Proper swidth and sunit for RAID 5 (does it matter)?

Proper swidth and sunit for RAID 5 (does it matter)?

[Justin Piszcz]

Dave,

What is the proper sunit and swidth for a 64, 256, and 1024 kilobyte chunk 
size with a 10-disk raid 5?

Also, in the majority of benchmarks it does not seem to matter whether the 
FS is stripe-aligned or not (with SW raid)- does it mainly/only affect HW 
raid?

Current settings (10 disks):

# xfs_info /dev/md3
meta-data=/dev/md3               isize=256    agcount=4, agsize=82417536 
blks
          =                       sectsz=4096  attr=2
data     =                       bsize=4096   blocks=329670144, imaxpct=25
          =                       sunit=0      swidth=0 blks, unwritten=1
naming   =version 2              bsize=4096
log      =internal               bsize=4096   blocks=32768, version=2
          =                       sectsz=4096  sunit=1 blks, lazy-count=1
realtime =none                   extsz=9437184 blocks=0, rtextents=0

Justin.

Re: Proper swidth and sunit for RAID 5 (does it matter)?

[David Chinner]

On Tue, Jan 15, 2008 at 12:29:14PM -0500, Justin Piszcz wrote:
> Dave,
> 
> What is the proper sunit and swidth for a 64, 256, and 1024 kilobyte chunk 
> size with a 10-disk raid 5?
> 
> Also, in the majority of benchmarks it does not seem to matter whether the 
> FS is stripe-aligned or not (with SW raid)- does it mainly/only affect HW 
> raid?

Affects both. If you are doing large I/O, both SW and HW raid will
avoid RMW cycles if you can do full stripe writes and that means
they go faster. The faster the RAID array, the bigger the difference
it will make.

If your tests are with small I/O or with a config that can't
do I/O large enough for full stripe writes, then you won't see
any difference as you're not avoiding RMW cycles.

> Current settings (10 disks):
> 
> # xfs_info /dev/md3
> meta-data=/dev/md3               isize=256    agcount=4, agsize=82417536 

Why 4 ags?

The low number of AGs is an optimisation for single disks,
not multi-disk arrays that have much more parallelism and
seek capacity available.

> blks
>          =                       sectsz=4096  attr=2
> data     =                       bsize=4096   blocks=329670144, imaxpct=25
>          =                       sunit=0      swidth=0 blks, unwritten=1

And you don't even have su/sw set here, so XFS won't be doing
any alignment optimisation at all.

Cheers,

Dave.
-- 
Dave Chinner
Principal Engineer
SGI Australian Software Group