RAID-6 support in kernel?

RAID-6 support in kernel?

[Roy Sigurd Karlsbakk]

hi all

I'n working on server setup with some 16 disks in RAID-5; one of them a 
spare. After a little reading, I find myself longing for support for 
RAID-6 support in kernel, giving the opportunity to allow for two failed 
drives without a chrash (see links about RAID-6 below if interested).

I am aware of that not all kernel hackers like such configurations, and
that some will rather see small RAID-configurations connected with VLM.  
I beleive there is a reason for using RAID-6, and RAID-controller vendors
(such as Compaq) are already using them, so why shouldn't linux do so
also? With a high number of cheap IDE drives, the chance of one failing is 
quite high, so why not RAID-6? At least for a system doing most reads...

thanks

roy

RAID-6 layout: http://www.acnc.com/04_01_06.html

-- 
Roy Sigurd Karlsbakk, Datavaktmester

Computers are like air conditioners.
They stop working when you open Windows.

Re: RAID-6 support in kernel?

[Derek Vadala]

On Mon, 3 Jun 2002, Roy Sigurd Karlsbakk wrote:

> I am aware of that not all kernel hackers like such configurations, and
> that some will rather see small RAID-configurations connected with VLM.  
> I beleive there is a reason for using RAID-6, and RAID-controller vendors
> (such as Compaq) are already using them, so why shouldn't linux do so
> also? With a high number of cheap IDE drives, the chance of one failing is 
> quite high, so why not RAID-6? At least for a system doing most reads...

See the following thread from March 2002 on linux-raid:
http://groups.google.com/groups?hl=en&lr=&safe=off&th=804941541a023c63&seekm=linux.raid.Pine.LNX.4.44.0203261239110.12942-100000

You can always fake this effect by combining two 8-disk RAID-5s into a
RAID-0. It's not technically RAID-6, but can withstand a 2-disk failure,
although not _any_ 2-disk failure. However, it's my understanding that
RAID-6 cannot withstand _any_ two disk failure either (see the above
thread). 

I also suspect that the use of dual RAID-5s combined with the CPU overhead
of ATA will kill most systems under any kind of load. For that matter, the
2x parity hit from RAID-6 probably wouldn't make you CPU too happy either,
even if there was a kernel driver that implemented it.

---
Derek Vadala, derek@cynicism.com, http://www.cynicism.com/~derek

Re: RAID-6 support in kernel?

[Roy Sigurd Karlsbakk]

> You can always fake this effect by combining two 8-disk RAID-5s into a
> RAID-0. It's not technically RAID-6, but can withstand a 2-disk failure,
> although not _any_ 2-disk failure. However, it's my understanding that
> RAID-6 cannot withstand _any_ two disk failure either (see the above
> thread). 

It'll waste 9 drives, giving me a total capacity of 7n instead of 14n. 
And, by definition, RAID-6 _can_ withstand _any_ two-drive failure.

> I also suspect that the use of dual RAID-5s combined with the CPU overhead
> of ATA will kill most systems under any kind of load. For that matter, the
> 2x parity hit from RAID-6 probably wouldn't make you CPU too happy either,
> even if there was a kernel driver that implemented it.

With a 1500MHz Athlon on a typical file server where there's not much 
writes, the CPU is sitting there chrunching RC5-64 som 99,95 % of the 
time. I don't think it'll make much differnce with today's CPUs

roy

-- 
Roy Sigurd Karlsbakk, Datavaktmester

Computers are like air conditioners.
They stop working when you open Windows.

Re: RAID-6 support in kernel?

[Derek Vadala]

On Mon, 3 Jun 2002, Roy Sigurd Karlsbakk wrote:

> It'll waste 9 drives, giving me a total capacity of 7n instead of 14n. 
> And, by definition, RAID-6 _can_ withstand _any_ two-drive failure.

This is certainly not true. 

Combining N RAID-5 into a stripe wastes on N disks. 

If you combine two it wastes 2 disks, etc.

That is, for each RAID-5 you waste a single disk worth of storage for
partiy. I don't know what equation you're using where you get 9 drives
from.

