[GE users] Machine Priority

Kirk Patton kpatton at transmeta.com
Tue Oct 19 15:34:08 BST 2004


On Tue, Oct 19, 2004 at 10:22:20AM -0400, Sean Dilda wrote:
> On Tue, 2004-10-19 at 09:58, Reuti wrote:
> > > I have a cluster running SGE 6.0  This cluster consists of a mixture of 
> > > 1Ghz, 2Ghz, and 3Ghz machines.  Is there a way to ensure that the 3Ghz 
> > > machines are always used first out of the pool of open resources?
> > 
> > You can set a different sequence number for each type of machine and 
> > change the scheduler to sort by seqno (qconf -msconf). This will give 
> > the intended result (for serial jobs).
> > 
> > For simpleness I suggest defining three hostgroups. Then you can set:
> 
> I have an almost identical problem to the original poster.  I had
> considered this, but from what I can tell, it means that load average
> will no longer be taken into account (except for alarm states) when
> scheduling jobs.  This seems like a bad thing to me.
> 
> I've also considered the load scaling.  Unfortunately, not matter what
> scaling you do, a load of 0.00 will always scale to a load of 0.00, and
> this doesn't help the problem any.
> 
> Does anyone have a better way to do this?  Perhaps somehow add a
> constant to the generated load of each host based on processor speed?

The way I approached this issue was to create a complex named 'relative_speed'.
Before I add a host to the cluster, I run a little script that generates a number
to rate the speed of the system.  The script just times how quickly the system can
gzip a random amount of data.

One I have this number, I attach it to each host.  In my scheduler config, I user
'qsort_sort_method load' and my load formula subtracts the 'relative_speed' from the
np_load_avg.

It seems to work o.k.

Kirk
 
:>qconf -ssconf 
algorithm                  default
schedule_interval          0:0:15
maxujobs                   0
queue_sort_method          load                          <===
user_sort                  true
job_load_adjustments       np_load_avg=0.50
load_adjustment_decay_time 0:7:30
load_formula               np_load_avg-relative_speed    <====
schedd_job_info            true

qconf -sc host
relative_speed   rspeed     INT    0               <=    YES         NO         1    



> 
> 
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: users-unsubscribe at gridengine.sunsource.net
> For additional commands, e-mail: users-help at gridengine.sunsource.net
> 

-- 
Kirk Patton
Unix Administrator
Transmeta Inc.
Tel. 408 919-3055

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