[GE users] Help: Slots Problem

Reuti reuti at staff.uni-marburg.de
Tue Sep 16 17:22:40 BST 2008


Am 16.09.2008 um 14:03 schrieb Lee Amy:

> 2008/9/15 Sean Davis <sdavis2 at mail.nih.gov>
> On Sun, Sep 14, 2008 at 9:06 AM, Lee Amy  
> <openlinuxsource at gmail.com> wrote:
> >
> >
> > 2008/9/14 Ravi Chandra Nallan <Ravichandra.Nallan at sun.com>
> >>
> >> Lee Amy wrote:
> >>>
> >>> Hello,
> >>>
> >>> I build a small cluster with 5 nodes, every node has 2 Opteron  
> 270 HE
> >>> processors(4 cores). There's a program I always called "TGICL"  
> can use the
> >>> total amount of processors on one machine. For example, when I  
> specific CPU
> >>> number of tgicl it will execute as the same number of threads.  
> However, SGE
> >>> still shows the job is taking up 1 slot. But I have set 4 cores  
> to run on
> >>> one node.
> >>>
> >>> So my problem is how to let SGE know the "real" taking-up slots?
> >>>
> >>> Thank you very much~
> >>>
> >>> Regards,
> >>>
> >>> Amy Lee
> >>>
> >> If you want SGE to know the no. of CPU's used by tgicl you could  
> use a new
> >> complex (qconf -sc). And for restricting the usage of CPUs per  
> tgicl program
> >> you may force to request the complex by your program. Then you  
> could set the
> >> complex value at host level and making it consumable should do  
> the job.
>
> Just to come back here, using a parallel environment with $pe_slots as
> the allocation method is an exact fit for the original post.  I don't
> think there is a need to involve complexes, etc.
>
> Sean
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: users-unsubscribe at gridengine.sunsource.net
> For additional commands, e-mail: users-help at gridengine.sunsource.net
>
> Should I use loose integration to replace tight integration?  
> Because I just enable the corresponding slots.

What do you mean with this? If you have e.g. 4 cores, you would need  
4 slots on this machine, and using a PE with allocation_rule  
$pe_slots you would always get 4 slots from one and the same machine.

This would be a Tight Integration which is always worth to be  
implemented. If you want to avoid serial jobs on the nodes: set  
"qtype NONE" in the queue definition and attach only the PE to it.

I agree with Sean, that you won't need any complexes for your setup.

-- Reuti

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