[Columbus] Columbus Digest, Vol 128, Issue 3

Thomas Mueller th.mueller at fz-juelich.de
Wed Mar 26 15:42:09 EDT 2025


!-------------------------------------------------------------------|
  This Message Is From an External Sender
  This message came from outside your organization.
|-------------------------------------------------------------------!

On Tue, 25 Mar 2025, Aleksandr Zaichenko wrote:

   AO-MO trafo is fine.
   DRT construction is fine

   The ci part may indicate some overflows probably because some
   variables are not initialized.

   I'd suspect that you created with this CASCI calculation a
   very special case which somehow crashes the memory management.

   What might be a good idea is to simply retain say the lowest
   two VOs and allow single and double excitations - this will
   increase your csf space by  a factor of 4 - so no problem.
   This is a standard case and should work; then remove the
   doubles and check again, finally try the CASCI.

   SOCI with CASCI(20o,6e) is not easy.
   With actinides the ground state is not well-separated,
   so you will likely need all spin multiplicities
   (7,5,3,1). The basis set size is the same as for a
   determinant basis. 6e in 40 spin orbitals
   gives rise to 40*39*38*37*36*35/720  determinants
   (and in this case also CSFs).
   Also you will have to compute a (possibly large)
   number of roots. Symmetry would help a lot here.

