Home Articles FAQs XREF Games Software Instant Books BBS About FOLDOC RFCs Feedback Sitemap
irt.Org

Request For Comments - RFC3754

You are here: irt.org | RFCs | RFC3754 [ previous next ]






Network Working Group                                           R. Bless
Request for Comments: 3754                            Univ. of Karlsruhe
Category: Informational                                        K. Wehrle
                                                      Univ. of Tuebingen
                                                              April 2004


         IP Multicast in Differentiated Services (DS) Networks

Status of this Memo

   This memo provides information for the Internet community.  It does
   not specify an Internet standard of any kind.  Distribution of this
   memo is unlimited.

Copyright Notice

   Copyright (C) The Internet Society (2004).  All Rights Reserved.

Abstract

   This document discusses the problems of IP Multicast use in
   Differentiated Services (DS) networks, expanding on the discussion in
   RFC 2475 ("An Architecture of Differentiated Services").  It also
   suggests possible solutions to these problems, describes a potential
   implementation model, and presents simulation results.

Table of Contents

   1.  Introduction . . . . . . . . . . . . . . . . . . . . . . . . .  2
       1.1.  Management of Differentiated Services. . . . . . . . . .  2
   2.  Problems of IP Multicast in DS Domains . . . . . . . . . . . .  3
       2.1.  Neglected Reservation Subtree Problem (NRS Problem). . .  4
       2.2.  Heterogeneous Multicast Groups . . . . . . . . . . . . . 12
       2.3.  Dynamics of Any-Source Multicast . . . . . . . . . . . . 13
   3.  Solutions for Enabling IP-Multicast in Differentiated
       Services Networks. . . . . . . . . . . . . . . . . . . . . . . 13
       3.1.  Solution for the NRS Problem . . . . . . . . . . . . . . 13
       3.2.  Solution for Supporting Heterogeneous Multicast Groups . 15
       3.3.  Solution for Any-Source Multicast. . . . . . . . . . . . 16
   4.  Scalability Considerations . . . . . . . . . . . . . . . . . . 16
   5.  Deployment Considerations. . . . . . . . . . . . . . . . . . . 17
   6.  Security Considerations. . . . . . . . . . . . . . . . . . . . 18
   7.  Implementation Model Example . . . . . . . . . . . . . . . . . 18
   8.  Proof of the Neglected Reservation Subtree Problem . . . . . . 19
       8.1.  Implementation of the Proposed Solution. . . . . . . . . 20
       8.2.  Test Environment and Execution . . . . . . . . . . . . . 21
   9.  Simulative Study of the NRS Problem and Limited Effort PHB . . 23



Bless & Wehrle               Informational                      [Page 1]



RFC 3754              IP Multicast in DS Networks             April 2004


       9.1.  Simulation Scenario. . . . . . . . . . . . . . . . . . . 24
       9.2.  Simulation Results for Different Router Types. . . . . . 26
   10. Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . 31
   11. References . . . . . . . . . . . . . . . . . . . . . . . . . . 31
       11.1. Normative References . . . . . . . . . . . . . . . . . . 31
       11.2. Informative References . . . . . . . . . . . . . . . . . 31
   12. Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . 33
   13. Full Copyright Statement . . . . . . . . . . . . . . . . . . . 34

1.  Introduction

   This document discusses the problems of IP Multicast use in
   Differentiated Services (DS) networks, expanding on the discussion in
   RFC 2475 ("An Architecture of Differentiated Services").  It also
   suggests possible solutions to these problems, describes a potential
   implementation model, and presents simulation results.

   The "Differentiated Services" (DiffServ or DS) approach [1, 2, 3]
   defines certain building blocks and mechanisms to offer qualitatively
   better services than the traditional best-effort delivery service in
   an IP network.  In the DiffServ Architecture [2], scalability is
   achieved by avoiding complexity and maintenance of per-flow state
   information in core nodes, and by pushing unavoidable complexity to
   the network edges.  Therefore, individual flows belonging to the same
   service are aggregated, thereby eliminating the need for complex
   classification or managing state information per flow in interior
   nodes.

