On Wed, 15 Jun 2005, Marco La Rosa wrote:
> Hi all,
>
> Is there any way to submit a job to a pbs cluster which DOES NOT have
> globus/LCG installed on any of the compute nodes?
Hi Marco,
what would be the point of it? A grid job should be able to use grid
functionality on the WNs, I would say...
> I know that globus-job-* with the pbs jobmanager works.
>
> However, it seems that the edg-job-* tools result in the creation of a
> wrapper script that uses globus-url-copy (this is in addition to the
> lcgpbs jobmanager which also uses it) . This script seems to be embedded
> in the job adaptor (i think) !?!
Indeed: the RB creates the job wrapper script that actually is submitted
to the CE, where the batch system adapter perl script then embeds it in
a wrapper that actually gets submitted to the batch system. The RB script
uses globus-url-copy to transfer the input/output sandbox directly from/to
the RB. The "lcg" job managers use globus-url-copy to sync the job area
on the WN with the corresponding area on the CE, in particular to copy a
renewed proxy from the CE to the WN. (In principle "scp" can be used
instead, but we would rather not depend on that, because we should move
toward making it impossible for a job on a WN to access the CE via ssh.)
The "lcg" job managers are really needed only when the CE and the WNs do
not share home directories. Any site can cook up their own job manager,
e.g. to interface to a batch system not directly supported by LCG.
> Is there a way to bypass this?
If your job does not have an input/output sandbox and does not use any
grid functionality, I think the WN does not need Globus/LCG. A typical
job submitted via an RB will at least have an input sandbox, though,
viz. to transfer the user's job script...
|