SATA controllers

[Chan Chung Hang Christopher]

Liam Proven wrote:
> On Sat, Mar 13, 2010 at 3:12 PM, Chan Chung Hang Christopher
> <christopher.chan at bradbury.edu.hk> wrote:
>> Liam Proven wrote:
>>> On Wed, Mar 10, 2010 at 3:19 PM, Dave Howorth
>>> <dhoworth at mrc-lmb.cam.ac.uk> wrote:
>>>> I'm speccing a new machine and as part of it I'd like to have a
>>>> linux-controlled 4-disk RAID 6 array using SATA 3 Gbps disks (aka STA 2).
>>> You want more than 4 disks for RAID6. Seriously.
>>>
>>> RAID6 uses 2 parity drives; this means you get the capacity of (N-2)
>>> drives, where N is the number of drives. Ergo, use 4 drives, you only
>>> get the capacity of 2. This is pointless, because if you lose half the
>>> capacity, you would get /much/ better performance from RAID10 (a
>>> mirror of stripes) or RAID 0+1 (a stripe set of mirror pairs).
>> md raid10 != md raid1+0. You can do a lot of fancy configs with the
>> raid10 module that is not quite the same as nested raid1+0.
> 
> Do tell...? Always keen to learn!

http://en.wikipedia.org/wiki/Non-standard_RAID_levels#Linux_MD_RAID_10

http://neil.brown.name/blog/20040827225440

>>> (I may have got the definitions of 0+1 and 10 transposed, but it's not
>>> really important at this point!)
>>>
>>> RAID levels 0+1 and 10 use simple mirroring and striping, requiring
>>> little CPU, therefore making them very fast, whereas with RAID6 you
>>> impose 2 sets of parity calculations on the system, making writes
>>> slow.
>> How much cpu is used is irrelevant because raid5 has been easily catered
>> for (given no really heavy duty cpu chewing service like
>> spamassassin+amavisd) since the AMD Duron.
> 
> Well, yes, the PC can take it & you won't notice much load, but it
> does negatively impact the write performance of the array, and also
> the rebuild time.

The point is that it is not necessarily because the cpu needs to do the 
calculation.


> 
>> The real performance breaker
>> for raid5/6 is the bus traffic necessary to read from disk before parity
>> calculations can be done and the bus contention. Therefore, hardware
>> raid cards that use any processor better than an Intel i960 with a
>> sufficiently large cache will blow Linux software raid out of the water
>> - sometimes even if the md array is raid1+0.
> 
> Hmm. May depend on the power of the controller. I have such a
> controller & it's *not* quick. I suspect an md RAID would be quicker.

An md raid5/6 array will definitely be slower unless the board is 
underpowered. What board are you talking about?


> But mine is a few years old.

What is it?


> 
>>> This means the *minimum* number of drives for a RAID6 is 5 drives,
>>> which will give you the capacity of 3× a single drive.
>> RAID5 = 3. I do not see why RAID6 has to jump to 5.
> 
>>> I don't know how smart the Linux software RAID system is, but if it
>>> has good rules built in, it won't let you create a RAID 6 out of <5
>>> drives and should outright block 4. It's an invalid config.
>> Outright block 3.
> 
> Well, OK. You do make a valid point about sensitivity to particular
> disks failing in 0+1/10. I still contend that the minimum /sensible/
> number for RAID6 is 5 disks, though.
> 

That I won't argue with. I wonder what you can get using the raid10 
module with 5 disks.../me looking for a guinea pig ;)