[GE users] MPI processes don't get killed with control_slaves

markhewitt mh613 at york.ac.uk
Wed Sep 2 14:34:18 BST 2009


Nothing to add except that I've follow the instructions for tight 
integration to the letter (checked it a dozen dozen times) and still the 
  processes do not get killed.

I'm using mvapich1 over infiniband.

Mark


craffi wrote:
> You might be seeing the difference in "tight" vs. "loose" integration  
> of your parallel environment:
> 
> http://gridengine.info/2005/09/19/parallel-environments-pes-loose-vs-tight-integration
> 
> Quite a bit of the implementation details depend on what MPI you are  
> using and how it was installed.
> 
> Regards,
> Chris
> 
> 
> 
> On Sep 2, 2009, at 4:46 AM, geralt wrote:
> 
>> Hello,
>>
>> I've been trying to limit the memory usage of MPI processes on our  
>> cluster by setting the vmem limit. The goal was to kill the job if a  
>> certain limit is exceeded.
>>
>> This works only partially - SGE kills the scheduled job but the  
>> spawned MPI processes are still running.
>>
>> I've tried to use the control_slaves option in the parallel  
>> environment. But something seems to be missing; according to the  
>> sge_pe manpage some "sophisticated PE interface" is necessary.
>>
>> Does anyone know an answer?
>>
> 
> ------------------------------------------------------
> http://gridengine.sunsource.net/ds/viewMessage.do?dsForumId=38&dsMessageId=215437
> 
> To unsubscribe from this discussion, e-mail: [users-unsubscribe at gridengine.sunsource.net].

------------------------------------------------------
http://gridengine.sunsource.net/ds/viewMessage.do?dsForumId=38&dsMessageId=215451

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



More information about the gridengine-users mailing list