[GE users] Setting minimal number of slots per job on some nodes

Emir Imamagic eimamagi at srce.hr
Fri Feb 3 11:08:33 GMT 2006


    [ The following text is in the "ISO-8859-2" character set. ]
    [ Your display is set for the "ISO-8859-10" character set.  ]
    [ Some special characters may be displayed incorrectly. ]

Reuti wrote:
> you can't specify a limit in the PE for now, but what about a forced- 
> complex for these nodes (or queue-instances)? So users submitting  jobs 
> have to specify this complex as a resource to get these nodes,  which is 
> only attached to the 8-core nodes. Of course, this wouldn't  prevent 
> anyone from requesting these nodes and submit still a serial  job (a 
> special PE attached only to these queue instances might also  be a way 
> [and no "BATCH" attribute set for them]).
> 
> Another thing I did to enforce the agreed 8-CPU limit for parallel  jobs 
> in my group: test in the PE start_proc_args the number of slots  you 
> got. I simply kill the job himself, if he got more than 8 slots.  So my 
> users learned this fast... (all nodes must be submit-hosts  though - or 
> put the job in error-state which is more convenient with  an exit code 
> of 100 in PE start [man sge_conf for details, under  FORBID_APPERROR]). 
> So you could test in a queue prolog, whether it's  on the forbidden 
> nodes, or in a start_proc_args whether it requests  at least 4 slots and 
> send it to error otherwise.
> 

Thanks for a quick response. Situation is pretty much as we expected. I 
just wanted to check if there is some clean solution that I'm not aware of.

For the beginning we will isolate SMP nodes in specific queue and 
disable BATCH jobs in that queue. We will create specific PE environemnt 
with allocation rule $pe_slots. That way SGE wont fill those nodes with 
serial jobs and small parallel jobs. Also we'll assume that users will 
behave.

We will monitor the usage and if users start abusing the system we will 
implement the prolog check that you suggested.

Thanks,
emir

---------------------------------------------------------------------
To unsubscribe, e-mail: users-unsubscribe at gridengine.sunsource.net
For additional commands, e-mail: users-help at gridengine.sunsource.net




More information about the gridengine-users mailing list