[e2e] Number of persistent connections per HTTP server?

[Jim.Gettys at hp.com]

> As for HTTP 1.0's multiplexing of single TCP connections - it seemed
> well-thought-out given the constraints.   But I wonder if the best scheme
> would have been to construct a protocol replacement for HTTP that was not
> built on TCP, but instead on datagrams, where the "new HTTP" implemented a
> TCP-compatible congestion control scheme on the overall datagram flow,
> while keeping state for longer than an individual transaction and allowing
> concurrent "GET" commands to proceed without artificial ordering constraints.
>
> This would avoid the "slow start" and connection establishment overheads
> (by combining the flows into a single bundle for the purpose of congestion
> management), while maximizing concurrency at the endpoints.  ECN, RED, and
> window-based congestion control would work fine, actually BETTER.
>
> Such a protocol would be extremely useful for things other than HTTP, too.

I agree.  It would have been the best solution, though it would have begged
a major deployment problem.

Unfortunately, the reality of current IETF organization is such that
the transport and applications areas are considered independently (and
there is a cultural gap, as well), so this was never seriously considered.
I suggested creating something analogous to the IPv6 directorate for the web
as a possible solution, but it never went anywhere.

I've noted a large lack of understanding in people working on Internet
transport/routing to the real needs of applications developers, to the
point that basic understanding is often (usually) missing.  I find this
very unfortunate, indeed, and currently completely prevents useful
discussions of systems like the web as a "system".

Example: Radia Perlman once suggested to me that the typical type, length
value binary protocols would be useful for the Web, not understanding that
much/most content is dynamic and therefore the length was not known in advance.

This is why my OSDI keynote some years ago was titled "The Blind Men and
the Elephant".  Most people don't know enough of the parts of the animal to
understand the web, and as our SIGCOM paper showed, you have to understand
even the role of things like CSS to understand the trends of what is likely
to happen on the wire.
                        - Jim