Increasing nb of cores per node degrade drastically the performance of cp2k.popt

[Axel]

>
>
> TRACE revealed that qs_forces,  qs_energies_scf,  scf_env_do_scf,   
> velocity_verlet,  qs_forces,  qs_energies_scf,   scf_env_do_scf and others 
> are 5 times slower in 2x8 than 2x2
>
>
> Do you have any suggections and ideas why it's happened?
>

this happens a lot with CPUs that have very limited memory bandwith.
the quickstep algorithm is very demanding in terms of memory bandwidth.


>
> CP2K version 2.2.262
> the lib is MKL-Scalapack
> the system is cluster of XEON nodes (8cores/node) with Infiniband switch
>

what type of xeon processors exactly? that makes all the difference.
the 56xx (westmere) and 55xx (nehalem) series ones for example have
_much_ more memory bandwidth than 54xx (harpertown) series ones.

cheers,
     axel.

 

> the compiler is Intel's
>
> Thank you in advance.
> Best regards, Maria.
>  
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20110712/17c353e8/attachment.htm>