Re: [mpls] working group last call ondraft-ietf-mpls-p2mp-sig-requirement-02.txt

Hi Eric,

Thank you for your detail comments.
We very much appreciate your continuous interest and support
of this draft.

 >Overall,  I think this is a  vast  improvement over last year's
 >version.

Thank you. I'm glad you think so.
Your input on the previous version was a great help.
And thanks to a sufficient discussion and inputs from
the community, we think we get stable requirements right now.

 >However, I still have a few issues with it.
 >
 >- The presumption that there is a low (indeed, very low) rate of churn
 >   seems to be essential to much of the document.  However, it is only
 >   mentioned in passing  towards  the  end.   As  this  is  a  major
 >   restriction  on the applicability of  the requirements, the presumption
 >   should  be stated near the front of the document; it should  be clear
 >   to anyone who has read only the introduction.  It  might be a good idea
 >   to mention  it in the abstract as well.

It is true that this requirement document assume solution to be tuned to
equip several important P2MP TE features such as FRR, Grafting/Pruning
rather than to be tuned to handle very high rate of churn of P2MP tree.
But we have no intention this requirement document to preclude a solution
and implementations to handle such a condition. We think it is up to
a solution and implementations whether to support high rate
(not so high as other technology) of churn of tree in addition to
support important P2MP TE features.

We also agree that discussion related to "rate of change of the tree" and
"rate of change of the network" is quite late in the document even though
we use almost a page in section 4.19.1.

We had taken it for granted that it is a fundamental part of the
application of Traffic Engineering that there is a low rate of churn
(whether low is actually very low, is subjective).

Nevertheless, there is no issue with highlighting this in the introduction
where we discuss some of the implications of Traffic Engineering.

We will make this change.

 >- There are a number of  references to requirements that are fully
 >   specified only in other documents, mostly from CCAMP.  The
 >   applicability of these is not necessarily obvious to those who don't
 >   follow CCAMP, e.g., crankback, hierarchy.
 >   I'd really  like to see this document  be more self-contained,
 >   so we have a better idea of what we are agreeing to even if we do
 >   not follow CCAMP.

I don't think that we should re-document any material that is present
in other I-Ds. Please note:

- The only requirements that specifically reference CCAMP work are
   in 4.21 and 4.23.
- Hierarchy is an MPLS WG draft and so we should be familiar with this.
- The requirements in this draft are applicable to MPLS TE and GMPLS,
    but we would not expect to re-document the requirements of FRR in
   case they were not familiar to the GMPLS community.
- It is better for everyone who is interested in MPLS TE to follow
   CCAMP.  For example, this is where inter-AS MPLS TE is being
   developed.

 > - There is some  stuff about non-IP payloads which, it  seems to me,
 >    clearly infringes upon  PWE3's scope.  The  stuff about non-IP
 >    payloads should be discussed with PWE3, and if  that stuff remains
 >    in the draft, then PWE3 needs to  review it.  However, since  PWE3
 >    has never  worked on multipoint PWs, I can imagine that this would
 >    be a very long discussion.

No. This discussion is intended to highlight that GMPLS includes
non-packet traffic types. It is not a discussion of carrying packetized
signals in pseudowires.

 > - The authors just can't resist putting in unnecessary plugs for
 >    RSVP-TE ;-)

We tried to remove them all, but maybe some slipped through.
Searching the document, I find one reference in 4.7 that can be removed.
Please note that the reference in section 5 is not an advertisement.
It is a warning!

 >Abstract
 >
 >    This document presents a set of requirements for the establishment
 >    and maintenance of Point-to-Multipoint (P2MP) Traffic Engineered (TE)
 >    Multiprotocol Label Switching (MPLS) Label Switched Paths (LSPs).
 >
 >    There is no intent to specify solution specific details nor
 >    application specific requirements in this document.
 >
 >    It is intended that the requirements presented in this document are
 >    not limited to the requirements of packet switched networks, but also
 >    encompass the requirements of Layer two Switching (L2SC), Time
 >    Division Multiplexing (TDM), lambda and port switching networks
 >    managed by Generalized MPLS (GMPLS) protocols. Protocol solutions
 >    developed to meet the requirements set out in this document must
 >    attempt to be equally applicable to MPLS and GMPLS.
 >
 >**** What is the  right group of people to review  this to determine
 >**** whether the  requirements  apply  to  all these  non-packet-switched
 >**** types of network?

The MPLS and CCAMP working groups have been invited to review the I-D.
And many of the authors are participants in both working groups.

 >**** This seems too vague and open-ended.

If you think "must attempt" is too vague and open-ended, please note
that this was said to cover the fact that it might turn out to be
impossible to have a single solution applicable to MPLS and GMPLS.

