Subject: Re: SMP on the other free Unices (was Re: core resignations)
To: Gandhi woulda smacked you <email@example.com>
From: Brian C. Grayson <firstname.lastname@example.org>
Date: 04/20/1999 10:29:03
On Tue, Apr 20, 1999 at 12:19:31AM -0700, Gandhi woulda smacked you wrote:
> What's the current state of FreeBSD SMP vs. NetBSD SMP vs. Linux SMP?
> Or even MP in general?
I've been forced to run FreeBSD and Linux on a dozen
quad-processor boxes here. Although FreeBSD and Linux support
SMP, they do not appear to be as stable as NetBSD on uni. The
following is My Experience Only, from around 2 years of
Linux-SMP, and a half-year or so of FreeBSD-SMP. YMMV.
Up until recently, Linux SMP really didn't like swapping much
--- if you swapped too much, it would panic. Supposedly
they've fixed that to some degree, but when I run a >1G process
on a quad machine with 1G RAM and a kernel from less than 1 month
ago, the whole system behaves slowly, not just the one
process/processor. And it appears to take many minutes, not seconds,
to swap a 300MB process out to disk (80MB/sec UltraWide disk).
I still get occasional several-minute freezes or panics if I
push the machines too hard -- like forking off 12 CPU-bound
simulations on a 4-proc machine. (Uh-oh, a load of 3 per CPU
-- time to call in the cavalry!) :)
Besides the SMP support, Linux is still vulnerable to all of
the traditional Linux vulnerabilities --- in particular, NFSv2,
and a not-good implementation at that. I've had random
characters dropped from occasional NFS writes.
FreeBSD is much more stable. It deals with swapping much more
gracefully, and performance while swapping appears much better
than Linux. However, I've had around 10 hangs in the last
two months on three quad boxes. No panic message, no apparent
cause besides high load (~10 for 4 processors). This is with
4.0-current (or whatever) code from around February. I've been
able to push these machines up to loads of 20 and 30 at times
with no ill effects, though.
For any new SMP boxes, my preference and my advice is to go
In my opinion, if NetBSD can make a really solid, stable SMP
system, it has a good chance to shine, even for folks like me
who are not doing multithreaded stuff at all (yet!), just batch
simulations. And even if we don't gain many converts :(,
hopefully our better stability will force the others to improve
their stability just to keep up with the Jones's.