   On the other hand, the reduced complexity in DS nodes makes it more
   complex to use those "better" services together with IP Multicast
   (i.e., point-to-multipoint or multipoint-to-multipoint
   communication).  Problems emerging from this fact are described in
   section 2.  Although the basic DS forwarding mechanisms also work
   with IP Multicast, some facts have to be considered which are related
   to the provisioning of multicast resources.  It is important to
   integrate IP Multicast functionality into the architecture from the
   beginning, and to provide simple solutions for those problems that
   will not defeat the already gained advantages.

1.1.  Management of Differentiated Services

   At least for Per-Domain Behaviors and services based on the EF PHB,
   admission control and resource reservation are required [14, 15].
   Installation and updating of traffic profiles in boundary nodes is
   necessary.  Most network administrators cannot accomplish this task
   manually, even for long term service level agreements (SLAs).
   Furthermore, offering services on demand requires some kind of
   signaling and automatic admission control procedures.



Bless & Wehrle               Informational                      [Page 2]



RFC 3754              IP Multicast in DS Networks             April 2004


   However, no standardized resource management architecture for
   DiffServ domains exists.  The remainder of this document assumes that
   at least some logical resource management entity is available to
   perform resource-based admission control and allotment functions.
   This entity may also be realized in a distributed fashion, e.g.,
   within the routers themselves.  Detailed aspects of the resource
   management realization within a DiffServ domain, as well as the
   interactions between resource management and routers or end-systems
   (e.g., signaling for resources), are out of scope of this document.

   Protocols for signaling a reservation request to a Differentiated
   Services Domain are required.  For accomplishing end-system signaling
   to DS domains, RSVP [4] may be used with new DS specific reservation
   objects [5].  RSVP provides support for multicast scenarios and is
   already supported by many systems.  However, application of RSVP in a
   DiffServ multicast context may lead to problems that are also
   described in the next section. The NSIS Working Group is currently
   defining new signaling protocols that may show a different behavior,
   but the WG has its current focus more on unicast flows than on
   multicast flows.

2.  Problems of IP Multicast in DS Domains

   Although potential problems and the complexity of providing multicast
   with Differentiated Services are considered in a separate section of
   [2], both aspects have to be discussed in greater detail.  The
   simplicity of the DiffServ Architecture and its DS node types is
   necessary to reach high scalability, but it also causes fundamental
   problems in conjunction with the use of IP Multicast in DS domains.
   The following subsections describe these problems for which a generic
   solution is proposed in section 3.  This solution is as scalable as
   IP Multicast and needs no resource separation by using different
   codepoint values for unicast and multicast traffic.

   Because Differentiated Services are unidirectional by definition, the
   point-to-multipoint communication is also considered as
   unidirectional.  In traditional IP Multicast, any node can send
   packets spontaneously and asynchronously to a multicast group
   specified by their multicast group address, i.e., traditional IP
   Multicast offers a multipoint-to-multipoint service, also referred to
   as Any-Source Multicast.  Implications of this feature are discussed
   in section 2.3.

   For subsequent considerations we assume, unless stated otherwise, at
   least a unidirectional point-to-multipoint communication scenario in
   which the sender generates packets which experience a "better" Per-
   Hop-Behavior than the traditional default PHB, resulting in a service
   of better quality than the default best-effort service.  In order to



Bless & Wehrle               Informational                      [Page 3]



RFC 3754              IP Multicast in DS Networks             April 2004


   accomplish this, a traffic profile corresponding to the traffic
   conditioning specification has to be installed in the sender's first
   DS-capable boundary node.  Furthermore, it must be assured that the
   corresponding resources are available on the path from the sender to
   all the receivers, possibly requiring adaptation of traffic profiles
   at involved domain boundaries.  Moreover, on demand resource
   reservations may be receiver-initiated.

2.1.  Neglected Reservation Subtree Problem (NRS Problem)

   Typically, resources for Differentiated Services must be reserved
   before they are used.  But in a multicast scenario, group membership
   is often highly dynamic, thereby limiting the use of a sender-
   initiated resource reservation in advance.  Unfortunately, dynamic
   addition of new members of the multicast group using Differentiated
   Services can adversely affect other existing traffic if resources
   were not explicitly reserved before use.  A practical proof of this
   problem is given in section 8.

