[Moabusers] Fairshare

Laurence Dawson larry.dawson at vanderbilt.edu
Thu May 18 09:04:55 MDT 2006


Dave,

I'll send the results of  mdiag-f in a non-public email but there are 
two things that are not as we expected:

- the interval setup seems to be ignored when we use "dedicatedps*", 
there is only one interval - this might just be because the data is not 
available, that's not a problem if that is the case.

- the fairshare usage totals for some accounts increases over their 
prior values...this is impossible in the model we were thinking of - are 
you distributing fairshare so that it will sum to 100%? (not what we 
were expecting)

Larry


Dave Jackson wrote:
> Larry,
>
>   Enabling 'percentage based fairshare' is documented in section 6.3.2.1
> of the online Moab Admin Manual.  Please confirm that the algorithm
> description meets your expectations and that recent versions properly
> deliver this capability to your satisfaction.
>
> Thanks,
> Dave
>
> On Wed, 2006-05-17 at 13:58 -0500, Laurence Dawson wrote:
>   
>> Dave,
>> The stability problems we were seeing here on the earlier version of 
>> moab 4.5 seem to be fixed with 4.5.0p4, and it is now running stably on 
>> our cluster with torque 2.1.0p0. We are ready to test the fairshare 
>> calculation change...how do I enable it?
>>
>> Larry.
>>
>> Dave Jackson wrote:
>>     
>>> Larry,
>>>
>>>   This capability was enabled in Moab 4.5.0 last week.  We are still in
>>> testing.  We will send you config details as soon as testing is
>>> complete.
>>>
>>> Thanks,
>>> Dave
>>>
>>> On Wed, 2006-02-01 at 12:51 -0600, Laurence Dawson wrote:
>>>   
>>>       
>>>> Yes, that's exactly what we need.
>>>> Thanks - I'll look forward to trying it out.
>>>> Larry.
>>>>
>>>> Dave Jackson wrote:
>>>>     
>>>>         
>>>>> Larry,
>>>>>
>>>>>   You are correct that Moab calculates fairshare usage as a percentage
>>>>> of 'delivered' cycles.  In cases where the cluster is full and a backlog
>>>>> exists, this is often deemed the most fair and effective method.
>>>>> However, in a case where the cluster is not fully utilized, this can
>>>>> lead to issues such as you are seeing.  
>>>>>
>>>>>   We propose to enable a new policy whereby fairshare usage can be
>>>>> calculated as a percentage of theoretical (or potential) cycles.  This
>>>>> would result in the same job for job fairshare usage regardless of
>>>>> actual cluster workload conditions.
>>>>>
>>>>>   Would this work for you?  If so, we will get this to you this week.
>>>>>
>>>>> Dave
>>>>>
>>>>> On Wed, 2006-02-01 at 10:35 -0600, Laurence Dawson wrote:
>>>>>   
>>>>>       
>>>>>           
>>>>>> We are working on cluster fairness and the way fairshare is calculated 
>>>>>> is not working very well for us. Fairshare is calculated on the basis of 
>>>>>> resource usage on the cluster and not on total resource availability. 
>>>>>> This means that users who submit jobs during very quiet periods are 
>>>>>> charged more against their fairshare than we want.
>>>>>> A simple example:
>>>>>> Day 1: 10 processors out of 1000 in use = 1%
>>>>>> Day 2: 10 processors out of 100 in use = 10%
>>>>>>
>>>>>> We want to be able to have the fairshare calculated on the basis of  
>>>>>> total resources available, so in both cases it would be (1200 total 
>>>>>> procs)/10. Is there a way of configuring this? Any ideas on alternative 
>>>>>> methods?
>>>>>> _______________________________________________
>>>>>> moabusers mailing list
>>>>>> moabusers at supercluster.org
>>>>>> http://www.supercluster.org/mailman/listinfo/moabusers
>>>>>>     
>>>>>>         
>>>>>>             
>>>>>   
>>>>>       
>>>>>           
>>>   
>>>       
>
>   


More information about the moabusers mailing list