 >    [RFC2702] also explains how MPLS is particularly suited to traffic
 >    engineering, and presents the following eight reason.
 >
 >       1. Explicit label switched paths which are not constrained by
 >          the destination based forwarding paradigm can be easily
 >          created through manual administrative action or through
 >          automated action by the underlying protocols.
 >       2. LSPs can potentially be efficiently maintained.
 >       3. Traffic trunks can be instantiated and mapped onto LSPs.
 >
 >**** Applicability of this to P2MP is not obvious.

We can add an explanation at the end of the list (where we say that all of
the points are applicable).

 >    Note that there is a separation between routing and signaling in
 >    MPLS TE. In particular, the path of the MPLS TE LSP is determined by
 >    performing a constraint-based computation (such as CSPF) on a
 >    traffic engineering database (TED). The contents of the TED may be
 >    collected through a variety of mechanism - extensions to the IGPs
 >    are a popular mechanism for P2P MPLS TE.
 >
 >**** The last part of this sentence (following the dash might best be
 >**** omitted.)

Yes. Just because it is used widely, doesn't mean it is popular :-)

 >    A P2MP TE LSP will be set up with TE constraints and will allow
 >    efficient packet or data replication at various branching points in
 >    the network. Note that the notion of "efficient" packet replication
 >    is relative and may have different meanings depending on the
 >    objectives (see section 4.2).
 >
 >**** The notion of efficient packet replication is a data plane notion,
 >**** not a signaling  notion.  The document purports to  be setting
 >**** requirements only on  the control plane mechanisms
 >**** ("signaling requirements"), yet clearly there are some requirements
 >**** on  the data plane as well. This is an issue of scope that really
 >**** needs to be clarified.

If there are different data plane efficiencies and objectives
(I think this is always a case with TE), it is the responsiblity of the
control plane (signaling) to install the LSP so that the service
requirements are met.
For example, a common "efficiency" dictates that only one copy of any
packet flows down any link. This can be seen as an issue of efficient
packet replication - the packet must be replicated at the branch point
not further up the tree.

 >1.1 Non-Objectives
 >
 >**** This section  seems to talk about mechanisms rather than
 >**** requirements.(It seems like it might have been cut and pasted
 >**** from a solutions doc.)

No, it was written specially to address your concerns in a previous
review. If it seems to concentrate on mechansims, this may be because
mechanisms are out of scope, and this section lists the items that
are out of scope.

 >    P2MP-ID (Pid):
 >
 >**** "Pid" has another meaning, not the best acronym here.

Agreed.
Would P2ID be OK?

 >    This document does not restrict the choice of signaling protocol
 >    used to set up a P2MP TE LSP, but it should be noted that [RFC3468]
 >    states
 >      ... the consensus reached by the Multiprotocol Label Switching
 >    (MPLS) Working Group within the IETF to focus its efforts on
 >    "Resource Reservation Protocol (RSVP)-TE: Extensions to RSVP for
 >    Label-Switched Paths (LSP) Tunnels" (RFC 3209) as the MPLS signaling
 >    protocol for traffic engineering applications...
 >
 >**** This statement does not seem to be within the scope of this
 >**** document.

It is very strongly our opinion that this document should not restrict
the choice of signaling protocol.

 >    The P2MP TE LSP setup mechanism MUST include the ability to
 >    add/remove egress LSRs to/from an existing P2MP TE LSP and MUST
 >    allow for the support of all the TE LSP management procedures
 >    already defined for P2P TE LSP such as the non disruptive rerouting
 >    (the so called "Make before break" procedure).
 >
 >**** I  don't  see the  point  of  requiring support  for  "ALL  the TE
 >**** LSP management procedures  already defined for P2P", and  then
 >**** mentioning a single  example of  one.  It's  also not  clear what
 >****  "already defined" means or will be taken to mean  years from now
 >**** when folks read the RFC.
 >**** I'd say  either explicitly list the  things for which  support is
 >**** being required, or reference specific RFCs.

You make a good point, but inclusive lists are notoriously hard to
achieve. (Perhaps they require extra thought by the authors.)

If you would like to suggest a list we will see if we can include it.
Otherwise we suggest to change the text to read...

      and MUST
      allow for the support of all the TE LSP management procedures
      already defined for P2P TE LSP. Further, when new TE LSP procedures
      are developed for P2P TE LSPs equivalent or identical procedures
      SHOULD be developed for P2MP TE LSPs.

 >4.1 P2MP LSP tunnels
 >
 >    The P2MP TE extensions MUST be applicable to the signaling of LSPs
 >    of different traffic types. For example, it MUST be possible to
 >    signal a P2MP TE LSP to carry any kind of payload being packet or
 >    non-packet based (including frame, cell, TDM un/structured, etc.)
 >
 >**** I don't understand what is being required here.
 >**** As  far  as  I  know,  these  non-packet things  are  carried  by
 >**** PWE3 pseudowires, and  there hasn't  been much work  on multipoint
 >**** PWs. Is direct  MPLS   encapsulation  (no  PWE3)  of   non-packet
 >**** stuff being envisioned?  I don't see why that should be required.
 >**** Perhaps this requirement infringes upon the scope of the PWE3 group.