As far as it's ability to withstand _any_ 2-disk failure... I'm not sure
what you mean by definition. RAID-6 implemations don't follow a standard
because there isn't one. Depending on how it's implemented, RAID-6 is not
necessarily able to withstand a filaure of any two disks. We can argue as
much as you want, but I'm not willing to invest the time.

> With a 1500MHz Athlon on a typical file server where there's not much 
> writes, the CPU is sitting there chrunching RC5-64 som 99,95 % of the 
> time. I don't think it'll make much differnce with today's CPUs

It's up to you to decide if the performance trade-off is worthwhile. I
merely trying to point out that system with 2 RAID-5 is likely to incur
the same CPU hit as a single RAID-6, implemented in the kernel. 


---
Derek Vadala, derek@cynicism.com, http://www.cynicism.com/~derek

Re: RAID-6 support in kernel?

[Vojtech Pavlik]

On Mon, Jun 03, 2002 at 02:25:22AM -0700, Derek Vadala wrote:
> On Mon, 3 Jun 2002, Roy Sigurd Karlsbakk wrote:
> 
> > It'll waste 9 drives, giving me a total capacity of 7n instead of 14n. 
> > And, by definition, RAID-6 _can_ withstand _any_ two-drive failure.
> 
> This is certainly not true. 
> 
> Combining N RAID-5 into a stripe wastes on N disks. 
> 
> If you combine two it wastes 2 disks, etc.
> 
> That is, for each RAID-5 you waste a single disk worth of storage for
> partiy. I don't know what equation you're using where you get 9 drives
> from.

He was thinking "mirror", not "stripe". Mirror of 2 RAID-5 arrays (would
be probably called RAID-15 (when there is a RAID-10 for mirrored stripe
arrays)), can withstand any two disks failing anytime. Even more for
certain combinations. But it is terribly inefficient.

> As far as it's ability to withstand _any_ 2-disk failure... I'm not sure
> what you mean by definition. RAID-6 implemations don't follow a standard
> because there isn't one. Depending on how it's implemented, RAID-6 is not
> necessarily able to withstand a filaure of any two disks. We can argue as
> much as you want, but I'm not willing to invest the time.
> 
> > With a 1500MHz Athlon on a typical file server where there's not much 
> > writes, the CPU is sitting there chrunching RC5-64 som 99,95 % of the 
> > time. I don't think it'll make much differnce with today's CPUs
> 
> It's up to you to decide if the performance trade-off is worthwhile. I
> merely trying to point out that system with 2 RAID-5 is likely to incur
> the same CPU hit as a single RAID-6, implemented in the kernel. 

-- 
Vojtech Pavlik
SuSE Labs

Re: RAID-6 support in kernel?

[Kasper Dupont]

Vojtech Pavlik wrote:
> 
> He was thinking "mirror", not "stripe". Mirror of 2 RAID-5 arrays (would
> be probably called RAID-15 (when there is a RAID-10 for mirrored stripe
> arrays)), can withstand any two disks failing anytime.

It can actually withstand any *three* disks failing anytime.

> Even more for
> certain combinations. But it is terribly inefficient.

-- 
Kasper Dupont -- der bruger for meget tid på usenet.
For sending spam use mailto:razor-report@daimi.au.dk
-
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

Re: RAID-6 support in kernel?

[Vojtech Pavlik]

On Mon, Jun 03, 2002 at 04:52:16PM +0200, Kasper Dupont wrote:

> > He was thinking "mirror", not "stripe". Mirror of 2 RAID-5 arrays (would
> > be probably called RAID-15 (when there is a RAID-10 for mirrored stripe
> > arrays)), can withstand any two disks failing anytime.
> 
> It can actually withstand any *three* disks failing anytime.

Yes, you're right.

> > Even more for
> > certain combinations. But it is terribly inefficient.

-- 
Vojtech Pavlik
SuSE Labs

Re: RAID-6 support in kernel?

[Gregory Leblanc]

On Mon, 2002-06-03 at 02:25, Derek Vadala wrote:
        On Mon, 3 Jun 2002, Roy Sigurd Karlsbakk wrote:
        
        > It'll waste 9 drives, giving me a total capacity of 7n instead of 14n. 
        > And, by definition, RAID-6 _can_ withstand _any_ two-drive failure.
        
