Creating a Temporary Cluster
 

I was wondering if anyone has experience in creating a system where say one has three computers that normally work separately, but on occasion they work as a cluster to solve larger problems.

The advantage for us would be that normally 3 people will have fairly powerful computers, but when required we can combine them together for a particularly large/long problem.

How difficult is this to do? I am using Ansys CFX.

Would I need to set up the computers to dual boot, once as an individual machine and on a cluster configuration? If possible, how long would it take to turnover from one configuration to the other.

Or am I just looking for trouble....

I do this at work with 5 computers and everything is seamless. All 5 users on the 5 machines can continue working normally while I and others can solve on the "cluster" of those 5 computers. No configuration switching needed, the machines can function as both 5 separate machines and a cluster simultaneously. I guess really it could be 5 clusters, as each machine can act as the head node and run the other 4 machines distributed, even at the same time.

Just follow the normal CFX distributed setup in the documentation.

I will expect followup questions on this standard CFX parallel process, it isn't simple the first time depending on you computer setup knowledge.

Thanks for the answer. Do you use Ethernet or infiniband for your clustering? I've heard with only a few computers infiniband doesn't really offer that much of an advantage, but if you have any experience I would be happy to hear it.

I use infiniband, since this cluster is also used for ANSYS mechanical FEA, which seems to need better node to node communication.

Though I ran one test using CFX and a large model using 1Gbps ethernet. Going from 1 machine to 2 machines showed almost perfectly linear speedup. (99.7%) this was using 4 cores on 1 machine vs 4 cores on each machine (8 total). So Yes, with just a few nodes, GigE shouldn't slow you down too much, if at all.

More nodes, or smaller models might not scale as well.