Calculation gets frozen during/after wannierization stage

[CHILLL]

Dear experts and users,

I have been using EPW for quite some time now, and I never had problem like this before.
I started a calculation of a material, and I just can't get past the wannierization step.
1) When I use EPW v.5.2.0 the calculation gets stuck immediately after the minimization is done. So here is the point it stops in the .wout file:

Code: Select all

     494      45.22087726      45.22008908       1.743E-05     30.44    <-- DIS
     495      45.22022771      45.21937839       1.878E-05     30.50    <-- DIS
     496      45.21953855      45.21861861       2.034E-05     30.56    <-- DIS
     497      45.21880374      45.21780166       2.216E-05     30.62    <-- DIS
     498      45.21801596      45.21691759       2.429E-05     30.68    <-- DIS
     499      45.21716624      45.21595404       2.681E-05     30.75    <-- DIS
     500      45.21624346      45.21489543       2.981E-05     30.81    <-- DIS

     <<< Warning: Maximum number of disentanglement iterations reached >>>
          <<< Disentanglement convergence criteria not satisfied >>>

        Final Omega_I    45.21489543 (Ang^2)

 +----------------------------------------------------------------------------+


 Writing checkpoint file 'prefix'.chk... done

 Time to disentangle           30.945 (sec)

After that the code is just running but nothing is happening.


2) When I use EPW v.5.1.0 the calculation does one step forward - compared to v.5.2.0 - and gets stuck after the disentanglement. The output of the EPW:

Code: Select all

     MMN
      k points =   144 in   28 pools
            1 of    6 on ionode
            2 of    6 on ionode
            3 of    6 on ionode
            4 of    6 on ionode
            5 of    6 on ionode
            6 of    6 on ionode
     MMN calculated
 
     Running Wannier90
     

Again, it is stuck, and the .ukk file is not printed.

Some additional checks I did:
1. I tried the same codes with some other structures, and everything works well.
2. I tried this particular structure in Wannier90 separately, and everything works well there too.
3. This problematic structure works in EPW if I use only one processor, so it is most likely some parallelization problem.

Do you have any suggestions?
Thank you.
Best,
Mikhail

[hlee]

Dear Mikhail:

Your issue is strange to me.

In your case, two steps, disentanglement and Wannierization should be done in library call to Wannier90 (W90).
In the case (I), the calculation is still running in W90; so basically this issue is related to W90.
In the case (II), you didn't show me the output of *.wout; could you confirm that both disentanglement and Wannierization finished successfully? In this case, the spread information should be printed out in epw.out, but I can't see it.

Last question:
Did you try the most recent version of EPW (v5.4)? First, I would like to know whether this issue is still present in EPW v5.4.

PS) I assume that you compiled W90 serially; currently, the parallel build of W90 doesn't work with EPW.

Sincerely,

H. Lee

[CHILLL]

Thank you for your response.
epw.out file ends with:

Code: Select all

  Reading data about k-point neighbours

      - All neighbours are found

     AMN
      k points =   144 in   64 pools
            1 of    3 on ionode
            2 of    3 on ionode
            3 of    3 on ionode

     AMN calculated

     MMN
      k points =   144 in   64 pools
            1 of    3 on ionode
            2 of    3 on ionode
            3 of    3 on ionode
     MMN calculated

     Running Wannier90

But from the .wout file it seems that W90 calculation

Code: Select all

         Spreads (Ang^2)       Omega I      =    45.214895430
        ================       Omega D      =     0.267904925
                               Omega OD     =     4.142041021
    Final Spread (Ang^2)       Omega Total  =    49.624841376
 ------------------------------------------------------------------------------

 Writing checkpoint file 'prefix'.chk... done

 Time for wannierise           28.624 (sec)
 *---------------------------------------------------------------------------*
 |                               PLOTTING                                    |
 *---------------------------------------------------------------------------*


 Calculating interpolated band-structure

 Time to calculate interpolated band structure       3.108 (sec)

 Time for plotting              3.157 (sec)
 Total Execution Time          62.013 (sec)

 *===========================================================================*
 |                             TIMING INFORMATION                            |
 *===========================================================================*
 |    Tag                                                Ncalls      Time (s)|
 |---------------------------------------------------------------------------|
 |kmesh: get                                        :         2         0.170|
 |kmesh: write                                      :         1         0.002|
 |overlap: allocate                                 :         1         0.000|
 |dis: main                                         :         1        30.112|
 |wann: main                                        :         1        28.622|
 |plot: main                                        :         1         3.157|
 *---------------------------------------------------------------------------*

 All done: wannier90 exiting

is successfully completed. So I don't know where the problem is.

I didn't try EPW v.5.4. I will have to install it on my cluster first. I don't really compile anything myself, so I am not sure if W90 is compiled serially. But since all the other structures worked well it can't be a compilation problem right?

Best,
Mikhail

[CHILLL]

UPDATE: the was no problem.
I made a mistake in the initial projections, so naturally the whole convergence of the disentanglement procedure was very bad. In 500 cycles it was still of the order to 10^(-1), but surprisingly it gave me very good bands (this actually confused me). But even with the bad convergence, I am still puzzled about why in that case W90 would freeze when it is running in parallel, but not when it is running on 1 processor.

I guess the message here: DOUBLE CHECK your projections/convergence of the disentanglement procedure.
Sorry for the confusion.