Scalability and parallel performance with hybrid functionals

Message

hipertrofia · #1 Post by **hipertrofia** » Mon Mar 26, 2012 3:51 am

Hi all,

I have been testing the performance of VASP on the cluster in my institute and have noticed a very poor speed up with the number of nodes for the particular problem I am tackling.

Basically from 1 to 4 cores I see a nice improvement, but beyond that the computing time might even increase with the number of nodes, I guess because of communication issues.

I would like to know what is the limiting factor in my case. I have a two atom unit cell and am doing a HSE06 type calculation with a varying number of kpoints, ranging from 4x4x4 to 10x10x10.

Is the speed-up bad because of the very low number of atoms? Is it an intrinsic issue related to HF-type calculations?

Thanks

<span class='smallblacktext'>[ Edited ]</span>

#2 Post by **admin** » Mon Mar 26, 2012 1:05 pm

Have you tried massive parallelization?
http://cms.mpi.univie.ac.at/vasp/vasp/P ... .html#4641

hipertrofia · #3 Post by **hipertrofia** » Mon Mar 26, 2012 9:19 pm

VASP won't allow me to set NPAR to anything different than the number of cores for HF calculations, so this is not an option. Changing LPLANE only makes things worse...