Hello all.

I use starHPC on 4 node linux machines. Master node has a 120Gb hard disk and 2Gb memory, and each slave node has a 40Gb hard disk and 2Gb memory. Each node communicates by mpich through 1000Base-T ethernet. NFS is used to share user's data on master node.

I can choose distributed data and shared data, but I don't know which should I use.

Does anyone know which is better(faster),distributed or shared data?

Thanks advance.

Well, I am not an expert. But I think you'll get much better performance with distributed data on a cluster, because traffic through network will be bottleneck for performance.

Distributed data will be faster. That way you are not tieing up the ethernet with I/O while you are running the job. Then the ethernet can be dedicated to message passing and all the I/O occurs only at the start and end of a run.