[Wien] memory configuration for k-point parallelization

[Peter Blaha]

Yes, XAS calculations for 100 atoms/cell should be possible on your computer.
But note, the required resources may vary a lot with the specific system (atoms,
nearest neighbor distances, structure (inversion symmetry).

With just 2 nodes and 4 cores in total, mpi-parallelization is most-likely
not the way to go (except you need it because of memory limitations).

For XAS you need a good DOS at the end, so you will need finally some k-points,
thus k-parallelization should be fine, for a large system probably only
1 job per node (but with OMP_NUM_PTHREAD = 2).

No, you don't need an explicit master node, except for NFS, i.e. your data
should be on one node, but via NFS also available on the other one.

Tomohiko Tasaka schrieb:
> Hello,
> 
> Thanks for your kind reply. if possible, could you advise me of the minimum
> system to analyze core-orbital spectroscopy (e.g. x-ray absorption spectrum)
> for a case with 100 atoms/cell?
> for example, how about 2 machines which has dual cores with 4 GB RAM (total
> 4 cores, 8 GB RAM)? is it sufficient on condition of mix parallelization?
> 
> Further, I'd like to confirm whether these parallel jobs require a single
> master (control) node to install much more requirement resources(e.g. RAM
> or any I/O bandwidth) than client nodes.
> 
> 
> Sorry, vague question.
> Tomo
> 
> -----Original Message-----
> From: wien-bounces at zeus.theochem.tuwien.ac.at
> [mailto:wien-bounces at zeus.theochem.tuwien.ac.at] On Behalf Of Peter Blaha
> Sent: Monday, September 01, 2008 10:24 PM
> To: A Mailing list for WIEN2k users
> Subject: Re: [Wien] memory configuration for k-point parallelization
> 
> For a case with 100 atoms/cell most likely you will have only 1 k-point,
> thus k-point parallel is not possible, only mpi-parallel.
> 
> In case you have a metallic (magnetic) case, you may still have a few
> k-points and than k-parallel is still most efficient, provided the
> memory (RAM) is large enough ON EACH NODE (has nothing to do with the
> master node).
> 
> If you have enough nodes available, one can also mix k-point and
> mpi-parallel
> mode, eg. when you have 4-kpoints, but you have 16 nodes, run 4-mpi-jobs
> per k-point.
> 
> mpi-version only with the most recent versions.
> 
> Tomohiko Tasaka schrieb:
>> Dear wien2k users,
>>
>>  
>>
>> I heard that the k-point parallel calculation with wien2k old version
>>
>> needs huge physical memory for a master node. The reason is that
>>
>> the master node (job submitting node) collects many distributed results
>>
>> which are calculated by each calculation node, at the same time.
>>
>> Therefore, there is difficulty in executing k-point parallel calculation
>>
>> for the large system (e.g. 100 atoms per unit cell).
>>
>>  
>>
>> is this correct regardless of the version? please advise.
>>
>>  
>>
>> Any information would be appreciated.
>>
>>  
>>
>> With best regards,
>>
>> Tomo
>>
>>  
>>
>>
>> ------------------------------------------------------------------------
>>
>> _______________________________________________
>> Wien mailing list
>> Wien at zeus.theochem.tuwien.ac.at
>> http://zeus.theochem.tuwien.ac.at/mailman/listinfo/wien
> 

-- 

                                       P.Blaha
--------------------------------------------------------------------------
Peter BLAHA, Inst.f. Materials Chemistry, TU Vienna, A-1060 Vienna
Phone: +43-1-58801-15671             FAX: +43-1-58801-15698
Email: blaha at theochem.tuwien.ac.at    WWW: http://info.tuwien.ac.at/theochem/
--------------------------------------------------------------------------