Tue, 19 Jul 2005 @ 13:01 -0400, Steven M. Bellovin said:

> In message <20050719160023.GA25276@bl.org>, Michael Parson writes:
> 
> > Newer IDE (ATA) is acting more and more like this.
> >Once upon a time, disk I/O with IDE disks was CPU intensive.
> 
> Especially before ATA DMA.  

(S)ATA still uses more CPU than SCSI though, it's just that modern CPUs
are so fast they hide a lot of issues like this.

For example, I still notice I/O stalls at times when the CPU is very
busy with something else, that I don't notice as much with SCSI.

I think it is mostly a factor of SCSI drives having longer and better
queues, and being able to function without bugging the CPU for longer
periods of time.  It seems more pronounced when you have more drives.

The other problem is that some (S)ATA controllers are horrible.

Aside:

I was reading an article the other day on modern servo-positioned SCSI
drives that don't use index data, and how they could make multiple head
movement and read operations in a single rotation.

One example was optimal ordering of a queue of around a dozen reads at
different head positions in a single rotation.  That's a head movement
for each read, all within on rotation.

I still remember manual skew ordering, and being happy when a system
could do that *once* per revolution, and on slow media!




-- 
shannon "AT" widomaker.com -- ["Consulting wouldn't be what it is today
without Microsoft Windows" -- Chris Pinkham]