TCPMUX vs. port numbers

[[e-mail]]

From Dave Bridgham's note:

>RFC931 is a TCP based protocol; I assume TAP is as well.  If we're
>going to talk at the principle level, as a matter of principle all new
>TCP protocols should use TCPMUX.  Allocating new port numbers, or
>reallocating old ones, is not an issue.

This is a controversial opinion that I don't agree with (although maybe
someone could persuade me to change my mind).  At any rate it isn't an
adopted policy or a consensus, at least not yet.

One argument against things like TCPMUX is that (as with OSI) application
filtering is painfully hard to do in network elements when there isn't an
application-specific token in each pdu (or perhaps more importantly, in
the first one).  This is also a controversial topic but there are many
people doing this, so a decision to break it shouldn't be taken lightly.

Dave Crocker (at least I think it was Dave! If not, apologies to Dave and
whoever) commented on this in plenary at San Diego IETF, responding
to what he had heard about Walt Lazear's presentation in NOOP WG about OSI
filtering.  Now, I might not be as worried as Dave was, but neither would
I flush the issue.

Another issue is 64k connections limit on a single well-known port, maybe
an issue with big hosts or if connections are frequent and short (this
is the TIME-WAIT issue).  More WKPs localizes the problem to a single
application (at a time).

Some people would say that the fact that TCPMUX resembles OSI more than
it does current TCP style is all the reason anyone needs to declare it a
bad thing, but I'm not that prejudiced.

Bottom line (IMHO) is that I wish the problem were that simple, but it isn't.

Bill Barns / MITRE / [e-mail]