As above. GMPLS (non-packet) LSPs carry non-packet traffic types.
We will try to tidy up this text.

 >    One example is the Steiner P2MP tree (Cost minimum P2MP tree)
 >    [STEINER]. This P2MP tree is suitable for constructing a cost
 >    minimum P2MP tree so as to minimize the bandwidth consumption in
 >    the core. To realize this P2MP tree, several intermediate LSRs must
 >    be both MPLS data terminating LSRs and transit LSRs (LSRs E, F, G,
 >    H, I, J and K in the figure 4). This means that the LSRs must
 >    perform both label swapping and popping at the same time. Therefore,
 >    the P2MP TE solution MUST support a mechanism that can setup this
 >    kind of bud LSR between an ingress LSR and egress LSRs. Note that
 >    this includes constrained Steiner trees that allow for the
 >    computation of a minimal cost trees with some other constraints such
 >    as a bounded delay between the source and every receiver.
 >
 >**** It seems to me that this  is not a signaling requirement (or not
 >**** merely a signaling requirement) but also a requirement on the data
 >**** plane.

Yes, there is a requirement on the data plane if this type of tree (with
buds) is to be supported. Another requirement on the data plane is that it
supports MPLS labels.

We could probably safely remove the statement that support of buds
"...means that the LSRs must perform both label swapping and popping at
the same time" because there are other possible ways of achieving this in
the data plane. The principal requirement remains, as stated here, that
there is a requirement on the control plane to be able to signal LSPs that
have buds. Clearly, this would only be used when the data plane supports
buds.

 >4.3 Explicit Path Loose Hops and Widely Scoped Abstract Nodes
 >
 >    A P2MP tree is completely specified if all of the required branches
 >    and hops between a sender and leaf LSR are indicated.
 >
 >    A P2MP tree is partially specified if only a subset of intermediate
 >    branches and hops are indicated. This may be achieved using loose
 >    hops in the explicit path, or using widely scoped abstract nodes
 >    such as IPv4 prefixes shorter than 32 bits, or AS numbers.
 >
 >**** I  think the  notion  of "widely  scoped  abstract nodes"  needs to
 >**** be defined and explicated.

OK. Do you feel that the examples of IPv4 prefixes shorter than 32 bits
and AS numbers are unclear, or you feel that we should give a more precise
definition or a reference to RFC3209?

 >    Protocol solutions MUST allow the P2MP tree to be completely
 >    specified at the ingress where sufficient information exists to
 >    allow the full tree to be computed.
 >
 >**** The usefulness  of complete specification  at the ingress  depends
 >**** both upon the  presence of  sufficient information and  on the
 >**** existence of suitable policies along the path.

Agreed.

 >    In all cases, the egress nodes of the P2MP TE LSP must be fully
 >    specified.
 >
 >**** Why?

We will clarify. I assume you have in mind a TE LSP that targets a group
address.

 >    In case of a tree being computed by some downstream LSRs (e.g. the
 >    case of hops specified as loose hops), the solution MUST provide
 >    the ability for the ingress LSR of the P2MP TE LSP to learn the full
 >    P2MP tree. Note that this requirement MAY be relaxed in some
 >    environments (e.g. Inter-AS) where confidentiality must be
 >    preserved.
 >
 >**** I think that what paragraph is  trying to say is that the solution
 >**** MUST provide protocol to enable the  ingress to learn the full
 >**** specification of the tree, but that this information may not always
 >**** be obtainable due to  policy  considerations.   But  then  perhaps
 >**** it  should  pose some requirement as to what happens in  cases where
 >**** some of the path remains confidential.

Yes, your interpretation is correct, and your wording is helpful.

We will add a note that where part of the path remains confidential it
MUST be reported through aggregation (for example, using an AS number).

 >4.4 P2MP TE LSP establishment, teardown, and modification mechanisms
 >
 >    The P2MP TE solution MUST support establishment, maintenance and
 >    teardown of P2MP TE LSPs in a scalable manner. This MUST include
 >    both the existence of very many LSPs at once, and the existence of
 >    very many destinations for a single P2MP LSP.
 >
 >**** This would  seem to  preclude any mechanism  in which add/removes
 >**** of a leaf  requires a  message  to be  sent  to the  ingress.
 >**** Is that the intention?  Or  is the "no churn"  presumption supposed
 >**** to rule out of scope any conditions under which this is not scalable?

Actually the presumption is for linear scalability.
However, if certain mechanisms are precluded by this requirement, then
that is the point exposing the requirement.

