[CPMD-list] how many nodes to use

[Mark Kosmowski]

AK> ouch. if the job uses 2.7GB per MPI thread (cpmd prints usage per
AK> process and not total usage) and you have only 4GB physically
AK> (which will be ~3.6GB = ~1.8GB/CPU) you will have heavy swapping
AK> going on, which might explain the discrepancy between cpu time
AK> and wall time.

I had thought that total memory was reported, not memory per process -
thank you for clearing this up.

As a follow-up, right now all three nodes have 4 Gb RAM.  If I were to
upgrade two nodes to have 8 Gb, and ran a job across all three nodes
where the total memory needed (neglecting for the moment OS overhead
and other non-ideal considerations) was 16 Gb, the low memory system
would use swap as opposed to the memory load being asymmetrically
distributed, correct?