Subject: Re: perhaps time to check our TCP against spec?
To: Jonathan Stone <jonathan@DSG.Stanford.EDU>
From: Stefan Grefen <>
List: tech-net
Date: 04/08/1998 21:33:49
In message <199804080251.TAA10822@Whisk.DSG.Stanford.EDU>  Jonathan Stone wrote:


> I am trying to make a legitimate techmical point.
> In_maxmtu will cause problems for people who want to use NetBSD on
> mobile multihomed hosts with heterogenous media.  
> Like, for example, laptops with Ethernet and Metricom radios.  those
> are very widespread on the Stanford campus.  The data in Stuart
> Cheshire's Hot Interconnects 96 paper explains why fragmentation will
> cause severre performance problems here-- immediately halving
> thorughput and increasing congestion which may in turn further reduce
> throughput.

The existing approach cause  performance problems on people with
other network topology ...

> I made the point that this change violates existing practice.  
> Kevin Lahey points out that there's a consensus on tcp-impl about
> what the conservative (i.e., right) thing to do is. the consensus
> is to not do in_maxmtu but to stick with the conservative approach.
> We have a point where one of the NetBSD Core Group is flaming that
> they intend to violate ``be conservative in what you send'', and they
> dont CARE what it breaks.

No, I don't read Jasons mail that way. I think there is consensus that
you can configure either behaviour and if I recall that correct  "threat-mail"
(just to identify it) states that we're talking about the default behaviour.
I can see your point and I think more people will suffer from the 
new thing than from the old way.
But I also think fighting for defaults doesn't warrant this flame war.
Everybody can tune the system they he/she wants. Either by compiling
with a different default or by using sysctl.
For the benefit of those people in a old enviroment I would recommend
to use the old-behaviour in Install-kernels.

> I think that's a dumb technical decision, and if we stick with it,
> then I will have to stop recommending NetBSD to people at Stanford
> altogether.

Face it, the world is full of dumb technical decisions, and most of them
are more severe than the default we're talking about here.
Think of 'strategic descions to use NT' which in one case leads to
replacing one working SUN mailserver with 1200 NT-exchange servers.
Yes that may be off-topic, but I think while we fight like crazy over
minor details, the rest of the world (or a big part of it) gets 
brainwashed to think a inferior system is the way to go.
How about using our energy to implement something realy exiting like
clustering and/or high-availbiltity stuff? 


Stefan Grefen                                Tandem Computers Europe Inc.                       High Performance Research Center
 --- Hacking's just another word for nothing left to kludge. ---