[gridengine users] "Distribution" of processes to nodes

Skylar Thompson skylar2 at u.washington.edu
Fri Mar 20 15:27:53 UTC 2015


Hi Christian,

I think you're looking for allocation_rule in sge_pe(5). It sounds like
you're set to $round_robin, when you want to be on $fill_up. It can be
tuned on a per-PE basis, so you don't need to reconfigure your existing
queues to take advantage of it.

On Fri, Mar 20, 2015 at 04:18:26PM +0100, Christian Schmidt wrote:
> Hello grindengine-users,
> 
> we're running SGE on a small cluster consisting of nodes offering 16 
> slots each.
> 
> one of our cluster users asked us to change the cluster configuration in 
> this way:
> 
> "Even when a job requests an integer multiple of 16 processes, these 
> tasks are distributed across many nodes with a range of tasks/node from 
> 16 to 1; obviously the total performance is reduced to that of the 
> slowest task, which is that of the isolated process for which all 
> communication goes over the infiniband.  The difference in performance 
> is serious.
> 
> Please can you reconfigure the queue to assign tasks to nodes such that 
> large parallel jobs will not be split up to fill in all small spaces on 
> the queue, but rather will be run on the minimum number of complete nodes?"
> 
> We are not sure how to achieve this. Could you please give a any hint?
> 
> TI & Kind Regards,
> Christian
> 
> -- 
> No signature available.
> _______________________________________________
> users mailing list
> users at gridengine.org
> https://gridengine.org/mailman/listinfo/users

-- 
-- Skylar Thompson (skylar2 at u.washington.edu)
-- Genome Sciences Department, System Administrator
-- Foege Building S046, (206)-685-7354
-- University of Washington School of Medicine



More information about the users mailing list