how to improve CP2K scaling for small systems

Matt W MattWa... at
Tue Mar 18 12:28:06 UTC 2014


unfortunately I can't see anything obvious - it just seems that the box 
gets overloaded and things slow can see that the FFTs aren't 
scaling at all for instance, but it seems to be hitting everything fairly 
equally, presumably the bottle neck is between the processors on the board.

Pragmatically, a couple of minor suggestions
- as Flo said, your cutoff is pretty high, which is making the FFTs etc 
very costly. Whilst this may well be needed for geometry optimization, a 
lower cutoff ~400 Ry or should be fine for molecular dynamics
- if you run with more nodes, the amount of data per node should drop 
significantly and the scaling should improve significantly. I'd guess you'd 
see much better scaling if you ran on 64-256 cores (which doesn't help if 
your infiniband is down, but in the longer term...).
- depending on what you are planning to do you could experiment with the 
GAPW method. It should require a much smaller cutoff.

Sorry for nothing more constructive, maybe someone else will have some 
bright ideas,


On Tuesday, March 18, 2014 10:46:57 AM UTC, labicia wrote:
> Hi,
> please find attached the file containing the scaling behavior in a single 
> machine with 4 processors:
> 1 MPI 
> 4 MPI (one MPI per different processor)
> 8 MPI (two MPI per processor)
> 16 MPI (saturated)
> please note that there is no network communications because the processors 
> are all on the same motherboard.
> Currently our multi node cluster with InfiniBand interconnections is down, 
> but we can assure that scaling across nodes is linear. 
> The non-linear scaling within a single node is what concern us! (It seems 
> a waste of resources) 
> Best regards
> Marco and Andrea
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <>

More information about the CP2K-user mailing list