[Pw_forum] about OMP_NUM_THREADS

[Lorenzo Paulatto]

In data 25 settembre 2009 alle ore 16:06:16, Jason Larkin  
<jmlarkin at andrew.cmu.edu> ha scritto:
> First, thanks very much for the great software package.

Welcome Jason, and thank you for providing all the necessary information  
 from the beginning. There are indeed a couple of issues with your input,  
let's have a look.

>     forc_conv_thr = 1.0d-5,
>     etot_conv_thr = 1.0d-5,

force and energy are not in the same unit of measure, I don't know if this  
is a sensible choice, but it looks like the force threshold is too strict  
compared to the energy: you are requesting the total energy to change less  
that 1mRy if you move one atom by one hundredth of bohr...

>     ecutwfc = 100, ecutrho = 800,

These thresholds are very high: in QE cutoffs are specified in Rydberg  
units (some less-free DFT-base software uses eV). You are using  
norm-conserving pseudopotential, I think 50-60Ry should be enough (you  
convergence should test for a small system), for the same reason you can  
use the default value of ecutrho (=4*ecutwfc).


> My problem (issue) is this; the simulation continues to time out before
> completing the relaxation.

You can always restart a simulation from the point you've left it (as long  
the new run has access to outdir), just set restart_mode='restart' in the  
control namelist.

> As such, I've been coninuously increasing
> the amount of time and processors that I am requesting to perform this
> calculation.  However, it has reached a point (32 processors and 3:30
> hours) where I'm almost certain that I am doing something wrong.

You have about 100 electrons in your system, with such a high cutoff it  
won't be able to do many iterations, furthermore the very strict threshold  
for force convergence could cause problems (all criteria has to be  
satisfied!).


> It appears that my force(s) are not consistent with the force threshold
> that I set. I'm guessing this is because my initial guess at the
> structure is quite far from the relaxed structure.  So, would it be
> reasonable to take the unit cell coordinates from the last completed
> self-consistent calcualtion and use those as starting coordinates for a
> brand new vc-relax run?

Yes, or you can just use the restart feature, which will also recover the  
partially-completed scf cycle.

> This would be an effort to continue the
> calculation, but not lose the 3:30 of run time I've already used.

If you want to be sure that you job is not killed when it's saving the  
restart data (which requires a boring manual restart) you can use the  
max_seconds variables. Be careful because setting it to close to the time  
limit will likely cause the effect it's meant to prevent!

Finally, you cannot use the restart feature if you change the cutoff,  
you'll have to do a manual restart by copying the last atomic position.

cheers

-- 
Lorenzo Paulatto
SISSA  &  DEMOCRITOS (Trieste)
phone: +39 040 3787 511
skype: paulatz
www:   http://people.sissa.it/~paulatto/

     *** save italian brains ***
  http://saveitalianbrains.wordpress.com/