Subject: Re: more on performance of "make build" in -current vs. 1.5
To: Hubert Feyrer <hubert.feyrer@informatik.fh-regensburg.de>
From: Matthew Jacob <mjacob@feral.com>
List: tech-kern
Date: 02/19/2001 10:23:04
> On Mon, 19 Feb 2001, Jason R Thorpe wrote:
> >  > shouldn't this result in *less* CPU use (as no IO needs to be scheduled)?
> > I/O generally doesn't cause CPU time unless it's non-DMA.
> 
> IO itself not, but telling the disk to do IO (setting up registers,
> ...) does as far as I understand. 

Register setup (command instruct) scales linearly with the amount of
I/O, but is fairly lightweight. The big issue is how many command completions
per interrupt you can get for your hardware.

Also- if you take something that requires care && feeding for I/O coherency,
it *really* is painful and requires a faster CPU. For example, the rough
amount of time a Sparc1 takes to VAC flush a 128K transfer is 1.7ms. Yes,
that's 'milliseconds', which is equal to the setup and transfer time of the
that I/O from the SCSI bus and disk itself.

-matt