 >    In addition to P2MP TE LSP establishment and teardown mechanism, it
 >    SHOULD implement partial P2MP tree modification mechanism.
 >
 >**** I  don't  believe  that   the  term  "partial  P2MP  tree
 >**** modification mechanism" has been defined yet.

Section 2.2.1 describes partial P2MP trees.

 >    For the purpose of adding sub-P2MP TE LSPs to an existing P2MP TE
 >    LSP, the extensions SHOULD support a grafting mechanism. For the
 >    purpose of deleting a sub-P2MP TE LSPs from an existing P2MP TE LSP,
 >    the extensions SHOULD support a pruning mechanism.
 >
 >    It is RECOMMENDED that these grafting and pruning operations do not
 >    cause any additional processing in nodes except along the path to
 >    the grafting and pruning node and its downstream nodes. Moreover,
 >    both grafting and pruning operations MUST not be traffic disruptive
 >    for the traffic currently forwarded along the P2MP tree.
 >
 >**** I'm  having a little  trouble putting  together "scalable  grafting
 >**** and pruning" with "explicit routing".

Yes, this is a complex subject. There is no assumption that the explicitly
routed P2MP LSP remains on an optimal path after several grafts and prunes
have occurred. Scalable refers to the signaling process within the context
of the P2MP TE LSP. The TE nature of the LSP allows that re-optimization
may take place from time to time.

 >4.5 Fragmentation
 >
 >    The P2MP TE solution MUST handle the situation where a single
 >    protocol message cannot contain all of the information necessary to
 >    signal the establishment of the P2MP LSP. It MUST be possible to
 >    establish the LSP in these circumstances.
 >
 >    This situation may arise in either of the following circumstances.
 >      a. The ingress LSR cannot signal the whole tree in a single
 >         message.
 >
 >      b. The information in a message expands to be too large (or is
 >         discovered to be too large) at some transit node. This may
 >         occur because of some increase in the information that needs
 >         to be signaled or because of a reduction in the size of
 >         signaling message that is supported.
 >
 >    The solution to these problems SHOULD NOT rely on IP fragmentation,
 >    it is RECOMMENDED to rely on some protocol procedures specific to
 >    the signaling solution.
 >
 >    It is NOT RECOMMENDED that fragmented protocol messages are
 >    re-combined at any downstream LSR.
 >
 >**** The term  is "reassembled", not  "recombined".  If IP  fragmentation
 >**** of protocol messages is  allowed, it hardly makes sense  to recommend
 >**** that the fragments not be reassembled!
 >**** I think  what this section is trying  to do is to  prohibit the
 >**** control messages from ever being fragmented, instead requiring that the
 >**** control protocol always generate messages that  will not need to be
 >**** fragmented.
 >**** If so, that should be made clear.  However, it's not completely
 >**** obvious to me that this requirement can be met in all environments.

Yes. Thank you for the clarifying words.
This requirement follows the requirements expressed in RFC3209. In
particular, it notes the problems for a protocol when one of the IP
fragments is lost.

 >**** With regard to the  recommendation that there be some
 >**** protocol-specific procedure  to allow  arbitrarily  large control
 >**** messages without ever incurring  IP  fragmentation, I  wonder  if
 >**** this is all really just infringing on the solution space.

Yes. it is very close.
Perhaps we can rephrase this requirement in terms of absolute requirements
without thinking about the solution.

 >    Note that application-specific requirement documents may introduce
 >    even more stringent requirement, such as no packet loss, as a
 >    trade-off for the relaxation of other requirements, such as increased
 >    bandwidth consumption.
 >
 >**** I  don't  understand whether  the  solution  is  supposed to  meet
 >**** the unspecified  requirements  of  future application-specific
 >**** requirement documents or  not.  If not,  why mention it?
 >****  But then does  that mean that  different  applications  are  going
 >****  to  end  up  with different solutions?

We will clarify.
The intention is to indicate that the requirements expressed here are
general and that the solution that meets them will therefore be general.
Specific applications may have more additional requirements or may want to
relax some other requirements. this may lead to variations in the
solution.

 >    The solution SHOULD also support the ability to meet other network
 >    recovery requirements such as bandwidth protection and bounded
 >    propagation delay increase along the backup path during failure.
 >
 >    A P2MP TE solution MUST support P2MP fast protection mechanism to
 >    handle P2MP applications sensitive to traffic disruption.
 >
 >    The report of the failure of delivery to fewer than all of the
 >
 >**** Report?  What report?

Any report.

We will tidy the text. We cannot say "failure to deliver" because it is
the knowledge that we have failed to delliver that is important.

 >    egress nodes SHOULD NOT cause automatic teardown of the P2MP TE LSP.
 >    That is, while some egress nodes remain connected to the P2MP tree
 >    it should be a matter of local policy at the ingress whether the
 >    P2MP LSP is retained.
 >
 >**** That would  be some  protocol that stopped  an entire  multicast
 >**** stream because one egress goes down!