   Thomas




> Dear Thomas,
> 
> Yes. It is weird in my opinion too. Initially I tried to compute the spin-orbit coupling in this active space and have encountered the problem, after
> that I tried just to compute this septet ground state and met the same issue. I am not sure that the MCSCF is a good alternative for my case, however
> I don't use the MR-CISD in parallel. I don't see any obvious problem in my calculation setups, moreover all beyond the active space is frozen and any
> additional excitations are impossible. I can send you the logs or inputs or boths, if you want. Can I have something wrong with the installation or
> something goes wrong in the program such as infinite loop arizing? I have really no idea how such memory demand can occur. Should I try something
> else? This problem does not look good.
> 
> Best,
> Aleksandr
> 
> вт, 25 мар. 2025 г. в 05:16, Thomas Mueller <th.mueller at fz-juelich.de>:
>       On Mon, 24 Mar 2025, Aleksandr Zaichenko via Columbus wrote:
>          Dear Aleksandr,
>
>          this looks quite weird. There is no way to produce a 1.2 TB diagonal
>          integral file ( this would amount to a basis of
>           approx. sqrt(1.2*10**12/(8))=38*10**4 ).
>
>          Since you have probably defined a reference space of 38760 CSFs,
>          the program should not try to use the standard dsyevx/dsyevd eigensolver
>          to extract the lowest root which would require (38760**2)*8 bytes of
>          memory plus say  20% overhead but instead use iterative reference space diagonalization.
>          The alternative is to just start from the unit vector with the smallest
>          diagonal matrix element - technically there is no difference.
>
>          CASCI is not really the target of the MR-CISD code - although of course
>          one can use it. The only charming feature is that the iterative CASCI
>          calculation is parallel (provided it is switched on).
>
>          BTW, for CASCI(20o,6e) without symmetry and S=0
>          the size of the CSF space should be
>          379050 CSFs (= 21!/(3!18!) * 21!/(4!17!) * 1/21
>
>          The alternative is to use the mcscf code, run a single macro iteration,
>          switch off orbital rotations (which ought also switch off the explicit
>          construction of the Hessian matrix components  B,C and M) and set to
>          iterative solution of the CI problem. This should also force the AO-MO
>          trafo step to just compute the all-internal MOs.
>
>          Beyond about 10**6 CSFs a CASCI calculation with MR-CISD is less
>          useful, since  the formula tape construction is getting more and more
>          dominant and  the matrix element construction is scalar (i.e. not fast
>          but with sufficient patience you can go to 10**9 CSFs - but imagine
>          running FCI with a general MR-CISD code is not that efficient).
>          The MCSCF code is somewhat better positioned regarding this special
>          application. Despite of all shortcomings DMRG (especially CheMPS2) does
>          a good job here.
>
>          The documentation in $COLUMBUS/docs/fulldoc/ciudg.txt,
>          $COLUMBUS/docs/fulldoc/mcscf.txt hopefully clarifies the various
>          input options.
>
>          Best,
>
>         Thomas
> 
>
>       > Dear Hans, Yes. Correct. It is the MR-CISD program. Excitations in the virtual space are 0, moreover all
>       > virtual orbitals as well as lower are frozen. The problem size from cidrtls is: total: 38760 and it seems
>       > correct. The growth of the memory
>       > Dear Hans,
>       >
>       > Yes. Correct. It is the MR-CISD program. Excitations in the virtual space are 0, moreover all virtual orbitals
>       > as well as lower are frozen.
>       >
>       > The problem size from cidrtls is:
>       >
>       > total: 38760
>       >
>       > and it seems correct. The growth of the memory demand occurs in the diagint file up to 1.2 Tb and calculation
>       > falls with SIGSEGV.
>       >
>       > It should not happen with the CI-problem of that dimension including only one root with SO, as I know. I can
>       > share listings, if you will need it. 
>       >
>       > Best,
>       > Aleksandr
>       >
>       >
>       >
>       > On Mon, Mar 24, 2025 at 6:46 PM <columbus-request at lists.osc.edu> wrote:
>       >       Send Columbus mailing list submissions to
>       >               columbus at lists.osc.edu
>       >
>       >       To subscribe or unsubscribe via the World Wide Web, visit
>       >               https://lists.osu.edu/mailman/listinfo/columbus 
>       >       or, via email, send a message with subject or body 'help' to
>       >               columbus-request at lists.osc.edu
>       >
>       >       You can reach the person managing the list at
>       >               columbus-owner at lists.osc.edu
>       >
>       >       When replying, please edit your Subject line so it is more specific
>       >       than "Re: Contents of Columbus digest..."
>       >
>       >
>       >       Today's Topics:
>       >
>       >          1. CASCI extraordinary memory demand (Aleksandr Zaichenko)
>       >          2. Re: CASCI extraordinary memory demand (Hans Lischka)
>       >
>       >
>       >       ----------------------------------------------------------------------
>       >
>       >       Message: 1
>       >       Date: Mon, 24 Mar 2025 15:36:14 -0400
>       >       From: Aleksandr Zaichenko <alex.zai90 at gmail.com>
>       >       To: columbus at lists.osc.edu
>       >       Subject: [Columbus] CASCI extraordinary memory demand
>       >       Message-ID:
>       >               <CAAJgZKtK7jg7Vij3mJuNcCrShToEQrxkgJZfQ=BKkgUz6Qvc9A at mail.gmail.com>
>       >       Content-Type: text/plain; charset="utf-8"
>       >
>       >       Dear Columbus Team,
>       >
>       >       I have a question about an ressources requirement issue with the CI module
