Problem with convergence test

Concerns any physical issues arising during the setup step (-i option). This includes problems with symmetries, k/q-point sets, and so on. For technical problems (running in parallel, etc), refer to the Technical forum.

Moderator: Daniele Varsano

Re: Problem with convergence test

Postby barnali » Mon Dec 01, 2014 6:12 am

email.rar
Respected Sir,
Thank you for kind help. Now I am facing another problem regarding convergence test. I have got the convergence for k-point, EXXRLvcs and BndsRnxp, but in case of GbndsRange, I have not got convergence up to 50 band. When I am trying to use 60 or 70 in
% GbndRnge
1 | 60 |
%
the job terminated without showing any error.
Is this problem due to insufficient memory or anything else?
I have already 50 GB free memory in my account. Here I have attached the input and report file with this mail.
Please help me to fix this problem. I am waiting for your valuable suggestion. Thanking you.

Sincerely

Barnali Bhattacharya
Ph.D student
Department of Physics,
Assam University, India
You do not have the required permissions to view the files attached to this post.
barnali
 
Posts: 25
Joined: Sat Nov 08, 2014 5:53 am

Re: Problem with convergence test

Postby Daniele Varsano » Mon Dec 01, 2014 8:27 am

Dear Barnali,
looking at your report:

Code: Select all
[RD./SAVE//s.db1]-------------------------------------------
  Bands                           :  50


I can see that you calculated 50 bands in your nscf calculations. This implies that you cannot use more than 50 bands. In order to raise this number you need to include more unoccupied bands in the nscf calculations.

Moreover, just a curiosity:
In your input file:
Code: Select all
%QPkrange                    # [GW] QP generalized Kpoint/Band indices
  1| 20|  1| 50|
%

are you interested in calculating the corrections for all these bands? Because if not, reducing this input variale to the range of interest only (gap, frontier bands etc..) wil save you a lot of computational time, avoifing to calculate quantity your are not interested in.

Best,
Daniele
Dr. Daniele Varsano
S3-CNR Institute of Nanoscience and MaX Center, Italy
MaX - Materials design at the Exascale
http://www.nano.cnr.it
http://www.max-centre.eu/
User avatar
Daniele Varsano
 
Posts: 1978
Joined: Tue Mar 17, 2009 2:23 pm

Re: Problem with convergence test

Postby barnali » Sun Dec 14, 2014 6:11 am

yambo.rar

Respected Sir,

Again I am facing another problem regarding convergence test of Response block size. When I have use 1 and 3 Ry in NGsBlkXp the job completed successfully, but when I have used NGsBlkXp= 6, 9 or 12 the job exit the time limit. I have found a folder named RESTART. I have restarted this job but that time also the job exit the time limit. Actually I am using a cluster where I can submit a job only for 24 hours and moreover I have installed only serial version of yambo there.
Could you please tell me how can I reduce the computation time but get the good result?
Here I have attached the input file, output file.
Please help me to fix this problem.
Thanking you.

Sincerely
Barnali Bhattacharya
Ph.D student,
Assam university
You do not have the required permissions to view the files attached to this post.
barnali
 
Posts: 25
Joined: Sat Nov 08, 2014 5:53 am

Re: Problem with convergence test

Postby Daniele Varsano » Sun Dec 14, 2014 9:16 am

Dear Barnali,
Increasing the block size, increase the computational load, and of course if you run with a 1 cpu only it will be very long to finish the calculation.
If you have access to a cluster , I strongly suggest you to compile and run the code in parallel. Otherwise the only way I can see is to restart the
calculation many times (Be careful to use always the same input file!).
Anyway, I can see again in your input:

Code: Select all
%QPkrange                    # [GW] QP generalized Kpoint/Band indices
  1| 20|  20|30|
%


As I told you in the previous post, do you need to calculate 200 QP corrections? I do not think so, moreover when checking for convergences?

If you change this line in:
[code]%QPkrange # [GW] QP generalized Kpoint/Band indices
1| 20| 24|25|
%[/code

so just the last occupied and first unoccupied bands, you save a factor 5 in time.
If you consider only the k point at the direct gap:
[code]%QPkrange # [GW] QP generalized Kpoint/Band indices
20| 20| 24|25|
%[/code


Check it, I'm not sure this is the right k point. In this case you save a factor 100 in your computational time.
Best,
Daniele
Dr. Daniele Varsano
S3-CNR Institute of Nanoscience and MaX Center, Italy
MaX - Materials design at the Exascale
http://www.nano.cnr.it
http://www.max-centre.eu/
User avatar
Daniele Varsano
 
Posts: 1978
Joined: Tue Mar 17, 2009 2:23 pm

Re: Problem with convergence test

Postby Khaled » Sat Apr 11, 2015 12:37 pm

Respected Dr. Daniele Varsano,

I am newbie with yambo. I have tried (many times) to make change the variable MaxGev by using the following command "yambo -i -V RL" in order to study sensitivity of change this parameter with respect QP energies and their bands and other parameters as well. after changing and run "yambo" and then testing that changing by using the command "yambo -D". I found the original values of MaxGev. Please, I am waiting for your available suggestion. Thank you in advance.
Khaled Shehata,
PhD Student,
School of Physics,
Witwatersrand University
Khaled
 