No. That would be a local policy that determined whether the service was
to deliver to a certain percentage of the intended recipients. This is why
the text says "SHOULD NOT cause automatic teardown" etc.

Note that the final "should" should be "SHOULD".

 >    When all egress nodes downstream of a branch node have become
 >    disconnected from the P2MP tree, and the some branch node is unable
 >    to restore connectivity to any of them by means of some recovery or
 >    protection mechanisms, the branch node MAY remove itself from the
 >    P2MP tree provided that it is not also an egress LSR. Since the
 >    faults that severed the various downstream egress nodes from the
 >    P2MP tree may be disparate, the branch node MUST report all such
 >    errors to its upstream neighbor. The ingress node can then decide
 >    to re-compute the path to those particular egress nodes, around the
 >    failure point.
 >
 >**** I  don't  understand just  what  is  being  required here.   Any
 >**** fault anywhere downstream  is supposed to  be reported to the
 >**** ingress? That seems to have scaling problems.

Nevertheless, this is in inherrant in TE. There may be a choice to be made
when deciding whether to deploy TE, but that is out of scope.

 >4.7 Record route of P2MP TE LSP tunnels
 >
 >    Being able to identify the established topology of P2MP TE LSP is
 >    very important for various purposes such as management and operation
 >    of some local recovery mechanisms like Fast Reroute [FRR]. A network
 >    operator uses this information to manage P2MP TE LSPs. Therefore,
 >    topology information MUST be collected and updated after P2MP TE LSP
 >    establishment and modification process.
 >
 >    The P2MP TE solution MUST support a mechanism which can collect and
 >    update P2MP tree topology information after P2MP LSP establishment
 >    and modification process. For example, the P2P MPLS TE mechanism of
 >    route recording could be extended and used if RSVP-TE was used as
 >    the P2MP signaling protocol.
 >
 >**** There's  a big  difference between  saying that  topology info  MUST
 >**** be collected and saying  that the solution MUST support  a mechanism
 >**** which can collect  topology information.  Which  is intended?  How  does
 >**** this interact with a  previous requirement stating some of  the topology
 >**** may be confidential?   Fragmentation and scaling issues  also seem
 >**** relevant here.

The intent is to say that the solution MUST include a mechanism to gather
and report the topology of the tree.

The confidentiality issue is handled in the note at the top of this email.

Fragmentation and scaling issues are indeed a concern and the solution
should address them. We can add this point.

 >    It is RECOMMENDED that the information is collected in a data format
 >    by which the sender node can recognize the P2MP tree topology
 >    without involving some complicated data calculation process.
 >
 >**** I'm not sure this is an objective requirement.  I'd hate to have the
 >**** WG deciding whether some calculation process is "complicated" or not.

:-)

We will rephrase.

 >4.12 Data Duplication
 >
 >    Data duplication refers to the receipt by any recipient of duplicate
 >    instances of the data. In a packet environment this means the
 >    receipt of duplicate packets - although this should be a benign (if
 >    inefficient) situation, it may be catastrophic in certain existing
 >    and deployed applications.
 >
 >**** It's hard to see how  duplication is both benign and catastrophic.
 >**** The suggestion here is that a benign situation is turned into a
 >**** catastrophe by  stupid applications.   But  I don't  think  that the
 >**** long-standing duplication of  a multicast  stream is just  benign;
 >**** it could  be quite damaging to the endsystems, no matter what the
 >**** application.

It is not suggested that it is both benign and catastrophic. It is
suggested that it should be benign, but may be catastrophic. It is not our
intention to be judgmental about any applications, simply to report on the
consequences.

You are correct that we should draw a distinction between the duplication
of a few packets (which can still wreck some applications) and long-term
duplication which may be damaging if for no other reason than it causes
additional pressures on throughput.

 >    Instead, the solution:
 >    - SHOULD limit to transitory conditions only the cases where
 >      network bandwidth is wasted by the existence of duplicate
 >      delivery paths.
 >    - MUST limit the cases of delivery of duplicate data to an
 >      application to error cases or rare transitory conditions.
 >
 >**** Limit duplication to error cases?  What does that mean exactly?

We cannot regulate against error cases.

 >**** Whether  a  transitory  condition  is  "rare"  or  not  is
 >**** notoriously difficult  to predict.   It might  be  better to
 >**** say that duplication should only occur in transitory  conditions
 >**** and only if the duration of the transitory condition can be
 >**** suitably bounded.

Agreed.

