On 01/17/2013 10:35 AM, Alexander 'Leo' Bergolth wrote:

If using some kind of container like lvm or dm-crypt that has a
data-offset inside my md device...
Do I need to align the data offset to the stripe size (chunk-size*number
of data-bearing disks) or just to the chunk size?

If aligning to the chunk size is sufficient, could anyone give an
explanation? Doesn't this break filesystem alignment inside the
container? (I.e. the --stripe-width option)

I've done some benchmarks using a raid-6 md-device (5 disks, 64k
chunksize) with lvm (different data-offsets) and ext4 on top.

The ext4 filesystem is always aligned to the array
(stride=16,stripe-width=48) but the data offset of the lvm physical
volume is either the default of 2048 sectors (which isn't a multiple of
the stripe size of 3*64k) or 384 sectors (3 data bearing disks * 64k
chunks).

The results are available at
http://leo.kloburg.at/tmp/raid-dataoffsets/

It looks like it doesn't make any difference if the data offset is a
multiple of the stripe size or just of the chunk size.

But why doesn't a wrong data-offset break the --stripe-width hint given
to mkfs.ext4???

I'd really appreciate if someone could shed some light on that.

Thanks,
--leo

P.S.: The benchmarks were done with bonnie++, details can be found in
bench.sh at the above URL.

-- 
e-mail   ::: Leo.Bergolth (at) wu.ac.at
fax      ::: +43-1-31336-906050
location ::: IT-Services | Vienna University of Economics | Austria