[Pw_forum] Parallel efficiency problem

[Axel Kohlmeyer]

On Tue, 29 Jul 2008, lihm wrote:

HL> Dear all,

dear haiming li,

HL>    I am a fresh user of PWscf. I run the parallel calculations to have a test.
HL> A single process test is OK:     
HL> 
HL> PWSCF        :  5m40.68s CPU time,     6m 0.14s wall time
HL> 
HL> However, I found that the walltime is too long for parallel calculations:
HL> 
HL>      PWSCF        :  1m 4.86s CPU time,    32m53.67s wall time 
HL> 
HL> I am sure that the CPU are not busy, or influenced by other jobs. 

yes, but what kind of a network do you have?
please see the many discussions on efficient use
of paralellism in QE calculations in the QE documentation
(wiki!) and the mailing list archives.

HL> The running command is:
HL> 
HL> /usr/local/mpich/bin/mpirun -np 4 -machinefile nodelist /bin/pw.x < ZnO.scf.inp > ZnO.scf.out

try adding -npools 4 after pw.x
your system is small but has many k-points. that flag should
enable parallelization over k-points which is less sensitive
to high communication latencies and slow networks.

the value of the 'disk_io' parameter can have an effect as well.

cheers,
   axel.

HL> My input file is
HL> 
HL>  &CONTROL
HL>                  calculation = 'scf' ,
HL>                 restart_mode = 'from_scratch' ,
HL>                       outdir = './' ,
HL>                       wfcdir = '/tmp/' ,
HL>                   pseudo_dir = './' ,
HL>                       prefix = 'ZnO' ,
HL>                      disk_io = 'minimal' ,
HL>                      tstress = .true. ,
HL>                      tprnfor = .true. ,
HL>  /
HL>  &SYSTEM
HL>                        ibrav = 4,
HL>                    celldm(1) = 6.213,
HL>                    celldm(3) = 1.610,
HL>                          nat = 4,
HL>                         ntyp = 2,
HL>                      ecutwfc = 40 ,
HL>                      ecutrho = 400 ,
HL>  /
HL>  &ELECTRONS
HL>                  mixing_mode = 'plain' ,
HL>                  mixing_beta = 0.7 ,
HL>              diagonalization = 'david' ,
HL>  /
HL> ATOMIC_SPECIES
HL>    Zn   65.39200  Zn.pbe-van.UPF 
HL>     O   16.00000  O.pbe-rrkjus.UPF 
HL> ATOMIC_POSITIONS crystal 
HL>    Zn      0.666666667    0.333333333    0.000000000    
HL>    Zn      0.333333333    0.666666667    0.500000000    
HL>     O      0.666666667    0.333333333    0.380000000    
HL>     O      0.333333333    0.666666667    0.880000000    
HL> K_POINTS automatic 
HL>   16 16 8   0 0 0 
HL> 
HL> Can anyone help me to fix the problem? Thank you.
HL> 
HL>     Best wishes!
HL> 
HL>                            Haiming Li
HL>                            2008-07-29
HL>  				
HL> --------------
HL> Haiming Li
HL> Beijing Synchrotron Radiation Facility
HL> Institute of High Energy Physics
HL> Chinese Academy of Sciences
HL> 19 Yu Quan Lu, 100049 Beijing, P.R. China
HL> Tel: 0086+10 8823 6437  /  0086+135 8190 2824                
HL> E-mail:lihm at ihep.ac.cn
HL> 
HL> _______________________________________________
HL> Pw_forum mailing list
HL> Pw_forum at pwscf.org
HL> http://www.democritos.it/mailman/listinfo/pw_forum
HL> 

-- 
=======================================================================
Axel Kohlmeyer   akohlmey at cmm.chem.upenn.edu   http://www.cmm.upenn.edu
   Center for Molecular Modeling   --   University of Pennsylvania
Department of Chemistry, 231 S.34th Street, Philadelphia, PA 19104-6323
tel: 1-215-898-1582,  fax: 1-215-573-6233,  office-tel: 1-215-898-5425
=======================================================================
If you make something idiot-proof, the universe creates a better idiot.