Lost in the debate surrounding net neutrality is the assumption that all 
end hosts play nice and ``fair share'' the limited network bandwidth.  
Traditionally, ``fair share'' is enforced through a combination of 
forces: there is a general consensus on the TCP congestion
control algorithm, operating system vendors implement such algorithm 
in the kernels and it cannot be changed, applications use TCP and use
only a small number of TCP connections, and perhaps most importantly, 
network operators can identify abusive applications and traffic-shape them.

Unfortunately, today these protective forces are being weakened.  The 
networking research community has proposed at least nine different
congestion control
algorithms, some more aggressive than others.  Popular kernels such as 
Linux have implemented all of the algorithms, and furthermore, allow users 
with proper privileges to modify the congestion control algorithm
arbitrarily via loadable kernel modules.  
Many P2P applications use four or more connections between
each host pair to improve their performance.  And finally, most proposals
of Net-Neutrality legislation prohibit network operators from traffic-shaping
applications, for good reasons.

Therefore, for net-neutrality laws to avoid causing congestion collapse on 
the Internet, the law must also define proper end-host behavior, i.e.
``fair-share'', and allow network operators to detect violations of fair-share 
and apply punishment.  This presents three immediate difficulties.  First,
TCP only provides per-flow fairness, so to achieve fair-share, one application
shouldn't use more flows than the other.  But having a law to limit the number 
of TCP connections that an application can open is draconian to say the 
least.  Second, the TCP congestion control algorithm is the key in
providing per-flow fair-share, but the algorithm itself is under active 
research, and legislation in this area could stymie innovation.
Finally, it's difficult to have cheap and yet accurate detection of fair-share 
violations in routers, and any approximation techniques risk penalizing
the innocent.

This paper sketches out a set of potential solutions out of this dilemma.  
A central theme of the proposal is that there should be a code of conduct
for ``fair-share'' involving the number of TCP connection used and
the TCP congestion control algorithms used, and there should be
a differentiation between hosts and applications that follow the 
code-of-conduct and those that do not.  
Those that follow the ``fair-share'' code of conduct receive a guarantee 
from network providers that their traffic always receives the fair-share,
Those that do not want to be constrained by the code of conduct
accept that their traffic will be shaped
from time to time as deemed appropriate by the network provider.  
New technologies need to be developed to monitor the TCP implementations of
the end-hosts and applications, so that those who pledge to follow the
code of conduct can be verified.