Dear Abraham,
first of all thank you very much for your reply, it was very usefull and I
was able to make my small test calculations converge.
But now I moved to larger systems and I get very anoying runtime errors
that make my calculations crash.
The runtime errors I get are the following:

forrtl: severe (174): SIGSEGV, segmentation fault occurred
Image              PC                Routine            Line        Source
mca_pml_ob1.so     00002B0077DBE1C8  Unknown               Unknown  Unknown
mca_pml_ob1.so     00002B0077DBE493  Unknown               Unknown  Unknown
mca_btl_sm.so      00002B0078C462AC  Unknown               Unknown  Unknown
libopen-pal.so.0   00002B007319D427  Unknown               Unknown  Unknown
libmpi.so.0        00002B0072C8B596  Unknown               Unknown  Unknown
libmpi.so.0        00002B0072CBA83E  Unknown               Unknown  Unknown
libmpi_f77.so.0    00002B0072A4744F  Unknown               Unknown  Unknown
siesta_mod2        00000000004BA2B0  redistribxy_              657  fft3d.F
siesta_mod2        00000000004ABAE3  fft_                      196  fft3d.F
siesta_mod2        00000000005D89DE  poison_                   109  poison.F
siesta_mod2        00000000004A1AD2  dhscf_                    535  dhscf.F
siesta_mod2        00000000005FEB57  m_setup_hamiltoni         214
setup_hamiltonian.F
siesta_mod2        000000000063D68A  m_siesta_forces_m          78
siesta_forces.F
siesta_mod2        000000000063D86C  MAIN__                     23  siesta.F
siesta_mod2        000000000044CEEC  Unknown               Unknown  Unknown
libc.so.6          00000030FFE1D994  Unknown               Unknown  Unknown
siesta_mod2        000000000044CDF9  Unknown               Unknown  Unknown
[compute-2-18.local:989144] [[59087,0],1]-[[59087,1],12]
mca_oob_tcp_msg_recv: readv failed: Connection reset by peer (104)
--------------------------------------------------------------------------
mpiexec has exited due to process rank 12 with PID 989149 on
node compute-2-18 exiting without calling "finalize". This may
have caused other processes in the application to be
terminated by signals sent by mpiexec (as reported here).
--------------------------------------------------------------------------

I already encoutered segmentation faults and I got through them by
increasing the number of cores, so that is what I am going to do now as
well... but this isn't something I can keep doing since now I am dealing
with only 72 atoms at a time but I will have to simulate much bigger
systems.
I hope I am not asking to much but have you ever encoutered such erros?
Have you any clue?
It looks like if the problem is in the fft routine so I will start
increasing and decreasing the MeshCutoff that now is set to 150 Ry for a
20x37x20 Angstrom supercell, what do you think?
Thank you very mch for any help you can give me.

With best regards,

Luigi.


2013/5/30 Abraham Hmiel <[email protected]>

> Hi Luigi,
>
> A few suggestions:
>
> 1) Your MaxCGsteps and MaxSCFIterations are both very large, convergence
> in many problems can be adequately approached by using a small number of CG
> updates, like 10-50 to perform a stepwise convergence, changing mixing
> parameters slightly between runs, observing how the forces change between
> CG steps and using UseSaveData = .true. to use saved DM's, XV's and so on.
> MaxSCFIterations could be around 100-250, much more than that and you're
> wasting your time, might I add.
> 2) Try using more Pulay convergence parameters. This will probably get you
> to convergence if you choose them right. You will have to modify them until
> you attain something that works. For starters, try:
> DM.NumberPulay          5
> DM.MixingWeight         0.002
> DM.NumberKick          12
> DM.KickMixingWeight    0.0766
> I've found that oftentimes the Pulay "kick" will knock your SCF cycle into
> a convergence basin, especially with charged systems or hard
> pseudopotentials. Look in the manual to see how these are defined.
> 3) I'm not sure why SimulateDoping would help you, it seems unphysical to
> me in a molecule rather than a solid. Your meshcutoff also seems low, but
> it won't help your convergence problem. Look into converging that value
> once you can close your SCF loop.
> 4) 30x30x30A cell is a big box for a small molecule, perhaps there may be
> a benefit to using a smaller cell, since a Madelung correction term is
> applied for molecules? If you go to a 20x20x20 cell does the convergence
> issue get worse?
>
> That's all I can think of for now, definitely look into the Pulay
> convergence acceleration though.
>
> Warm regards,
>
> *Abraham Hmiel*
> Katherine Belz Groves Fellow in Nanoscience
> Xue Group, College of Nanoscale Science and Engineering at SUNY Albany
> http://abehmiel.net/about
>
>

Responder a