[GE users] how can configure SGE for different request of jobs

reuti reuti at staff.uni-marburg.de
Sun Aug 29 14:57:10 BST 2010


Am 29.08.2010 um 09:10 schrieb mrostaee:

> Thx for your reply
> another question:
> nodes on cluster have 4 slots.
>> in that senario, can i prevent oversubscription for cores of nodes at that queue? ( nodes have 4 slots, but a user request pe_slot PE that 48 slots needs and i want to use just one node)
>> can i limit users or hosts with condition by RQS?

What is your queue and PE definition in detail and what do you observe? 

-- Reuti


> --- On Sat, 28/8/10, reuti <reuti at staff.uni-marburg.de> wrote:
> 
>> From: reuti <reuti at staff.uni-marburg.de>
>> Subject: Re: [GE users] how can configure SGE for different request of jobs
>> To: users at gridengine.sunsource.net
>> Date: Saturday, 28 August, 2010, 5:12 PM
>> Hi,
>> 
>> Am 28.08.2010 um 12:51 schrieb mrostaee:
>> 
>>> I defined a queue on 5 nodes.
>>> I want a job  that request slots <48 just use
>> one host  (no distribution, just one host).
>>> and for slots>48, jobs can be distributed on all
>> hosts of defined queue.
>>> 
>>> how can i configure SGE for that user?
>> 
>> with two different PEs like mpi_pe and mpi_rr with
>> allocation_rule $pe_slots resp. $round_robin the user can
>> use the appropriate one for his intended number of slots.
>> 
>> Although it's an RFE having the resource quotas working the
>> other way round (i.e. as minimum requirements for certain
>> conditions), there is nothing built in like this for now to
>> make the correct request for the PEs automatically. But this
>> we can also do by other means:
>> 
>> Step 1) By defining a JSV (job submission verifier) the
>> users can request e.g. "-pe mpi* 77" and the JSV will have
>> to replace the "mpi*" with the correct one. The JSV also
>> needs to check, whether the user is trying to cheat the
>> system by requesting illegal combinations like "-pe mpi_rr
>> 4".
>> 
>> Optional step 2) In addition a fake PE called only "mpi"
>> with zero slots set in its definition can avoid, that users
>> have to quote "mpi*" all the time to avoid expansion of the
>> asterisk by the shell*). In the plain PE "mpi" no jobs will
>> ever run of course, but the JSV will replace this with the
>> best suited one anyway.
>> 
>> -- Reuti
>> 
>> *) I find it somewhat inconsequent, that the bash will on
>> the one hand replace "mpi*" with the list of filenames, in
>> case there are any starting with "mpi" . But on the other
>> hand forward the plain asterisk in case there are none - the
>> list should simply be empty then.
>> 
>> ------------------------------------------------------
>> http://gridengine.sunsource.net/ds/viewMessage.do?dsForumId=38&dsMessageId=277689
>> 
>> To unsubscribe from this discussion, e-mail: [users-unsubscribe at gridengine.sunsource.net].
>> 
> 
> ------------------------------------------------------
> http://gridengine.sunsource.net/ds/viewMessage.do?dsForumId=38&dsMessageId=277891
> 
> To unsubscribe from this discussion, e-mail: [users-unsubscribe at gridengine.sunsource.net].
>

------------------------------------------------------
http://gridengine.sunsource.net/ds/viewMessage.do?dsForumId=38&dsMessageId=277945

To unsubscribe from this discussion, e-mail: [users-unsubscribe at gridengine.sunsource.net].



More information about the gridengine-users mailing list