Re: Milestones document

[Havoc Pennington <hp redhat com>]

On Tue, 7 Dec 1999, Jim Gettys wrote:
> 
> Since the early 90's, on RISC systems, and since about 1995 on Intel, your
> code runs faster if it is SMALLER.
> 

A valuable point - and don't I know it, comparing my P333 with 128K cache
to my old P233 with 2M cache!

As Elliot points out though, a common case where we get a size/speed
tradeoff is a cache for something that requires I/O - and I/O is Super
Slow compared to almost anything else. Another common thing you can do to
reduce size is to "compress" data in various ways - not necessarily
gzip-style, but just recalculating some data in some way each time you use
it. This doesn't decrease memory bandwidth because you still have to deal
with the full-size data, you just don't keep it around all the time.
Another nasty tradeoff is abstraction vs. size/speed, where making
something more robust and maintainable can damage your efficiency... no
free lunches in the world.

What it comes down to - a lesson you taught us with respect to an X server
slowdown we were talking about a while back - we need to profile GNOME.
Complicating this task, GNOME is a multi-process composite where a given
user-visible app can be several CORBA servers and the X server, all taking
some CPU time... :-(

Havoc