Subject: Re: Removing "rows" from the RAIDframe driver..
To: Daniel Carosone <dan@geek.com.au>
From: Greg Oster <oster@cs.usask.ca>
List: tech-kern
Date: 12/20/2003 15:46:37
Daniel Carosone writes:
> On Sat, Dec 20, 2003 at 12:17:56PM -0600, Greg Oster wrote:
> >
> > The "row" code in RAIDframe would behave, in theory, much like doing
> > a concatenation ccd of RAID sets.
>
> This can be very very useful from a performance and reliability point
> of view. If I have, say, 28 disks, then making a single raid5 of all
> of them (if I even am allowed to by any implementation) will truly
> suck for any filesystem usage because the read-modify-write cycle is
> horrendous. (Using them as a pure sequential device might be better,
> just). And, of course, I'll need several spares and I'll be highly
> exposed to a second failure while rebuilding to a spare.
Right.
> > I forgot to mention, but had the row functionality actually worked in
> > RAIDframe, we could *still* replace it with a concatenation ccd and
> > regular "1D" (one-row) RAID sets.
>
> There's even an example of this in the manpage (RAID on RAID) using RF
> RAID0 rather than ccd(4).
Not quite... the 2D stuff is a "concatenation" of the RAID sets, not
a stripe across the RAID sets.
> Would there be any other (hypothetical) functionality lost doing this
> vs multi-row? One example that comes to mind would be in dealing with
> spares.
Spares could be one issue, yes.
> RF doesn't (seem to) nicely support the hot-spare-pool
> concept, where a spare device can be used by whichever of several sets
> fails first, at least without manual intervention (assiging the disk
> as spare to that set after failure).
"yet" :)
> If removing rows makes adding shared spares easier, then all the
> better :)
It will. The "extra dimension" of components made handling spares
even more of a hassle.. (never mind that spares were only ever
assigned to the first row of the array!). I'm hoping these changes
will make a lot of things easier to fix. (In fact, I know they will,
cause I'm already using a tree with the changes to start work on
other stuff, and it's been quite nice having the row stuff out of the
way :) )
Later...
Greg Oster