[FLASH-USERS] Re: Cluster scripts.

On Wed, 29 Jan 2003, Anshu Dubey wrote:

> There is no scaling script specific to the clusters. However, there are
> a couple of scripts that come with the distribution which you can use to
> create one. The relevant scripts are :
>
> 1. $FLASHHOME/setups/scaling_test/doscalings ---- is a script for
> scaling on the IBM machines. You should be able to use most
> of this script and the setup, except the job submission
> part. It is also fairly well commented to make adaption easy.
>
> 2. $FLASHHOME/source/sites/chiba.mcs.anl.gov/flashrun ---- is the script
> for job submission on Chiba.
>
> If you need more help with them, please let me know.

=========================================================

Hi, Anshu

We have been working for a while now at getting our scripts to run on
Chiba, but have met with mixed success:

(1)
Our equivalent of flashrun (ie, for a single run with N nodes) appears to
work fine, so that is something.

(2)
We cannot get the doscalings script to work. I have attached our version.
Specifically, we are crashing at the starting gate with the following
lines:

____________________
cat > scale.template.chiba <<EOF
#!/bin/csh
#PBS -l walltime=1800
#PBS -A mcs-flash
#ctrans in
/homes/mweber/FLASH2.2_init_1d/setups/coronal_loop/asch_100.0_39.8_as
ym/

echo "*** nodefile ***"
cat $PBS_NODEFILE
____________________

which gives this error:

PBS_NODEFILE: Undefined variable

Do you have any ideas?

(3)
We don't understand the "@@FOO@@" notation. I have asked our local syshelp
people, and nobody knows if this notation is from csh, PSUB, or what. I
don't think it is causing any of our current problems, but it makes it
harder to try to figure out what the scripts are doing.

(4) The FLASH2.2/setups/scaling_test/scale script says "# setup and run
the constant work per processor scaling test". This is the plot I am
trying to recreate for our setup. However, I can't tell from the scripts
which information in the results or flash.log file is to be interpreted as
the "constant work per processor". Is it just a plot of the "evol" time,
or something more sophisticated?

Regards,
Mark Weber