On Fri, Dec 19, 2003 at 03:48:04PM -0500, Pascal Renauld wrote:
> That driver despite its *bloated* architecture so well decried
> on the board is on average 10 to 15% faster on NFS tests we ran in
> our labs and in some occasions up to 35%. Tuning of the current
> driver seems a very desirable task to pursue.

I think we all agree that tuning is desirable.  I'm frankly not
too surprised at a 10-15% improvement.  I know that when I last
looked at tuning for one chip (the i82544, before we had some of
the others, I think) em(4) from FreeBSD had a couple of default
values that were (if I recall correctly) completely undocumented
magic constants.  Adopting those made a huge difference at the
time--easily on the order of 10-15%.  I'm sure that wm(4) could
stand to be tuned further/again.

My experience with porting drivers over from FreeBSD has not been
stellar.  Perhaps this is because I'm mostly using non-ia32 systems,
but the drivers often need byte-order and/or bus_dmamap_sync() work
before they'll even work.  This isn't really surprising, nor is it
to say anything about the FreeBSD coding practices--it's just
naturally much more ia32-centric.  And we're not.

It just seems to make more sense to fix known issues (are there
any PRs filed against the missing hardware support in wm(4)?  Or
the performance issues?) in a generally-known driver rather than
porting largely unknown code from elsewhere that's not known to
necessarily work on a reasonable subset of systems.

-allen

-- 
 Allen Briggs                     briggs@wasabisystems.com
 Wasabi Systems, Inc.             http://www.wasabisystems.com/