classiccmp server hardware

Rob O'Donnell classiccmp.org at irrelevant.fsnet.co.uk
Mon Oct 11 13:43:27 CDT 2004


At 16:12 11/10/2004, Fred N. van Kempen wrote:

>On Mon, 11 Oct 2004, Jochen Kunz wrote:
>
>
> > It depends on the load that the server already has and how much overhead
> > software RAID1 causes. Typically RAID1 doesn't cost that much CPU. To
> > the past - present comparision: CPUs got a _lot_ more faster then disk
> > dirves in the past 20 years. See the thoughts behind the BSD LFS.
>I have never trusted OS-level disk redundancy schemes, and never
>will.  Promise controllers have strange ways of doing things, and
>most certainly do require OS-level software assistance.  Disks
>attached to a Promise RAID1 controller can be taken off the Promise,
>and attached to a regular IDE port, and will work as-is.  Typical
>RAID controllers that do hardware-supported RAID store state and
>config info on the attached drives, so they usually claim a few
>sectors for that.


In my experience, running servers purely as a hobby, I've found the version 
of RAID1 as presented by cheap ATA controllers on FreeBSD more than 
adequate, a considerable speed up than a single drive, and definitely 
better than not having a raid, when a drive goes down!!

I've used various cards (mostly SiL 0680 chipset) on FreeBSD without 
problem.  At least on the 0680, they do write something to the disc, as 
taking the drives out of one machine and putting them on a different card 
in a different machine caused them to be spotted as a RAID pair before the 
OS had even had a chance to boot.  Drives can be removed, and they do work 
on their own however too, so I don't know what gets stored where?

You can mirror on FreeBSD using either vimum, which is pure software raid 
and will work across any drives no matter what types or controllers they 
are on (even a mix) or for most ATA IDE RAID cards, the ata driver has 
built in support (`man atacontrol`) which I presume negotiates with the 
card bios and sorts out who does what.

Oh, as far as the SiL0680 specifically is concerned, from personal 
experience FreeBSD 4.10 works fine... 5.2 works, but 5.2.1 panics on boot 
(even booting the install).  Later snapshots are OK again though, but for 
the moment 4.10 is still recommended for production servers, at least until 
5.3 is out, which is apparently only a couple of weeks off.  5.3 apparently 
offers some new facilities (geom_vinum and gmirror) but I've not used them, 
or anything later than 5.2 yet, just seen them discussed while sorting out 
beta bugs on the freebsd-current list.


Rob.






More information about the cctalk mailing list