Author Topic: aoforce parallelized?  (Read 6965 times)

Matteo Guglielmi

  • Jr. Member
  • **
  • Posts: 20
  • Karma: +0/-0
aoforce parallelized?
« on: March 12, 2009, 04:28:10 PM »
I'm currently running my first vibrational frequencies
calculation and noticed that aoforce does not run in
parallel... does it?

I've checked the manual but the aoforce "limitations"
do not mention anything about the parallel vs serial
implementation of this module.

My system is pretty big... 176 atoms at def2-SVP
basis set level for which I've already obtained a well
converged geometry.

Should I switch to a numforce calculation instead?

Thanks,
MG.

Matteo Guglielmi

  • Jr. Member
  • **
  • Posts: 20
  • Karma: +0/-0
Re: aoforce parallelized?
« Reply #1 on: March 25, 2009, 04:24:32 PM »
...
My system is pretty big... 176 atoms at def2-SVP
basis set level for which I've already obtained a well
converged geometry.

Should I switch to a numforce calculation instead?

Thanks,
MG.

Ok, let's put it this way:

does anyone of you have a rough idea on how long
such a serial calculation could take?

Currently I'm running this job on a single core of an
Intel Xeon Woodcrest and I have assigned 15GB of
RAM to it ($maxcor=15000).

The output file says:

Quote
PREPARING NUMERICAL INTEGRATION
 
 Remaining core memory for DFT =       11269 MB
 
 Memory needed per atom    =       33887 KiB
 
     i. e.         176 atoms per loop

... so I suppose that at least it will compute all the vibrational
frequencies in one single loop instead of looping over more
cycles with a smaller number of atoms per cycle.

The output file hasn't been update since a week or so already
and the single core is always 100% burning.

So, from your experience, do I have any chance to get this job
done in a reasonably amount of time?

Do vib. frequencies depend strongly from being calculated via
analytical/numerical second derivatives?

Thanks.
« Last Edit: March 25, 2009, 04:27:25 PM by matteo »

christof.haettig

  • Global Moderator
  • Sr. Member
  • *****
  • Posts: 262
  • Karma: +0/-0
    • Hattig's Group at the RUB
Re: aoforce parallelized?
« Reply #2 on: March 27, 2009, 02:42:02 PM »
The difference be numerical and analytic vib. frequencies depend on the step size (delta) for the numerical differentiation
and on the convergence setting. But at least the numerical calculation of vibrational frequencies can be trivially parallelized
(check NumForce -h for some hints). For DFT the weight derivatives option should be specified in the control file (see documentation).

Christof

Matteo Guglielmi

  • Jr. Member
  • **
  • Posts: 20
  • Karma: +0/-0
Re: aoforce parallelized?
« Reply #3 on: April 07, 2009, 12:54:39 AM »
The difference be numerical and analytic vib. frequencies depend on the step size (delta) for the numerical differentiation
and on the convergence setting. But at least the numerical calculation of vibrational frequencies can be trivially parallelized
(check NumForce -h for some hints). For DFT the weight derivatives option should be specified in the control file (see documentation).

Christof

Yes, I've strictly followed what's written in the manual indeed...
and dropped the aoforce calculation since it was taking too
much time already.

So now comes a correlated question, what grid size should I
use then?

I've used 'm4' so far... (I've also found this info reading the
manual somewhere) would that be enough for such a big
system? ($scfconv = 8, delta = 0.02 au)