DFT-D4 issues

[mayank...@gmail.com]

Hi,

I am using CP2K-2024.1 (git:96bff0e) installed via toolchain as

./install_cp2k_toolchain.sh --with-gcc=install --with-openmpi=system --with-openblas=system --with-sirius=no --with-gsl=no --with-spfft=no --with-spla=no --with-spglib=no --with-spla=no --with-dftd4=install

I was testing the dft-d4 module in CP2K for DFT NVT BOMD of 64 water molecules at 300K. The input file is a modified version of H2O-64.inp, which is attached below. For reference, when using the GGA revPBE-D3(BJ) functional, the simulation runs with no issues.

598          299.000000         0.299941422       330.589518254     -1104.202651786     -1103.888925173         6.335769739
599          299.500000         0.301541403       332.352985832     -1104.203091753     -1103.888929962         6.343083243
600          300.000000         0.304829862       335.977460517     -1104.206316541     -1103.888980814         6.294001573
601          300.500000         0.309924126       341.592257168     -1104.211543588     -1103.889060265         6.321073636
602          301.000000         0.316551623       348.896953508     -1104.217560779     -1103.889142976         5.964969439
603          301.500000         0.322822897       355.809028278     -1104.223085700     -1103.889206038         6.120740931
604          302.000000         0.327538416       361.006380378     -1104.227137720     -1103.889238373         5.925923496

The computation time per step here is ~6 sec on a 32-core Intel(R) Xeon(R) Silver 4210. The temperature increase of ~60K is expected due to DFT functional change. However if I switch to dft-d4, then 

77           38.500000         0.862684099       950.833395090     -1103.274695245     -1102.386378043        91.672277213
78           39.000000         0.881930821       972.046751741     -1103.282855444     -1102.374307902        90.513970271
79           39.500000         0.882748826       972.948340710     -1103.287092185     -1102.380456423        90.920328229
80           40.000000         0.896756672       988.387512978     -1103.287828468     -1102.367459596        91.310964988
81           40.500000         0.958453301      1056.388319792     -1103.284416570     -1102.302414268        90.900099128
82           41.000000         1.038365692      1144.466180877     -1103.270000131     -1102.208050576        91.362470460
83           41.500000         1.079534923      1189.842095377     -1103.244176025     -1102.140472954        91.729197830
84           42.000000         1.097126299      1209.230962278     -1103.212461950     -1102.090361488        90.937799281

85           42.500000         1.105902625      1218.904056354     -1103.178147196     -1102.046945536        90.644135027
        86           43.000000         1.101998771      1214.601305186     -1103.141896114     -1102.013002638        90.902946438

I can notice the energy divergence and the computation time is stable at around ~91 sec, which in my experience, is more than for the equivalent revPBE0-D3(BJ) ADMM simulation. 

The relevant section in the input files are :

DFT-D3

       &VDW_POTENTIAL
         POTENTIAL_TYPE  PAIR_POTENTIAL
         &PAIR_POTENTIAL
           R_CUTOFF     1.0000000000000009E+01
           TYPE  DFTD3(BJ)
           PARAMETER_FILE_NAME ./dftd3.dat
           REFERENCE_FUNCTIONAL revPBE
           CALCULATE_C9_TERM  T
           REFERENCE_C9_TERM  T
           LONG_RANGE_CORRECTION  F
         &END PAIR_POTENTIAL
       &END VDW_POTENTIAL

DFT-D4

       &VDW_POTENTIAL
         POTENTIAL_TYPE  PAIR_POTENTIAL
         &PAIR_POTENTIAL
           R_CUTOFF     1.0000000000000009E+01
           !TYPE  DFTD3(BJ)
           TYPE  DFTD4
           !PARAMETER_FILE_NAME ./dftd3.dat
           REFERENCE_FUNCTIONAL revPBE
           !CALCULATE_C9_TERM  T
           !REFERENCE_C9_TERM  T
           LONG_RANGE_CORRECTION  F
         &END PAIR_POTENTIAL
       &END VDW_POTENTIAL

I haven't used DFT-D4 before, so I am not sure if I need to enter any addtional parameters in the dispersion block. But the 15:1 computation time ratio between dft-d4 and dft-d3bj suggest there is some issue in the simulation setup. Can you suggest any relevant changes I need to make resolve this situation?

