Subject: Re: motherboard recommendations?
To: Curt Sampson <cjs@cynic.net>
From: David Laight <david@l8s.co.uk>
List: netbsd-users
Date: 08/07/2002 11:58:24
On Wed, Aug 07, 2002 at 06:21:54PM +0900, Curt Sampson wrote:
> On Tue, 6 Aug 2002, Wojciech Puchar wrote:
> 
> > but using master/slave config is bad idea and will slow down everything.
> 
> This is not necessarially true. Testing with modern 7200 RPM IBM IDE
> drives on a UDMA-100 controller, I found that sequential reads and
> writes were basically unaffected. I.e., with one drive I got about 30
> MB/sec; with two drives on the same controller I got about 30 MB/sec
> from each, when being read or written simultaneously.
> 
> What *was* affected was the number of disk transactions I could get
> per second; with one disk or two, I could do no more than about 90
> transactions per second. So if you do a lot of small reads and writes,
> this might be a problem; if you're not sustaining a high transaction
> rate, it may make no difference.

IIRC the 'problem' with the master/slave scheme is that there is
only 1 interrupt line.  This means that the software can't (easily)
be waiting for either driver to finish an action [1].

However doing a request on the slave should be no slower than
doing one on the master - you just can't (easily) interleave requests.
This means you don't really get a speed benifit from a raid.

	David

[1] It struck me while writing this that you could get h/w to
drive a square wave at (say) 1MHz onto the master/slave select
line....

-- 
David Laight: david@l8s.co.uk