The ABINIT Users Mailing List ( CLOSED )

[Mohua Bhattacharya <mohua.simoom@gmail.com>]

Hello,

I am trying to do band fft parallelization using version 5.3.5 and 8 processors.

I found an example input file t_bandfft.in (attached below) and try to run it . I am including the input file and snippets of the log file .   I was expecting the log file to contain some information showing that band parallelization is activated. I don't know what's going on. I don't see that. I really need help with this.

Thank you very much for your time

Regards

Mohua

input file

# Gold with one vacancy (107 atoms of gold). 

 nstep 15

 ecut   24

 ngfft 108 108 108

#The following parameters are completely irrealistic even for parallel testing

#They were used to check that the file can run correctly in sequential 

# (they lead to a 5-minute sequential run).

#To activate these, comment the three lines above.

#nstep 1    nline 1   ecut 2.5

iscf 7 npulayit 10

useylm 0 

#For the parallelisation

npband  4 #27 

npfft  2    #4

timopt -1

fftalg 401 wfoptalg 4 fft_opt_lob 2 

iprcch 0 intxc 0

 nband 648

 acell 3*23.01

 occopt 3 tsmear 0.002 

 enunit 2    #sorties dans toutes les unites

 nkpt 1 istwfk 1

 rprim  1. 0. 0.

          0. 1. 0. 

          0. 0. 1. 

 toldfe 1.e-7

 znucl  79.0

 natom  107 ntypat 1 typat  107*1

#total number of atoms= 107

...

Snippets of the log file

inkpts : istwfk preprocessed, gives following first values (max. 6): 1

  distrb2: enter 

  mpi_enreg%parareel=           0

  mpi_enreg%paralbd=           0

  mpi_enreg%paral_compil_respfn=           0

 distrb2: WARNING -

  nproc=   8 >= nkpt=   1* nsppol=   1

  The number of processors is larger than nkpt. This is a waste.

  distrb2: exit 

 invars1: mkmem  undefined in the input file. Use default mkmem  = nkpt

 invars1: With nkpt_me=    1 and mkmem  =     1, ground state wf handled in core.

 invars1: mkqmem undefined in the input file. Use default mkqmem = nkpt

 invars1: With nkpt_me=    1 and mkqmem =     1, ground state wf handled in core.

 invars1: mk1mem undefined in the input file. Use default mk1mem = nkpt

 invars1: With nkpt_me=    1 and mk1mem =     1, ground state wf handled in core.

 Symmetries : space group Pm -3 m (#221); Bravais cP (primitive cubic)

 inkpts: Sum of    1 k point weights is    1.000000

 inkpts : istwfk preprocessed, gives following first values (max. 6): 1

 chkneu : initialized the occupation numbers for occopt=    3

.......

 For input ecut=  2.400000E+01 best grid ngfft=     108     108     108

       max ecut=  2.717841E+01

 input values of ngfft(1) =108 ngfft(2) =108 ngfft(3) =108 are alright and will be used

 getng: value of mgfft=     108 and nfft=     1259712

 getng: values of ngfft(4),ngfft(5),ngfft(6)     109     109     108

 getmpw: optimal value of mpw=   68315

 getdim_nloc : deduce lmnmax  =   8, lnmax  =   2,

                      lmnmaxso=   8, lnmaxso=   2.

 memory : analysis of memory needs

================================================================================

 Values of the parameters that define the memory need of the present run

   intxc =         0  ionmov =         0    iscf =         7     ixc =         1

  lmnmax =         2   lnmax =         2   mband =       648  mffmem =         1

P  mgfft =       108   mkmem =         1 mpssoang=         3     mpw =     68315

  mqgrid =      3001   natom =       107    nfft =   1259712    nkpt =         1

  nloalg =         4  nspden =         1 nspinor =         1  nsppol =         1

    nsym =        48  n1xccc =      2501  ntypat =         1  occopt =         3

================================================================================

P This job should need less than                    1117.178 Mbytes of memory.

  Rough estimation (10% accuracy) of disk space for files :

  WF disk file :    675.480 Mbytes ; DEN or POT disk file :      9.613 Mbytes.

================================================================================

 Biggest array : cg(disk), with    675.4799 MBytes.

-P-0000  leave_test : synchronization done...

 memana : allocated an array of    675.480 Mbytes, for testing purposes.

 memana : allocated    1117.178 Mbytes, for testing purposes.

 The job will continue.