>       >       for the CASCI calculation. A calculation with active space CAS(6,20)
>       >       without any external excitations (all lower orbitals are frozen, all higher
>       >       orbitals are removed) contains 38760 CI-vectors (as should be) and I should
>       >       get an CI-only solution for only one first root. Most QC programs do this
>       >       calculation without any problem, but the Columbus CI module requires more
>       >       than 1 Tb memory for that and it seems very demanding. Why is it so? Do I
>       >       understand correctly, that the program tries to construct a full-CI problem
>       >       for all of the 38760 roots despite 1 root being required?
>       >
>       >       Best,
>       >       Aleksandr
>       >       -------------- next part --------------
>       >       An HTML attachment was scrubbed...
>       >       URL: <http://lists.osu.edu/pipermail/columbus/attachments/20250324/b257d5cf/attachment-0001.html >
>       >
>       >       ------------------------------
>       >
>       >       Message: 2
>       >       Date: Mon, 24 Mar 2025 17:45:30 -0500
>       >       From: Hans Lischka <hans.lischka at univie.ac.at>
>       >       To: columbus at lists.osc.edu
>       >       Subject: Re: [Columbus] CASCI extraordinary memory demand
>       >       Message-ID: <bc527a53-1a8d-4bea-b08a-30e2d1f800d1 at univie.ac.at>
>       >       Content-Type: text/plain; charset="utf-8"; Format="flowed"
>       >
>       >       Hi Aleksandr,
>       >       I am not sure what kind of calculation you really did. The CI module is
>       >       meant to do a MR-CISD calculation. I assume you took the CAS(6,20) as
>       >       reference? How did you proceed further? The CI program will want to
>       >       create all SD excitation from the 38760 configurations. Did you opt for
>       >       that? This would explaining the 1TB of memory required. Look into the
>       >       cidrtls file. There you will see the CI dimension. If you want the CASCI
>       >       calculation you should input zero excitations (into the virtual space).
>       >       Please let e know what you did.
>       >       Best regards, Hans
>       >
>       >       On 3/24/2025 2:36 PM, Aleksandr Zaichenko via Columbus wrote:
>       >       > Dear Columbus Team, I have a question about an ressources requirement
>       >       > issue with the CI module for the CASCI calculation. A calculation with
>       >       > active space CAS(6,20) without any external excitations (all lower
>       >       > orbitals are frozen, all higher orbitals
>       >       > Dear Columbus Team,
>       >       >
>       >       > I have a question about an ressources requirement issue with the CI
>       >       > module for the CASCI calculation. A calculation with active space
>       >       > CAS(6,20) without any external excitations (all lower orbitals are
>       >       > frozen, all higher orbitals are removed) contains 38760 CI-vectors (as
>       >       > should be) and I should get an CI-only solution for only one first
>       >       > root. Most QC programs do this calculation without any problem, but
>       >       > the Columbus CI module requires more than 1 Tb memory for that and it
>       >       > seems very demanding. Why is it so? Do I understand correctly, that
>       >       > the program tries to construct a full-CI problem for all of the 38760
>       >       > roots despite 1 root being required?
>       >       >
>       >       > Best,
>       >       > Aleksandr
>       >       >
>       >       > _______________________________________________
>       >       > Columbus mailing list
>       >       > Columbus at lists.osc.edu
>       >       > https://lists.osu.edu/mailman/listinfo/columbus 
>       >       -------------- next part --------------
>       >       An HTML attachment was scrubbed...
>       >       URL: <http://lists.osu.edu/pipermail/columbus/attachments/20250324/06c785be/attachment.html >
>       >
>       >       ------------------------------
>       >
>       >       Subject: Digest Footer
>       >
>       >       _______________________________________________
>       >       Columbus mailing list
>       >       Columbus at lists.osc.edu
>       >       https://lists.osu.edu/mailman/listinfo/columbus 
>       >
>       >       ------------------------------
>       >
>       >       End of Columbus Digest, Vol 128, Issue 3
>       >       ****************************************
>       >
>       >
>       >
>
>       -----------------------------------------------------------
>       Dr. Thomas Mueller
>       Institute for Advanced Simulation (IAS)
>       Juelich Supercomputing Centre (JSC)
>
>       Phone:  +49-2461-61-3175
>       Fax:    +49-2461-61-6656
>       E-mail: th.mueller at fz-juelich.de
>       WWW:    https://urldefense.com/v3/__http://www.fz-juelich.de/jsc__;!!KGKeukY!xCBXMV2jPeoP7F5gPTDXsuyrzNVKsKREsD9qJziaDYzpzg3yM-MUAxCpJb9Mk2Aj7vQUfn78co3cB8S6AtDSgsKRQ-onPWo$ 
>
>       JSC is the coordinator of the
>       John von Neumann Institute for Computing (NIC)
>       and member of the
>       Gauss Centre for Supercomputing (GCS)
>       -----------------------------------------------------------
> 
> 
>

-----------------------------------------------------------
Dr. Thomas Mueller
Institute for Advanced Simulation (IAS)
Juelich Supercomputing Centre (JSC)

Phone:  +49-2461-61-3175
Fax:    +49-2461-61-6656
E-mail: th.mueller at fz-juelich.de
WWW:    https://urldefense.com/v3/__http://www.fz-juelich.de/jsc__;!!KGKeukY!xCBXMV2jPeoP7F5gPTDXsuyrzNVKsKREsD9qJziaDYzpzg3yM-MUAxCpJb9Mk2Aj7vQUfn78co3cB8S6AtDSgsKRQ-onPWo$ 

JSC is the coordinator of the
John von Neumann Institute for Computing (NIC)
and member of the
Gauss Centre for Supercomputing (GCS)
-----------------------------------------------------------


More information about the Columbus mailing list