[gridengine users] mem_free being treated as "per core" on certain queues

William Hay w.hay at ucl.ac.uk
Thu Mar 1 07:03:38 UTC 2012


On 29 February 2012 22:14, Joe Whitney <jwhitney at gmail.com> wrote:
> Hello,
>
> I am having a simple problem where the behaviour of mem_free resource
> requests are being treated differently on two different queues (actually,
> separate installations of SGE).
>
> For context, the hosts servicing queue.A have 32G/4cores available, while
> those servicing queue.B have 24G/4cores available. All queues/nodes are
> idle/empty when I perform the following tests.
>
> On queue.A, the following jobs get scheduled and run as expected:
>
> $ echo ls | qsub -q queue.A -pe smp 1 mem_free=20G #runs right away
> $ echo ls | qsub -q queue.A -pe smp 2 mem_free=20G #also runs right away
>
> On queue.B, only the first gets run:
>
> $ echo ls | qsub -q queue.A -pe smp 1 mem_free=20G #runs right away
> $ echo ls | qsub -q queue.A -pe smp 2 mem_free=20G #sits in 'qw' state
> forever
>
> Conjecture: mem_free is being treated as a "per-core" request for queue.B
> but not for queue.A.
>
> Question 1: is this possible?
> Question 2: how do I fix queue.B so that it performs like queue.A?
As you say they are two different SGE installations mem_free is
probably marked consumable
in the installation where it is treated per core.  The "fix"  would be
to edit the complex configuration
and mark it non consumable.   It is possible that it is also
consumable in installation/queue A but per JOB.

If you didn't do the installations you should try to work out why they
are set up this way unless you are
certain it is accidental

qconf -sc on each cluster to see how the various resources are configured


William



More information about the users mailing list