[Wien] parallel calc. without NFS

Peter Blaha pblaha at zeus.theochem.tuwien.ac.at
Mon Jul 14 10:08:53 CEST 2003


> > 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.
> >
>
> That is what I'm doing at present. The problem is that $PBS_JOBFS (jobspace
> path) is different on every node, so I can't use more than one node (4 cpus)
> for one calculation.

Yes you can. All the "sed",... is done only on the "first" node, and all
files are completely local to this node.

The other nodes in a     x lapw1 -p    jobstep  do NOT do any of those
commands, but just call a fortran program. Of course there is some I/O,
but as was mentioned, this should be buffered by the operating system and
with the "SCRATCH" -trick of Thorsten only relatively small amount of
data must be actually transfered.

PS: We had quite some NFS troubles with Suse-Linux 8.0 AND the ifc 7.0
version of the compiler. Programs like lcore or mixer ran only with 1-5%
of cpu-usage when running over an NFS mounted directory.
Using ifc 7.1 this improved significantly!! (although it seems that our
old pgi compiler version still does a better NFS-read/write, so we
actually mix: lapw0,1,2 compiled with ifc 7.1/mkl6.0; lcore,mixer,sumpara
with pgf90.

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




More information about the Wien mailing list