[CPMD-list] how many nodes to use

[Mark Kosmowski]

Dear CPMD Community:

I use CPMD primarily to perform geometry optimizations and vibrational
analysis - I run on a modest personal cluster of 3 dual-Opteron
(single core) workstations each with 4 Gb RAM over dedicated gigabit
ethernet using OpenMPI (I have not updated any software versions since
almost exactly one year ago, Feb, 2007).  All nodes are running
OpenSUSE 10.2 64-bit.

My most recent job indicated a peak memory of 800 Mb for the geometry
analysis and the cpu and elapsed times were approximately equivalent.
However, the vibrational analysis had a peak memory of 2,700 Mb, took
30 hours of cpu time and 69 hours of elapsed time.  I ran both of
these jobs using two nodes at 2 cpu's for each node.  The NFS share is
on the other node, which was idle during the geometry optimization and
running a different job during the vibrational analysis.

Is this difference in cpu vs. elapsed time typical?

Is it fastest to have a "head node" / file server not involved in
calculations?

     Can this be approximated by only using 1 of 2 processors on the
NFS share machine for calculations?

Would it be faster to run jobs taking less than ~ 3 Gb on one node?

     Should both processors on each node be used?

Thank you,

Mark Kosmowski