 >4.13 IPv4/IPv6 support
 >
 >    Any P2MP TE solution MUST be equally applicable to IPv4 and IPv6.
 >
 >**** It might be  better just to say that any P2MP  TE solution MUST
 >**** support both IPv4 and  IPv6.  I don't like the term
 >**** "equally applicable", if a solution supports  v4 and v6
 >****  I don't want  to have to ask  whether the
 >****  solution is also "equally applicable" to each of them.

Yes.

 >4.14 P2MP MPLS Label
 >
 >    A P2MP TE solution MUST support establishment of both P2P and P2MP
 >    TE LSPs and MUST NOT impede the operation of P2P TE LSPs within the
 >    same network.
 >
 >**** This could  be interpreted as meaning  that P2P TE  LSPs get
 >**** preference for the bandwidth.  I don't think that is what is meant.

Agreed.

 >    A P2MP TE solution MUST be specified in such a way
 >    that it allows P2MP and P2P TE LSPs to be signaled on the same
 >    interface. Labels for P2MP TE LSPs and P2P TE LSPs MAY be assigned
 >    from shared or dedicated label space(s). Label space shareability is
 >    implementation specific.
 >
 >**** I would  strike the last  two sentences.  They  may end up  coming
 >**** into conflict with  the two recent  drafts on MPLS multicast
 >**** codepoints and upstream-assigned labels.  The  issues of multicast/
 >**** unicast codepoints, label spaces, etc.,  are really left up to
 >**** the solutions. Whether they are implementation-specific  is a
 >**** solutions issue, not  a requirements issue.

The two recent drafts on MPLS multicast do not mention TE.
Nevertheless, there seems to be no great need to make this point and we
can drop it if that makes people more comfortable.

 >4.15 Routing advertisement of P2MP capability
 >
 >    Several high-level requirements have been identified to determine
 >    the capabilities of LSRs within a P2MP network. The aim of such
 >    information is to facilitate the computation of P2MP trees using TE
 >    constraints within a network that contains LSRs that do not all have
 >    the same capabilities levels with respect to P2MP signaling and data
 >    forwarding.
 >
 >    These capabilities include, but are not limited to:
 >
 >    - the ability of an LSR to support branching.
 >    - the ability of an LSR to act as an egress and a branch for the
 >      same LSP.
 >    - the ability of an LSR to support P2MP MPLS-TE signaling.
 >
 >    It is expected that it may be appropriate to gather this information
 >    through extensions to TE IGPs (see [RFC3630] and [IS-IS-TE]), but
 >    the precise requirements and mechanisms are out of the scope of this
 >    document. It is expected that a separate document will cover this
 >    requirement.
 >
 >**** I  don't like when  a document  says, "we  don't require  X, but  it
 >**** is expected  that  X will  be  required  by  some other  document".
 >**** This paragraph should just be stricken.

Agreed. This implies a preference for a particular solution and should not
be present.

 >4.16 Multi-Area/AS LSP
 >
 >    P2MP TE solutions SHOULD support multi-area/AS P2MP TE LSPs.
 >
 >    The precise requirements in support of multi-area/AS P2MP TE LSPs is
 >    out of the scope of this document. It is expected that a separate
 >    document will cover this requirement.
 >
 >**** Again, it makes no sense to say that the solutions are required to
 >**** meet some requirements which are going to be specified in some
 >**** document that doesn't exist yet.  This section is out of scope
 >**** and should be stricken.

Yes. You make a good point.
We came under strong pressure to say something about multi-area and
multi-AS support because this is of great interest to several people at
the moment. But this is a non-objective and should be removed.

 >4.17 Multi-access LANs
 >
 >    P2MP MPLS TE may be used to traverse network segments that are
 >    provided by multi-access media such as Ethernet. In these cases, it
 >    is also possible that the entry point to the network segment is a
 >    branch point of the P2MP LSP.
 >
 >    Two options clearly exist:
 >
 >     - the branch point replicates the data and transmits multiple
 >       copies onto the segment
 >     - the branch point sends a single copy of the data to the segment
 >       and relies on the exit points to discriminate the reception of
 >       the data.
 >
 >    The first option has a significant scaling issue since all
 >    replicated data must be sent through the same port and carried on
 >    the same segment. Thus, a solution SHOULD provide a mechanism for a
 >    branch node to send a single copy of the data onto a multi-access
 >    network and reach multiple (adjacent) downstream nodes.
 >
 >**** The second option also has  scaling issues, of course.  In addition,
 >**** it has some label distribution issues.

Yes. The issues for the second option are control plane rather than data
link issues. We will add a note.

 >4.18 P2MP MPLS OAM
 >
 >    Management of P2MP LSPs is as important as the management of P2P
 >    LSPs.
 >
 >**** This sentence has no real content and should be stricken.

Does it do any harm?

