Return to BSD News archive
Path: euryale.cc.adfa.oz.au!newshost.carno.net.au!harbinger.cc.monash.edu.au!msunews!agate!ihnp4.ucsd.edu!munnari.OZ.AU!news.ecn.uoknor.edu!news.ysu.edu!news.radio.cz!newsbastard.radio.cz!mr.net!news-peer.gsl.net!news.gsl.net!newsfeed.nacamar.de!news1.best.com!nntp1.ba.best.com!not-for-mail
From: dillon@flea.best.net (Matt Dillon)
Newsgroups: comp.unix.bsd.freebsd.misc,comp.unix.bsd.bsdi.misc,comp.sys.sgi.misc
Subject: Re: no such thing as a "general user community"
Date: 27 Mar 1997 13:47:33 -0800
Organization: BEST Internet Communications, Inc.
Lines: 87
Message-ID: <5heptl$8ir@flea.best.net>
References: <331BB7DD.28EC@net5.net> <5gn3ig$83d@flea.best.net> <5goqrq$5ak$1@news.clinet.fi> <5hd29s$e7t@fido.asd.sgi.com>
NNTP-Posting-Host: flea.best.net
Xref: euryale.cc.adfa.oz.au comp.unix.bsd.freebsd.misc:37897 comp.unix.bsd.bsdi.misc:6485 comp.sys.sgi.misc:29464
:In article <5hd29s$e7t@fido.asd.sgi.com>,
:Larry McVoy <lm@slovax.engr.sgi.com> wrote:
:...
:>one smart cookie, I'd love to hire him. Ditto for the FreeBSD team.
:>But make no mistake there is more to an OS than a simple minded file
:>system or a context switch. SGI's OS scales to 100s of CPUs, supports
:>more devices, and has more stuff that people with $$$ want. This is
:>serious software with serious people working on it. Do you want IRIX for
:>a desktop to read news? Hell no, install Linux or FreeBSD or whatever
:>on your 386 and you are done. If you have real work to do you may find
:>that you have to pay for your tools.
:>--
:>---
:>Larry McVoy lm@sgi.com http://reality.sgi.com/lm (415) 933-1804
IRIX scale to 100s of cpus ? Not in its current incarnation. There
is nothing magical about XFS and there is nothing magical about striping
or mirroring (in fact, SGI's mirroring implementation isn't even all
that good... we've lost mirrored disks on our L, and it not only takes
the machine down, it also gets *real* confused on reboot). There is
nothing magical about doing 500 MBytes/sec of disk throughput in a
test (though I will admit the limitations of PCI prevent a FreeBSD
system from being able to do more then 1/5 of that rate, there is
virtually nobody on this earth who needs 500 MBytes/sec of disk throughput
from a UNIX box). There is also nothing magical about running what is
essentially a single program in an SMP configuration. But there
IS something to be said for keeping utility programs up to date, fixing
serious bugs, and keeping the friggin kernel up to date. There is
something to be said for using modern paging algorithms and for paying
attention to degenerate conditions and cascade failures. There is
something to be said for paying attention to SECURITY. What I see when
I look at IRIX is an extremely hacked circa 1985 sysV kernel that needs
a serious workover. It needs a rewrite.
The kernel, basically, works just fine if you don't stress it or run
too many I/O transactions (disk and network), but the moment you
start doing anything real with it the fireworks start. Open too many
files, start paging, run too many processes, or do too much with your
network and watch out! Scale to 100 cpu's ? Maybe if all 100 cpu's
were running the same program you could, but otherwise, no. It can't
even scale to 4 or 8.
Now I am sure there are plenty of companies out there who will happily
throw money to oufit an SGI box to the point where none of these
failure points are actually reached, just to be able to run certain
software that nobody else has. I would leave you with this warning,
however: The same thing occured with IBM mainframes and you can see what
happened to them. While it can be argued that IBM is actually recovering
its mainframe business somewhat, the potential business that they
have LOST to other, smaller boxes, can be measured in the trillions
of dollars. Not billions, trillions. SGI is in a much tighter race.
I happen to like MIPS cpus. I think they are just dandy compared to
the pentium monstrosity. But our sysadmin work load has been cut by
an order of magnitude with the FreeBSD boxes. We don't have to struggle
with them the way we have to struggle with IRIX boxes... things just
work. There are very few degenerate cases or cascade failures when
a box is attacked over the internet or a user does something dumb.
The security is an order of magnitude better... FreeBSD is reasonably
secure 'out of the box', verses IRIX with it's hundreds of completely
unsecure suid-root programs, out of date security mechanisms, and the half
dozen completely unsecured and internet-accessible accounts that come
with the machine.
There are lots of little things that just make life easier all around,
such as chmod/chown/chgrp -R doing the right thing with softlinks, such
as /bin/sh properly using O_APPEND mode. Such as major system daemons
not barfing, or utmp getting lost. When something goes wrong on a BSD
box, it's *easy* to find the source of problem. At least we can *count*
on what performance we do get with FreeBSD... the curve doesn't have
any major glitches in it until it the box is DEEP into memory starvation.
It runs relatively flat whether there are 1000 vnodes or 10,000 vnodes,
whether it's not paging at all or has a VM size that is double the
available physical memory, whether the network is operating normally or
under attack. With IRIX, any one glitch turns the machine into a sludge
pile and the curve drops precipitously the moment any one resource
begins to get starved. The only way to safely run an IRIX box is if
you never push it.
Scale to a 100 cpus? Not unless you can actually run 100 cpu's worth of
load, and right now you cannot. I know I am being somewhat unfair,
after all FreeBSD is essentially a mono-cpu OS (not counting chickens
before the hatch)... but if you ARE going to create a multi-processor
kernel, you have to do it RIGHT... from the ground up.
-Matt