[GE users] dedicated master (admin/submit) host

Rayson Ho rayrayson at gmail.com
Wed Jan 18 15:55:43 GMT 2006


    [ 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. ]

If sharetree (or other complex scheduling algorithms) is on, then the
scheduler needs a bit more CPU cycles... but like Chris said, most
server class machines can easily handle 1000 jobs.

Use "top" to get an estimate of how busy are the SGE daemons... or if
you really want to dig into it, turn on qmaster/scheduler profiling
;-)

Rayson



On 1/18/06, Chris Dagdigian <dag at sonsorol.org> wrote:
>
> Hi Bill,
>
> The admin/submit functions are relatively small resource consumers.
> The biggest resource usage is by the node that hosts sge_schedd and
> sge_qmaster. The number of active and pending jobs in a system will
> determine how much CPU and Memory SGE is grabbing from the host
> system. For 1000 pending/active jobs this should be trivial on most
> server class systems.
>
> For a 15 node cluster with 1000 jobs pending it should be no problem
> to have the SGE master be on the same machine where people are doing
> submit/admin tasks. At most you may want to bump up the physical
> memory if you need to.
>
> For smaller clusters, the big pain point comes when people run a
> central NFS server on the same machine as the SGE master. In those
> cases it becomes beneficial to separate out fileservice and SGE
> service onto separate hosts.
>
> Regards,
> Chris
>
>
>
>
>
> On Jan 18, 2006, at 9:36 AM, Bill Knebel wrote:
>
> > I had a question that I am sure has been asked in some way.  We
> > have a small cluster ( 7 nodes) that performs serial jobs only. The
> > nodes are dual Xeon's. We limit the jobs so that only  one job runs
> > per CPU becuase they are  primarily CPU bound/intensive. We will be
> > adding an additional 8 nodes (dual Opterons).  On many occasions we
> > have 500 - 1000 jobs pending but at other times the cluster is
> > empty. We run jobs on all nodes (all nodes are set up as execute
> > nodes and one is both an exec host and a master (submit/admin)
> > host) .  I was wondering at what point, if ever, it becomes
> > necessary to dedicate one cpu or  one node (2 cpu's) to the SGE
> > admin/submit host functions and remove it as an execute host.  Can
> > anybody shed some light on this given their experience or point me
> > in the direction of some info?
> >
> > Bill
> >
> > --
> > Bill Knebel, PharmD, Ph.D.
> > Principal Scientist
> > Metrum Research Group
> > 2 Tunxis Road
> > Suite 112
> > Tariffville, CT 06081
> > email: billk at metrumrg.com
> > tel: (860) 930-1370
> >
>
> ---------------------------------------------------------------------
> 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