I've tried the prolog/epilog way and it starts from the jobs workdir
so I'll try the JSV way: I do not want to bother the users with this
kind of "feature" ;)

Many tanks to all.

Stefano

On Tue, Feb 28, 2012 at 12:52 PM, Reuti <[email protected]> wrote:
> Am 28.02.2012 um 12:35 schrieb William Hay:
>
>> On 28 February 2012 11:02, Stefano Bridi <[email protected]> wrote:
>>> Hi list, I have a problem on a SGE setup where the home directory are
>>> shared trough glusterfs and some job failed to start because of a
>>> latency on the filesystem propagation between the login node and the
>>> compute node.
>>> What happen is that a script create a workdir with some support files,
>>> "cd" inside and then qsub a script, sometime the script start to run
>>> on the compute node too quickly and the "workdir" is not  yet visible
>>> on that node. I know it is a glusterfs problem that must be resolved
>>> elsewhere but in the meantime, where can I put a "sleep"?
>>> Does exist a prerun hook that I can use for that? For other use
>>> (copying files around and cleanup) does exist a similar postrun hook?
>>>
>> Had another thought.
>> Set up a load sensor for a >= complex that reports the current time
>> (seconds since 1970). Add a request to the qsub (via jsv if you don't
>> want to make the submission process more complex)
>> for that complex with a value greater than now+fudge factor.
>
> The idea with the set time is good, but maybe it can be done without a load 
> sensor:
>
> Set -a <date/time> in the qsub submission for the starttime, either in the 
> script or as suggested in an JSV.
>
> -- Reuti
>
>
>> William
>>
>> _______________________________________________
>> users mailing list
>> [email protected]
>> https://gridengine.org/mailman/listinfo/users
>>
>

_______________________________________________
users mailing list
[email protected]
https://gridengine.org/mailman/listinfo/users

Reply via email to