Thought on Serial IP links


Barry Shein (bzs@bu-cs.bu.edu)
Tue, 22 Apr 86 02:01:36 EST


I have been thinking about the problem of attaching computers from,
typically, homes to, for example, a campus. Perhaps a little out-loud
thinking would move us all closer to a solution as, I agree, the
emergence of internet protocols on PCs does present some new and
interesting problems. Further, I am presuming in my model that an
ethernet exists at the 'campus' end. Sites that do not fit this
model could at modest cost, or perhaps other similar solutions
could be made to fit (the framework is really more general.)

Problems

1. A PC will typically connect to a campus over a serial line via
a phone connection to/from a pair of modems. A typical scheme is
to dial-up a capable host as a user and then request that the line
be set into a special mode to pass packets, that is, a point to
point link (p-p) analogous to ones currently in use.

The primary problem with this model is that the host in question is
being used as a packet switch if (and this is likely) the primary
target of packets is not the host which was dialed. Further, this
yields a milking-machine problem, although this may be unavoidable
in the near future putting it directly on the host is not the best
solution.

2. The PC needs an IP address and an internet hostname (even if one
only locally recognized.) Two obvious solutions are either the PC
stores it and announces it to the host it is p-p'd to or it uses
something analagous to RARP as suggested on this list.

The worst problem with this is that in most current schemes each
p-p link requires a NETWORK address. Various software hacks analogous
to subnetting can be used, but read on.

3. Links and identities are temporary, unlike most internet hosts
we would expect that most of these p-p links are usually NOT on-line.
In the proposed p-p model this is not a major problem, but is worth
noting for the ensuing discussion.

One easy way to solve this problem would be to put an ethernet in your
home and use one of the commercially available level 2 bridges, level
3 would be overkill here as most likely all packets on the remote
('home') side probably want to travel to the local ('campus') end
anyway.

The problem with this is the cost of providing the pieces, they are
quite simply additive to the original cost as you still need two modems
to effect the bridge, so I will reject it but note that it leads to the
next proposal.

-> What you really want is an ethernet drop cable which extends from your
house to the campus.

This could be accomplished by a magic box which had a modem at one
end, a transceiver to the campus ethernet at the other and received
packets over the serial link and simply transmitted them into the
ethernet. Further, I believe this box could be made quite
economically out of something like the Bridge or Annex TCP
multiplexors. The only requirement would be that they be modified
quite a bit at the software level (actually, simplified, as they are
simply now IP level links) and hardware would have to either support
multiple ethernet addresses (one for each serial line) or multiple
ethernet interfaces (that is, either logical or physical translation
from each serial port to an ethernet transceiver.)

I am not enough of an electrical engineer to judge how hard it would
be to provide such a box, hopefully someone on this list could.

The general idea would be that the PC would behave as if it were
attached to an ethernet but use its serial port to pass data back and
forth. Thus, it would be responsible to answer ARP requests etc.
Conceivably this could be moved into the local multiplexor, but that
is simply an efficiency hack. I believe at the software level this
would be quite simple in the PC, mostly re-using the existing serial
and ethernet software already (typically, as in PC/IP) provided.

One problem with this is the assignment of IP addresses and their
mapping to ethernet hardware addresses. Another problem would be ARP
caches in remote hosts as this mapping would change as people hung up
the phone and others re-used the ethernet address.

Thinking about this, I would require that the PC not have an IP
address at all but use a RARP protocol. I would administer this by
having the user apply at the hosting site for a username, password,
hostname and IP address. The magical box would, on attach, allow the
user to present a username/password pair and then query a server (we
assume this box has no significant disk space, and we need a more
universal database anyhow) for the hostname/IP address to return which
will then be bound to the link.

The next problem is the cache'd IP->ethernet addresses that most hosts
would keep. One could take the attitude that it really doesn't much
matter. If a packet were sent from that host with the 'wrong' IP
address to that ethernet address the PC, realizing it is not a
gateway, would ignore the packet and the ARP cache would die away soon
enough, for security's sake the magic box could even provide this
filtering.

Another solution perhaps more appealing to those concerned with
robustness would be the addition of another protocol to the ARP suite,
namely a Flush ARP broadcast packet type. The magic box, upon
realizing that the line has hung up could broadcast the ethernet/IP
address pair with a field 'ARP_FLUSH' indicating that hosts who
currently cache this pair should immediately forget it. This, I
believe, presents little burden to anyone involved.

The niceties of this approach are 1) A user could use their
IP/hostname from any PC (or have one for each, their choice) by simply
specifying a user/password pair after dialing in 2) the PC appears to
be a part of the ethernet already in place and 3) the hardware
involved seems simple, at least to me, and cost effective, and puts
the milking-machine problem in the right place (away from a host and
into the net directly.)

Although I realize there are some large 'ifs' here, primarily the
mythical box, I do believe a scheme like this is a far better approach
than I have seen so far. I would project that the additional cost
(beyond the modems already assumed) would be on the order of $250/port
all at the 'campus' end (which is attractive, as it is re-usable and
thus easier to justify.) The major unique thought here is the assignment
of an ethernet address to each serial link.

I await your judgement. I will name it the Remote Transceiver Protocol
(RXP.) It obviously needs some more fleshing out (such as whether SLIP
is robust enough for such a link both in providing error handling and
flow control) but I think this is enough to see if it stimulates any
basic judgement on the merits of the idea.

        -Barry Shein, Boston University



This archive was generated by hypermail 2.0b3 on Thu Mar 09 2000 - 14:36:06 GMT