Quoting Seren Soner <[email protected]>:

>
> Hey Moe,
>
> On Wed, Feb 15, 2012 at 12:37 AM, Moe Jette <[email protected]> wrote:
>>
>> Hi Seren,
>>
>> I am not sure exactly what a "submit host" is in SGE, but in SLURM any
>> node with SLURM installed and with the matching configuration files
>> can submit to your cluster. That would typically include compute nodes
>> (where the parallel jobs run) and the nodes where users log in. If you
>> want to prevent users from submitting jobs from some node(s), a
>> job_submit plugin could be used for that purpose. See
>> http://www.schedmd.com/slurmdocs/job_submit_plugins.html (the Lua
>> version would be simplest to use).
> I fiddled out a bit more, and found out that what I was looking for
> was exactly AllocNodes. Basically, the users can submit batch jobs
> from these jobs.

Excellent.

>> The mail processing logic can be found in src/slurmctld/agent.c
>> mail_job_info(). If you decide to work on this, I would suggest that
>> you post ideas and any work you do to the mailing list so that it can
>> be incorporated into the code for everyone to use.
> I think that mail_job_info() could be improved highly with some little
> effort. We can change the function so that, the user can add such a
> line to their batch script:
> #SBATCH  
> --mail-info:submittime,starttime,endtime,processors,nodes,nodelist,jobid,jobname,account
> .. etc
>
> We can parse the user's requests and in the message itself, we can
> include all of the requested variables.
>
> What would your opinions be on that ?

Good idea. Add another option to srun/salloc/sbatch, update the man  
pages, add another field to the job submit request (data structure and  
RPC), the slurmctld's job structure (including save/restore of the  
field on restarts), perform some parsing in mail_job_info(), and that  
should work fine. It's not very complicated, but quite a few  
components are effected. If you send a patch to the list, it could be  
added to SLURM version 2.4 (we are only adding bug fixes to version  
2.3, although you are free to modify v2.3 for your own use).

>>
>> Thanks,
>> Moe Jette
>> SchedMD
>>
>> Quoting Seren Soner <[email protected]>:
>>
>>>
>>> Dear all,
>>>
>>> I have two little questions that I'm sure there are really quick
>>> answers to, just couldn't figure them out, so I decided I should
>>> finally send a question to the list :)
>>>
>>> First of all, can I set some nodes as "submit hosts", as in SGE ? I
>>> want batch jobs to be submitted from not only front-end but also the
>>> compute nodes that I am using. When I try that, I get, "sbatch: error:
>>> Batch job submission failed: Access/permission denied". I've looked
>>> into sacctmgr manual, but failed to find anything related.
>>>
>>> Secondly, what is the easiest way to modify sbatch mailer, i.e. the
>>> mail that we receive when a job begins, completes etc. It only has a
>>> subject, and no text in it, and I think there could be some more
>>> features that could be added to the mail. I thought of writing a
>>> jobcomp script to do that, but this won't work in the job begin case.
>>> Any ideas on that end ?
>>>
>>> Thank you all in advance,
>>> Seren Soner
>>> Bogazici University
>>
>
>
>
> --
> Thanks,
> Seren Soner

Reply via email to