[OpenAFS] batch/cron jobs writing to afs file systems

Harald Barth haba@kth.se
Fri, 20 Apr 2012 17:39:59 +0200 (CEST)

> I'm looking for any help and/or advice or better yet an example of a
> working batch/cron system for writing to an afs file system.

I think the credentials should be long enough for worst job wall time
length at job submission time. So the credentials should be "stuffed
away" during qsub and then when the job is started the job start part
should forward the credentials into all batch nodes. That's how our
old and modified (hackish) easy does it. But it is old and modified 
and should be replaced.

> I did come across a project called AUKS
> (http://auks.sourceforge.net/) which looks like it could do the job,
> but it's not clear to me how to actually use it.



What I don't like with auksd is actually on slide 14

>>  Single addressless credential per user
>>    Stored in Auks Memory Cache
>>    Provided to requesters without KDC interaction
>>    Forwarding to thousands of peers without KDC interaction

I think I want KDC interaction when forwarding tickets and issuing service tickets.

> For info we use a mixture of scientific linux 5 and 6 machines in
> the group and use torque/maui for controlling the batch system. The
> afs servers are all running openafs 1.6.0 on scientific linux 6.2
> boxes.

I have looked at torque. Both at the gssapi branch and the cgroups
branch. I was not happy with what I saw. So I will continue to look at
slurm instead. Unfortunately other things pop up all the time, so I
can't say how much time I can spend on this. We mostly run Centos 5
and some 6 (yet another Redhat recompile like SC, but without the cool
"scientific" in its name).