Re: To RAID or not to RAID (...or how to RAID)

Ed Stevens <spamdump_at_nospam.noway.nohow> wrote:
>
>Well, I've learned something new today. (Can I go home now!?)
>
>Sounds interesting. Sounds like it would require LOTS of drives. Given that
>each R-5 set would require a minimum of 3 drives, and I presume the R-0 layered
>over that would require 3 R5 sets (certainly would require at least 2) we're
>talking about a minimum of 6 or 9 drives. How many drives worth of capacity is
>'lost' in this configuration? Certainly one out of each R5 set. Any more when
>the R0 is layered on?

None more for the RAID-0 layer, just the one drive from each RAID-5 set is "lost", or rather used to stop the business-critical system from going offline _when_ a failure occurs.

...

Totally coincidently, just after I wrote that the sysadmin came in to tell me that we'd had a disk fail over the weekend, but there wasn't any problem because the array had taken the hot spare online. So, that's probably more "lost" capacity - a hot spare for every RAID-5 array, reducing the double-failure risk, and degraded performance, to a hour-long window.

>Not that I believe that this 'lost' capacity is a concern -- you're buying speed
>and resiliancy, and that seems to me to be a good thing. But I find that a
>tough (no, impossible) sell to the guys that control the hardware purchasing and
>configuration. All they can see is the total capacity they are purchasing, and
>feel that RAID-5 is resiliant enough, even though we just dodged the bullet on
>that one a couple of months ago. And as devices keep getting larger and faster,
>it becomes a tougher sell. They about s**t a brick when I asked for a test
>server to be configured Raid 0+1 and they 'lost' half of the drives.

Ask them to calculate the cost of the system being offline for however many hours a full restore takes? Disks are mechanical devices, they will fail. (N.B. Last time I tried this, it didn't actually change the decision).

I sometimes wish disks were sold by 8kb IO/s, with capacity mentioned somewhere in the small-print. When I spec. a benchmark storage array, it's always first by IO rate.

-- 
Andrew Mobbs - http://www.chiark.greenend.org.uk/~andrewm/