At Fri, 23 Feb 2007 08:51:36 -0800,
jonathan@dsg.stanford.edu wrote:
> 
> 
> I think we *need* to do that, to continue good performance on machines
> where the full switch is prohibitive.  If we collectively buy that,
> then that decision drives us into a "top half" / "bottom half" model.

Indeed.

On a perhaps slightly related note, and at the risk of comparing apples
and oranges, I was trying FreeBSD-6.2 on my PE2650 to see if it had any
better luck seeing both the built-in PERC and the add-on HP NetRAID
controllers I have on it.  I'm not sure what the results were there yet,
as I had a power failure in the middle of my test last night.  However I
did have time to observe that on a plain raw read of the ld0d device the
FreeBSD aac(4) driver achieved a rate nearly twice that of the NetBSD
driver.  The FreeBSD driver is, if I understand correctly, split into a
top/bottom model already, with the top half running in a kernel thread.

-- 
						Greg A. Woods

H:+1 416 218-0098 W:+1 416 489-5852 x122 VE3TCP RoboHack <woods@robohack.ca>
Planix, Inc. <woods@planix.com>       Secrets of the Weird <woods@weird.com>