how to improve CP2K scaling for small systems

[Florian Schiffmann]

Hi,

could you give information about the interconnects between the nodes? This 
the poor scaling sounds to me a bit like a standard gigabit ethernet. A 
system with the dimensions you discribe should scale on several hundered 
MPI ranks with CP2K if the internode connection can transfer data quick 
enough. If my assumption is correct there is little you can do as 
communication speed is crucial to cp2k.
The funny (superlinear) bit about the hybrid scaling is that the integrals 
calculation (most expensive part) dosn't require communication therefore 
for this bit scaling should be perfect. The superlinear scaling happens if 
your nodes don't have enough memory to store all integrals. By using more 
nodes you allow for more memory -> less integrals to be computed (perfectly 
parallel) on the fly -> superlinear scaling. (using ADMM 276 atoms for 
hybrids is no problem).

Flo

 
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20140314/2f910b53/attachment.htm>