[prev in list] [next in list] [prev in thread] [next in thread] 

List:       lustre-discuss
Subject:    [Lustre-discuss] Re: adding space
From:       nathan () clusterfs ! com (Nathaniel Rutman)
Date:       2006-10-24 13:49:56
Message-ID: 453E6E61.2060602 () clusterfs ! com
[Download RAW message or body]

Oops.  The free-space stripe weighting (what we call "stripe QOS"), 
although present in all the 1.6 betas, was inadvertently set to give a 
priority of "0" to the free space (versus trying to place the stripes 
"widely" -- nicely distributed across OSSs and OSTs to maximize network 
balancing).
This priority can be adjusted via the proc file 
/proc/fs/lustre/lov/lustre-mdtlov/qos_prio_free
The default in the future will be 90%.
You can set this permanently on existing betas with this command on the MGS:
lctl conf_param <fsname>-MDT0000.lov.qos_prio_free=90

Note that setting the priority to 100% just means that OSS distribution 
doesn't count in the weighting, but the stripe assignment is still done 
via a weighting -- if OST2 has twice as much free space as OST1, it will 
be twice as likely to be used, _not_ guaranteed to be used.

Also note that stripe QOS doesn't kick in until two OSTs are imbalanced 
by more than 20%.  Until then, a faster round-robin stripe allocater is 
used.  (The new round-robin order also maximizes network balancing.)


Nathaniel Rutman wrote:
> David Bernick wrote:
>> Nathaniel Rutman wrote:
>>  
>>> Indeed.  Lustre 1.6 includes a weighting function to try to fill
>>> less-used OSTs preferentially.
>>> You can tell Lustre to assign a file to a particular set of OSTs using
>>> "lfs setstripe" (stripe start).
>>> You can "setstripe" a directory and everything (new) in that dir 
>>> will be
>>> written using those stripe
>>> parameters.
>>>     
>>
>> Is this the weighting function present in Lustre 1.5.95? Or will it be
>> added to 1.6 further down the line?
>>
>>   
> It's in all the betas, bug 9862 in ChangeLog.
>
> _______________________________________________
> Lustre-discuss mailing list
> Lustre-discuss@clusterfs.com
> https://mail.clusterfs.com/mailman/listinfo/lustre-discuss
>

[prev in list] [next in list] [prev in thread] [next in thread]