 -outvars: echo values of preprocessed input variables --------

     acell    2.3010000000E+01  2.3010000000E+01  2.3010000000E+01 Bohr

       amu    1.96966540E+02

      ecut    2.40000000E+01 Hartree

    enunit         2

    fftalg       401

 fft_opt_l         2

    iprcch         0

istwfk      1

P    mkmem         1

     natom       107

     npfft         2

    npband         4

  npulayit        10

     nband       648

     ngfft       108     108     108

      nkpt         1

     nstep        15

      nsym        48

    ntypat         1

 getdim_nloc : deduce lmnmax  =   8, lnmax  =   2,

                      lmnmaxso=   8, lnmaxso=   2.

  distrb2: enter 

  mpi_enreg%parareel=           0

  mpi_enreg%paralbd=           0

  mpi_enreg%paral_compil_respfn=           0

 distrb2: WARNING -

  nproc=   8 >= nkpt=   1* nsppol=   1

  The number of processors is larger than nkpt. This is a waste.

  distrb2: exit 

.......

pspatm: atomic psp has been read  and splines computed

   1.14368199E+07                                ecore*ucvol(ha*bohr**3)

-P-0000  wfconv:   648 bands initialized randomly with npw= 68315, for ikpt=     1

-P-0000  leave_test : synchronization done...

 newkpt: loop on k-points done in parallel

 pareigocc : MPI_ALLREDUCE

On Wed, Feb 17, 2010 at 10:32 AM, Mohua Bhattacharya <mohua.simoom@gmail.com> wrote:

Dear Francois Bottin,

Thank you very much for your inputs. The version that I am using in 5.3.5.
I changed the necessary variables to

paral_kgb 1
npband 2
npfft 4

but it has the same problem. I was hoping if you could tell me where to find the paral tests that you mentioned .
It would be really helpful.

Thank you very much for your time,
Regards
Mohua

On Wed, Feb 17, 2010 at 3:04 AM, BOTTIN Francois <francois.bottin@cea.fr> wrote:

It's strange, paral_kgb 1 seems to have no meaning. Your parallelization is
performed on k-points only, with 2 k-points on each processors and 7 processors at all.
What is the version?
I suggest you to launch the tests P, R, T, X & Y of the paral tests series to
check the ability of your code to perform calculations in parallel using paral_kgb 1.

I would also mention that your input variables are not set properly. See:
http://www.abinit.org/documentation/helpfiles/for-v5.8/input_variables/varpar.html#npband
and all the related variables.
For example, nproc=npband*npfft=4x4=16 in your input file and not 8 as you want (set 4x2 in your case).
In addition, npband has to be a divisor of nband (10 in your case).

Best regards,
Francois

Mohua Bhattacharya a écrit :

Hello,
 I am trying to optimize bulk Nb structure with abinip. I have 14 k points and 10 bands and 8 processors. I want to use both k pt and band FFT parallelization. (This is just a test that I am running before going into larger systems). So I set my parallelization variables as

paral_kgb 1
npband 4
npfft 4
npkpt 1

wfoptalg 4
fft_opt_lob 2
fftalg 401
iprcch 0
istwfk 14*1

The output file shows


P newkpt: treating     10 bands with npw=    1190 for ikpt=   1 by node    0
P newkpt: treating     10 bands with npw=    1191 for ikpt=   2 by node    0
P newkpt: treating     10 bands with npw=    1192 for ikpt=   3 by node    1
P newkpt: treating     10 bands with npw=    1205 for ikpt=   4 by node    1
P newkpt: treating     10 bands with npw=    1190 for ikpt=   5 by node    2
P newkpt: treating     10 bands with npw=    1191 for ikpt=   6 by node    2
P newkpt: treating     10 bands with npw=    1196 for ikpt=   7 by node    3
P newkpt: treating     10 bands with npw=    1194 for ikpt=   8 by node    3
P newkpt: treating     10 bands with npw=    1195 for ikpt=   9 by node    4
P newkpt: treating     10 bands with npw=    1220 for ikpt=  10 by node    4
P newkpt: treating     10 bands with npw=    1196 for ikpt=  11 by node    5
P newkpt: treating     10 bands with npw=    1192 for ikpt=  12 by node    5
P newkpt: treating     10 bands with npw=    1204 for ikpt=  13 by node    6
P newkpt: treating     10 bands with npw=    1212 for ikpt=  14 by node    6



and the log file has

mkrho: loop on k-points and spins done in parallel

It looks like band parallelization is not even invoked. Could someone please help me with this.
I am confused about how I should instruct abinip to use certain number of processors  for band parallelization and some others for k pt parallelization.

Thank you very much for your time,
Regards

Mohua

--
##############################################################
Francois Bottin                    tel: 01 69 26 41 73
CEA/DIF                            fax: 01 69 26 70 77
BP 12 Bruyeres-le-Chatel         email: Francois.Bottin@cea.fr
##############################################################