New Liaison Statement, "IETF CCAMP Response to Liaison on MFA CNI"

[Matthew Bocci (IETF) <matthew.bocci@alcatel.co.uk>]

Title: IETF CCAMP Response to Liaison on MFA CNI
Submission Date: 2006-11-30
URL of the IETF Web page: https://datatracker.ietf.org/public/liaison_detail.cgi?detail_id=293 

From: Matthew Bocci(IETF) <matthew.bocci@alcatel.co.uk>
To: MFA Forum(Rao Cherukuri <cherukuri@juniper.net>)
Cc: CCAMP Mailing list <ccamp@ops.ietf.org>
Ross Callon <rcallon@juniper.net>
Bill Fenner <fenner@research.att.com>
Adrian Farrel <adrian@olddog.co.uk>
Deborah Brungard <dbrungard@att.com>

Reponse Contact: Adrian Farrel <adrian@olddog.co.uk>
Deborah Brungard <dbrungard@att.com>
Technical Contact: Adrian Farrel <adrian@olddog.co.uk>
Deborah Brungard <dbrungard@att.com>

Purpose: In response 
Body: Dear Rao,

Thank you very much for giving the CCAMP working group the opportunity to
comment on the MFA's MPLS CNI specification.

At this stage we have the following high-level comments that we request the
MFA to consider and answer. Some of these comments are for clarification as
we try to ramp-up on the requirements you are addressing. Others concern
technical issues that we feel you should address before advancing this work.

Best regards,
Adrian Farrel and Deborah Brungard
CCAMP Working Group Co-Chairs

===

We note that you have opted to define a new RSVP object to support a
multi-class LSP following the rules for vendor private assignment as
described in section 2.2 of RFC3936. We believe that you may have
misinterpreted the purpose of vendor private extensions since such
extensions are specifically not intended to interoperate, but you are
attempting to define a specification directly for the purpose of
interworking devices from different vendors. In your case it would seem to
make more sense to define a standardized extension to the protocol.

Should you decide that a standardized extension is better able to deliver
the functionality that you require, we should like to draw your attention to
draft-andersson-rtg-gmpls-change-06.txt that defines how other SDOs may
influence the development of MPLS and GMPLS protocols within the IETF, and
which is currently in IETF last call. The (G)MPLS suites of protocols have
become popular among multiple SDOs resulting in a need for IETF to clarify
it's role as the responsible SDO for (G)MPLS protocol extensions so as to
prevent unnecessary replication of functionality and the resulting
interoperability problems.

1. The document is marked as Straw Ballot Text. Can you tell us
   what that means the status of the work is?

2. We think that your use of terminology may be a little loose.

   In many cases, where you say "MPLS" you are probably referring
   to the data plane, and specifically a packet-switching data
   plane with an MPLS encapsulation.

   But in other cases, "MPLS" and "MPLS-TE" are synonymous and
   refer to a signaling/routing control plane using the MPLS-TE
   extensions to RSVP and to the two IGPs OSPF and IS-IS.

   In many cases you say "GMPLS-TE" which is not something we
   specifically recognise although we can assume that you mean
   simply "GMPLS". Sometimes, where you are trying to
   distinguish a TE LSP established using MPLS-TE from one set
   up using GMPLS you may intend to say "GMPLS TE-LSP" rather
   than "GMPLS-TE LSP".

   We feel that close attention to the terminology may help
   clarify the document.

3. Section 1.1 states:
     The purpose of this specification is to define an MPLS-based
     Client to Network Interconnect (CNI) for establishing GMPLS
     Traffic Engineered (TE) Label Switched Paths (LSPs).
   Can we assume that this means that the client-to-client LSPs
   are established using GMPLS protocols, but that the signaling
   within the core network is out of scope? Especially since
   section 1.2 states:
     At the CNI, it is not desirable to have the client equipment
     participate in the internal control protocols of the MPLS
     network.

4. Can you clarify why you have selected GMPLS protocols and not
   MPLS-TE protocols on which to build your CNI. We are not
   opposed to this, but are seeking to understand the choice.
   Perhaps the main reason is the requirement for bidirectional
   LSPs.

5. Can you clarify whether the core network is assumed to be
   PSC only? That is, for example, if the CNI encoding is POS,
   would it be acceptable for the PE and the rest of the core
   network to switch the LSP as TDM until the remote PE or even
   CE, or do you require that the PE must perform packet
   switching? If the PE must perform packet switching, is it
   still acceptable for the core LSP (PE-PE) to be switched at
   some other technology?

6. Section 1.3 states:
     Where the network uses MPLS-TE signaling, the PE routers are
     expected to perform the translation.
   It is our opinion that this translation is non-trivial and may
   be impossible for some of the GMPLS services that are
   available at the CNI. For example, supporting a bidirectional
   service over an MPLS-TE signaling network requires additional
   coordination between the end-points that is currently not
   available in the MPLS-TE extensions to RSVP-TE.
   From the following text in section 7.1, we assume that the PE
   may refuse a CNI request if it is unable to provide the
   required level of function.
      The transport network in the provider network is a GMPLS or
      MPLS-TE based packet switched network that must support
      request for uni-directional LSPs and may support requests
      for bi-directional LSPs

7. Section 2.1
   The correct expansion of "GMPLS" is "Generalized Multiprotocol
   Label Switching". In view of you chosen expansion of "MPLS",
   you may prefer to show this as "Generalized Multi Protocol
   Label Switching".

