[Pw_forum] Large difference between the CPU and wall time in the c_bands and sum_bands routines

[Harsha Vardhan]

Dear QE users,

I am running an scf calculation using the pw.x executable for a distorted
graphene sheet, using the ultrasoft pseudopotential file
C_pbe_v1.2.uspp.F.UPF. I am using a large supercell (128 atoms).

I am using the following values for the energy cutoffs: 60.0 ecutwfc, 500.0
ecutrho, with gaussian smearing and Davidson diagonalization. I am using a
9x9x1 Monkhorst-Pack k-point grid.

I am running the calculations on a node in a cluster, which contains 16
processors. I am running it with a 16-process MPIrun.


I have observed that the c_bands and sum_bands routines are taking up a
huge amount of wall time, as compared to the CPU time. I am attaching the
time report for the completed calculation below:

     init_run     :   1327.31s CPU   1554.01s WALL (       1 calls)
     electrons    : 130443.79s CPU 771483.34s WALL (       1 calls)

     Called by init_run:
     wfcinit      :   1299.34s CPU   1524.78s WALL (       1 calls)
     potinit      :      5.56s CPU      5.65s WALL (       1 calls)

     Called by electrons:
     c_bands      :  97779.04s CPU 441468.41s WALL (      29 calls)
     sum_band     :  28359.60s CPU 311790.57s WALL (      29 calls)
     v_of_rho     :   3421.30s CPU   4610.39s WALL (      30 calls)
     newd         :     58.45s CPU   1180.47s WALL (      30 calls)
     mix_rho      :    559.21s CPU   9112.55s WALL (      29 calls)

     Called by c_bands:
     init_us_2    :    157.78s CPU   6714.76s WALL (    2950 calls)
     cegterg      :  72929.24s CPU 101098.23s WALL (    1463 calls)

     Called by sum_band:
     sum_band:bec :      4.65s CPU     39.29s WALL (    1450 calls)
     addusdens    :    546.07s CPU  26711.20s WALL (      29 calls)

     Called by *egterg:
     h_psi        :  49261.91s CPU  58153.26s WALL (    7375 calls)
     s_psi        :   5325.93s CPU   5437.34s WALL (    7375 calls)
     g_psi        :     65.05s CPU     72.50s WALL (    5862 calls)
     cdiaghg      :   3142.76s CPU   3801.60s WALL (    7325 calls)

     Called by h_psi:
     add_vuspsi   :   5378.26s CPU   5626.14s WALL (    7375 calls)

     General routines
     calbec       :  10792.25s CPU  15224.73s WALL (    8825 calls)
     fft          :   4271.25s CPU   4843.82s WALL (     920 calls)
     ffts         :    337.97s CPU   1110.36s WALL (     118 calls)
     fftw         :  42573.09s CPU  47318.16s WALL ( 2685986 calls)
     interpolate  :    598.65s CPU   3372.52s WALL (     118 calls)
     davcio       :      0.02s CPU    442.60s WALL (      50 calls)

     Parallel routines
     fft_scatter  :  37432.00s CPU  42982.58s WALL ( 2687024 calls)

     PWSCF        :     1d   13h 3m CPU        9d    3h33m WALL


   This run was terminated on:   4:55:36  26Jun2017

=------------------------------------------------------------------------------=
   JOB DONE.
=------------------------------------------------------------------------------=



As it can be seen from the report, there is a large discrepancy between the
CPU, and wall time.

When I run the same calculation, with a smaller k-point grid (6x6x1), and
smaller ecutwfc = 40.0 and ecutrho= 400.0, the CPU time is only slightly
lesser than the wall time (just 1 hour or so difference between the two).

Kindly inform me what might be causing the large delay between the CPU time
and wall time, and how I might minimize it.

Thanking You,
Yours Sincerely,
M Harshavardhan
Fourth Year Undergraduate
Engineering Physics
IIT Madras
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.quantum-espresso.org/pipermail/users/attachments/20170629/85d833ad/attachment.html>