Hi,

Just to follow up this, at the end what I did was to recreate the sacctmgr
accounts that reset everything to 0 again. Thanks for your suggestions.

Regards,
Daniel

2011/7/20 Daniel Adriano Silva M <[email protected]>

> Hi,
>
> Thanks for your suggestions. I tried the approach:
>
> sacctmgr modify account <xxx> cluster=<yyy> set RawUsage=0
>
> Of course it reset the Rawusage to 0, but after Firshare was
> auto-recalculated the statistic are still "historic" because the "Effective
> Usage" still reflects the historic usage, so the Fairshare is more/less the
> same than before. Additionally,  something funny is that the RawUsage got
> some numbers (different to 0) when the statistics were recalculated. Any
> ideas to reset the whole thing completely?
>
> Daniel
>
> 2011/7/19 Danny Auble <[email protected]>
>
>> **
>>
>> Along Don's sacctmgr suggestion you should be able to specify "root" as
>> the account which will flush the entire fairshare usage. I would avoid using
>> the first method unless all else fails.
>>
>>
>> On a side note, using the sacctmgr method it appears you need to specify
>> the cluster as well (i.e. sacctmgr modify account <xxx> cluster=<yyy> set
>> RawUsage=0). I'll see about getting that to default to the local cluster in
>> 2.3.
>>
>>
>> Danny
>>
>>
>> On Monday July 18 2011 8:15:43 AM you wrote:
>>
>> > Daniel,
>>
>> >
>>
>> > The brute force approach to zero out all usage would be to stop the
>> slurmctld, remove the StateSaveLocation/assoc_usage and assoc_usage.old
>> files, and restart the slurmctld.
>>
>> >
>>
>> > You can reset the usage on a specific account using "sacctmgr modify
>> account <xxx> set RawUsage=0".
>>
>> >
>>
>> > Don
>>
>> >
>>
>> > From: [email protected] [mailto:
>> [email protected]] On Behalf Of Daniel Adriano Silva M
>>
>> > Sent: Monday, July 18, 2011 12:08 AM
>>
>> > To: slurm-dev
>>
>> > Subject: [slurm-dev] How to completely reset the Fairshare stats?
>>
>> >
>>
>> > Hi,
>>
>> >
>>
>> > I have a problem here! Previously we were using the multifactor plugin
>> by some months but just based in the QOS (i.e. I put 0 to all the other
>> factors). But now it become evident tat we really need the fairshare to keep
>> a better balance of the resources allocation, so two days ago I activated
>> the other multifactors like this:
>>
>> >
>>
>> >
>>
>> > #MultiFactorPriority for Jobs
>>
>> > PriorityType=priority/multifactor
>>
>> > PriorityDecayHalfLife=7
>>
>> > PriorityCalcPeriod=5
>>
>> > PriorityFavorSmall=NO
>>
>> > PriorityMaxAge=7-0
>>
>> > PriorityWeightAge=1000
>>
>> > PriorityWeightFairshare=5000
>>
>> > PriorityWeightJobSize=1000
>>
>> > PriorityWeightPartition=1000
>>
>> > PriorityWeightQOS=5000
>>
>> > PriorityUsageResetPeriod=NONE
>>
>> >
>>
>> > Now the problem is the next, as the history of the database is OLD the
>> users got a very small Fairshare factor (near to 0). I tried the reset trick
>> ( PriorityUsageResetPeriod=NOW) to reset the stats, it worked, but as soon
>> as I returned it to "NONE" and restarted the controller the sats went again
>> near to 0. So the question is, how to really reset the fairshare thing? This
>> in order to give a "fresh start" to the users.
>>
>> >
>>
>> > JOBID PRIORITY AGE FAIRSHARE JOBSIZE PARTITION QOS
>>
>> > 14057 6319 256 0 63 1000 5000
>>
>> > 14064 6191 168 2 21 1000 5000
>>
>> > 14065 6191 168 2 21 1000 5000
>>
>> > 14066 6191 168 2 21 1000 5000
>>
>> > 14067 6191 168 2 21 1000 5000
>>
>> > 14068 6191 168 2 21 1000 5000
>>
>> > 14069 6191 168 2 21 1000 5000
>>
>> > 14070 6191 168 2 21 1000 5000
>>
>> > 14071 6191 168 2 21 1000 5000
>>
>> > 14072 6191 168 2 21 1000 5000
>>
>> > 14093 6141 120 0 21 1000 5000
>>
>> > 14094 6141 119 0 21 1000 5000
>>
>> > 14105 6085 21 0 63 1000 5000
>>
>> > 14106 6085 21 0 63 1000 5000
>>
>> > 14107 6085 21 0 63 1000 5000
>>
>> > 14108 6085 21 0 63 1000 5000
>>
>> > 14114 6061 6 12 42 1000 5000
>>
>> > 14115 6061 6 12 42 1000 5000
>>
>> > 14116 6061 6 12 42 1000 5000
>>
>> >
>>
>> >
>>
>> > BTW. SLURM is fantastic. Good Job!
>>
>> >
>>
>> > Thanks,
>>
>> > Daniel
>>
>
>

Reply via email to