   IP Multicast packet replication usually takes place when the packet
   is handled by the forwarding core (cf. Fig. 1), i.e., when it is
   forwarded and replicated according to the multicast forwarding table.
   Thus, a DiffServ capable node would also copy the content of the DS
   field [1] into the IP packet header of every replicate.
   Consequently, replicated packets get exactly the same DS codepoint
   (DSCP) as the original packet, and therefore experience the same
   forwarding treatment as the incoming packets of this multicast group.
   This is also illustrated in Fig. 1, where each egress interface
   comprises functions for (BA-) classification, traffic conditioning
   (TC), and queueing.

            Interface A        IP Forwarding        Interface B
           +-----------+     +--------------+      +-----------+
   MC-flow |           |     | replication  |      |  egress   |
      ---->|  ingress  |---->|------+-------|----->|(class.,TC,|---->
           |           |     |      |       |      | queueing) |
           +-----------+     |      |       |      +-----------+
                             |      |       |
                             |      |       |       Interface C
                             |      |       |      +-----------+
                             |      |       |      |  egress   |
                             |      +-------|----->|(class.,TC,|---->
                             |              |      | queueing) |
                             +--------------+      +-----------+

        Figure 1: Multicast packet replication in a DS node





Bless & Wehrle               Informational                      [Page 4]



RFC 3754              IP Multicast in DS Networks             April 2004


   Normally, the replicating node cannot test whether a corresponding
   resource reservation exists for a particular flow of replicated
   packets on an output link (i.e., its corresponding interface).  This
   is because flow-specific information (e.g., traffic profiles) is
   usually not available in every boundary and interior node.

   When a new receiver joins an IP Multicast group, a multicast routing
   protocol (e.g., DVMRP [6], PIM-DM [7] or PIM-SM [8]) grafts a new
   branch to an existing multicast tree in order to connect the new
   receiver to the tree.  As a result of tree expansion, missing per-
   flow classification, and policing mechanisms, the new receiver will
   implicitly use the service of better quality, because of the "better"
   copied DSCP.

   If the additional amount of resources which are consumed by the new
   part of the multicast tree are not taken into account by the domain
   resource management (cf. section 1.1), the currently provided quality
   of service of other receivers (with correct reservations) will be
   affected adversely or even violated.  This negative effect on
   existing traffic contracts by a neglected resource reservation -- in
   the following designated as the Neglected Reservation Subtree Problem
   (NRS Problem) -- must be avoided under all circumstances.  Strong
   admission control policies at the domain boundary will not help to
   prevent this problem either, because the new flow that inadmissibly
   consumes resources has its origin inside the domain.

   One can distinguish two major cases of the NRS Problem.  They show a
   different behavior depending on the location of the branching point.
   In order to compare their different effects, a simplistic example of
   a share of bandwidth is illustrated in Fig. 2 and is used in the
   following explanations.  Neither the specific PHB types nor their
   assigned bandwidth share are important; however, their relative
   priority with respect to each other is of importance.

             40%                 40%               20%
   +--------------------+---------------------+------------+
   |Expedited Forwarding| Assured Forwarding  | Best-Effort|
   +--------------------+---------------------+------------+
   ---------------------------------------------------------->
                                      output link bandwidth

        Figure 2: An example bandwidth share of different
                  behavior aggregates








Bless & Wehrle               Informational                      [Page 5]



RFC 3754              IP Multicast in DS Networks             April 2004


   The bandwidth of the considered output link is shared by three types
   of services (i.e., by three behavior aggregates): Expedited
   Forwarding, Assured Forwarding, and the traditional Best-Effort
   service.  In this example, we assume that routers perform strict
   priority queueing, where EF has the highest, AF the middle, and
   Best-Effort the lowest assigned scheduling priority.  Though not
   mandatory for an EF implementation, a strict non-preemptive priority
   scheduler is one implementation option as described in section 5.1.1
   of RFC 3247 [15].  Were Weighted Fair Queueing (WFQ) to be used, the
   described effects would essentially also occur, but with minor
   differences.  In the following scenarios, it is illustrated that PHBs
   of equal or lower priority (in comparison to the multicast flow's
   PHB) are affected by the NRS problem.

