Page 1 of 1

RF calculation (iscf=-3) is very slow...

Posted: Sun Dec 09, 2018 8:55 am
by Hiroki
Dear ABINIT developers/users

I tried to calculate the response functions with iscf = -3 and previous SCF result(WFK file).
The calculation terminated successfully, but took a lot of time (3-4 times longer than SCF cycle). I think this is a problem.
Looking the log file to survey the cause, I found below comments:

Code: Select all

--- !WARNING
src_file: mpi_setup.F90
src_line: 672
message: |
    Your number of spins*k-points (=4) and bands (=400) will not distribute      correctly
    with the current number of processors (=144).
    You will leave some empty.
    YOU ARE STRONGLY ADVICED TO ACTIVATE AUTOMATIC PARALLELIZATION!
    PUT "AUTOPARAL=1" IN THE INPUT FILE.
...

 -   nproc =  144   -> not optimal: autoparal keyword recommended in input file

In this run, however, I put "autoparal=1" in my input file.
I can't figure out what this message means...

And I tried to calculate without autoparal, set parameters with respect to parallelization using paral_kgb.
But I got the following comment again:

Code: Select all

--- !WARNING
src_file: mpi_setup.F90
src_line: 672
message: |
    Your number of spins*k-points (=4) and bands (=400) will not distribute      correctly
    with the current number of processors (=144).
    You will leave some empty.
    YOU ARE STRONGLY ADVICED TO ACTIVATE AUTOMATIC PARALLELIZATION!
    PUT "AUTOPARAL=1" IN THE INPUT FILE.

--- !WARNING
src_file: mpi_setup.F90
src_line: 240
message: |
    For non ground state calculation, set bandpp, npfft, npband, npspinor npkpt and nphf to 1
...

So, according to error comments, I need to use autoparal...
What should I do to optimize the parallelization for RF calculation?

The attached files are what I used in this calculation.

Sincerely yours,

Re: RF calculation (iscf=-3) is very slow...

Posted: Fri Dec 14, 2018 4:25 pm
by ebousquet
Dear Hiroki,
You should indeed adapt the number of CPU w.r.t. the number of k-points and the number of bands, etc.
Please have alook here on advises to do so here:
viewtopic.php?f=8&t=3837

For a convergence problem in this calculation, I also advise you to set nline to a larger value than the default, something like 8 to 12.

Let us know if this fix your problems,
Best wishes,
Eric