Folding on clustered machines. Is it possible, is it worth it?

[Jimborae]

I just read an interesting article on a superduper clustered machine, but thing that really interested me was a comment that said you could basically do the same thing with normal pc's.

That got me thinking as to whether it would be possible to cluster my 3 dedicated folding rigs to create a 6ghz folding monster. All based on Athlon XPs, each with a gig of ram in them.

So my question is does anybody know how to do this, can folding@home be run on a clustered set up & is there any advantage to be gained in doing this as opposed to just running it on individual machines??

[mmonnin]

No, FAH doesnt work on clusters. FAH is a sequential program. One frame must be done before the next one can start. So having 10 100MHz machines in a beowulf cluster wont work. If anything it will run at 100MHz and not 1GHz.

[drasnor]

Short answer:
It is possible to do (I'm planning to do it on my cluster when it's ready) but the only advantage is being able to have a cluster that can fold AND do cluster tasks. You won't see any speedup vs. running the clients on each machine individually.

Long answer:
As mmonnin stated FAH is not parallelized at all. It is worth mentioning that there are many kinds of clusters other than Beowulf (message passing / threadless) clusters. My cluster will be an OpenMOSIX cluster where the OS on each node can divert running threads to unoccupied nodes to balance the load across the cluster. It's simpler for me to implement than an MPI/Beowulf cluster and can load balance normal software (basically the cluster acts like one big SMP machine). As such, I have a couple of options: if I have eight or less processors in the cluster, I can run up to eight instances of FAH on one machine and let the OS distribute the FAH threads across the cluster; if I have more than eight it would be better if I had each node run FAH like you would normally set up your folding farm and then let the OS sort out load balancing anything else you choose to run on your cluster. I'm probably going to implement the latter solution.

-drasnor :fold:

[Jimborae]

thanks guys, oh well it was worth a thought :)