        This is certainly not true. 
        
        Combining N RAID-5 into a stripe wastes on N disks. 

Hot spares are quite a nice way to increase the reliability of your
arrays, somewhat.  You can still be in trouble if a second disk fails
before the resync finishes, but at that point you're probably talking
about something of a more catastrophic failure, perhaps outside of the
machine itself.  
        

        > With a 1500MHz Athlon on a typical file server where there's not much 
        > writes, the CPU is sitting there chrunching RC5-64 som 99,95 % of the 
        > time. I don't think it'll make much differnce with today's CPUs
        
        It's up to you to decide if the performance trade-off is worthwhile. I
        merely trying to point out that system with 2 RAID-5 is likely to incur
        the same CPU hit as a single RAID-6, implemented in the kernel. 

The issue isn't so much CPU load, but latency.  I'm too lazy to go read
a summary on RAID 6, but with RAID 5, blocks to be written as part of a
stripe often need to be read from the disk in order to generate the
parity.  Parity calculations are pretty trivial on modern CPUs, but disk
latency certainly isn't.  HTH,
	Greg

-- 
Portland, Oregon, USA.
Please don't copy me on replies to the list.

Re: RAID-6 support in kernel?

[Ross Vandegrift]

>         > It'll waste 9 drives, giving me a total capacity of 7n instead of 14n. 
>         > And, by definition, RAID-6 _can_ withstand _any_ two-drive failure.
>         
>         This is certainly not true. 
>         
>         Combining N RAID-5 into a stripe wastes on N disks. 
> 
> Hot spares are quite a nice way to increase the reliability of your
> arrays, somewhat.  You can still be in trouble if a second disk fails
> before the resync finishes, but at that point you're probably talking
> about something of a more catastrophic failure, perhaps outside of the
> machine itself.  

This could become a lot less of an issue.  I recall Neil Brown recently
mentioning that he was thinking about journalling RAID code.  This would
do away with long resyncs much like journalling filesystems did away
with long fscks.  Obviously, I'm not sure if it's something he'll
decide to do or not, but it would really increase the viability of hot
spares.

Ross Vandegrift
ross@willow.seitz.com

Re: RAID-6 support in kernel?

[Roy Sigurd Karlsbakk]

On Monday 03 June 2002 16:52, Kasper Dupont wrote:
> Vojtech Pavlik wrote:
> > He was thinking "mirror", not "stripe". Mirror of 2 RAID-5 arrays (would
> > be probably called RAID-15 (when there is a RAID-10 for mirrored stripe
> > arrays)), can withstand any two disks failing anytime.
>
> It can actually withstand any *three* disks failing anytime.

still - I don't want to waste that money

-- 
Roy Sigurd Karlsbakk, Datavaktmester

Computers are like air conditioners.
They stop working when you open Windows.

Re: RAID-6 support in kernel?

[Pavel Machek]

Hi!

> > > It'll waste 9 drives, giving me a total capacity of 7n instead of 14n. 
> > > And, by definition, RAID-6 _can_ withstand _any_ two-drive failure.
> > 
> > This is certainly not true. 
> > 
> > Combining N RAID-5 into a stripe wastes on N disks. 
> > 
> > If you combine two it wastes 2 disks, etc.
> > 
> > That is, for each RAID-5 you waste a single disk worth of storage for
> > partiy. I don't know what equation you're using where you get 9 drives
> > from.
> 
> He was thinking "mirror", not "stripe". Mirror of 2 RAID-5 arrays (would
> be probably called RAID-15 (when there is a RAID-10 for mirrored stripe
> arrays)), can withstand any two disks failing anytime. Even more for

RAID-1 over two RAID-5s should withstand any three failures, AFAICS.

You could do RAID-5 over RAID-5. That should survive any 2 failures and
still be reasonably efficient.
								Pavel
-- 
Philips Velo 1: 1"x4"x8", 300gram, 60, 12MB, 40bogomips, linux, mutt,
details at http://atrey.karlin.mff.cuni.cz/~pavel/velo/index.html.

Re: RAID-6 support in kernel?

[Pavel Machek]

Hi!

