[gridengine users] Exclusive somewhere non-exclusive elsewhere
reuti at staff.uni-marburg.de
Tue Oct 30 11:25:02 UTC 2012
Am 30.10.2012 um 08:58 schrieb William Hay:
> On 29 October 2012 18:33, Reuti <reuti at staff.uni-marburg.de> wrote:
>> Am 20.08.2012 um 09:32 schrieb William Hay:
>>> On 18 August 2012 22:23, Reuti <reuti at staff.uni-marburg.de> wrote:
>>>> Am 17.08.2012 um 16:40 schrieb William Hay:
>>>>> We have a heterogenous cluster with several different types of node.
>>>>> We also have a couple of policies:
>>>>> i)Jobs won't run across multiple nodes of a given type if they can fit
>>>>> into a single node of that type.
>> With $fillup as allocation _rule and requesting excluisve access you should get the minimum amount of nodes.
> Doesn't help if you have partially full nodes though.
>>>>> ii)Jobs that run across multiple nodes have exclusive access to those
>>>>> nodes while jobs that run within a single
>>>>> node share if sufficient resources are available.
>>>>> A consequence of this is that some jobs get exclusive access to some
>>>>> nodes but not to others. It's not a simple case of some nodes being
>>>>> bigger in all respects than others either.
>>>> Looking at ii) I don't understand this. Does a job spanning several nodes have exclusive access to *all* nodes of this job, or only when they fill all slots on this node? If there are leftover slots, they have to share the node with another job?
>>> Exclusive access to all nodes it runs on. That's both desired and the
>>> effect of requesting an exclusive resource associated with a host.
>>> The idea here is that we don't want a multi-host job to run slow
>>> because it is being slugged by another job on a single host. On the
>>> other hand we don't want to devote a whole 12 core node to a piddly
>>> single slot job.
>> You mean, if set up as outlined above, that the very last process of a parallel job would go to a node and maybe executed there alone. And for this very last one, you would allow additional jobs to be scheduled to?
> No multi-host jobs always get exclusive access to all nodes they run on.
God, then it's only necessary to request exclusive access with the parallel jobs. But this would be too simple. I still miss the configuration issue here.
More information about the users