Re: Draft Agenda for IETF 56

On Tue, 11 Mar 2003, Andre Beck wrote:

> I believe the consensus at the conference call was that we do away
> with the transport protocol negotiation mechanism and instead
> suggest that the callout transport protocol be selected based on the
> application protocol used between data consumer and provider.
Great! That's what I would vote for as well.

> Section 3.12:
>
> "The callout protocol MUST NOT negotiate the transport protocol to be
> used for callout connections.  The callout protocol MAY, however,
> specify that a certain application message protocol (e.g.  HTTP [5],
> RTP [8]) requires the use of a certain transport protocol (e.g.  TCP
> [6], SCTP [7])."
Sorry for missing that paragraph.

> On Tue, 11 Mar 2003, Reinaldo Penno wrote:
>
> Yes, I understand the current requirements, but since HTTP is the
> protocol in-scope at the moment and there is a push in the IETF (the
> chairs/ADs can correct me if I'm wrong) to use congestion aware
> protocols, we thought using SCTP might be a good compromise for
> those that think TCP is not appropiate.
>
> We had the same debate in another WG that is chossing an transport
> protocol and although some people wanted UDP the AD said the
> protocol must be congestion aware. Anyway, we can revisit this issue
> if people think it does not cater to the needs of OPES deployments.
I suspect that for many OPES applications it would be perfectly fine
to use TCP for the callout protocol while the application protocol is
using something like UDP. Most OPES transformations are per
application message (e.g., a UDP packet) so we should be fine as long
as the callout protocol includes appropriate per-application-message
timeout mechanisms. The predraft documents already contains some
wording in that direction.

If Reinaldo agrees with Andre's recollection of the conference call,
then there seems to be no conflict regarding support for non-TCP
application transport. This mini-thread started when jfc asked whether
OPES transport must be TCP, and I think my initial response remains
valid.


Alex.





> -----Original Message-----
> From: Alex Rousskov [mailto:rousskov(_at_)measurement-factory(_dot_)com]
> Sent: Tuesday, March 11, 2003 2:16 PM
> To: OPES Group
> Subject: RE: Draft Agenda for IETF 56
>
>
>
> On Tue, 11 Mar 2003, Reinaldo Penno wrote:
>
> > If you are running a media session with a friend, let's say RTP over
> > UDP, and the opes dispatcher (between you two) decides that this media
> > needs adaptaion because you are on a cell phone.
> >
> > In this case the callout protocol should (may?) run over UDP. This is
> > a simple effort in order to guess the best transport protocol to use
> > in the callout protocol for this specific session.
> >
> > Is this clearer?
> Yes, thank you! You are saying that both application protocol and callout
> protocol should (may?) use the same transport protocol, right?
>
> This means that either we cannot support application protocols that use UDP
> (and other lossy/etc protocols), or we will be violating current callout
> protocol requirements (that demand reliable congestion-aware transport):
>
> 3.1 Reliability
>
>    The OPES callout protocol MUST be able to provide ordered reliability
>    for the communication between OPES processor and callout server.
>    Additionally, the callout protocol SHOULD be able to provide
>    unordered reliability.
>
>    In order to satisfy the reliability requirements, the callout
>    protocol SHOULD specify that it must be used with a transport
>    protocol which provides ordered/unordered reliability at the
>    transport-layer, for example TCP [6] or SCTP [7].
>
> 3.2 Congestion Avoidance
>
>    The OPES callout protocol MUST ensure that congestion avoidance that
>    matches the standard of RFC 2914 [4] is applied on all communication
>    between OPES processor and callout server.  For this purpose, the
>    callout protocol SHOULD use a congestion-controlled transport-layer
>    protocol, presumably either TCP [6] or SCTP [7].
>
>
> IMO, callout protocol has sufficiently different purpose to deserve its own
> transport binding, which MAY be the same as application transport for some
> applications, but does not have to be. This approach would let us support
> adaptation of application protocols that do not satisfy OPES transport
> requirements.
>
> Thank you,
>
> Alex.
>
> > -----Original Message-----
> > From: Alex Rousskov [mailto:rousskov(_at_)measurement-factory(_dot_)com]
> > Sent: Tuesday, March 11, 2003 1:31 PM
> > To: OPES Group
> > Subject: RE: Draft Agenda for IETF 56
> >
> >
> >
> > On Tue, 11 Mar 2003, Reinaldo Penno wrote:
> >
> > > If I remember correctly the decision reached some time ago during a
> > > conference call was that the callout transport protocol would be the
> > > same one used between user<----...OPES box....--->consumer
> > Reinaldo,
> >
> >     I am confused: Callout protocol is the one OPES dispatcher uses to
> > talk to the callout server. Why do you have "user", "OPES box", and
> > "consumer" in your picture if you are talking about the callout
> > protocol? The callout protocol is used _inside_ the OPES box, not
> > outside it, AFAIK.
> >
> > Did you mean application protocol instead? If so, would it be accurate
> > to say that the conference call resulted in a decision to prohibit
> > application protocol conversion by OPES (e.g., converting HTTP to
> > FTP)?
> >
> > Please clarify or point to the relevant documents. BTW, were the
> > conference call minutes posted on the list?
> >
> > Thank you,
> >
> > Alex.
> >
> >
> > > -----Original Message-----
> > > From: Alex Rousskov [mailto:rousskov(_at_)measurement-factory(_dot_)com]
> > > Sent: Tuesday, March 11, 2003 12:06 PM
> > > To: OPES Group
> > > Subject: Re: Draft Agenda for IETF 56
> > >
> > >
> > >
> > > On Mon, 10 Mar 2003, jfcm wrote:
> > >
> > > > My main question is about the entry and the ending points of the
> > > > OPES concept. If it is accepted that not only HTTP but any
> > > > protocol is OK, and that alternatively both end points can be
> > > > user/producer or that we may have a peer to peer support, it will
> > > > mean that any entry/output protocol is OK, including a
> > > > keyboard/user file/application etc, towards a screen, a file, etc.
> > > >
> > > > This means there is no pseudo-synchoronous (like in hhtp)
> > > > obligation (or as options). This means that we have a general
> > > > purpose concept of a filter/transformer system on a flow. Which
> > > > will standardize everywhere.
> > > >
> > > > This means that, the filter/transformer relation (callout
> > > > protocol) can be anything from realtime to soap. Am I correct?
> > > I doubt we can design a good OPES system that can support adaptation
> > > of anything from highly optimized realtime protocols to
> > > structure-rich, super extensible XML protocols. There will be
> > > tradeoffs that we will have to resolve one way or the other. That is
> > > why it is important to agree on scope and use cases up-front.
> > >
> > > > Because this means that we will have a good general framework for
> > > > a protected/acked balanced protocol able to support one to many or
> > > > even many to many relations.
> > > We should strive for that, but IMO we will come short as things get
> > > more specific.
> > >
> > > > Then the next question will be: will it always be under TCP?
> > > Transport protocol binding wise, TCP is probably not the only
> > > option. We do need a reliable, order-preserving, congestion-aware
> > > protocol though (see protocol requirements draft for details). These
> > > features may already exclude certain realtime optimizations where
> > > losing packets is not only acceptable but desirable in the presence
> > > of unexpected delays or congestion.
> > >
> > > An important design question is whether the OPES protocol should
> > > allow multiple transport bindings/encodings or just one.
> > >
> > > Alex.