Parallel Jobs: How to select select optimum number of c.p.u. s. ?

Problems running VASP: crashes, internal errors, "wrong" results.


Moderators: Global Moderator, Moderator

Post Reply
Message
Author
gpkmohan

Parallel Jobs: How to select select optimum number of c.p.u. s. ?

#1 Post by gpkmohan » Mon Jul 14, 2008 2:25 pm

Dear List

My calculation in VASP involves a large super-cell and 17 atoms on it. With a single CPU it takes days, so the natural way to run VASP is in parallel mode i.e. try with different combinations of cpus ranging from 2 to 16 etc.

machine:
http://www.sara.nl/userinfo/lisa/description/index.html

But now I have some trivial doubts:

1. What part of the VASP really takes the advantageous of the parallel calculation?
Is it the evaluation of the time consuming integrals? or 3D-FFT ?

I “guess� if it is the case of integrals:
The number of CPUs should be the factor of the irreducible k-points, as we can see in the OUTCAR file.

In the case of FFT: http://cms.mpi.univie.ac.at/vasp/guide/ ... 0000000000
The more cpus I use the more efficiency will be the result [is it ?]

I don’t know that my guesses are true or false:
Can anyone suggest how we can efficiently choose an optimum number of cpus?

2. NPAR
I also hope: I can use default setting for the NPAR and NPLANE in different combinations of the cpus.

But if I use say, 4 nodes(or more say, X) can I use the following?
NPAR=4(X)
LPLANE = .TRUE.

Does it fasts the calculation?


With thanks in advance and best regards

Krishna Mohan G P

Theor Chem
Universitiet Leiden
Last edited by gpkmohan on Mon Jul 14, 2008 2:25 pm, edited 1 time in total.

Danny
Full Member
Full Member
Posts: 201
Joined: Thu Nov 02, 2006 4:35 pm
License Nr.: 5-532
Location: Ghent, Belgium
Contact:

Parallel Jobs: How to select select optimum number of c.p.u. s. ?

#2 Post by Danny » Mon Jul 14, 2008 4:56 pm

I don't know about the rest, but vasp is not parallelized over Kpoints, so more K-points slows you down, more CPU's speeds you up, but there is no need to get a relation between the number of IBZKPT and CPU's

About the Scaling, the only thing I noticed was that on the Huygens power5 machine the optimum number of CPU's was about: NBANDS/(20 to 25) Any Less and the communications starts to kill your gain.

Cheers
Danny
Last edited by Danny on Mon Jul 14, 2008 4:56 pm, edited 1 time in total.

gpkmohan

Parallel Jobs: How to select select optimum number of c.p.u. s. ?

#3 Post by gpkmohan » Mon Jul 14, 2008 5:10 pm

Hi Danny
bedankt for the mail!

this is a good thing that VASP is NOT parallelized over KPOINTS

time to play with cpus !

km
Last edited by gpkmohan on Mon Jul 14, 2008 5:10 pm, edited 1 time in total.

Post Reply