[gridengine users] Some generic questions: binding, parallel, over-subscription
reuti at staff.uni-marburg.de
Fri Nov 1 13:37:28 UTC 2013
Oh dear, having the checkpoint/h_rt issue yesterday I found this without a reply.
Am 14.12.2012 um 17:43 schrieb Dave Love:
> Reuti <reuti at staff.uni-marburg.de> writes:
>>> I'd have said they're not the same generally. The reserved granted
>>> resources are a subset of the requested ones. Even an h_vmem request,
>>> say, could be soft. (This is relevant for resource reservation as well
>>> as advance reservation.)
>> True for soft requests of booleans/strings (where is it set e.g. to
>> prefer a certain type of machine, but others would be accepted too),
>> but the soft request of h_vmem, h_cpu and h_rt are different, as they
>> are more treated like to generate a warning if passed.
> Could you explain that?
The s/h_vmem, s/h_cpu and s/h_rt are observed by SGE and set up as two stages: a) a warning when passing the soft limit, b) a termination when passing the hard limit.
(The s/h ulimits are handled differently: the process can change the soft limit as long as it is below the hard limit, but if the soft limit is passed it's killed instantly - there is no "warning facility" here.)
> Experimentally it seems to work as I thought, e.g. the job gets the
> h_vmem value defined by a queue when submitted with a soft request for a
> larger value, if h_vmem hasn't been made consumable.
Yes, but this way the warning will never take place. Requesting s_vmem I would expect a warning when this value is passed. Not like "I want 2 GB but I'm also happy with 1 GB".
>> And consumables can't be soft.
> Community Grid Engine: http://arc.liv.ac.uk/SGE/
More information about the users