> > > It'll waste 9 drives, giving me a total capacity of 7n instead of 14n. 
> > > And, by definition, RAID-6 _can_ withstand _any_ two-drive failure.
> > 
> > This is certainly not true. 
> > 
> > Combining N RAID-5 into a stripe wastes on N disks. 
> > 
> > If you combine two it wastes 2 disks, etc.
> > 
> > That is, for each RAID-5 you waste a single disk worth of storage for
> > partiy. I don't know what equation you're using where you get 9 drives
> > from.
> 
> He was thinking "mirror", not "stripe". Mirror of 2 RAID-5 arrays (would
> be probably called RAID-15 (when there is a RAID-10 for mirrored stripe
> arrays)), can withstand any two disks failing anytime. Even more for

RAID-1 over two RAID-5s should withstand any three failures, AFAICS.

You could do RAID-5 over RAID-5. That should survive any 2 failures and
still be reasonably efficient.
								Pavel
-- 
Philips Velo 1: 1"x4"x8", 300gram, 60, 12MB, 40bogomips, linux, mutt,
details at http://atrey.karlin.mff.cuni.cz/~pavel/velo/index.html.

Re: RAID-6 support in kernel?

[Bill Davidsen]

On Sun, 2 Jun 2002, Derek Vadala wrote:

> You can always fake this effect by combining two 8-disk RAID-5s into a
> RAID-0. It's not technically RAID-6, but can withstand a 2-disk failure,
> although not _any_ 2-disk failure. However, it's my understanding that
> RAID-6 cannot withstand _any_ two disk failure either (see the above
> thread). 

I think (hope) you meant 1+5, which will stand any three disk failure, and
up to 1+N/2 if just the right drives fail. They never do, of course.
 
> I also suspect that the use of dual RAID-5s combined with the CPU overhead
> of ATA will kill most systems under any kind of load. For that matter, the
> 2x parity hit from RAID-6 probably wouldn't make you CPU too happy either,
> even if there was a kernel driver that implemented it.

I doubt it. Unless you run a system with heavy CPU demand there are lots
of cycles for this stuff. I run 0+1 several places and I don't see serious
CPU load. I would be very interested in RAID-6 in the kernel, but I have
the feeling that RAID-6 means diferent things to diferent people, judging
from posts here and articles online. I haven't found the performance info
you, I assume I will.

-- 
bill davidsen <davidsen@tmr.com>
  CTO, TMR Associates, Inc
Doing interesting things with little computers since 1979.

Re: RAID-6 support in kernel?

[Jakob Østergaard]

On Mon, Jun 03, 2002 at 10:33:24AM -0700, Gregory Leblanc wrote:
> On Mon, 2002-06-03 at 02:25, Derek Vadala wrote:
>         On Mon, 3 Jun 2002, Roy Sigurd Karlsbakk wrote:
>         
>         > It'll waste 9 drives, giving me a total capacity of 7n instead of 14n. 
>         > And, by definition, RAID-6 _can_ withstand _any_ two-drive failure.
>         
>         This is certainly not true. 
>         
>         Combining N RAID-5 into a stripe wastes on N disks. 
> 
> Hot spares are quite a nice way to increase the reliability of your
> arrays, somewhat.  You can still be in trouble if a second disk fails
> before the resync finishes, but at that point you're probably talking
> about something of a more catastrophic failure, perhaps outside of the
> machine itself.  

What often happens (in my experience) is, that a number of disks build up bad
blocks.  One day, you hit one of those bad blocks, and that one disk is kicked
from the array.

When you re-sync, you *will* hit the remaining bad blocks on the other disks,
causing the array to fail completely.

Using hot-spares will "automate" this failure - meaning that an administrator
may not be anywhere near the system when this total failure happens.

Not using hot-spares is less "automatic" in the lucky case where everything
works, but it also assures that an administrator actually is near the system
when the total failure is likely to occur.

-- 
................................................................
:   jakob@unthought.net   : And I see the elder races,         :
:.........................: putrid forms of man                :
:   Jakob Østergaard      : See him rise and claim the earth,  :
:        OZ9ABN           : his downfall is at hand.           :
:.........................:............{Konkhra}...............:
-
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

Re: RAID-6 support in kernel?

