[lug] RAID installation on Fedora 6 Zod

[Nate Duehr]

> On Fri, 11 May 2007 11:20:15 -0400 (EDT)
> steve at badcheese.com wrote:
>
>> By flakey when talking about linux's software raid is when a drive
>> fails. Since the IDE controller (underneath the software raid) isn't
>> raid-aware, if a drive fails, you'll get timeouts and your machine
>> will be either super-slow or hang due to HD I/O.  The IDE controller
>> will give those DMA timeouts and keep trying to read/write to the
>> drive when it's failing. If you pull the bad drive and reboot, then
>> linux will be happy and operate in degraded-mode and be ok, just
>> needs some manual intervention.  That's mainly what I meant by flakey.
>
> Odd. I have never seen this behavior with linux software raid.
>
> When a drive gets an error, linux notices, removes the drive from the
> raid, and disables I/O to that drive and continues in degraded mode.
> I don't think the IDE controller needs to be aware of anything. The
> kernel just simply stops sending any I/O to the dead drive...

I've seen a failure where the IDE bus itself was killed by the dead drive.
 Keeping mirrors (if you're doing RAID 1) on separate controllers is not
only smart, but virtually required when using any form of software RAID.

Personally I think software RAID stinks... but I've been working on
Solaris boxes with both Veritas managing RAID and hardware RAID, and find
most Linux solutions fairly laughable these days... compared with the ease
of replacing a drive in the commercial systems.

>From what I hear, Sun's ZFS is even better... it doesn't even care if
deport the filesystem, remove and scramble all the disks, and put them
back in as ANY device... it stores metadata on the disks themselves, so
you just run the import after reconnecting all the drives and it "figures
it out" for you and gets your data back online automatically.

They have a cute video out there of their German research group demo'ing 
this with 12 x 1GB flash drives and three USB hubs.

Linux needs to catch up to that!

-- 
Nate Duehr, nate at natetech.com