[linux-lvm] Data alignment

[Mike Snitzer]

On Fri, Mar 19 2010 at  6:21pm -0400,
Mike Snitzer <snitzer at redhat.com> wrote:

> On Fri, Mar 19 2010 at  5:27pm -0400,
> Phillip Susi <psusi at cfl.rr.com> wrote:
> 
> > On 3/19/2010 3:25 PM, Mike Snitzer wrote:
> > > Any reason why you'd like to use DOS partitioning (first primary
> > > partition starting at 63rd sector)?
> > 
> > Convention.  I am looking into using GPT instead though.
> > 
> > > Anyway, if you have a recent kernel (e.g. 2.6.33) you'll be in good
> > > shape regardless.  If you create a partition on the device (using DOS
> > > partitions) the kernel _should_ be aware of where the partition starts
> > > and tell you how far you'd have to shift the start of your LVM2 PV data
> > > area to get it perfectly aligned relative to the underlying physical
> > > block size.  Check for example:
> > > # cat /sys/block/sda/sda1/alignment_offset
> > > 
> > > But if your device is using 512b physical_sector_size you'll just have a
> > > 0 for alignment_offset.  Check physical_block_size with:
> > > # cat /sys/block/sda/queue/physical_block_size
> > 
> > 0 alignment, physical block size 512.
> > 
> > > Also verify that your SSD device is naturally aligned (aka
> > > alignment_offset=0); I'd wager it is naturally aligned:
> > > # cat /sys/block/sda/alignment_offset
> > 
> > Kernel thinks so, and based on performance tests it appears so.
> > 
> > > All said, even if you have an older kernel, to manually get what you
> > > want (shift start to account for DOS partition at 63rd sector, align PV
> > > pe_start on a 512K boundary), please try:
> > > # pvcreate --dataalignmentoffset 512b --dataalignment 512K ...
> > 
> > Won't that just add one sector to the start, placing it at sector 1025?
> >  How does an alignment offset of 1 sector account for the partition
> > starting on sector 63?
> 
> The --dataalignmentoffset acts as padding.  I was focused on getting you
> to a power of 2 start (sector 64).  Then from there --dataalignment
> governs the start boundary for the PV data area.
> 
> So you're right, 64*512b + 512K isn't a multiple of 512K :)
> 
> This will get you what you want:
> # pvcreate --dataalignmentoffset 512b --dataalignment 480K /dev/sda1

Alternatively you could do (as long as sysfs has alignment_offset=0):
# pvcreate --dataalignment 480.5K /dev/sda1

or

# pvcreate --dataalignment 961S /dev/sda1