Re: [Condor-users] Standard Universe and Job Hooks (condor_starter vs condor_starter.std)

[Todd Tannenbaum <tannenba@xxxxxxxxxxx>]

Joan J. Piles wrote:
> Hi all,
>
> We have a hook that must be called for each job running in our cluster, 
> an instance of  xxxxx_HOOK_JOB_EXIT. In the Vanilla universe (the one 
> most of our jobs use), there is no problem, and it works almost as 
> expected (I say almost because the exit reason is shown as "evict" even 
> when "condor_rm" is used, but that's not an important problem for us).
>
> We have recently found that this hook is completely ignored for Standard 
> universe jobs. According to the documentation it should work, and it is 
> condor_starter's job to run the hooks. However, there seem to be two 
> condor_starter executables, one for most jobs, and another one 
> (condor_starter.std) for Standard universe jobs. Furthermore, in the 
> sourece code there are two completely different implementations, and the 
> Standard universe one seems to have no hook capability at all, so I 
> don't know if this is a bug or a feature ;-)
>
> What are our options for implementing hooks for Standard Universe jobs? 
> Is this being worked upon (in development versions), or we should find a 
> workaround? We already tried ditching condor_starter.std, but the 
> default condor_starter doesn't seem to be able to start Standard 
> Universe jobs.
>
> Thanks in advance,
>
> Joan

Hi Joan -

You are correct, standard universe has its own shadow/starter pair that 
does not support a bunch of mechanisms found in the newer shadow/starter 
pair that supports other universes like Vanilla, Java, etc.  Besides 
hooks, other features like ssh_to_job and CCB do not work in standard 
universe for this reason.

We are currently actively looking at moving some functionality from the 
standard universe starter/shadow into the newer starter/shadow.  ( For 
some details, see some thinking we did on this a couple weeks ago at 
https://condor-wiki.cs.wisc.edu/index.cgi/tktview?tn=1956,67 ).

Question: do you primarily use standard universe for checkpointing, or 
do you rely on remote system calls as well?  I ask because another 
option we are considering is to add support to the vanilla universe to 
easily handle standalone checkpointing where some signal is sent 
periodically to create a ckpt file in the vanilla job's output sandbox, 
whether the executable is linked w/ Condor's standalone checkpointing 
library or some other one.

regards,
Todd