluster computer.

http://home.attmil.ne.jp/a/jm/

The other factors to consider in judging the speed difference bewteen clusters and SMP machines are the number of CPUs, the speed of the CPUs, and the overhead involved in breaking up your application so that it can be solved in parallel.

If instead of a single CPU intensive application you want to run several independent CPU intensive applications simultaneously then the SMP would be far easier to set up than the cluster.

___________________________________
Be prepared. Create a LifeBoat CD.
http://users.rcn.com/srstites/LifeBo...home.page.html

Steve Stites

here are some typical results using multi-threaded code with LAM-MPI 6.5.6:

# nodes------Gflops
1---------------3.4
2---------------5.9
4---------------9.6

this implies that performance scales with ~0.75(#nodes) for a small cluster running appropriate code. thus the proposed shuttle cluster would outperform a dual-cpu board with proper coding and equivalent processors, but you will get better performance/$ with the single-machine approach.