[CP2K:3146] Re: A strange problem when running CP2K

Urban Borštnik urban.b... at gmail.com
Fri Mar 18 08:26:08 UTC 2011


Hello,

On Thu, 2011-03-17 at 01:09 -0700, Crashlinger wrote:
> [...] and somebody said that it is better to use number of nodes
> multiplied by 8 : 8, 16, 24, 32, 40, 64.

Rules of thumb for number of MPI processers to get the best CP2K
parallel performance, roughly in decreasing order of importance:
     1. use a square (256),
     2. use multiple of the number of cores/node (combined with rule 1:
        8 cores/node => 256 = 16²; 12 cores/node => 576 = 24², ...), and
     3. use a power of two.

* Depending on hardware, some of these are mutually exclusive; specifics
depend on the actual calculation.

Best regards,
Urban.




More information about the CP2K-user mailing list