Best Regards,

Mayank

[Jürg Hutter]

Hi

CP2K uses an external library for D4. Currently, this library is not running in parallel and the
number of parameters that can be adjusted is minimal. You can change the cutoff radius (D4_CUTOFF)
and the coordination cutoff D4_CN_CUTOFF for possible speedups.
In the CP2K internal implementation for D3, there are some additional features used to speed up
the calculation (besides parallelization), e.g. constant C6 terms for the 3-body terms (REFERENCE_C9_TERM).

regards
JH

________________________________________
From: cp...@googlegroups.com <cp...@googlegroups.com> on behalf of mayank...@gmail.com <mayank...@gmail.com>
Sent: Friday, August 9, 2024 6:17 PM
To: cp2k
Subject: [CP2K:20540] DFT-D4 issues

--
You received this message because you are subscribed to the Google Groups "cp2k" group.
To unsubscribe from this group and stop receiving emails from it, send an email to cp2k+uns...@googlegroups.com<mailto:cp2k+uns...@googlegroups.com>.
To view this discussion on the web visit https://groups.google.com/d/msgid/cp2k/49eaaddb-c3f9-45ed-ac49-c91bdf6c73d7n%40googlegroups.com<https://groups.google.com/d/msgid/cp2k/49eaaddb-c3f9-45ed-ac49-c91bdf6c73d7n%40googlegroups.com?utm_medium=email&utm_source=footer>.

[jgh]

Hi

I forgot to mention that the D4 library allows for OpenMP parallelism. You can probably get

some speedups by running with 4 or 8 OpenMP threads.

regards

JH

[Johann Pototschnig]

As Juerg said it should be faster with less MPI processes, but more OpenMP threads.

I have run your input with 96bff0e and get a non-divergent temperature, see attached files.

        77           38.500000         0.299839326       330.476990188     -1103.969230863     -1103.657189273        17.857932705
        78           39.000000         0.307346340       338.751073029     -1103.977498885     -1103.657531080        18.164155295
        79           39.500000         0.314567900       346.710533754     -1103.983238691     -1103.657830726        17.741509376
        80           40.000000         0.317580114       350.030536001     -1103.986168791     -1103.658073710        18.011582585
        81           40.500000         0.317927662       350.413597467     -1103.986493519     -1103.658253882        17.798200017
        82           41.000000         0.316336645       348.660009515     -1103.984850934     -1103.658375196        18.344547175
        83           41.500000         0.313430498       345.456911738     -1103.982150123     -1103.658446039        17.976945195
        84           42.000000         0.310384529       342.099704129     -1103.979372481     -1103.658476537        17.793025376
        85           42.500000         0.308448632       339.965996766     -1103.977407540     -1103.658479115        18.275814425
        86           43.000000         0.307327612       338.730430643     -1103.976918117     -1103.658470852        18.830695687
        87           43.500000         0.309050165       340.628994637     -1103.978149971     -1103.658471897        18.735918805
        88           44.000000         0.311776515       343.633923974     -1103.980768492     -1103.658492562        18.843529105
        89           44.500000         0.314976909       347.161335500     -1103.983874737     -1103.658524800        17.970038306
        90           45.000000         0.318084374       350.586321635     -1103.986269644     -1103.658549288        17.953601286
        91           45.500000         0.318574343       351.126356036     -1103.986913690     -1103.658546746        18.015417816
        92           46.000000         0.317332281       349.757380146     -1103.985352951     -1103.658512315        17.927721675
        93           46.500000         0.313819386       345.885535701     -1103.981876150     -1103.658460342        17.877082366
        94           47.000000         0.309418485       341.034950355     -1103.977405356     -1103.658419075        17.874089465
        95           47.500000         0.305410022       336.616901517     -1103.973131047     -1103.658418859        17.901049895
        96           48.000000         0.302510382       333.420975189     -1103.970125214     -1103.658479706        17.981440485
        97           48.500000         0.301197664       331.974123573     -1103.969074798     -1103.658605806        17.807046037
        98           49.000000         0.302529523       333.442072165     -1103.970206549     -1103.658788096        17.922524827
        99           49.500000         0.305364537       336.566768459     -1103.973353332     -1103.659013102        18.066427635
       100           50.000000         0.309864212       341.526221524     -1103.978088370     -1103.659269171        18.189648336

I think you problem is that you compile your own gcc, but use openmpi and openblas from the system, a combination that is prone for errors.

best, Johann