Questions about MD 2-drive RAID-10

Questions about MD 2-drive RAID-10

[Nathan Eisenberg]

Hi there,

It's my first post here!  For some background, I'm a sysadmin at a datacenter provider.  I have a large number of MD devices under my control, and a fair amount of experience using MD and the majority of its RAID drivers.  Just so I've said it, I did quite a lot of searching before deciding to post, in the hopes that I could find my answer documented somewhere.  If I've missed a good article where my question was clearly answered, then I do apologize (and feel free to mock me for it)!

Until a few hours ago, I was under the impression that a RAID-10 required at least four disks, in pairs that form the underpinning mirrors for the stripe.  A few days ago, a customer of mine requested a two drive RAID-10, and was surprised by my insistence that it was impossible.  Imagine my shock when I stumbled across http://neil.brown.name/blog/20040827225440.

http://en.wikipedia.org/wiki/Non-standard_RAID_levels#Linux_MD_RAID_10
This article talks about MD's RAID-10 driver's support for a 2-disk RAID-10, in both 'near' and 'far' modes.  It also mentions the potential performance advantages (over a RAID-1) of using the RAID10 driver in the 'far' configuration on a 2-disk array, but it doesn't say much about the 'near' mode.

My question is this: what differences, if any, are there between building a RAID-1 with two disks, and building a RAID-10 with two disks in the 'near' configuration?  Does this result in the OS 'striping' its reads across the mirrored disks, improving read (but presumably not write) access?  Do I correctly interpret the article's suggestion that the 'far' configuration will result in increased performance for both reads and writes (again, over a RAID-1), or am I crazy?

Thank you for your time!

Nathan

Re: Questions about MD 2-drive RAID-10

[NeilBrown]

[-- Attachment #1: Type: text/plain, Size: 2492 bytes --]

On Thu, 12 Jan 2012 23:27:06 +0000 Nathan Eisenberg <nathan@atlasnetworks.us>
wrote:

> Hi there,
> 
> It's my first post here!  For some background, I'm a sysadmin at a datacenter provider.  I have a large number of MD devices under my control, and a fair amount of experience using MD and the majority of its RAID drivers.  Just so I've said it, I did quite a lot of searching before deciding to post, in the hopes that I could find my answer documented somewhere.  If I've missed a good article where my question was clearly answered, then I do apologize (and feel free to mock me for it)!
> 
> Until a few hours ago, I was under the impression that a RAID-10 required at least four disks, in pairs that form the underpinning mirrors for the stripe.  A few days ago, a customer of mine requested a two drive RAID-10, and was surprised by my insistence that it was impossible.  Imagine my shock when I stumbled across http://neil.brown.name/blog/20040827225440.
> 
> http://en.wikipedia.org/wiki/Non-standard_RAID_levels#Linux_MD_RAID_10
> This article talks about MD's RAID-10 driver's support for a 2-disk RAID-10, in both 'near' and 'far' modes.  It also mentions the potential performance advantages (over a RAID-1) of using the RAID10 driver in the 'far' configuration on a 2-disk array, but it doesn't say much about the 'near' mode.
> 
> My question is this: what differences, if any, are there between building a RAID-1 with two disks, and building a RAID-10 with two disks in the 'near' configuration?  Does this result in the OS 'striping' its reads across the mirrored disks, improving read (but presumably not write) access?  Do I correctly interpret the article's suggestion that the 'far' configuration will result in increased performance for both reads and writes (again, over a RAID-1), or am I crazy?
> 

It works best if questions are one-per-line so that I can write the answers
between the questions.

1/
There is no differences in where the data gets written to the devices.
There are no differences in how data is written.
There might be differences in how data is read.  Each try to 'balance' reads
between the drives, but might do it slightly differently.

2/ no

3/ 'far' is faster for reads and slower for writes in my measurements.  You
should do your own measurements on you own hardware with your own
load-generation to see how it works for you.

4/  no, I don't think you are crazy.   But the jury is still out....

NeilBrown

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 828 bytes --]