[GE users] Help: How SGE dispatches jobs?

Chris Dagdigian dag at sonsorol.org
Fri Jul 25 16:37:40 BST 2008


    [ The following text is in the "WINDOWS-1252" character set. ]
    [ Your display is set for the "ISO-8859-10" character set.  ]
    [ Some special characters may be displayed incorrectly. ]

Hi Amy,

This may help:

"Scheduler Policies for Job Prioritization in the N1 Grid Engine 6  
System"
http://www.sun.com/blueprints/1005/819-4325.html

There is a lot of stuff that happens within every SGE scheduling  
interval:

In every scheduling cycle, Grid Engine hast to:

1 Gather info about all jobs currently running
2 Gather info from hosts via queue execd daemons
3 Load/utilization reports, slot info
4 Recheck status of all global resources and consumables
5 Remember some things about past usage (if share tree is active)

Then it needs to ...

- Filter pending job list for those that CAN be started
- Check hold status, dependency requirements, etc.
- Compute priority for all pending jobs
- Filter prioritized list for jobs that CAN be run
- Availability of hard resource requests & hosts
- Dispatch into available job slots in order of priority
- Queue instance selection sort
- Sort among available queue instances
- Hard requests 1st, soft requests 2nd
- If multiple queue instances are acceptable; subsort on load or seqno
- Dispatch to remote host

Rinse, repeat ?

At the end of this process the SGE scheduler ends up with a list of  
one or more exechosts that are capable of satisfying the needs of the  
next job to be dispatch.

By default SGE will sort the "suitable" hosts by load to pick the  
"final" host that gets the job - this behavior is overriden by turning  
on SEQNO-based queue sorting.

It boils down to: "In the default configuration, Grid Engine will  
select the *least busy* system that is capable of providing all the  
resources required by the job"

-Chris







On Jul 25, 2008, at 11:24 AM, Lee Amy wrote:

> Hello,
>
> I'm in curiosity with SGE that how SGE dispatch jobs to the nodes in  
> a big cluster? Based on Load Average or anything else?
>
> Can anyone show me some hints?
>
> Thank you very much~
>
> Regards,
>
> Amy Lee


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