[GE users] $round_robin vs $fill_up for LAM/MPI Integration for SGE

Reuti reuti at staff.uni-marburg.de
Fri Jan 20 08:58:32 GMT 2006


Depends on the application. Having some programs doing heavy memory I/ 
O but only low interprocess-communication via MPI, they would have to  
share one memory controller - in Intel dual dual-core it would be  
four to share the memory bandwidth. Even with AMD's Opterons, the  
dual-core type might be only second choice for some application, as  
here again they share one memory controller to the CPU-attached memory.

As dual-core Opterons have a lower power consumption (and so also  
good to buy because of the saved energy), idea might be to have per  
dual dual-core machine two parallel slots for the high-memory-IO  
jobs, and two slots for low-memory-IO jobs. Until some kind of root- 
only CPU affinity is available (and requestable in SGE), you can only  
hope that the internal Linux scheduler puts the jobs on the CPUs in  
the best way.

For other types of jobs of course, having a 8 CPU dual-core solution  
might be better suited (although more expensive) and use $fill_up there.

-- Reuti

Am 20.01.2006 um 04:39 schrieb Bernard Li:

> Currently the default allocation_rule for the LAM/MPI Integration for
> SGE is $round_robin.  I was wondering why that is the case and  
> wouldn't
> $fill_up be a better way to go?
>
> In that case you make sure that cpus that are closest to each other  
> (on
> the same node for SMP servers) are allocated first.
>
> Thoughts?
>
> Thanks,
>
> Bernard
>
> ---------------------------------------------------------------------
> 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