[gridengine users] making core binding useful for MPI

Reuti reuti at staff.uni-marburg.de
Tue Jun 7 14:26:41 UTC 2011


Am 07.06.2011 um 15:28 schrieb Dave Love:

> Reuti <reuti at staff.uni-marburg.de> writes:
> 
>> Hi,
>> 
>> Am 25.05.2011 um 17:30 schrieb Dave Love:
>> 
>>> It seems that core binding currently isn't typically useful with MPI
>>> jobs.  The pe method needs exclusive node access and a fixed  
>>> allocation
>>> rule that you need to know in advance, but then you could have the mpi
>>> do it anyhow (e.g. openmpi paffinity_alone) which works more  
>>> generally.
>>> (See past discussion.)  There seem similar restrictions on the set
>>> method in practice -- at least needing a fixed allocation rule, as you
>>> have to specify a fixed core count.
>>> 
>>> It looks straightforward to implement binding a number of cores  
>>> equal to
>>> the host-local slot count.  Does ayone see a problem with that or have
>>> strong feelings how to specify it?  I'd be inclined to let bare
>>> `-binding set' do it, but otherwise maybe a new type like `setslots'  
>>> analogous
>>> to `set'.
>> 
>> what would be the difference to the current implementation?
>> 
>> -- Reuti
> 
> Binding a number of cores equal to the allocated slot count for the node
> rather than a fixed number.  E.g. you have 16-core nodes with $fill_up
> and 24-slot (process) job, you'd probably want to bind 16 on one and 8
> on another and you can't currently specify that.

I see, thx. - Reuti



More information about the users mailing list