[parted-devel] [PATCH] libparted: fix optimal IO alignment

[Phillip Susi]

Looks like this one fell off the radar too, could you apply it now?

On 3/5/2012 9:17 PM, Phillip Susi wrote:
> -----BEGIN PGP SIGNED MESSAGE-----
> Hash: SHA1
>
> On 03/01/2012 12:50 PM, Jim Meyering wrote:
>>
>> Hi Phillip,
>>
>> Can you describe, or better still, tell me how to demonstrate the
>> problem this is fixing?  Or can you at least give a few details, like
>> which kernel you used, which command invoked and which values you got
>> for the two *_io variables?
>
> I have a server with a 4 disk mdadm raid5 that has an optimal value of
> 1.5 MiB as a result of the 512k stripe factor.  The current default
> stripe factor is 512k but it used to be 64k.  If you used a 64k stripe
> factor that would give an optimal value of 192k with a 4 disk raid5.
>
>> I think your description was somehow negated, since it's the reverse
>> actually: it accepts a nonzero kernel alignment A only if 2^20 % A == 0.
>> I.e., that requires A<= 1024^2.  A kernel alignment value larger
>> than 1024^2 is ignored.
>
> You are right, the comment was wrong.  Try this instead:
>
> There were several apparently incorrect tests that would cause the
> kernel supplied optimal io size to be discarded in favor of the default
> 1 MiB alignment.  As written, it would only accept values less than the
> parted default ( 1 MiB ) that were also an even power of two.  For
> example, 128k or 512k would be accepted, but 192k or 1.5 MiB would be
> ignored.