Posts: 21
Joined: Fri Apr 10, 2015 10:11 am
Location: South Africa

Re: Problem with convergence test

Postby Daniele Varsano » Sat Apr 11, 2015 4:33 pm

Dear Khaled,
First please note that the correct name of the variable is MaxGvec and not MaxGev.
This variable is used once for all in the setup (yambo -i -V RL) as you correctly reported. Next it means that the setup is done with the
max number of gvectors you provide in input Next you cannot use more Gvectors.
If you do not have problem of memory with your system there is not reason to change it, as it is not strictly a convergence parameter, it tells you the maximum number of G vectors you can use in the following runs. Best practice is to keep the maximum value (the default).
Anyway if you want to change it, you need to delete the ./SAVE/ndb.gops database and generate a new one with the MaxGvecs you provided in input. You can store all the gvec and then use a lower value in the exchange self energy and block in the dielectric matrix which are parameters that have to be brought to convergence.
Hope it help ,
Best,
Daniele
Dr. Daniele Varsano
S3-CNR Institute of Nanoscience and MaX Center, Italy
MaX - Materials design at the Exascale
http://www.nano.cnr.it
http://www.max-centre.eu/
User avatar
Daniele Varsano
 
Posts: 1978
Joined: Tue Mar 17, 2009 2:23 pm

Re: Problem with convergence test

Postby Khaled » Sun Apr 12, 2015 2:17 pm

Dear Dr. Daniele Varsano,

Really, I appreciate your fast respond. Thank you for your effort and interest. everything now clear, but the time of calculation is one of my target, I'd like to include it in my work and according to the informations I got from the one of tutorial (Basics concept of GW approximation) that the Parameter MaxGvec is related to or depend on the the time of calculation, please clarify this one because I am confused.
Do you meant by the block in dielectric matrix parameter is NGsBlK parameter?, I have read a bout this parameter in the variables section and documentation as well but I don't get the actual meaning and the range of the values i have to go through in this parameter. Thank you in advance


Best regards,
Khaled.
Khaled Shehata,
PhD Student,
School of Physics,
Witwatersrand University
Khaled
 
Posts: 21
Joined: Fri Apr 10, 2015 10:11 am
Location: South Africa

Re: Problem with convergence test

Postby Daniele Varsano » Mon Apr 13, 2015 10:07 am

Dear Khaled,
MaxGvec is not an important variable to touch if you do not have memory problem, as it does not affect at all the convergence procedure.
It just fix the number of G vector you plan to use in the future do describe your wavefunctions.
Contrary NGsBlK is a parameter that have to be brought to convergence. It is needed to describ local fields, please have a look in the documentation at:
http://www.yambo-code.org/theory/docs/doc_Xd.php
where the Random Phase approximation is described. It essentially controls the size of the reciprocal space matrix of the response function.
As all convergence studies, start with a small value and increase it until convergence is reached. Of course the bigger the value, the bigger the matrix the longer the cpu times needed to carry on the calculation.
Best,
Daniele
Dr. Daniele Varsano
S3-CNR Institute of Nanoscience and MaX Center, Italy
MaX - Materials design at the Exascale
http://www.nano.cnr.it
http://www.max-centre.eu/
User avatar
Daniele Varsano
 
Posts: 1978
Joined: Tue Mar 17, 2009 2:23 pm

Re: Problem with convergence test

Postby Khaled » Fri Apr 17, 2015 11:53 am

Respected Dr. Daniele Varsano,

Thank you a lot Sir for your respond. I have some more questions. I have changed the parameter NGsBlK by using the following values (from 1 to 12), when I setted with values from 1 to 9, I found in r and o output files the value of NGsBlK was changes automatically, and become 9 for that range. Again when also I setted it with the values from 10 to 12 , I found the value of NGsBlk = 15 in the outputs files. I don't know why ?. I have enclosed r-files with this message.

the second question is that. I have changed BndsRn and GbndRnge parameters, when I increased more than the max- value (I got it from the number of Bands, which I got it from yambo -D command). in case of BndsRn, the calculation done with putting this value at the max- value. in case of GbndRnge, the calculation was stopped and when I looked at the r-file, I didn't get any information told me anything. I am asking you because I need to know, if I find behaviour like that, what should do I do ?
if I don't know the reason. Thank you in advance.

Best,
Khaled.
You do not have the required permissions to view the files attached to this post.
Khaled Shehata,
PhD Student,
School of Physics,
Witwatersrand University
Khaled
 
Posts: 21
Joined: Fri Apr 10, 2015 10:11 am
Location: South Africa

Re: Problem with convergence test

Postby Daniele Varsano » Fri Apr 17, 2015 12:09 pm

Dear Khaled,
First question: this is because in Yambo G vectors are ordered in shells. It means Gvectors which same modulus. Simple example in one dimension, if you want to include a Gvectors G you need to include also the -G.
So the code uses a number of G vector that close a shell, the more similar to the one requested in input.
Second question, I'm not sure I well understood. In any case does not make any sense to put in input more bands than the ones you calculated in your ground state calculations, as they are not stored in the database (in your case 10(. If you want to increase that number you need to calculate more bands in the nscf calculation.
I do not know the system you are considering, anyway your parameters both in Gblk and bands seems to me very low.

Best,
Daniele
Dr. Daniele Varsano
S3-CNR Institute of Nanoscience and MaX Center, Italy
MaX - Materials design at the Exascale
http://www.nano.cnr.it
http://www.max-centre.eu/
User avatar
Daniele Varsano
 
Posts: 1978
Joined: Tue Mar 17, 2009 2:23 pm

Previous

Return to Initialization

Who is online

Users browsing this forum: No registered users and 1 guest

cron