[Kasper Dupont]

Pavel Machek wrote:
> 
> Hi!
> 
> > > > It'll waste 9 drives, giving me a total capacity of 7n instead of 14n.
> > > > And, by definition, RAID-6 _can_ withstand _any_ two-drive failure.
> > >
> > > This is certainly not true.
> > >
> > > Combining N RAID-5 into a stripe wastes on N disks.
> > >
> > > If you combine two it wastes 2 disks, etc.
> > >
> > > That is, for each RAID-5 you waste a single disk worth of storage for
> > > partiy. I don't know what equation you're using where you get 9 drives
> > > from.
> >
> > He was thinking "mirror", not "stripe". Mirror of 2 RAID-5 arrays (would
> > be probably called RAID-15 (when there is a RAID-10 for mirrored stripe
> > arrays)), can withstand any two disks failing anytime. Even more for
> 
> RAID-1 over two RAID-5s should withstand any three failures, AFAICS.
> 
> You could do RAID-5 over RAID-5. That should survive any 2 failures and
> still be reasonably efficient.

It will actually survive any 3 disk failures. It is reasonable
if you have a lot of disks. It requires at least 9 disks and
I would prefere at least 25 disks.

RAID-4 and RAID-5 are very similar. And it happens to be the
case that if you only use two disks RAID-1, RAID-4, and RAID-5
are all identical. And each of them can survive a single disk
failure.

Any two of these RAIDs on top of each other can survive three
disk failures. That is true because it takes four disk failures
to loose data. On the upper most RAID you must loose two of the
lower level RAIDs, each of these two must have lost two disks.

RAID-4 on top of RAID-4 is actually just a two-dimentional
parity. RAID-5 on top of RAID-5 is very similar. 

-- 
Kasper Dupont -- der bruger for meget tid på usenet.
For sending spam use mailto:razor-report@daimi.au.dk
-
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

Re: RAID-6 support in kernel?

[Luca Berra]

On Tue, Jun 04, 2002 at 10:20:53PM +0200, Jakob Østergaard wrote:
> What often happens (in my experience) is, that a number of disks build up bad
> blocks.  One day, you hit one of those bad blocks, and that one disk is kicked
> from the array.
> 
> When you re-sync, you *will* hit the remaining bad blocks on the other disks,
> causing the array to fail completely.
> 
> Using hot-spares will "automate" this failure - meaning that an administrator
> may not be anywhere near the system when this total failure happens.
> 
> Not using hot-spares is less "automatic" in the lucky case where everything
> works, but it also assures that an administrator actually is near the system
> when the total failure is likely to occur.
well, what we could do to prevent this. if you don't have or trust S.M.A.R.T.
is having a 'consistency check' function in md, that would read from all disks
and even compare data or calculate parity for raid5, it could be scheduled
to run periodically with a very very low priority.

L.

-- 
Luca Berra -- bluca@comedia.it
        Communication Media & Services S.r.l.
 /"\
 \ /     ASCII RIBBON CAMPAIGN
  X        AGAINST HTML MAIL
 / \
-
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

Re: RAID-6 support in kernel?

[Roy Sigurd Karlsbakk]

> RAID-1 over two RAID-5s should withstand any three failures, AFAICS.
>
> You could do RAID-5 over RAID-5. That should survive any 2 failures and
> still be reasonably efficient.

Sure, but still: Is there a good reason why RAID-6 shouldn't be implemented 
in linux?

-- 
Roy Sigurd Karlsbakk, Datavaktmester

Computers are like air conditioners.
They stop working when you open Windows.

Re: RAID-6 support in kernel?

[Roy Sigurd Karlsbakk]

> RAID-4 on top of RAID-4 is actually just a two-dimentional
> parity. RAID-5 on top of RAID-5 is very similar.

er. RAID-4 is as RAID-5, but with dedicated parity disk.
RAID-6 is with two-dimentional parity

-- 
Roy Sigurd Karlsbakk, Datavaktmester

Computers are like air conditioners.
They stop working when you open Windows.

Re: RAID-6 support in kernel?