   The correct expansion of "FEC" is "Forwarding Equivalence
   Class".

8. Section 7.2 states:
     The CE and PE nodes are inter-connected by point-to-point
     interfaces.  The signaling channel is "in-band", i.e., the
     labeled packets share the same access connection as the
     RSVP-TE signaling.
   This is an acceptable, but not required, method of deploying
   GMPLS-based signaling. It is our suspicion reading this very
   short section that it is your intention to forbid the use of
   the IF_ID RSVP_HOP Object at the CNI. Can you confirm or deny
   this?

9. Section 7.3 states:
     A client need only know its own address, a reachable address
     of the adjacent PE-node, and know the address of any other
     client to which it wishes to connect.  The addresses listed
     above must be configured on each client.

     A PE need only know (and track) the addresses on interfaces
     attached to clients, as well as the Node IDs of these
     attached clients. In addition, the IP/MPLS network needs to
     know reachability to the interface addresses and Node IDs of
     other PEs to which an attached client is permitted to
     connect.

   This appears to miss the fact that the client will address a
   CNI connection request to a remote client address. The local
   PE must, therefore, know how to route to these client
   addresses that are outside the core routing domain.

   Perhaps the final sentence should say CE not PE?

   But in 9.1.2 you have:
     When a PE receives a Path message from a client that
     contains no ERO indicating transit network selection, then
     the PE is responsible for progressing the Path message
     toward the destination.  The progression of the Path
     message is beyond the scope of this specification.

   While the details are clearly out of scope, it *is* relevant
   to the definition of the CNI how the core acquires and
   distributes the client-side addressing information that is
   necessary for routing across the core. You will observe that
   the problem you are solving (including the fact that the
   client addresses may come from an address space that overlaps
   with the core address space) is similar to the L3VPN problem.

10. What is the expected behavior from the core network when an
    E-LSP is requested at the CNI?

    Can we assume that the expectation is that an appropriate
    E-LSP will also be established across the core so that
    Diffserv behavior will be performed along the whole length
    of the client-client connection, or is this not a
    requirement?

    If core Diffserv behavior is required, how will the core
    handle the presence of multiple classes?

11. You are correct to observe that the ERO is optional in GMPLS
    implementations (sections 9.1.1 and 9.1.2), however, since
    you are specifying a profile for use at the CNI, and since
    both the CE and the PE must be CNI-aware (i.e., you cannot
    simply use legacy implementations) you may find it convenient
    to mandate support of the ERO at the CNI.

    We believe that in practice all implementations support ERO.

12. In section 9.1.1 you have:
      The client populates the ERO object with only one sub-object
      containing an Autonomous System Number (ASN) representing a
      transit network beyond the originating service provider.
      The client equipment must set the ASN sub-object 'L' bit to
      1, indicating a loose route.
    It is not completely clear what is meant by 'the originating
    service provider', but we assume that this refers to the
    network that the ingress PE belongs to. In this case, this
     ERO is malformed and will be rejected. The first sub-object
    of a received ERO must always define an abstract node that
    the receiver is a member of. See RFC 3209, section 4.3.4.1,
    point 1).

13. In section 9.4:
      PE next to a client receives a PathErr with
      Path_State_Removed from the network, it may in turn
      generate either a ResvTear or PathTear, whichever is
      applicable, to be sent to the client.
    There are no circumstances in which a PE receiving a PathErr
    with Path_State_Removed from the network would send PathTear
    to the client.

    It is unclear to us why you would specify that the CNI built
    on GMPLS might not use this standard GMPLS procedure.

14. In section 9.5.3:
      The Extended Classtype object is signaled in the Path
      message, and saved in the Path State Block (PSB) at every
      hop.
    We recommend that you simply state that the state is stored
    as every hop. The existence of a PSB is implementation-
    specific.

    Can you please clarify "at every hop". Are you expecting
    nodes in the core network to store this information. If so,
    you should note that the core nodes will not recognize the
    object class and will reject any messages carrying it.

    We also suggest that before progressing your own extensions
    for multi-class DSTE LSPs you should look at the existing
    work within the IETF:
    http://www.ietf.org/internet-drafts/draft-minei-diffserv-te-multi-class-02.txt.
    If this work is not adequate for your requirements, we
    would encourage you to work with its authors to produce
    a single standardized solution within the IETF.

15. In 9.5.4:
      An LSR that recognizes the Extended Classtype object and
      that receives a Path message which contains the Extended
      Classtype object but which does not contain a Label Request
      object or which does not have a session type of
      LSP_Tunnel_IPv4, must send a PathErr message towards the
      sender with the error code 'extended-classtype Error' and
      an error value of 'Unexpected Extended Classtype object'.
      These are defined below:
    Why do you define new parsing behavior for the absence of a
    Label Request object (by the way, you should say Generalized
    Label Request object, since this is GMPLS)? The absence of
    mandatory objects is already covered in RFC2205.

16. The error code defined in 9.5.4 is conformant with RFC 3936.
    You may wish to look at draft-swallow-rsvp-user-error-spec
    in case this gives you the ability to handle more detailed
    private error codes.

Finally, we would like to refer you to
draft-kumaki-ccamp-mpls-gmpls-interwork-reqts-01.txt and
draft-ietf-ccamp-mpls-gmpls-interwork-fmwk-01.txt for the latest state of
discussions in CCAMP with respect to interworking MPLS and GMPLS networks.
Attachment(s):
No document has been attached