[Pw_forum] network problem for parallel calculations

Mithra Chan chan.mithra at gmail.com
Mon Oct 18 10:41:58 CEST 2010


Dear all,

I am suffering the problem of network caused by transferring data for
parallel runnings on PC clusters. It seems the pw.x occupies a lot of
network from nodes to the server. If I run several pw.x jobs, the network of
cluster got stuck. I noted that this is due to the pw.x write data (such as
wave functions) to my HOME directory, the server, but not the temporary
directory. My job file like

#$ -N SCF
#$ -q quad0
#$ -pe orte 16
#$ -j n

INDIR=`pwd`
echo $TMP    > $INDIR/tmpdir
cd $TMP
cp $INDIR/JOB.scf.in .
mpirun -np 16 $PATH_TO_PW/pw.x -npool 4 < JOB.scf.in > scf.out
cd $INDIR
cp $TMP/* .

Then I copied the pw.x to the temporary directory, and run it locally.
However, the pw.x still write files to my home directory. In this step my
job file is :

#$ -N SCF
#$ -q quad0
#$ -pe orte 16
#$ -j n

INDIR=`pwd`
echo $TMP    > $INDIR/tmpdir
cd $TMP
cp $INDIR/JOB.scf.in .
cp $INDIR/pw.x .
mpirun -np 16 pw.x -npool 4 < JOB.scf.in > scf.out
cd $INDIR
cp $TMP/* .

How can I make the pw.x just writes data to the $TMP and after its running
copies the resulting files to my home directory, i.e. the server?
Any suggestion would be appreciated. Thanks a lot.

Mithra Chan

-- 
Dr. Mithra Chan,
Department of Physics, National University of Singapore, Singapore
chan.mithra at gmail.com
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.quantum-espresso.org/pipermail/users/attachments/20101018/35c478b3/attachment.html>


More information about the users mailing list