What is advantage and disadvantage of stripe-mirror and mirror-stripe?

[nataris]

What is advantage and disadvantage of stripe-mirror and mirror-stripe? How do they different?

Please help me to clarify.

Thanks

[VK2COT · Accepted solution]

Hello,

These days, most of valuable data is kept on
SAN. Hence, RAID-1+0 and RAID-0+1 are
not often used at Volume Manager level.

I assume your question was related to
volume managers and these RAID levels.
I am sure you are aware that
RAID-1+0 and RAID-0+1 can be implemented
in software (volume managers) and
hardware (smart controllers, SAN, and
similar).

The two special RAID levels were designed
to provide additional data-protection and
performance benefits. They do it at an expense of
"doubling the number of disks required for normal operation".

The answer to your question cannot be
provided in simple terms because there are
other elements that influence the solutions.
For example, are all disks on one or multiple
controllers, do you use smart controllers,
and so on...

Here is a simple digram to visualise these
two RAID levels (as you know, as of HP-UX 11.31 Sep 2008,
you can now use these RAID levels with LVM too):

http://docs.hp.com/en/5187-1369/ch01s06.html#i1359720

If you search this Forum, you will find
some other references, like:

http://forums13.itrc.hp.com/service/forums/postanswer.do?threadid=1291033&forumId=1&closedFlag=0

From that thread, I quote a nice, short
response by Doug de Werd:

BEGIN QUOTE
This is the difference. Let's assume 8 drives. for simplicity,
look at it like this:

1 3 5 7
2 4 6 8

In a RAID 0+1, the odd number drives would
be striped first (essentially a RAID 0 with
drives 1,3,5,7), and then the stripe set
mirrored to the even number drives (another
RAID 0 with 2,4,6,8). If you lost drive 3,
you would still be OK, but if you lost drive
3 and drive 6, you would be hosed because
both RAID 0 mirror sets would be broken.

In RAID 1+0, drives 1-2 are mirrored, 3-4,
5-6, and 7-8, then striped. You can lose any
combination - up to 4 drives total, as long
as they are both not a mirrored pair. So you
could lose 1, 4, 5, and 8 and still be
running. But if you lost any of the mirrored
pairs (like 1-2, 3-4, 5-6, or 7-8) you would
be hosed.
END QUOTE

Cheers,

VK2COT

[Steven E. Protter]

Shalom,

mirror then strip give you better write performance as there are two places for writes to go instead of a bunch with the alternative.

It also gives you the possibility, done right of surviving the failure of a single disk.

stripe-mirror gives none of these advantages.

SEP

[Suraj Singh_1]

Apart from the comments above, recovery of RAID 1+0 volume would be much faster than the recovery of RAID 0+1 volume, because of the fact that in the latter case the whole stripe will have to be recovered from its mirror, while in RAID 1+0, only one mirror will need to be recovered. ( This scenario mentioned here assumes a single disk failure scenario).

[Hein van den Heuvel]

Nataris,

It depends some on the place where the striping and mirroring is happeling, but typically stining mirrors is preferred.
Is all work in software, all in hardware (controllers?), or one in hardware (typically the mirror), the other in software (striping accross controllers/fibres).

SEP>> mirror then strip give you better write performance as there are two places for writes to go instead of a bunch with the alternative.

I beg to differ. Both in the 1+0 and 0+1 each modified block has to be written to exactly 2 spindles.

SEP>> It also gives you the possibility, done right of surviving the failure of a single disk.

Both methods give you basic single disk survival. Both methods give you selected multi disk survival, but the level of survival is different.
Striping mirrors allows the mirroring to go partially forward. With one drive failing in a mirrored stripe set, typically that whole stripeset is taken out of the mirror, and as Suraj indicates, the all the blocks for all the drives will have to be copied to re-intergrate a repaired member. Typically the mirror software does not know which specific drive had failed, and has stopped mirroring the whole set.

Btw... at least one (old) controller I know, the MYLEX, had a cute raid + 1 algoritme which allowed for an odd member count. If would rotate chunks and their copies over all drives one after an other.
So if there were spindles 1, 2 and 3 and the chunks were A and its copy a, B and its copy b, then the data would look like:

...> more chunks
1 A b D e G
2 a C d F g
3 B c E f .
:
v

more drives if available

Cheers,
Hein.