Cell Relay Archive

[Jean-Philippe Martin-Flatin <jpmf@tcom.epfl.ch>]

In article <870276108.4917@dejanews.com>,manfredi@arl.bna.boeing.com
wrote:
> 
> [cut]
> What the discussion was about was whether the ATM underlying
> network is best described as "connection oriented" or as "circuit
> switched." I would simply suggest to you that IP running over an
> ATM net, especially an ATM net which uses only PVCs, is better
> described as IP running over a circuit switched network than
> anything else. Even if the PVCs are UBR, what sits under that IP
> layer is simply a set of pipes - circuits.

This last sentence makes crystal clear why many people cannot agree with
you: your terminology is somewhat different from the one usually
accepted, which you can find in:

    Martin de Prycker, ATM: Solution for Broadband ISDN, Prentice
    Hall, London, UK, 3rd edition, pp. 53-63.

For you, a circuit is just a pipe: you do not care whether the transfer
in this pipe is synchronous or asynchronous.

For most people though, circuit-switching is based on time division
multiplexing (TDM), so it is necessarily synchronous: a given circuit
will be given a periodic time slot. Conversely, packet-switching is not
based on TDM, so there is no periodicity in the time slots: you have
queues, contention, and new issues like jitter introduced by queue
lengths varying in time.

In that sense, widely accepted by the network community, ATM is
packet-switched, not circuit-switched as you claim.

As for connections, your definition of a circuit is exactly the
traditional view of a Virtual Path Connection (VPC): anything can go in
it, it's just a pipe (we use precisely this image for our students).

> Since IP over ATM always works essentially this way, i.e. even if
> SVCs are used, they are long-lived pipes rather than dynamic
> connections made between end stations, I would suggest that
> describing ATM as circuit switched makes a lot more sense than
> confusing the issue by making it sound sort of TCP-like (as in,
> "connection oriented ATM").
> 
> Note the definition of "connection" in Douglas Comer's book
> _Internetworking with TCP/IP_:

I agree with Qiwei Xiao's answer in article
<5rp2dq$6a1@mozo.cc.purdue.edu>.

You're mixing up connections at different layers here. The ATM model
doesn't map easily onto the OSI 7-layer model, but in a first
approximation, we can consider that IP-over-ATM can be modelled as
follows:

    layer 4: TCP, UDP
    layer 3: subdivided into:
      layer 3.2: IP
      layer 3.1: AAL
    layer 2: ATM connection (PVC, SVC)

TCP is connection-oriented (layer 4), IP is connectionless (layer 3),
and ATM is connection-oriented (layer 2). Whether you have a connection
or not at layer N is independent from what you have at other layers. To
come back to my initial question, if the routing tables change (an event
occurring at layer 3), then:

- With IP-over-ATM, new packets will take the new route immediately
- With plain ATM (i.e. VCs in VPs), rerouting of existing long-lived
  connections won't happen automatically: cells in old VCs will use
  the old route.

Jean-Philippe

-----------------------------------------------------------------------
Jean-Philippe Martin-Flatin                    Email: jpmf@tcom.epfl.ch
Research and teaching assistant                Fax: +41-21-693-2683
Telecommunications Laboratory
EPFL (Swiss Federal Institute of Technology at Lausanne)
1015 Lausanne
Switzerland
-----------------------------------------------------------------------