[Wien] parallelisation problem in optic

[Pavel Ondračka]

Dear prof. Blaha,

thank you very much, the "x lapw1 -p -d" command was just the thing I
was looking for, it now works like a charm.

Best regards
Pavel Ondračka

Peter Blaha píše v Pá 30. 01. 2015 v 11:23 +0100:
> You have first to understand how parallelization works in wien2k:
> 
> The file   .machines   (which should be generated automatically by a 
> script which you submit to the queuing system) is used by the "lapw1" 
> step to get the parallelization. At the same time, lapw1 produces a 
> .processes file, which is used in other steps (like lapw2, optic,...)
> 
> When you want to regenerate this .processes file for a different set of 
> processors, you can use:
> 
> x lapw1 -p -d
> 
> which will create .processes from the new .machines file.
> 
> In your case, note, however, that for the optic step you need the same 
> number of k-parallel jobs, so could use a .machines of
> 1:node1
> 
> 
> The command to "regenerate" a new parallelization
> 1:node1
> 1:node1
> 1:node1
> 1:node1
> 
> and you would use only ONE of your nodes.
> Am 30.01.2015 um 10:31 schrieb Pavel Ondračka:
> > Dear Wien2k mailing list,
> >
> > I'm calculating optical properties on some large systems and I'm using a
> > combination of k-point and mpi parallelisation. After converging the
> > calculations would like to run optics with different .machines file.
> >
> > Motivation behind this is for example if I run calculation with 4
> > kpoints on 4 nodes with 4 cores each node, e.g. something like this:
> >
> > 1:node1:4
> > 1:node2:4
> > 1:node3:4
> > 1:node4:4
> >
> > then for optics (which to my knowledge can't be parallelized on mpi or
> > MKL level, right?) I use only 1 core on each node thus wasting CPU-hours
> > of my allocation (I can only allocate a full node). In the optimal case
> > I would supply a new .machines file and calculate optics just on 1 node
> >
> > 1:node1
> > 1:node1
> > 1:node1
> > 1:node1
> >
> > however optic -p command doesn't regenerate the .machine1...x files.
> >
> > This is also a problem if I want to do just
> >
> > x lapw2 -p -fermi
> > x optics -p
> >
> > in already converged case. This is because I'm generating my .machines
> > files based on current node allocation, however neither "x lapw2 -p" nor
> > "x optics -p" refresh .machine1...x files and hence I need to rerun "x
> > lapw1 -p" again, just to get the right .machine1...x files, which is a
> > huge waste of CPU time.
> >
> > One solution is of course to generate my own .machine1...x on the fly
> > with some script, but this is a nuisance and error prone.
> >
> > Is it possible to force x script to just generate right .machine1...x
> > files and do nothing else?
> >
> > The optimal way would be a switch, something like
> >
> > x -genmachines
> >
> > so that I can just continue with:
> >
> > x lapw2 -p -fermi
> > x optic -p
> >
> > Any ideas?
> >
> > Best regards
> > Pavel Ondračka
> >
> > _______________________________________________
> > 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
> >
>