It depends on the purpose. Parallelisation over ethernet is not very useful, communication slows down the calculation. We use MPI on multicore computers so that ethernet communication is avoided. We just use computers with shared /home via NFS. On one cluster Grid Engine was used. It serves our purpose but may not be sufficient for CFD.
Thank you for the response. I want to use for running FE solver on cluster. I am thinking to build cluster (about 10 pc). Any suggestion how to set cluster for beginner would be appreciated.
Some Ethernet implementation's might be good if you can distribute a larger workload package to clusters and avoid constant Ethernet communication, i.e. clusters are supplied with formulas, you send arguments, wait for them to many times and come up with a fraction of the total result and send it back to the main controller.
This way you avoid making a live cluster chatter and choking the network.
However, this can be dangerous, because such setups have a single weak point. The switch, which might be very powerful and many Gbps fast, could eventual become the single point of failure and block your entire cluster.
In case your algorithm or formula cannot be split in a calculable packages, the shared memory approach would be best. High Availability systems for clusters can be found all over the Web, and I know of MuleSotf's implementation.
You could consider taking a look at their solution.
I have no experience with FE but it maybe you can consider using nVidia and CUDA instead of a cluster. The normal nVidia cards have limitations, you would almost certainly have to use a card manufactured for calculation with CUDA. The price of such a card will be about 600 USD but it may be comparable with the price of the cluster plus a switch. The internal bus on the nVidia card is faster than ethernet. Anyway, it would be better to find an answer from an FE expert.