We are considering building a 8-10 node Linux cluster (maybe Beowulf) to run CFX5. We have come across many hardware issues and questions along the way.

Our current one is the question over processors in a node. Should it be 2 or one (for PC based setups)? The bottle neck is always the network so is there any advantage over using a Dual processor motherboard, baring in mind they can be twice as expensive.

Are there any people out there with experience or that have investigated this. We would be greatful if you would share youre knowledge.

Thanks in advance

Hi

My recommondation is to use only one processor per node. The network is not always the bottle neck. Actually the main bottle neck would be the access to the memory in a dual processor pc, since you only got way to transfer the data.

Also when using single processor pc would allow you to change the computer more often because the low prices on pc's.

This is oure experiance.

Regards

Jens

Agreed.

A dual processor PC will acheive a speedup of about 1.4 to 1.7, depending on processor type and operating system (Linux does better than Windows). Whereas two PC's can acheive nearly perfect speedup, 1.9 to 2.0. This is a result of the fact that PC's were never designed to run in parallel and the CPU's end up competing for memory. Since CFD is very memory intensive, this causes more problems than the occasional communication which must occurr over the network between partitions.

Don't bother spending a lot of money to build a true "Beowulf" (actually there really is no clear definition of what a Beowulf cluster is, but I digress...). A simple 100 BaseT switched network is all you need. A switchv(NOT a hub) will run you $2k to $5k, depending on the options. I certainly recommend sheltering the cluster from your network (and vice-versa) as you do not want to be competing for bandwidth.

The networking is not the bottleneck unless your partitions drop below ~20k nodes, although there may be some slowdown below ~100k nodes. Generally, meshes are large enough to ensure this.

Good luck, and let us know how things work out.

Regards, Robin

Yeah, the main issue is your intranet speed, the faster it is, it'll be better for you. However about the speedup issue, one thing is to compare the CPU time between the machine nodes. You'll notice some lovely speedup's. But what about the time spent over data transferring? It is almost costly as the solution. The CPU time doesn't account this data transferring time, or am I misunderstanding something?


Regards, cfd guy.

Hi CFD Guy,

I disagree. The level of communicaiton is quite small. With partitions of 100k nodes or greater, a faster network will not help.

I do agree that CPU time is not the most important. Rather, you should look at the wall clock time.

Robin