Using XFS on Raid

Using XFS on Raid

[James Woolliscroft]

[-- Attachment #1.1: Type: text/plain, Size: 2017 bytes --]

Hi,

 

I am new to XFS, I have spent a significant time searching for information
and have not found any documentation relating to the use of XFS over RAID
with N Drives where the Stripe Unit >256K.

 

To be clear, using (for example) 3 Disks with RAID 5 with a Stripe Unit of
1MB and Stripe Width of 2 would give 2MB Data and 1M Parity.

 

I have found through extensive testing of RAW RAID volumes with various
chunk sizes that larger chunks tend to produce the best performance in most
cases whether Sequential, Random, Read or Write, or in combinations. I am
keen therefor to perform similar tests with XFS to determine the best
performance and would like to test the widest range of larger chunks/Stripe
Units possible, including those larger than XFS apparently permits. I am
aware that XFS performance will not necessarily mirror that of performance
tests of RAW RAID volumes but this needs to be verified by testing.

 

It appears that XFS on RHEL7 and clones is limited to a maximum Stripe Unit
of 256KB.

 

Since it is not possible to specify the correct parameters (SU=1024K, SW=2),
Is it possible and valid to specify an adjusted value where SU is reduced
and SW is increased by the same factor.

 

So using SU=256K, SW=8 would be the same as SU=1024K, SW=2 and give a total
of 2MB Data.

 

Is this valid? Are there any performance or other consequences of using
different dimensions? Is there any issue with using other permutations e.g.
SU=32K, SW=64?

 

While I understand that XFS is intended to select optimum defaults, I am
concerned and less than enthusiastic about the default settings particularly
with larger RAID chunks.

 

I also would like to find some information on how to correctly interpret the
statistics output when formatting a volume with XFS, particularly as the
process may use other settings which may require further investigation?

 

If you cannot answer these questions directly, can you point me to a
resource where I might find this information?

 

Regards

 

JW


[-- Attachment #1.2: Type: text/html, Size: 4449 bytes --]

[-- Attachment #2: Type: text/plain, Size: 121 bytes --]

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

Re: Using XFS on Raid

[Dave Chinner]

On Thu, Feb 05, 2015 at 12:23:53AM +0000, James Woolliscroft wrote:
> Hi,
> 
>  
> 
> I am new to XFS, I have spent a significant time searching for information
> and have not found any documentation relating to the use of XFS over RAID
> with N Drives where the Stripe Unit >256K.

Use the defaults - they do the right thing.

> It appears that XFS on RHEL7 and clones is limited to a maximum Stripe Unit
> of 256KB.

That's not correct. mkfs.xfs prior to v3.2.2 (i.e the latest) throw
a warning about the default log stripe unit not being able to be set
larger than 256k, but it then sets the lsu to a default of 32k and
continues.

The current version (which I think will hit RHEL7.1) omits
the warning and just sets the lsu to 32k - defaults should just do
the right thing and not bother users who are using defaults.

Either version, however, will still result in the data stripe
unit/width is still going to be 1m/2m in your case.

> I also would like to find some information on how to correctly interpret the
> statistics output when formatting a volume with XFS, particularly as the
> process may use other settings which may require further investigation?

They are a reflection of the command line parameters, or if not
specified, the default CLI values. So, read the mkfs.xfs man page as
it explains what everything means.

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com

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