Author Topic: Parallel ridft???  (Read 7377 times)

idet2

  • Jr. Member
  • **
  • Posts: 21
  • Karma: +0/-0
Parallel ridft???
« on: April 29, 2007, 07:35:13 PM »
Hello to all!

I know that it is possible to run RIDFT in parallel for SinglePoint and Geometry Optimization calculations.
The thing that I 'd like to ask is if there is a way to run it in semi-direct mode. I am asking this because
I have seen that when Turbomole is running DSCF in Parallel DIRECT mode the results are not so good
(in terms of time). It requires 4CPUs to reduce the time in half. On the other hand when it is in semi-direct
mode the reduction is almost linear (although I cannot verify this for b-lyp functional). The thing is that DSCF
needs a statistical run before the parallel implentation. Is that required for RIDFT to? If not is there any other
way to use both the RI approximation and semi-direct mode? Furthermore, I 'd like to know if there is
any list with the available functionals for the parallel ridft?


Regards
« Last Edit: April 30, 2007, 12:06:34 AM by idet2 »

uwe

  • Global Moderator
  • Sr. Member
  • *****
  • Posts: 486
  • Karma: +0/-0
Re: Parallel ridft???
« Reply #1 on: April 30, 2007, 10:21:43 AM »
Hello,

in a semi-direct run, dscf saves the 'two electron four-centre' integral files to disk which are needed for the Coulomb and the exchange part of the Fock matrix. ridft uses RI-J for the Coulomb part and does not need those types of integrals. It computes 'two electron three-centre' integrals instead and stores as many of them as possible in the RAM (the amount of memory used for storage is controlled by the $ricore keyword). Writing them to disk would not speed up the calculation, because the I/O for reading and/or writing would be slower than the calculation itself.

Hence, ridft does semi-direct calculations, but it uses RAM rather than disk space.

By the way, it is always a good idea to switch on $marij by default (just add the keyword to the control file or use the last menu in define). The multipole expansion of the RI integrals is very efficient, saves a lot of time for large systems, and, if you are using the default settings, has an error that is much smaller than the RI error itself.

For the HF-exchange part things are more complicated, but for non-hybrid functionals, the HF-exchange is not needed at all.
Currently the exact HF-exchange in ridft as well as the RI-JK approximation are not parallelized at all, so you will not be able to use a hybrid functional (BH-LPY, B3-LYP, TPSSh, PBE0) for parallel ridft calculations - there you will have to use dscf.

Regards,

Uwe

idet2

  • Jr. Member
  • **
  • Posts: 21
  • Karma: +0/-0
Re: Parallel ridft???
« Reply #2 on: April 30, 2007, 02:41:20 PM »
So, is there any meaning in the "$statistics ridft" keyword?

When I have this keyword ridft does not finish properly (not even the statistics run), but instead, it gives a Segmantation fault.

Besides that I do not see the linearity when implementing parallel dscf even in semi-direct mode. For example,

1Proc. ~ 2h6min
2Proc. ~ 1h59min

Do you have any suggestions? What should I run to be sure that it is correctly parallelized?