how to improve CP2K scaling for small systems

[labicia]

Hi!

We have tried both threaded and non-threaded libraries. There is no 
difference.

We have new evidences that the problem responsible for the "bad" incore 
scaling behavior is the limited RAM-cache-processor communication:
using a single-channel RAM configuration to run the ZIF8 system we get a 
slow down of 30% with respect to a dual-channel set up (for the same 
system).
For this reason, in our future purchases we will move to quad-channel RAM 
configurations.

If someone has already a quad-channel system, he can try to study the CP2K 
scaling against number of RAM channels.

Best regards,
Marco and Andrea
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20140324/e3195505/attachment.htm>