Hello,

> Most comes from the filesystem I think: par of this is that the FS is
> mounted async by default on linux (you can do so on NetBSD, but softdep
> is a much better solution); part of this is their dynamically sized buffer
> cache (something that UBC should solve, for now you can tweak NBUF/BUFPAGES).
> On my desktop, using softdep and tweaking NBUF/BUFPAGES to use the RAM I
> don't need for VM make NetBSD close to linux.

The newer Linux kernels are starting to move toward using page cache for
more things, to prevent penalizing processes (or giving them
artifical priorities) based upon accesses to inodes and other file
system metadata.  It would be nice to see the buffer cache expanded
(or eliminated) in the *BSD's.

> Now, I can tell you that under load, NetBSD is far better than linux,
> especially when it start swapping.

Lack of physical page aging is the killer here.  2.4-test8+ is doing some
multi-queue page laundering and page aging, but the page aging
is by VMA and not by physical page.  It will probably be awhile
before Linux catches up in this arena.

> > Additionally, they seem to be able to have better performance talking to
> > the graphics cards than we do.
> 
> Which graphic cards ? Maybe it's a X-related issue ?

In general I think driver support is better for Linux.  When I
look to load *BSD on a box, I choose the hardware I know is
best supported.  Whether this will change in the future, I
don't know.

As for raw I/O, that's in Linux 2.4, but not 2.2.  And 2.4
uses a 50ms rather than the standard 100ms time quantum,
which gives the _appearance_ of running faster under load
(but it's an illusion, things are just more responsive when
you're interacting with them)

I'm not interested in joining the "which is better" debate
on OS's.  That's for high school kids with no knowledge
and too much time on their hands.  I use FreeBSD,
NetBSD, and Linux, and I develop software for the Linux
kernel.  I think each OS has its intended audience,
and each does a good job at what it does best. :)

--
Eric Lowe
Software Engineer, Systran Corporation
elowe@systran.com