[Wien] machines file and klist splitting

[Fred Nastos]

Peter and Stefaan, thanks for your responses.  All those options seem handy 
for various situations.  Thanks.

On October 16, 2004 03:28 am, Peter Blaha wrote:
> Use the   extrafine    directive in .machines.
> This will split the remaining 2 k-points into 2 jobs with 1 k point each.
> This should be more efficient than your suggestion to use 11,11,10,10
> k-points, because it will allow for evential load ballence.
>
> If you want to force your distribution, select specific weights in your
> .machines file like:
> 11:node1
> 10:node1
> 11:node2
> 10:node2
>
> Since you mentioned you mach have more nodes (but don't want to use them
> at the moment), you may also try the     residue:nodexx   directive.
> This would submit the extra 2 k-points to some "residue" machine (which
> could be slower or more busy, but since it should only do 2 k points it
> should be done fast enough.
>
> > I setup a .machines file, and I'm suspicious of the how the klist is
> > being split.  It seems non-optimal, so I suspect I did something wrong.
> > The .machines file I made is:
> > ## begin .machines file ##
> > granularity:1
> > 1:node001:1
> > 1:node001:1
> > 1:node002:1
> > 1:node002:1
> > ## end .machines file ##
> >
> > I have 42 kpoints.  testpara_lapw tells me that it will split the list
> > into 5 Extrafine unset
> > 1 : node001(10) 10k
> > 2 : node001(10) 10k
> > 3 : node002(10) 10k
> > 4 : node002(10) 10k
> > 5 : node001(10) 2k
> > ### end testpara output ###
> >
> > Indeed when I run the calculation, it seems to send 10 kpoints to each
> > cpu, and then 2 kpoints to node01.
> >
> > Would it not make more sense to have the list split into 4 sublists, and
> > send 11 points to two cpus and 10 points?  That is, it seems like a
> > distribution of:
> >
> > 1: node001(10) 11k
> > 2: node001(10) 11k
> > 3: node002(10) 10k
> > 4: node002(10) 10k
> >
> > would be more optimal.  No?
> >
> > Thanks