Re: wish for Linux MD mirrored raid types

["Keld Jørn Simonsen" <keld@keldix.com>]

On Fri, May 06, 2011 at 12:54:28PM +0200, David Brown wrote:
> On 06/05/2011 12:05, Jonathan Tripathy wrote:
> >
> >>>>RAID1 is traditionally a mirror only setup (ok, some RAID
> >>>>implementations may do some load-balancing of some sort). So a RAID1
> >>>>with 4 disks is one data set copied onto 4 disks. Bandwidth is roughly
> >>>>the same as a single disk (ignoring any load balancing).
> >>>>RAID10 is mirror and stripe. A RAID10 with 4 disks is similar to a 2
> >>>>disk RAID0 (double bandwidth with data split in half across both
> >>>>disks),
> >>>>but with each disk having a mirror (which brings the total up to 4
> >>>>drives).
> >>>>
> >>>>Additionally, a RAID1 disk (at least using MD) can be accessed just
> >>>>like
> >>>>a normal disk (good for recovery etc.) however a single disk out of a
> >>>>RAID10 array is next to useless.
> >Just so I can sleep at night, is my understanding of RAID10 and RAID1
> >above correct?
> >
> 
> It's mostly right (assuming, of course, that /I/ am correct here...).
> 
> RAID1 is traditionally a two-way mirror on two disks (or possibly on two 
> other raid sets, as in raid0+1).  So if you talk about RAID1 with 4 
> disks, you should probably qualify it more precisely - otherwise people 
> will wonder what you mean, or think it is impossible (many other RAID1 
> solutions, hardware or software, don't support more than two-way 
> mirrors).  /I/ would certainly say that a 4-disk RAID1 is a four-way 
> mirror as you described - but some people might think of a standard 
> layout RAID1+0.
> 
> As you say, RAID10,near on four disks is pretty much identical to 
> RAID1+0 - i.e., a stripe of two normal RAID1 pairs.
> 
> A single md RAID1 disk can be accessed like a normal disk, /if/ it uses 
> metadata format 0.90 which is put at the end of the drive.  If you have 
> later metadata formats that are at the beginning, then that will cause 
> trouble if you try to view the disk without using md.  A single disk 
> from a RAID10 is, as you say, useless without md.
> 
> However, assuming your recovery PC supports md raid, then you can 
> assembly your single RAID10 disk as a degraded RAID10 array.  After all, 
> it wouldn't be very redundant if you only had access to your data when 
> all the disks were working!

I agree with what David Brown says here.

Just some further remarks:

You can have Linux MD raid10 arrays with only 2 disks. 
This is different from RAID1+0 that requires at least 4 disks.

You can also with Linux MD have an odd number of disk, which is
impossible with RAID1+0.

Linux MD raid10 is a kernel driver that handles everything in
one module, while RAID1+0 is handled in two drivers - raid0 over raid1.
The latter is called nested RAID, and the Linux MD raid10 is not a
nested RAID, while RAID1+0 is.

With Linux MD raid10,near and superblock 0.90 you can also 
boot it from old grub, lilo etc, as each partition can be seen as a normal
file system.

As David pointed out, "a single disk out of a RAID10 array is next to
useless" is not true. Linux MD RAID10 is designed to always be able to 
hold all data intact if one disk is out. You just need Linux MD raid10
aware software. 

I apologise for being abrupt in my first answer. But the thing is that
I am getting tired of explaining and correcting all these misconceptions
again and again. And my proposal is an attempt to stop at least some
of the need for explaining, while also bringing the best practise forward to 
the users of our technology, for the benefit of the users.

best regards
keld
and then