[Wien] question about mixing scf calculations with different NUME in param.inc at compilation
Peter Blaha
peter.blaha at tuwien.ac.at
Tue Dec 10 15:56:38 CET 2024
I don't know what kind of cluster you have and how many openMP cores can
be used there efficiently, but for such big systems a 64 core cluster
(eg. 4 nodes with 16 cores each) with mpi might be 10 times (or more)
faster, if you have more cores, an even larger speedup can be obtained.
The mpi-version (with ELPA !) is very efficient for such big systems.
Am 10.12.2024 um 15:28 schrieb Tomas Kana:
> Dear professor Blaha,
> Thank you very much for your answer.
> For sure I first use clean_lapw before starting
> new iterations. I am using the sequential version (really many
> calculations for many different structures).
> Fortunately, I have faced no crash of clusters on the
> governmental supercomputing center up to now. (They allow
> even running disruptive tasks).
> With best regards
> Tomas Kana
>
> For sure you should not run calculations with these 2 warning.
> You cannot expect to get the same results with RKmax = 7 or 6, the mixer gets confused by that if you do not delete old broyden files.
> Also the NUME warning is important, although from your 6720 e, it should be sufficient and can probably be neglected.
>
> On the other hand, I do not quite understand why the NMATMAX reduction
> occurs at all. NMATMAX is used primarily for sequential runs to avoid
> paging.
> I suppose you are using the mpi-parallel version ? While this needs
> twice as much memory, it divides it to all your cores, so the effective
> NMATMAX is increased by sqrt(n-cores). So with 64 cores, the effective
> NMATMAX=400000, and I doubt that your matrices are larger than that.
>
> PS: Compilation with NMATMAX=500000 is potentially VERY dangerous, as it may "crash" you computer when it runs out of memory.
>
> Am 10.12.2024 um 14:21 schrieb Tomas Kana:
>> Dear Wien2k community,
>>
>>
>> I am calculating a Mg23Al30 structure as in input for Phonopy code.
>> I have build a supercell with 636 atoms in the unit cell.
>> case.scf gives a total number of 6720 electrons.
>> For the 14 scf iterations I used proper version
>> wien2k/21.1.matmax500k (the name
>> given by our system admins) that was compiled using
>> NMATMAX= 500000
>> NUME= 50000
>> in param.inc.
>> After running out of computer time, I continued the scf calculation
>> but I did not pay attention and used the version wien2k/21.1 with
>> NMATMAX= 50000
>> NUME= 6000
>> in param.inc
>> I know from previous posts that
>> NUME > "number of electrons"/2 + 20-30%
>> When analyzing the file case.scf
>> it has two warnings:
>> :WARN : WARNING: RKmax reduced due to NMATMAX
>> :WARN : WARNING: NE limited by NUME in param.inc
>> I do not mind the first warning, it only reduced RMT times Kmax
>> from 7.0 to 5.94.
>> I tried to analyze the second warning and found that
>> after 14 "right" cycles the interstitial (grep :NPC case.scf)
>> charge decreased from 742 to 711 (I give only integers, not the
>> numbers after decimal points).
>> I now plan tostop the calculations and use the density
>> calculated using wien2k/21.1 as starting density for calculation
>> using wien2k/21.1.matmax500k.
>> Is this correct?
>> Thank you for your answer.
>>
>>
>>
>>
>> _______________________________________________
>> Wien mailing list
>> Wien at zeus.theochem.tuwien.ac.at
>> http://zeus.theochem.tuwien.ac.at/mailman/listinfo/wien
>> SEARCH the MAILING-LIST at: http://www.mail-archive.com/wien@zeus.theochem.tuwien.ac.at/index.html
>
--
-----------------------------------------------------------------------
Peter Blaha, Inst. f. Materials Chemistry, TU Vienna, A-1060 Vienna
Phone: +43-158801165300
Email: peter.blaha at tuwien.ac.at
WWW: http://www.imc.tuwien.ac.at WIEN2k: http://www.wien2k.at
-------------------------------------------------------------------------
More information about the Wien
mailing list