[GE users] How to schedule just on slots, and ignore load values?

Andreas.Haas at Sun.COM Andreas.Haas at Sun.COM
Thu Jul 19 11:37:46 BST 2007


    [ 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 Jeff,

On Wed, 18 Jul 2007, Beadles, Jeff wrote:

>
> I'd like to see what sort of scheduling performance boost I can get by
> just having the schedule dispatch jobs to systems with slots available,
> ignoring whatever load may be on the system.  I have a couple of
> clusters that are only used via sge,
>
> Is there an easy way to do this?  I tried changing the scheduler
> configuration for "load_formula" to "NONE", but was told;
>
> The keyword "none" is not allowed in "load_formula"

Surely, after all Grid Engine needs some load index ;-)

Actually I wouldn't expect that load formula costs so much, 
but you could try '0' as load formula.

> I can set the queue "load_thresholds" to "NONE", but I would rather just
> have it not even check.

Disabling load thresholds is at all events advisable for nodes 
whose workload is entirely managed by Grid Engine.

> Any suggestions?  I'm currently running 6.0 u8, and can't change anytime
> real soon.

As for boosting scheduler times upgrading to u9/u10/u11 is the
way to go. Note, u8 was the last update without fixes for these

    http://gridengine.sunsource.net/issues/show_bug.cgi?id=2093
    http://gridengine.sunsource.net/issues/show_bug.cgi?id=2094

with the scenarios I ran to test these improvements scheduler
dispatch times could significantly be reduced.

If upgrading is not doable in short-term an improvement (up to 50%) of 
the dispatch time is achievable by utilizing the fact, that -q matching 
is always quicker than -l matching. Reason is that -l is more indirect 
than -q.

Needless to say that -l can not always be substituted with -q,
but in setups with more than a single cluster queue some means
is often needed as to direct jobs into the cluster queue where
they shall run.

Scheduler dispatch can be monitored with PROFILE=true as sched_conf(5) 
param and doing grep for "job dispatching took" in scheduler messages 
file.

Regards,
Andreas

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