Harware vs. Software RAID (WAS: Re: [nycbug-talk] Fwd: Adaptec AAC raid support)

Jay Savage daggerquill
Mon Mar 21 16:09:18 EST 2005


On Mon, 21 Mar 2005 14:04:54 -0500, Marco Scoffier <marco at metm.org> wrote:
> On Mon, Mar 21, 2005 at 01:15:46PM -0500, Isaac Levy wrote:
> >With that stated, as the dust is settling with this issue- and myself
> >being a bit outside of the daily-use OpenBSD camp, what is everyone
> >going to use for RAID on OpenBSD?  (I'm particularly interested in sata
> >hardware raid...?!?)
> >
> 
> What I wonder is if hardware is really the way to go.  What's the difference in
> overhead btw say a hardware raid5 and putting the card in jbod mode and running
> raidctl.  [ I have not googled this lately ]
> 
> The huge advantages of software raid are :
> 
>   1) a consistent interface across different hardware (useful that one time in
>   three years that a disk needs to be replaced in some server and you don't
>   remember the ctrl-F1 or whatever combo to get into a particular hardware
>   vendor's raid software.)
> 
>   2) tested code, the software raid code in any of the BSDs is probably a lot
>   better tested that whatever is built into the raidcontroller-du-jour.
> 
> I am not an expert, and certainly welcome others opinions, but I was convinced
> by the above arguement about a year ago and have been setting up software raids
> since.  I would sacrifice performance for stability, especially since my bet is
> that the performance hit is not so bad.

It depends on how big you need your RAID to be, among other things,
and what hardware you want to build it with.  On IDE/ATA, you need one
controller per disk or quickly run out of bandwidth.  So your RAID is
effectively limited to the number of PCI slots in your box.  SCSI you
can push a little farther.  One thing harware RAID accomplishes for
many people is to act as an IDE/SCSI bridge, taking advantage of the
SCSI bus between the processor and the RAID controller, and then using
cheaper IDE disks to actually build the RAID.  and although I've never
tried it, I suspect that if you tried to build, say, a multiterabyte
software array with a dozen disks, you'd notice the overhead.




More information about the talk mailing list