Hi Rob, If I were you at the beginning of the project I would run several tests increasing the number of processors. You have to pay attention not only on the number of orbitals per core but also on the number of grid points per core (obtainning this number can be tricky, in some versions of siesta is printed on the standard output). When you perform this tests you don't have to wait until the self-consistency (SCF) is achieved. With just three or fours SCF cycles is enough (don't use one as the first is the most expensive, it includes the calculation of the hamiltonian matrix elements which don't vary during the SCF). Also take into account that those numbers are system and cpu + network interconnect dependent. Also, please could you post your results to the mailing list? They can be interesting. Good luck, Eduardo
On Jun 23, 2010, at 5:07 PM, Koch, Robert J wrote: > Hey everyone, > > Quick question. Is there some lower limit of orbitals per processor where it > becomes more efficient to use _fewer_ processors? For example, I assume for > 12 orbitals its more efficient to run on 6 nodes rather than 12, but in > general how far can this be taken? > > I realize that the answer to this is largely system specific, but can anyone > give me an order of magnitude ballpark? is this threshold say, 30 or 3,000? > I soon will be doing calculations with over 10,000 orbitals and I'm not sure > how to distribute it. Thanks in advance! > > Best, > Rob Koch
