XFS over SSD

XFS over SSD

[Raz]

I am thinking of using XFS over a SSD disk.
1. Can I separate xfs meta data ( not just the logging) from the SSD ?
can I put the meta on a different disk ?
2. When does a file's meta ( its block maps ) is being flushed to disk
? can I control it ?

thank you
raz

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs

Re: XFS over SSD

[Justin Piszcz]

On Mon, 8 Dec 2008, Raz wrote:

> I am thinking of using XFS over a SSD disk.
I am currently using XFS on an Intel X25-E with the logdev internal, using 
the defaults.
I see the point for putting the journal elsewhere; however, that may 
defeat the purpose of an SSD, if its latency is 0.01 or 0.001 and the HDD 
is 8-14ms, then it may defeat the purpose for writes or fast access 
anyhow?

> 1. Can I separate xfs meta data ( not just the logging) from 
the SSD ? > can I put the meta on a different disk ?
Yes, see the logdev option on the mkfs.xfs manpage.

> 2. When does a file's meta ( its block maps ) is being flushed to disk
> ? can I control it ?
I am not sure you can control this, will let the XFS devs/others take on 
this question.

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs

Re: XFS over SSD

[Eric Sandeen]

Raz wrote:
> I am thinking of using XFS over a SSD disk.
> 1. Can I separate xfs meta data ( not just the logging) from the SSD ?
> can I put the meta on a different disk ?

Are you talking about just the log (see the mkfs man page for external
logs, as Justin suggested) or all metadata?  For the latter, using the
realtime subvolume does accomplish this (data on one volume, metadata on
the other) but that's not used very often.

-Eric

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs

Re: XFS over SSD

[Raz]

On Mon, Dec 8, 2008 at 4:46 PM, Eric Sandeen <sandeen@sandeen.net> wrote:
> Raz wrote:
>> I am thinking of using XFS over a SSD disk.
>> 1. Can I separate xfs meta data ( not just the logging) from the SSD ?
>> can I put the meta on a different disk ?
>
> Are you talking about just the log (see the mkfs man page for external
> logs, as Justin suggested) or all metadata?  For the latter, using the
> realtime subvolume does accomplish this (data on one volume, metadata on
> the other) but that's not used very often.
>
> -Eric
>
I am referring to all the meta data. 128K of erase block for some block map
update is a big penalty. I do not like much rt volumes. I tried that and it is
cumbersome. UBIFS cannot handle 80GB Flash disks ( well, they say it is up to
16GB in MTD web site) . I am about to start benchmarking the SSD with XFS
( versus raw access ) and see how performance degrades, in read and writes.
If there was a way to set XFS meta data ( superblocks, allocation groups...
on a different device) it would have been nice, since we plan to use the SSD
as a fast IO device and data persistence is not the main thing here.

we use XFS on all our SATA based servers, we tweak it ( extents and
raid awareness).
XFS proved to be the fastest file system for appliances that use
Multimedia files
and big IOs ( 1MB).

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs

Re: XFS over SSD

[Eric Sandeen]

Raz wrote:
> On Mon, Dec 8, 2008 at 4:46 PM, Eric Sandeen <sandeen@sandeen.net> wrote:
>> Raz wrote:
>>> I am thinking of using XFS over a SSD disk.
>>> 1. Can I separate xfs meta data ( not just the logging) from the SSD ?
>>> can I put the meta on a different disk ?
>> Are you talking about just the log (see the mkfs man page for external
>> logs, as Justin suggested) or all metadata?  For the latter, using the
>> realtime subvolume does accomplish this (data on one volume, metadata on
>> the other) but that's not used very often.
>>
>> -Eric
>>
> I am referring to all the meta data. 128K of erase block for some block map
> update is a big penalty. I do not like much rt volumes. I tried that and it is
> cumbersome. UBIFS cannot handle 80GB Flash disks ( well, they say it is up to
> 16GB in MTD web site) . I am about to start benchmarking the SSD with XFS
> ( versus raw access ) and see how performance degrades, in read and writes.
> If there was a way to set XFS meta data ( superblocks, allocation groups...
> on a different device) it would have been nice, since we plan to use the SSD
> as a fast IO device and data persistence is not the main thing here.
> 
> we use XFS on all our SATA based servers, we tweak it ( extents and
> raid awareness).
> XFS proved to be the fastest file system for appliances that use
> Multimedia files
> and big IOs ( 1MB).
> 

I have yet to play with xfs on ssd, but I would imagine that setting up
the fs geometry to match the ssd preferred IO sizes and/or erase block
sizes might at least help.

-Eric

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs

Re: XFS over SSD

[Dave Chinner]

On Mon, Dec 08, 2008 at 05:11:58PM +0200, Raz wrote:
> On Mon, Dec 8, 2008 at 4:46 PM, Eric Sandeen <sandeen@sandeen.net> wrote:
> > Raz wrote:
> >> I am thinking of using XFS over a SSD disk.
> >> 1. Can I separate xfs meta data ( not just the logging) from the SSD ?
> >> can I put the meta on a different disk ?
> >
> > Are you talking about just the log (see the mkfs man page for external
> > logs, as Justin suggested) or all metadata?  For the latter, using the
> > realtime subvolume does accomplish this (data on one volume, metadata on
> > the other) but that's not used very often.
> >
> > -Eric
> >
> I am referring to all the meta data. 128K of erase block for some block map
> update is a big penalty.

That's not an issue for smart SSDs like the intel one. They
internally use a log-based structure that means that small writes
don't end up requiring an entire erase block to be rewritten.  That
is, they have a low write amplification factor.

Such SSDs don't really care that much about the filesystem structure
and random write patterns. Older/cheaper SSDs have a much larger
write amplification factor and so are unsuited to traditional
filesystem structures.....

Google "SSD write amplification" for more information.

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs

Re: XFS over SSD

[Emmanuel Florac]

Le Mon, 8 Dec 2008 14:42:34 +0200 vous écriviez:

> I am thinking of using XFS over a SSD disk.
> 1. Can I separate xfs meta data ( not just the logging) from the SSD ?
> can I put the meta on a different disk ?
> 2. When does a file's meta ( its block maps ) is being flushed to disk
> ? can I control it ?

I've made some testing on a bunch of MTRON SSDs. Actually, the best bet
is to use one SSD to store the XFS log for an array of SATA or SAS
drives, that's the highest performance gain ; it offers tremendous IO
performance.


-- 
--------------------------------------------------
Emmanuel Florac               www.intellique.com   
--------------------------------------------------

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs