Since my paper was rejected, I did not attend the middlebox workshop.
But I reading the back and forth on Facebook about it, and having spent
much of last week reading original sources on the architecture, I realized
that a lot of the problem seems to be the confusion between the use of
Internet Protocol in a network and an Inter-Network.
Today we typically use the term 'Internet' in a very broad fashion to refer
to all devices that run IP. That is a fine use of the term but it is not
the sense in which it is used in making the end-to-end argument. On the
contrary, in that era there was a sharp distinction between the local
network and the inter-network and very few machines in a university campus
had a direct inter-network connection.
When I read arguments from folk saying middleboxes should be eliminated,
they seem to all be making arguments for the Inter-Network. We certainly
want to keep the Inter-Network free of middleboxen. And with the rare
exception of boxes designed to perform mass surveillance deployed in Iran,
Syria, Russia, Saudi Arabia and other dictatorial regimes, we have pretty
much succeeded.
The question is not whether there should be middleboxen in the
Inter-Network. The question is how to provide control of the local network.
And here my college tutor, Tony Hoare was very insistent on collecting
security related functions into one single control point that can be
properly coded and audited within an operating system. I think the same
principle holds for a network.
It does not hold for an inter-network because the definition of an
Internetwork is that there is no central control point. Which in turn means
that we can't implement certain security functions in the Internet (though
there are some functions such as traffic analysis defense that can only be
implemented there).