> -----Original Message-----
> From: Testbed Support for GridPP member institutes [mailto:TB-
> [log in to unmask]] On Behalf Of Matt Doidge
>
> Sorry I'm late to the party on this, but I thought $TMPDIR was set by the
> job shepherd in SGE, that you didn't need to set it explicitly anywhere
> else. Any given job would have a slightly different $TMPDIR like
> /scratch/tmp/`jobid`.`some_gumpf`
>
> Not that (I think) there's any harm in also setting this in the grid-
> env.sh, but I don't think it's needed. At least not for SGE (or LSF).
>
I believe that's correct. The problem is essentially that
Sussex's GE was setting it to a location on each worker nodes'
small /tmp, and not a location on their nice big fast Lustre.
The WLCG standard on this is that a job should run in ${TMPDIR}
if the variable is set, or in its initial working directory if
the variable is not set. It's fairly common (though clearly, nothing
like universal) WLCG practice to leave it unset, and just have the
job stay where it's put (which generally makes it run in a per-job
subdirectory of the pool account's home). That's easy if you're
using torque, but seems not to be a straightforward option with GE,
which appear to insist on setting TMPDIR to something.
Setting GE's base directory to a suitable location on the Lustre
should work fine, provided such a location exists.
Ewan
|