[Wien] parallel calc. without NFS

[Peter Blaha]

> > What do you use as $SCRATCH ?  If that is a directory that exists locally on
> > every node (e.g. /tmp ), then the large scratch files do not need to go over
> > the network, which reduces wien-traffic significantly.
> >
>
> Local jobspace is allocated dynamically by the PBS and each node has its
> own jobspace and path (accesible only to processes that are part of the
> job and are running on that node), so there is no good choice for SCRATCH.
> Besides, reading/writing large chunks of data are not a problem. It's
> "trickly" I/O that many of the scripts are doing (eg. sed, awk) that are
> causing the problems. Things like writing a large array one element at a
> time, or reading a file line by line.

PBS must give you some local jobspace and you must be able to figure this out
(probably in some variable like $PBS_JOBSPACE  or so).

Use this jobspace and copy the main files into a new "local" directory
at the beginning of your job-script. Then all the "sed" is done locally.

cd ~/case
cp  case.struct *.in* case.klist case.kgen case.clmsum .machines $localdir/case
mkdir $localdir/case
cd $localdir/case
run_lapw -p
cp case.dayfile case.clmsum case.scf     ~/case

Eventually, you may want to copy back more files (like output*, vectors*,...),
but often this is not necessary. For spin-polarization add clmup/dn.

                                      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/
--------------------------------------------------------------------------