[Jakob Østergaard]

On Wed, Jun 05, 2002 at 09:57:01AM +0200, Luca Berra wrote:
> On Tue, Jun 04, 2002 at 10:20:53PM +0200, Jakob Østergaard wrote:
> > What often happens (in my experience) is, that a number of disks build up bad
> > blocks.  One day, you hit one of those bad blocks, and that one disk is kicked
> > from the array.
> > 
> > When you re-sync, you *will* hit the remaining bad blocks on the other disks,
> > causing the array to fail completely.
> > 
> > Using hot-spares will "automate" this failure - meaning that an administrator
> > may not be anywhere near the system when this total failure happens.
> > 
> > Not using hot-spares is less "automatic" in the lucky case where everything
> > works, but it also assures that an administrator actually is near the system
> > when the total failure is likely to occur.
> well, what we could do to prevent this. if you don't have or trust S.M.A.R.T.

I have and use SMART, and because of my experiences with it I do not trust it  :)

> is having a 'consistency check' function in md, that would read from all disks
> and even compare data or calculate parity for raid5, it could be scheduled
> to run periodically with a very very low priority.

This definitely belongs in user-space.  But you are right, and I've discussed
it with colleagues before as well - it would be a good thing to have such a
tool.

In root's crontab you could easily just put a 'dd if=/dev/hda of=/dev/null' and
so on for all your drives.   No need for kernel extensions here...

-- 
................................................................
:   jakob@unthought.net   : And I see the elder races,         :
:.........................: putrid forms of man                :
:   Jakob Østergaard      : See him rise and claim the earth,  :
:        OZ9ABN           : his downfall is at hand.           :
:.........................:............{Konkhra}...............:
-
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

Re: RAID-6 support in kernel?

[Luca Berra]

On Wed, Jun 05, 2002 at 12:53:41PM +0200, Jakob Østergaard wrote:
> > is having a 'consistency check' function in md, that would read from all disks
> > and even compare data or calculate parity for raid5, it could be scheduled
> > to run periodically with a very very low priority.
> 
> This definitely belongs in user-space.  But you are right, and I've discussed
> it with colleagues before as well - it would be a good thing to have such a
> tool.
> 
> In root's crontab you could easily just put a 'dd if=/dev/hda of=/dev/null' and
> so on for all your drives.   No need for kernel extensions here...
> 
well, yes and no
a kernel thread could do compare and check parity and also
when we have the support for this in md try to do a write and
see if the disk error correction code can relocate the bad sector.
also a 'dd if=/dev/hda of=/dev/null' would just sit a production
machine, while we could give it a minumum and maximum bandwidth
to use like we do in reconstruction.

L.


-- 
Luca Berra -- bluca@comedia.it
        Communication Media & Services S.r.l.
 /"\
 \ /     ASCII RIBBON CAMPAIGN
  X        AGAINST HTML MAIL
 / \
-
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

background scanning for media defects (was Re: RAID-6 support in kernel?)

[Friedrich Lobenstock]

Luca Berra wrote:
> well, yes and no
> a kernel thread could do compare and check parity and also
> when we have the support for this in md try to do a write and
> see if the disk error correction code can relocate the bad sector.
> also a 'dd if=/dev/hda of=/dev/null' would just sit a production
> machine, while we could give it a minumum and maximum bandwidth
> to use like we do in reconstruction.

The 3Ware IDE-Raid controllers have that kind of background scanning
functionality if their diskmanager daemon is running. I guess it
just sends the controller chip the start scanning command, which then
does its job.

Would be good if linux md had that too.

-- 
MfG / Regards
Friedrich Lobenstock

Re: RAID-6 support in kernel?

[Derek Vadala]

On Tue, 4 Jun 2002, Bill Davidsen wrote:

> On Sun, 2 Jun 2002, Derek Vadala wrote:
> 
> > You can always fake this effect by combining two 8-disk RAID-5s into a
> > RAID-0. It's not technically RAID-6, but can withstand a 2-disk failure,
> > although not _any_ 2-disk failure.
> 
> I think (hope) you meant 1+5, which will stand any three disk failure, and
> up to 1+N/2 if just the right drives fail. They never do, of course.

