LAMMPS WWW Site - LAMMPS Documentation - LAMMPS Mailing List Archives
Re: [lammps-users] Knight-landing best practices?
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [lammps-users] Knight-landing best practices?


From: "Moore, Stan" <stamoor@...3...>
Date: Thu, 15 Jun 2017 14:57:29 +0000

Are you weak or strong scaling? There is no ReaxFF support in the USER-INTEL package, so including that package won’t help. I’d recommend trying the USER-OMP package since a new threaded version of ReaxFF was merged into master GitHub today--you’ll need to make sure your version is updated though. You could also try the KOKKOS package, I’d recommend the MPI-only version for KNL. I’d also try using only 64 MPI per node instead of 68 MPI per node. Could also try using 128 MPI or 256 MPI per node if you aren’t using threading since each KNL core has 4 hardware threads.

 

Stan

 

From: Smith, Micholas D. [mailto:smithmd@...79...]
Sent: Thursday, June 15, 2017 8:21 AM
To: lammps-users@lists.sourceforge.net
Subject: [EXTERNAL] [lammps-users] Knight-landing best practices?

 

Dear lammps users,

 

I am trying to test some scaling of my reax system (water+zinc) on a number of KNL (intel knights landing) machines and was wondering if any one else has had any experience on this architecture. I can't seem to scale over 2 nodes. My input currently looks like this:

 

srun -n $(($nodes*68)) -N $nodes --cpu_bind=cores lmp_knl -in in.ZnOH2 -log log.lammps -pk intel 0 -sf intel

 

which is similar to the recipes provided by intel at https://software.intel.com/en-us/articles/recipe-lammps-for-intel-xeon-phi-processors

 

Any advice would be appreciated.

 

-Micholas