 >4.19 Scalability
 >
 >    Scalability is a key requirement in P2MP MPLS systems. Solutions
 >    MUST be designed to scale well with an increase in the number of any
 >    of the following:
 >
 >    - the number of recipients
 >    - the number of branch points
 >    - the number of branches.
 >
 >    Both scalability of performance and operation MUST be considered.
 >
 >**** I think I know what "performance" is, but what is "operation"?

We will clarify.

 >    Key considerations MUST include:
 >    - the amount of refresh processing associated with maintaining
 >      a P2MP TE LSP.
 >    - the amount of protocol state that must be maintained by ingress
 >      and transit LSRs along a P2MP tree.
 >    - the number of protocol messages required to set up or tear down a
 >      P2MP LSP as a function of the number of egress LSRs.
 >    - the number of protocol messages required to repair a P2MP LSP
 >      after failure or perform make-before-break.
 >    - the amount of protocol information transmitted to manage
 >      a P2MP TE LSP (i.e. the message size).
 >    - the amount of potential routing extensions.
 >
 >**** I think  this should be  the "number" of potential  routing
 >**** extensions, but I'm not really sure what is meant here.

It is meant to be quantative, but a simple count of the number of
extensions. One new TLV in the TE opaque LSA is only one extension, but if
it carries 10,000 bytes for each TE link it is going to be a problem.

We will attempt to clarify.

 >4.19.1 Absolute Limits
 >
 >    In order to achieve the best solution for the problem space it is
 >    helpful to clarify the boundaries for P2MP TE LSPs.
 >
 >    - Number of recipients.
 >      A P2MP TE LSP MUST reduce to similar scaling properties as a P2P
 >      LSP when the number of recipients reduces to one.
 >
 >**** I don't see why.  Frankly, who cares about the case of  a P2MP LSP
 >**** with one recipient?

I think many people will care.
But this is not the point. The intention is to describe the limiting bound
on the scaling.

 >      It is important to classify the problem as a Traffic Engineering
 >      problem.
 >
 >**** I don't  understand.  What problem?   Why is it important  "to
 >**** classify the problem"?
 >
 >      It is anticipated that the initial deployments of P2MP TE
 >      LSPs will be limited to a maximum of around a hundred recipients,
 >      but that medium term deployments may increase this to several
 >      hundred, and that future deployments may require significantly
 >      larger numbers.
 >
 >      An acceptable solution, therefore, is one that scales linearly
 >      with the number of recipients.
 >
 >**** If you don't  do any P2MP stuff at all, but  have the ingress
 >**** replicate everything  and do  one P2P  tunnel per  receipient,
 >**** then  doesn't that scale linearly with the number of recipients?
 >**** Generally you want your multicast-specific stuff to scale much
 >**** better than that.

This sets a top bound.
You are right that a target is better than linear in typical network
topologies.
We will add text.

 >      Solutions that scale worse than linear (that is, exponential or
 >      polynomial) are not acceptable whatever the number of recipients
 >      they could support
 >
 >**** missing text?

Missing period.

 >    - Number of branch points.
 >      Solutions MUST support all possibilities from one extreme of a
 >      single branch point that forks to all leaves on a separate branch,
 >      to the greatest number of branch points which is (n-1) for n
 >      recipients. Assumptions MUST NOT be made in the solution regarding
 >      which topology is more common, and the solution MUST be designed
 >      to ensure scalability in all topologies.
 >
 >**** Is this saying  that the scalability must be  independent of the
 >**** number of  branch points?  Given  the need  to replicate  the packet
 >**** for each branch point, that seems unlikely.  (The number of packet
 >**** transmissions scales better with more branch points, for example.)

Recall that this is a signaling requirements draft, and not data plane
requirements.
It says that the solution (i.e. the signaling) must scale well in all
branching scenarios. Since the branching scenario is a function of the
path computation algorithm and constraints, control of the branch points
is outside the control of the signaling protocol.

 >    - Dynamics of P2MP tree.
 >      Recall that the mechanisms for determining which recipients should
 >      be added to an LSP, and for adding and removing recipients from
 >      that group are out of the scope of this document. Nevertheless, it
 >      is useful to understand the expected rates of arrival and
 >      departure of recipients since this can impact the selection of
 >      solution techniques.
 >      Again, it must be recalled that this document is limited to
 >      Traffic Engineering, and in this model the rate of change of
 >      recipients may be expected to be lower than in an IP multicast
 >      group.
 >
 >**** I'd like  to see  an explanation of  this point.   I wonder if  this
 >**** is anything more than a tautology, in the  sense that if you have a
 >**** lot of churn traffic engineering is too inefficient to use.

This is correct some case. Although it can be seen the other way around
as well. For example, you do not typically have a high rate of churn in
VPN sites.
And I do not think TE is inefficient for a lot of churn traffic because
I could assume some SPs would introduce P2MP TE within a backbone area
to stabilize the network by terminating such a churn at the area boundary
with expecting enough Bandwidth in the backbone.