I did mean RAID-0 combined with RAID-5. You can search for RAID-50 for
more info. The configuration you describe (RAID-5s combined into a mirror)
would have a disk overhead that is worse than RAID-10/RAID-0+1. For two
5-disk RAID-5s combined into a RAID-1 you end up using six of your disks
for parity and disk mirroring:

  RAID-1 --------> RAID-5 (D0,D1,D2,D3,P0)
              |--> RAID-5 (D0,D1,D2,D3,P0)
   (four disks used for data, only one from each RAID-5 can fail)

With RAID-10:

  RAID-0 --------> RAID-1 (D0,D0)
              |--> RAID-1 (D1,D1)
              |--> RAID-1 (D2,D2)
              |--> RAID-1 (D3,D3)
              |--> RAID-1 (D4,D4)
   (five disks used for data, one from each mirror can fail)

With RAID-50:

  RAID-0 --------> RAID-5 (D0,D2,D4,D6,P0)
              |--> RAID-5 (D1,D3,D5,D7,P0)

   (two disks wasted only one from each RAID-5 can fail)

I believe that I/O performance would be similar for each
configuration. I'll try to run some tests in the next few days.

> I doubt it. Unless you run a system with heavy CPU demand there are lots
> of cycles for this stuff. I run 0+1 several places and I don't see serious
> CPU load. I would be very interested in RAID-6 in the kernel, but I have

Mirroing doesnt hit the CPU nearly as much as RAID-5 does. I suspect
RAID-6 would incur greater overhead because of its double parity blocks.
But, there's no point in arguing about kernel RAID-6 without data to back
it up.

---
Derek Vadala, derek@cynicism.com, http://www.cynicism.com/~derek

Re: RAID-6 support in kernel?

[Kasper Dupont]

Derek Vadala wrote:
> 
>   RAID-1 --------> RAID-5 (D0,D1,D2,D3,P0)
>               |--> RAID-5 (D0,D1,D2,D3,P0)
>    (four disks used for data, only one from each RAID-5 can fail)

Wrong, any three disks can fail. If the one RAID has only
one faulty disk, the other RAID can have any number of
faulty disks without loosing data.

> 
> With RAID-10:
> 
>   RAID-0 --------> RAID-1 (D0,D0)
>               |--> RAID-1 (D1,D1)
>               |--> RAID-1 (D2,D2)
>               |--> RAID-1 (D3,D3)
>               |--> RAID-1 (D4,D4)
>    (five disks used for data, one from each mirror can fail)
> 
> With RAID-50:
> 
>   RAID-0 --------> RAID-5 (D0,D2,D4,D6,P0)
>               |--> RAID-5 (D1,D3,D5,D7,P0)
> 
>    (two disks wasted only one from each RAID-5 can fail)
> 
> I believe that I/O performance would be similar for each
> configuration. I'll try to run some tests in the next few days.

I'd guess that depends on the access patterns.

-- 
Kasper Dupont -- der bruger for meget tid på usenet.
For sending spam use mailto:razor-report@daimi.au.dk
-
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

Re: RAID-6 support in kernel?

[Helge Hafting]

Kasper Dupont wrote:
> 
> Derek Vadala wrote:
> >
> >   RAID-1 --------> RAID-5 (D0,D1,D2,D3,P0)
> >               |--> RAID-5 (D0,D1,D2,D3,P0)
> >    (four disks used for data, only one from each RAID-5 can fail)
> 
> Wrong, any three disks can fail. If the one RAID has only
> one faulty disk, the other RAID can have any number of
> faulty disks without loosing data.
> 
This is a bit excessive, you waste more than half your disks
for 3-disk safety.  Consider raid-5 on top of raid-5.

You're still safe from any-3 failure, and the overhead
in percent can be made arbitrarily small. 
Using nxm disks, you get (n-1)(m-1) data disks, and
(n+m-1) parity disks.
(The overhead (n+m-1)/((n-1)(m-1)) approach
0 as noth n and m grows towards infinity.

Using many few-disk arrays to build the big array
gives good chances to survive 4-disk failures too,
as few of the many 4-disk combinations take out 
two small arrays simultaneously.

I'm not sure about the write performance for such
a beast, but it should be fine for reading.  I.e.
a safe archive.

Helge Hafting