Re: RAID5 reconstruction ?

On Sun, May 31, 2009 at 12:24 AM, Goswin von Brederlow
[off-list ref] wrote:

Redeeman [off-list ref] writes:

quoted

On Sat, 2009-05-30 at 14:35 +0100, John Robinson wrote:

quoted

On 30/05/2009 06:44, SandeepKsinha wrote:

quoted

Hi all,

Say If I have a RAID 5 array of 50GB of five disks of 10GB each.

I have data of 5GB. When a disk fails and replaced with a spare disk.
Will the reconstruction happen only for the 5GB allocated disk blocks
or it will happen for the whole disk size.

The whole disc size, for now anyway; md does not currently note which
blocks have been used by its client (the filesystem, LVM, whatever).

quoted

Is it possible to make  reconstruction intelligent enough to keep it optimized ?

This has been discussed in combination with supporting SSD drives' TRIM
function, and would mean md had to keep track of used chunks or possibly
even sectors using a bitmap or something like that, but whether anyone's
working on it I don't know.

I would say it should be possible to 'query' the filesystem for that
information. Obviously this will only work if you run a filesystem on it
which supports it, but it would seem like a nicer solution than a bitmap
for it.

You have put a big constraint here of "filesystem which supports it".
In general, I have not known any file system work as yet which
leverages the underlying device topology to optimize its block
allocation policies for enhaced I/O, etc.
Or for any other reasons.

Having a bitmap can surely have lot of other benefits too. Looking at
the drive sizes in recent times, think of situation where you have to
do a reconstruction or resysnc. It might take months for them to
complete. Also, in the meanwhile you will have degraded I/O's.

Just in worst case, if your drive has most of allocated blocks, it
will be a penalty.

On the other hand checking a bitmap is quick. You could use the bitmap
not only for reconstruction but also for reads. If the bitmap say the
block is unused you can skip the read and just zero fill the
buffer. This would speed up reads and writes that don't cover the full
stripe. And compared to a raid with current bitmap there shouldn't be
any real slowdown for the extra "unused" bit.

The only special case for resync would be that if all data blocks in a
stripe are unused then the parity blocks can be marked unused too.

MfG
       Goswin

-- 
Regards,
Sandeep.





 	
“To learn is to change. Education is a process that changes the learner.”
--
To unsubscribe from this list: send the line "unsubscribe linux-raid" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html