 >**** This is rather late in the document  to be saying, "by the way, none
 >**** of these requirements apply in some  of the most common cases of
 >**** multicast use".

We were requested to remove all comparisons with IP multicast.
I think it is widely understood that core Traffic Engineering is not
applicable to individual IP flows, but we will highlight this point in the
Introduction.

 >      Although the absolute number of recipients coming and going is the
 >      important element for determining the scalability of a solution,
 >      it may be noted that a percentage may be a more comprehensible
 >      measure but that this is not as significant for LSPs with a small
 >      number of recipients.
 >      A working figure for an established P2MP TE LSP is less than 10%
 >      churn per day. That is, a relatively slow rate of churn.
 >      We could say that a P2MP LSP would be shared by multiple multicast
 >      groups and dynamics of P2MP LSP would be relatively small.
 >      Considering applicability that P2MP LSP to use L2 multi-access
 >      path technology, we can consider stable P2MP L2 path even when we
 >      transfer IP multicast traffic over the path.
 >
 >      Solutions MUST optimize around such relatively low rates of change
 >      and are NOT REQUIRED to optimize for significantly higher rates
 >      of change.
 >
 >**** In light of this, I think  that "low rate of churn" should be
 >**** mentioned at the front of the document  as one of the fundamental
 >**** assumptions for traffic  engineering.   This  would   then  provide
 >**** some  context for statements like "there  has to be a scalable
 >**** method for adding/pruning receivers", which otherwise seem questionable.

I agreed to incorporate some text to explain our optimization policy
(we are much interested in optimizing solution to have important
P2MP TE features and not interested in optimizing solution to handle
high rate churn traffic at the price of decreasing these feature)
the front of the document. See top of this email.

 >4.21 GMPLS
 >
 >    The requirement for P2MP services for non-packet switch interfaces
 >    is similar to that for PSC interfaces. Therefore, it is a requirement
 >
 >**** This is  the first occurrence  of the term  "PSC" in this  document.I
 >**** wonder what it means.

"PSC" stands for Packet-Switch Capable.

 >    that reasonable attempts must be made to make all the features/
 >    mechanisms (and protocol extensions) that will be defined to provide
 >    MPLS P2MP TE LSPs equally applicable to P2MP PSC and non-PSC TE-LSPs.
 >    If the requirements of non-PSC networks over-complicate the PSC
 >    solution a decision may be taken to separate the solutions. This
 >    decision must be taken in full consultation with the MPLS and CCAMP
 >    working groups.
 >
 >**** Requirements on  the decision making  process are out of  scope,
 >**** aren't they?

We can remove this text if you like.
Looking at our notes, it was added becuase of a comment you made in your
previous review.

Eric> For all
Eric> I know,  GMPLS might introduce  a variety of  additional
Eric> complications that aren't necessary in MPLS.  If this is so,
Eric> I would not necessarily support using the same solution in
Eric> both cases.

We have added text to indicate that this separation may be necessary, and
also to describe the process needed to make the separation.

 >4.22 Requirements for Hierarchical P2MP TE LSPs
 >
 >    [LSP-HIER] defines concepts and procedures for P2P LSP hierarchy.
 >
 >    The P2MP MPLS-TE solution SHOULD support the concept of region and
 >    region hierarchy (PSC1<PSC2<PSC3<PSC4<L2SC<TDM<LSC<FSC).
 >
 >    Particularly it SHOULD allow a Region i P2MP TE LSP to be nested
 >    into a region j P2MP TE LSP or multiple region j P2P TE LSPs,
 >    providing that i<j.
 >
 >    The precise requirements and mechanisms for this function are out of
 >    the scope of this document. It is expected that a separate document
 >    will cover these requirements.
 >
 >**** I think  this section  should be removed,  as it  provides
 >**** insufficient information to determine what the requirements are.
 >**** You can't put in a requirement saying  that the solution  must meet
 >**** the  requirements that are to be specified in some other document
 >****  which is forthcoming.

Agreed to defer the discussion of this requirement if other also want to
remove it.

 >4.23 P2MP Crankback routing
 >
 >    P2MP solutions SHOULD support crankback requirements as defined in
 >    [CRANKBACK]. In particular, they SHOULD provide sufficient
 >    information to a branch LSR from downstream LSRs to allow the branch
 >    LSR to re-route a sub-tree around any failures or problems in the
 >    network.
 >
 >**** I think  this section  either needs to  be omitted  or else made  a
 >**** lotlonger.

There doesn't seem to be anything else to say unless we repeat the
(protocol independent) requirements from another I-D.

Thank you once again for spending time and supporting this I-D. We will
include your ideas in a new revision when the last call completes.

Best Regards,
Seisho 

_______________________________________________
mpls mailing list
mpls@lists.ietf.org
https://www1.ietf.org/mailman/listinfo/mpls