[e2e] Number of persistent connections per HTTP server?

[Christian Huitema]

> >> TCP _assumes_ per-connection is the definition of fair. That's
> >> incorrect when fair doesn't map to 'per connection'. Striping,
> >> pipelining, overlays, caching, etc. all examine areas of the design
> >> space where per-connection deliberately doesn't map to per-user or
> >> per-application.
> >
> > Is there a document describing TCP where it says "TCP assumes fair
is
> > per-connection"?
> 
> The original spec didn't really address fairness per se; Van's 88
paper
> is the earliest that made that assumption (somewhere around page 10).
> The implication is that per-connection is defined as fair.
> 
> It would have been more precise to say "most studies of TCP fairness
> assume per-connection is the definition of fair".

There were actually a lot of debates on "what is fair". There are
arguments for defining fairness per class of user, per network, per
machine, or per connection. There is also a discussion as to whether
enforcing this kind of fairness belongs in an end-to-end algorithm,
versus a network based tool. All of the definitions have failure modes:

- A per connection definition ties the throughput of a node to the
number of active connections, which will tend to advantage some
applications against others,
- A per computer definition treats a large shared server the same way as
an individual workstation,
- A per network definition treats large networks the same as small ones,
- A per user definition does not differentiate between classes of
users...

The only thing we are really sure off is that, in situation of
congestion, connections should slow down; it is also desirable that if
all other parameters are the same, different connections receive
approximately the same treatment. But justice ought to be enforced by
other means; indeed, today, the available bandwidth is mostly enforced
by the parameters of one's connection to the Internet. TAINSTAFL...

-- Christian Huitema