[ERROR]Incomplete Parallel Index Filling

[arlonne]

Dear all,

Does anyone experience an error something like "Incomplete Parallel Index Filling"?
I have tested several calculations:
1) RPA with 8x8x8,unoccupied bands 200 is ok
2) same RPA as 1), but include E-RIM with random k-points of 2000. Everything's fine!
3) RPA with 12x12x12,unoccupied bands 200; also fine!
4) same as 3), but E-RIM calculation with random k-points of 2000. The title error shown up!
"[ERROR] STOP signal received while in :[05] Optics
[ERROR]Incomplete Parallel Index Filling
"

Any idea about this error?

Thank you,
arlonne

[Daniele Varsano]

Dear Arlonne,
when posting about errors please post your inputs/outputs/reports and all that could be useful to spot the problem.
Best,
Daniele

[arlonne]

Dear Daniele,

Thank you for your attention.

I have tested two k-mesh: 8x8x8 and 12x12x12, with 2000 random k points in E-RIM method.
The input file for RPA calculation is exact the same. The only different thing is that the random k-grid is generated (ypp -r) different for k=8x8x8 and k=12x12x12.
And the strange error occurred only in k=12x12x12 with E-RIM method. Do you have any clue for this error?

Thank you,
arlonne

[Davide Sangalli]

Dear Arlonne,
unfortunaly the error shows up only on a bug run, which makes it difficult to reproduce.
From what I see, I suspect there is something wrong in your run when yambo enters the subroutine X_os.F

My guess is that the error shows up when the code execute line 185 of X_os.F:
--> call PARALLEL_index(px,(/cg_npts/))
due to a bad value of cg_npts.
Can yo try to prin it adding a line as write(*,*) cg_npts before line 185 ? What do you obtain ?
Also is the code interrupting at line 185 (i.e. if you put a write after that line do you obtain any result? )

Last thing. Can you try to do the run on the 12x12x12 *without the RIM* but setting the input variable
CGrdSpXd= 0.0000
Do you obtain the same error ?

Best regards,
Davide

[arlonne]

Dear Davide,

My guess is that the error shows up when the code execute line 185 of X_os.F:
--> call PARALLEL_index(px,(/cg_npts/))
due to a bad value of cg_npts.
Can yo try to prin it adding a line as write(*,*) cg_npts before line 185 ? What do you obtain ?
Also is the code interrupting at line 185 (i.e. if you put a write after that line do you obtain any result? )

The print cg_npts is 13824000, which is same to 12x12x12 without the RIM. And the code stop at "call PARALLEL_index".
P.s. This subroutine is at line 168 of X_os.F in yambo-3.4.1

Last thing. Can you try to do the run on the 12x12x12 *without the RIM* but setting the input variable
CGrdSpXd= 0.0000
Do you obtain the same error ?

Yes. I obtained the same error. It seems that you have know where the error come from. Do you have any suggestions?

Thank you,
arlonne

[Davide Sangalli]

Well, I've got an idea, but I'm not sure how to look for the exact problem.

What happens in X_os, which computes the independent particle (KS) response function, is that on a standard run degenerate poles are groupd to reduce the dimension of the loop. When you also use the RIM, this is not done anymore, because degenerate poles could be associated with poles at different energies in the random grid. Also Without RIM this behaviour can be controlled also with the input variable CGrdSpXd. If you set the variable to zero degenerate poles are not grouped. In this case the total number of poles in your run seems to be 13.824.000, which is a quite big number.

The lines which compute the total numer of poles are the 119-120 in X_os.F (call FREQUENCIES_coarse_grid)
As you can see in case force_bare_X_G=.true. (which is the case when you have the RIM) the call is always with 0. instead of the input CGrdSpXd (defaul is 100%).

When the subroutine PARALLEL_index (you find it inside src/mod/mod_interfaces.F) tries to split the poles among the processors (to do that it fill the vector px%element_1D, which is then cheked at line 195 of X_os) for some reason it fails...
If you see at line 95 of the module there is the call which is generating the error. Since I did not code that subroutine I'm not sure where the problem is.

You should try to understand if the error comes from a problem in FREQUENCIES_coarse_grid or from a problem in PARALLEL_index
To better exploint the problem you can try to compile yambo with some debug option. For gfortran I set
FCFLAGS="-g -pg -Wall -fbounds-check -fbacktrace -Wuninitialized -O0 "
when I run the configure before the compilation. After that do again the same run (You do not need to recompute the dipoles so it should be reasonably fast). Maybe you will get some out of bounds message ...

Also your system seems quite big, so I do not promise anything, but try to attach your pwscf input file and I'll see if I can repoduce it.

Davide

P.S.: sorry for the wrong line number in my previous email, I gave you the one of my devel version.

[arlonne]

Dear Davide,

I will try to debug. The attachment is the main input of pwscf. The full relativistic NCPP for Sn is in the tarball, other NCPPs can be found as follows:
H: http://www.quantum-espresso.org/wp-cont ... mt_fhi.UPF
C: http://www.quantum-espresso.org/wp-cont ... mt_fhi.UPF
N: http://www.quantum-espresso.org/wp-cont ... mt_fhi.UPF
I:http://www.quantum-espresso.org/wp-cont ... mt_fhi.UPF

Thanks,
arlonne