[prev in list] [next in list] [prev in thread] [next in thread] 

List:       evms-devel
Subject:    Re: [Evms-devel] Hardsector sizes
From:       Rik van Riel <riel () conectiva ! com ! br>
Date:       2001-07-23 16:25:53
[Download RAW message or body]

On Sat, 21 Jul 2001, Greg Lehey wrote:
> On Friday, 20 July 2001 at 13:38:55 +1000, Andrew Clausen wrote:
> > On Fri, Jul 20, 2001 at 11:40:44AM +0930, Greg Lehey wrote:
> >>> So what?  The CPU may be in demand for other things.
> >>
> >> So this doesn't happen very often.  If you're performing, say, 5000
> >> transfers a second, that might translate to 10,000 divisions.  On even
> >> the slowest ia32, this doesn't make it out of the noise.
> >
> > * clausen wouldn't know.  Rik seems to disagree ;)
>
> I suppose I should look at the code before I say too much more.  I'd
> be interested in Rik's opinion.

The divides are probably lost in the noise, except on
Alpha CPUs where the CPU doesn't have an instruction
for integer divide and a software library takes care
of divisions, this could give some icache pollution and
cache misses ;)

(then again, even that may be lost in the noise)

> > BTW: it's not "per transfer"... presumebly you need the math if you
> > access the (buffer) cache too.

Well, since you're already dealing with lookup overhead,
system call overhead, etc. I guess a divide is lost in
the noise even in this scenario.

Remember that even a divide cannot be as expensive as a
cache miss. People will be reading the data in larger
chunks.


> Seriously, there are so many more significant ways to improve
> performance than to eliminate division.

Couldn't agree more.  As Rusty's signature read:

"Premature optmiztn s rot of all evl"

cheers,

Rik
--
Executive summary of a recent Microsoft press release:
   "we are concerned about the GNU General Public License (GPL)"


		http://www.surriel.com/
http://www.conectiva.com/	http://distro.conectiva.com/

[prev in list] [next in list] [prev in thread] [next in thread]