Not reassigning host slots

[sosborne76]

I have noticed a general behaviour with my Qube installation. For instance if there were a number of jobs being processed the most recently submitted job could complete but its slots would not be reassigned to older still running jobs. Why is this? Is it a configuration issue or normal behaviour.

It makes sense to me that it works in a FIFO manner especially in a college environment.

Any thoughts?

[shinya]

Hi sosborne76,

It sounds like a bug, and we've had a few customers run into a similar issue.

May I ask what version your supervisor is? In particular, what does a "qbping" on the command prompt say? 

Also, what is your supervisor platform?

You might remove the 'stub_optimize' flag from the supervisor_flags setting; this will disable some scheduling optimizations, but will allow for more strict adherence to the first-in/first-out and priority scheduling.
You'll need to restart the supe once you make the change.

[sosborne76]

Its version 6.0.1 and its on a relatively new Win 2008 R2 build.

Although to be honest i did notice this behaviour on a version 5.5/Win 2003 R2 build.

[shinya]

Thanks for the version info.
Please do try removing the "stub_optimize" flag, and let us know how it goes.

BTW, we're about to release (early next week, if not this week) 6.1.1 and 6.0.2 maintenance releases, and they do have some improvements in this very area, so you may want to try it out when it's out. If you're going to go to 6.1, you'll need a new license (free upgrade if you're on support), so please be aware.

-shinya.

[sosborne76]

I have 6.0.1 installed on the supervisor, does this license not cover 6.1 (and 6.x for that matter)?

I will try removing that flag.

[sosborne76]

I changed the flag and performed a basic test when the farm was idle (locked a couple of workers, submitted big job, submitted small job to locked workers, unlocked them, small job completed, idle workers were then pulled into big job). So it seems to work but I will be interested to see it in a proper live environment. I will keep you posted.

By the way do I need a new license for the supe (which is 6.0.1) if I want to run 6.1?

[jburk]

You will need to request a 6.1 license.

[mim]

just FYI:

i had the same problem this night running on Qube 6.4-0 bld-2012.08.27.18.44.30

only 5 out of 10  nodes received the "old" job after the previous finished. workers stayed online but idle. restarting the workers didn't help, but resubmitting the job did.

i'll try removing the stub_optimize flag and keep you posted.

[mim]

stub doesn't change a thing......

still have the issue that high-priority jobs won't assign even to idle workers which is pretty annoying.
cluster/groups and so on are all correct but the idle workers will only accept the job when i resubmit it.