[GE users] Parallel jobs on Heterogenous clusters

Gerald Ragghianti geri at utk.edu
Tue Oct 28 15:08:15 GMT 2008


    [ The following text is in the "ISO-8859-1" character set. ]
    [ Your display is set for the "ISO-8859-10" character set.  ]
    [ Some special characters may be displayed incorrectly. ]

Hi John,
I can see that those two options would work, but it seems that the users 
should not need to make these sort of decisions if it is possible for 
the system to do so. I want my users to simple say "-pe mpi 16" and the 
system should be able to pick between the available sets of machines.  
In that respect, the pe wildcard solution is the best I've found so 
far.  The problem with using complexes as you suggest is that it 
requires the user to choose the complex at submit time.  But if the job 
cannot be scheduled immediately the user has no way to know which 
complex (hosts) would become available first.

- Gerald

John Hearns wrote:
>
>
> 2008/10/28 Gerald Ragghianti <geri at utk.edu <mailto:geri at utk.edu>>
>
>     I there a simple way to prevent parallel jobs from running across
>     certain sets of machines?  For instance, this would be important
>     if you have hosts with widely different processor speeds or if you
>     have multiple Infiniband fabrics. 
>
>  
> You can do this (if you want to do it this way) by defining a complex 
> (qconf -mc)
> Let's call them  'type1' and 'type2' These are Boolean resources.
> Then for each host, set complex_values  type1=true   (or type2=true)
> USers then   qsub -l type1   to select type1 machines
> (or qsub -l infiniband or whatever)
>  
>  
>  
> Actually, you are better to use hostgroups. Just create two 
> hostgroups, called @type1 and @type2
> Add the type1 and 2 machines to each hostgroup (qconf -ahgrp) and the 
> users then request a specific hostgroup on the qsub line.
>  
> Hostgroups would be the cleanest way to do this.
>  


-- 
Gerald Ragghianti
IT Administrator - High Performance Computing
http://hpc.usg.utk.edu/
Office of Information Technology
University of Tennessee
Phone: 865-974-2448
E-mail: geri at utk.edu


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