Subject Re: [IB-Architect] Disk Bandwidth was License Question
Author Jan Mikkelsen
>From: Jim Starkey <jas@...>
>At 11:40 AM 3/26/00 +1000, you wrote:
>>From: "Jan Mikkelsen" <janm@...>
>>There is a RAID kit for Linux which supports striping, mirroring and
>>in software. It isn't perfect (mirroring doesn't do proper read

>>for example), but the work is being done. We've been using various
>>for about 18 months.
>Good. Excellent things, RAIDs. But golly, machines are so fast
>and cheap now, wouldn't it be nice to move the intelligence to
>the OS for a lower cost high performance solution? Smart disk
>controllers are always going to have a tighter integration with
>the drivers and will probably always be the high end solution.
>But striping and mirroring can (quote in theory unquote) be done
>by the host OS. Sure would be nice. On the other hand, the big
>win in controller based mirroring is the ability to direct a read
>to a drive not in use or a drive with the arm on the target track
>is a screamer.

It is done in the operating system. No smart disk controller required. You
can do mirroring, striping, and RAID-5 with the two IDE channels on a $100
motherboard. You can easily do RAID-5 parity generation two orders of
magnitude faster than the disk speed. Windows NT will also do some RAID off
the shelf (striping only in workstation, mirroring and striping in server).

As you pointed out in another post, CPU cycles are getting really cheap.
The additional cost of putting in a smart controller isn't a win when
compared to just getting a slightly faster processor (or two processors).
And the processors (and memory subsystems) on the cards typically aren't
fast enough to beat the main CPU. Main memory bandwidth is really the big

I've seen mirroring and striping done in the operating systems for a long
time. You also get nice benefits like being able to use independent busses
and independent controllers. Necessary to build real fault tolerance: What
if your hardware RAID controller breaks?

Jan Mikkelsen.