Subject: Re: FreeBSD Bus DMA
To: Ted Lemon <mellon@hoffman.vix.com>
From: John S. Dyson <dyson@freebsd.org>
List: tech-kern
Date: 06/11/1998 19:31:48
Ted Lemon said:
> 
> John, I've never seen any honest benchmark numbers for FreeBSD/i386
> vs. NetBSD/i386, much less for FreeBSD/Alpha vs. NetBSD/Alpha.   I
> doubt you have either, or they'd be up on your web site and you'd be
> pointing us to the URL.
>
I don't post benchmarks, sorry :-).

>
>   Everything I've seen suggests that the two
> operating systems perform quite similarly.
>
That isn't what I have seen, and that has nothing to do with
benchmarks but compiler loads, and real user productivity in that
environment.  One of our teams decided that BSD really performed
badly, based upon their NetBSD experience.  We did experience
a 3X speedup, and paging performance had little to do with it.
I didn't run those tests, and they were done by a skeptic :-).

>
>   NetBSD had a problem with
> thrashing with Mach VM in high-memory-consumption situations, but that
> problem has been corrected with the introduction of UVM, which means
> that your old 3X number is now obsolete, and even when it was valid,
> it was valid for a corner case that didn't matter to many users.
> 
Try it yourself, I actually  haven't seen much of a real performance
improvement in NetBSD, but YMMV, I guess.  Anything is better than
the MACH stuff, and we fixed acknowleged and fixed the 1st order
problems a long time ago.  Frankly, the end conditions aren't a
major reason for FreeBSD's peformance.

The 1st order VM loading issues are to get rid of the guaranteed
thrash of clock and clock derivates (most those are totally
lame under load.)  There is tons more than just paging algorithm
to make the VM system work well.

-- 
John                  | Never try to teach a pig to sing,
dyson@freebsd.org     | it just makes you look stupid,
jdyson@nc.com         | and it irritates the pig.