CCSD(T) of embedded cluster, parallel run

Use this forum for general discussions of topics that don't fit in the other forums
Post Reply
AlexMazheika
Posts: 3
Joined: 24 Oct 2014, 16:01
First name(s): Aliaksei
Last name(s): Mazheika
Affiliation: Fritz-Haber Institut der MPG
Country: Germany

CCSD(T) of embedded cluster, parallel run

Post by AlexMazheika » 24 Oct 2014, 16:17

Hello,

I am very interested in Dalton code, in particular in CCSD(T) implementation. I want to perform UHF single-point CCSD(T) calculations of an embedded cluster containing O, Mg, Ni and C. The number of atoms is till 20.

Could anybody tell me about several issues?:
Is massive parallel run of CCSD(T) possible, i.e. under more than 1 node with many CPU's?
Is it in general possible to perform calculations of embedded cluster in the field of point charges and effective core potentials in unrestricted CCSD(T) with your code taking into account memory limitations? I mean, because CC needs usually quite large RAM that can be solved by parallelization, if it is the case.

Best regards,
Alex

taylor
Posts: 522
Joined: 15 Oct 2013, 05:37
First name(s): Peter
Middle name(s): Robert
Last name(s): Taylor
Affiliation: Tianjin University
Country: China

Re: CCSD(T) of embedded cluster, parallel run

Post by taylor » 25 Oct 2014, 03:34

I am afraid that the answer to most of your questions is "no". It is explicitly pointed out in the Dalton manual that the code can do restricted open-shell SCF: it cannot do UHF. Further, it is explicitly pointed out that the CC code is able to handle closed-shell reference functions only. This does not preclude (e.g.) calculating triplet states via triplet response CC from a closed-shell reference, but the CC code itself deals only with closed-shell reference functions. Since there is no code to deal with non-particle-number-conserving operators in a response calculation there is no way to deal with systems like doublets or quartets with an odd number of unpaired spins.

As I understand it the CC code is heavily matrix-arithmetic oriented and it should thus be able to take advantage of threaded libraries as a level of parallelism. It is (again, as far as I understand) not parallelized in any "scalable" or "massively parallel" sense, such as using MPI. Any parallelism will therefore be limited to (most likely) the number of processor cores you have on a single node. There is no mechanism for taking advantage of multiple nodes.

If your system is closed-shell (and your posting suggests that it is not, so this may not be useful) it may be worth looking at the linear scaling code LSDalton. LSDalton \emph{is} parallelized in various ways, including not only MPI but also OpenMP/thread-level parallelism, and both approaches to parallelization can be used simultaneously. However, the CC capability in 2013.4 is somewhat limited and (since I'm not an LSDalton developer) I am not sure what the situation will be with Dalton2014. Irrespective of what CC functionality is included, however, I believe it will be limited to closed-shell systems.

Best regards
Pete

AlexMazheika
Posts: 3
Joined: 24 Oct 2014, 16:01
First name(s): Aliaksei
Last name(s): Mazheika
Affiliation: Fritz-Haber Institut der MPG
Country: Germany

Re: CCSD(T) of embedded cluster, parallel run

Post by AlexMazheika » 25 Oct 2014, 23:31

Thank you very much for such comprehensive answer!

janus
Posts: 82
Joined: 27 Aug 2013, 22:05
First name(s): Janus
Middle name(s): Juul
Last name(s): Eriksen
Affiliation: Institut für Physikalische Chemie, Johannes Gutenberg-Universität Mainz
Country: Germany

Re: CCSD(T) of embedded cluster, parallel run

Post by janus » 26 Oct 2014, 07:49

You should consider using the CFOUR program (http://www.cfour.de/) instead. In CFOUR, CCSD(T) calculations can be run in parallel using RHF and UHF references. However, this implementation is only parallel, not massively parallel.

The CCSD(T) method is not part of the public release of LSDalton yet, I'm afraid. We hope that it will be for next year's release, but this will anyway still be for RHF references only.

AlexMazheika
Posts: 3
Joined: 24 Oct 2014, 16:01
First name(s): Aliaksei
Last name(s): Mazheika
Affiliation: Fritz-Haber Institut der MPG
Country: Germany

Re: CCSD(T) of embedded cluster, parallel run

Post by AlexMazheika » 26 Oct 2014, 12:00

Thanks janus! Cfour is what I am defenetly going to try.

Post Reply

Who is online

Users browsing this forum: No registered users and 1 guest