[GE users] allocation rules

Reuti reuti at staff.uni-marburg.de
Wed Apr 23 18:29:48 BST 2008


Am 23.04.2008 um 16:53 schrieb Maxime Kinet:

> Is it possible to specify an upper/lower limit on the number of  
> slots per job on a defined queue? I cannot figure out how to do it.

Unfortunately: no. I would also like to have it: http:// 
gridengine.sunsource.net/issues/show_bug.cgi?id=2148

-- Reuti


> Thanks
>
> Max
>
>
>
> On 21 Apr 2008, at 12:29, Reuti wrote:
>> Hi,
>>
>> Am 21.04.2008 um 11:05 schrieb Maxime Kinet:
>>
>>> we have a cluster of 20 machines, each one made of 8 cores,  
>>> running SGE 6.0. The jobs are either parallel, running on more  
>>> than one core, or serial jobs using only one core. I would like  
>>> to have jobs running on less than eight processors to be  
>>> dispatched on the same machine, in order to minimize the  
>>> communication time. Right now, when jobs are submitted, the  
>>> scheduler choose an empty node, if there is one, regardless the  
>>> number of nodes required.
>>>
>>> Here is an example to make things clearer. If I have two  
>>> machines, that is sixteen nodes, and I start by submitting two  
>>> one-processors jobs. The first one will be dispatched to  
>>> machine1, and the second one to machine2. Then, if I launch a 8  
>>> processors job, it will be spread between the two nodes (7 procs  
>>> on machine1 and one on machine2). While it could have choosed to  
>>> run the two serial jobs on machine1, and the parallel job on  
>>> machine2.
>>>
>>> Therefore, I think the allocation rule I choosed for the parallel  
>>> environment ('Fill up') is obviously not the good one. Is there  
>>> another allocation rule that wou,ld more apropriate to the kind  
>>> of tasks at the hand? Should I create several queues for the  
>>> different kinds of jobs?
>>>
>>> Thanks a lot for any answers/explanation.
>>
>> This might help: http://blogs.sun.com/sgrell/entry/ 
>> grid_engine_scheduler_hacks_least (maybe with "allocation_rule  
>> $pe_slots").
>>
>> In addition (or as a more simple standalone setup), you could  
>> define two queues (one "qtype none" for parallel jobs only), and  
>> change the seq_no for both queues in the opposite way. I.e. serial  
>> jobs will first use machine 1, then 2. While parallel ones will  
>> first use machine 2, and then 1(the scheduler must be set to  
>> "queue_sort_method seqno").
>>
>> -- Reuti
>>
>> ---------------------------------------------------------------------
>> To unsubscribe, e-mail: users-unsubscribe at gridengine.sunsource.net
>> For additional commands, e-mail: users-help at gridengine.sunsource.net
>>
>


---------------------------------------------------------------------
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