[GE users] Please Review: Non-Multiplied Consumable Requests for Parallel Jobs

Bradford, Matthew matthew.bradford at eds.com
Fri Sep 26 10:50:21 BST 2008


>-----Original Message-----
>From: Andy Schwierskott [mailto:andy.schwierskott at sun.com] 
>Sent: 25 September 2008 14:46
>To: users at gridengine.sunsource.net
>Subject: RE: [GE users] Please Review: Non-Multiplied 
>Consumable Requests for Parallel Jobs
>> The ability to request exclusive nodes would be really useful for us.
>> One thing, we use the ACCT_RESERVED_USAGE and 
>> so that users get charged for the number of slots that they 
>are using 
>> up (NSlots x WallTime). It would be good if requesting an exclusive 
>> job would mean that a user gets charged for using up all the 
>slots in 
>> the containing queue, as we wouldn't want a user to request a single 
>> slot on a 16 core machine, but with exclusive access and not 
>be charged for it.
>I understand you need: you fear your users could abuse this 
>feature and try to get nodes exclusively but will not get 
>accounted for it properly. However there are use cases where 
>this features is needed to get accesss exclusively to a 
>machine because there's some special hardware in it - in this 
>case the user should not be charged for the remaining slots. 
>You certainly could argue that if you have job types in your 
>cluster which require exclusive access and this leaves some of 
>the computational resources on a machine unused you should not 
>charge the user for this problem.


I guess this is moving away from the discussion, but what about having
something like a charging/usage parameter set on the nodes which can be
optionally enabled system wide or on a per host basis. It would only
parameters were set, . Then, if a user requests exclusive access to a
node, then that user is charged correctly for using the entire node, so
instead of the usage being recorded as (Nslots x wall_time) it would be
(Charge_Parameter X wall_time).

Just an idea, and there may be (and probably will be) implication to
other aspects which I haven't thought about.


>You say a user should be accounted for the remaining slots in 
>the queue instance in that host(s). However this exclusive 
>access also makes all other queue instances on that hosts 
>unusable. Even no "debug" "low priority" or other jobs could 
>run on that host.

That's right, but it is a consequence of the SGE configuration, and it
is up to the SGE Administrator to design the queueing system
accordingly. If the configuration is not set to 1 slot per core, then
that is the admins decision, and the accounting may not reflect how many
cores a job is actually reserving.
>Regards "documentation" purposes: I would not change the 
>"accounting" file - it would overstretch what this simple 
>single line record can do. Rather this information 
>could/should be somehow reflected for the "reporting" file - 
>just like information for Advance Reservations in the 
>"reporting" enable an admin to see who has requested resources 
>(which possibly had not been used).
>To unsubscribe, e-mail: users-unsubscribe at gridengine.sunsource.net
>For additional commands, e-mail: users-help at gridengine.sunsource.net

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