xfs
[Top] [All Lists]

Re: XFS shrinking planned?

To: Spelic <spelic@xxxxxxxxxxxxx>, xfs@xxxxxxxxxxx
Subject: Re: XFS shrinking planned?
From: Eric Sandeen <sandeen@xxxxxxxxxxx>
Date: Wed, 29 Oct 2014 08:03:07 -0500
Delivered-to: xfs@xxxxxxxxxxx
In-reply-to: <5450B560.6000208@xxxxxxxxxxxxx>
References: <544FC202.1000200@xxxxxxxxxxxxx> <544FD4C1.4020004@xxxxxxxxxxx> <5450B560.6000208@xxxxxxxxxxxxx>
On 10/29/14 4:37 AM, Spelic wrote:
> On 28/10/2014 18:39, Eric Sandeen wrote:
>> Not formally planned, there are bits and pieces out there (i.e. the inode
>> mover) which are part of what it might take to achieve a shrinker.
>>
>> Another option, rather than fs shrinking, is to use the dm-thinp target, 
>> which
>> would allow you to allocate a large-but-sparse block device, create a very
>> large filesystem on that, and add or remove storage as needed.
>> (At least I think you can remove it...!)
>>
>> -Eric
> 
> Thanks for your reply Eric
> 
> Interesting technique, but for enforcing a maximum size (smaller than
> the very large allocated thin device) I would have to rely on quotas,
> which probably decreases performance.

"probably"

> Then using thinp would mess up
> all the disk layout, basically replacing the XFS allocator, which
> most likely would decrease performances significantly.

"most likely"

> And then the
> thinp code itself is a medium performance thing and I don't think it
> can keep up with XFS performances, so that would presumably be a hard
> bottleneck.

"presumably"

> All this would result in a performance almost certainly
> lower than ext4.

"almost certainly..."
 
All possibilities, but possibly also worth testing to find out.  ;)
It's true that today the thinp allocator will impact XFS allocation
patterns to some degree.

Anyway, shrink has been on the radar for years, it's just never really
been a priority.  It might happen some day...

-Eric

> Thanks
> S.
> 

<Prev in Thread] Current Thread [Next in Thread>