3 disk raid-5 without parity

3 disk raid-5 without parity

[Jurriaan]

I am trying to convince my boss our new database-server wants raid-0+1,
not raid-5, and I got an idea while reading endless articles about
raid-5 being slow when writing and management not listening.

suppose you make a 3-disc raid-5 without parity:

data    disc1 disc2 disc3
A       A     A     B
B       B     C     C

How would that perform compared to raid-5 and raid-0+1?

As I understand, the performance problem with raid-5 when writing is
that you may need to read old data to recompute the parity block, and
the write the parity block and the data in parallel.

So in this case, you can read straight away from all 3 disks (with
raid-5 one of the disks will have parity information) and you can write
without reading old data.

There is a problem with extending this: you need groups of 3 disks.
Then again, compared to raid-0+1 you need fewer disks.

Has this ever been implemented? Even better, benchmarked?

Just in case:
[If this is a genuinly new idea, I hereby place it under the GPL.]

Jurriaan
-- 
"And in the future, if we come across any other water holes, I think
we'll drop a grenade down it first, and check the quality of the
water afterwards."
	Simon R Green - Hellworld
Debian (Unstable) GNU/Linux 2.6.7-rc2-mm2 2x6078 bogomips load 2.97

Re: 3 disk raid-5 without parity

[Brad Campbell]

Jurriaan wrote:
> I am trying to convince my boss our new database-server wants raid-0+1,
> not raid-5, and I got an idea while reading endless articles about
> raid-5 being slow when writing and management not listening.
> 
> suppose you make a 3-disc raid-5 without parity:
> 
> data    disc1 disc2 disc3
> A       A     A     B
> B       B     C     C
> 
> How would that perform compared to raid-5 and raid-0+1?
> 
> As I understand, the performance problem with raid-5 when writing is
> that you may need to read old data to recompute the parity block, and
> the write the parity block and the data in parallel.
> 
> So in this case, you can read straight away from all 3 disks (with
> raid-5 one of the disks will have parity information) and you can write
> without reading old data.

What you propose is not quite raid-5, but appears to be a striping hybrid between raid-0 and raid-1.
If I get what you are saying, you are writing each block twice, and ensuring each block is written 
to at least 2 different drives. Interesting idea, but not really efficient. In addition, you are 
doubling your write bandwidth requirements (not unlike pure raid-1).

Yes, raid-5 random write performance is limited by read-modify-write cycles when they are required. 
The raid-5 md driver does do a pretty good job of avoiding this where possible though.

I have not really benchmarked my raid-5 write performance, but on reads I sit around 90MB/s across 
my 10 drives. (Even my SATA raid-0 with 2 7200RPM drives has trouble getting much quicker than this 
as I keep saturating the PCI bus)

Regards,
Brad

Re: 3 disk raid-5 without parity

[Neil Brown]

On Monday June 14, thunder7@xs4all.nl wrote:
> I am trying to convince my boss our new database-server wants raid-0+1,
> not raid-5, and I got an idea while reading endless articles about
> raid-5 being slow when writing and management not listening.

You want numbers, not abstract arguments.

Configure your server with raid5 and do some performance measurements
- preferably with your database suite.
Then reconfigure with raid 0+1 and test again.

Show the numbers to your boss.  You get to choose which numbers to
show :-)

> 
> suppose you make a 3-disc raid-5 without parity:
> 
> data    disc1 disc2 disc3
> A       A     A     B
> B       B     C     C
> 
> How would that perform compared to raid-5 and raid-0+1?

Should be slightly better than raid1 of 2 drives, and slower than raid
0+1 on 4 drives.  How it compares with raid-5 depends largely on load
characteristics.  With only 3 drives, some loads will very often
provide raid5 with both data blocks in a stripe, and so no pre-reading
will be needed.

I hope to release a "raid10" module for 2.6 within a couple of weeks.
raid10 is basically a combination of raid1 and raid0 all in one module
with some interesting geometry possibilities.  This particular
geometry is one of the possibilities.

> 
> There is a problem with extending this: you need groups of 3 disks.
> Then again, compared to raid-0+1 you need fewer disks.

There is no problem extending this. 
With 4 discs it would be
  A A B B
  C C D D
which is a lot like raid0 over raid1
With 5 disks is would be
  A A B B C
  C D D E E
or, if you wanted more redundancy:
  A A A B B
  B C C C D
  D D E E E

The more discs you have, the faster it should be.

> 
> Has this ever been implemented? Even better, benchmarked?

Implemented:  I'm fairly sure it has.  Not in Linux just yet.
Benchmarks: I don't know of any details.

NeilBrown

Re: 3 disk raid-5 without parity

[Jurriaan]

From: Neil Brown <neilb@cse.unsw.edu.au>
Date: Mon, Jun 14, 2004 at 09:45:21PM +1000
> On Monday June 14, thunder7@xs4all.nl wrote:
> > I am trying to convince my boss our new database-server wants raid-0+1,
> > not raid-5, and I got an idea while reading endless articles about
> > raid-5 being slow when writing and management not listening.
> 
> You want numbers, not abstract arguments.
> 
> Configure your server with raid5 and do some performance measurements
> - preferably with your database suite.
> Then reconfigure with raid 0+1 and test again.
> 
> Show the numbers to your boss.  You get to choose which numbers to
> show :-)

yes - unfortunately, corporate IT gets to decide which hardware to test
on. Guess what? They just got a shiny new SAN with 140 Gb disks,
configured in raid-5. Not quite the raid-0+1 over 36 Gb disks I had in
mind :-(
Anyway, that was just an introduction - nu use regurgling that on the
list.

> I hope to release a "raid10" module for 2.6 within a couple of weeks.
> raid10 is basically a combination of raid1 and raid0 all in one module
> with some interesting geometry possibilities.  This particular
> geometry is one of the possibilities.
> 
I'm looking forward to that!

Kind regards,
Jurriaan
-- 
Anyone can build a fast processor. The trick is to build a fast system.
	Seymour Cray
Debian (Unstable) GNU/Linux 2.6.7-rc2-mm2 2x6078 bogomips load 2.49