   The Neglected Reservation Subtree problem appears in two different
   cases:

   o  Case 1: If the branching point of the new subtree (at first only a
      branch) and the previous multicast tree is a (egress) boundary
      node, as shown in Fig. 3, the additional multicast flow now
      increases the total amount of used resources for the corresponding
      behavior aggregate on the affected output link.  The total amount
      will be greater than the originally reserved amount.
      Consequently, the policing component in the egress boundary node
      drops packets until the traffic aggregate is in accordance with
      the traffic contract.  But while dropping packets, the router can
      not identify the responsible flow (because of missing flow
      classification functionality), and thus randomly discards packets,
      whether they belong to a correctly behaving flow or not.  As a
      result, there will no longer be any service guarantee for the
      flows with properly reserved resources.




















Bless & Wehrle               Informational                      [Page 6]



RFC 3754              IP Multicast in DS Networks             April 2004


    Sender
     +---+
     | S |                 DS domains
     +---+                  /       \
      .||...............   /         \   ................
     . ||               .<-           ->.                .
    .  ||                .             .                  .
    . +---+   +--+     +--+     *)    +--+    +--+      +--+    +------+
    . |FHN|===|IN|=====|BN|###########|BN|####|IN|######|BN|####|Recv.B|
    . +---+   +--+     +--+\\         +--+    +--+      +--+    +------+
    .   \\       \        . \\         .         \        .
    .  +--+     +--+      .  \\        .          \       .
    .  |IN|-----|IN|      .   \\        .          +--+  .
    .  +--+     +--+      .    \\        ..........|BN|..
    .   ||        \      .     +------+            +--+
     .  ||         \    .      |Recv.A|
      .+--+        +--+.       +------+
       |BN|........|BN|
       +--+        +--+
        ||

    S: Sender
    Recv.x: Receiver x
    FHN: First-Hop Node
    BN: Boundary Node
    IN: Interior Node
    ===: Multicast branch with reserved bandwidth
    ###: Multicast branch without reservation
    *) Bandwidth of EF aggregated on the output link is higher than
       actual reservation, EF aggregate will be limited in bandwidth
       without considering the responsible flow.

         Figure 3: The NRS Problem (case 1) occurs when Receiver
                   B joins

      In figure 3, it is assumed that receiver A is already attached to
      the egress boundary node (BN) of the first domain.  Furthermore,
      resources are properly reserved along the path to receiver A and
      used by correspondingly marked packets.  When receiver B joins the
      same group as receiver A, packets are replicated and forwarded
      along the new branch towards the second domain with the same PHB
      as for receiver A.  If this PHB is EF, the new branch possibly
      exhausts allotted resources for the EF PHB, adversely affecting
      other EF users that receive their packets over the link that is
      marked with the *).  The BN usually ensures that outgoing traffic
      aggregates to the next domain are conforming to the agreed traffic
      conditioning specification.  The egress BN will, therefore, drop
      packets of the PHB type that are used for the multicast flow.



Bless & Wehrle               Informational                      [Page 7]



RFC 3754              IP Multicast in DS Networks             April 2004


      Other PHBs of lower or higher priority are not affected adversely
      in this case.  The following example in Fig. 4.  illustrates this
      for two PHBs.

   +------------------+---------------------+--------------+------+
   | Expedited Forw.  | Expedited Forw.     | Assured Forw.|  BE  |
   |                  |                     |              |      |
   | with reservation | excess flow         | with reserv. |      |
   |                  | without reservation |              |      |
   +------------------+---------------------+--------------+------+
   | EF with and without reservation share  |    40 %      |  20% |
   | 40% of reserved EF aggregate.          |              |      |
   | -> EF packets with reservation and     |              |      |
   |    without reservation will be         |              |      |
   |    discarded!                          |              |      |
   +------------------+---------------------+--------------+------+

               (a) Excess flow has EF codepoint

   +------------------+---------------------+--------------+------+
   | Expedited Forw.  | Assured Forwarding  | Assured Forw.|  BE  |
   |                  |                     |              |      |
   | with reservation | excess flow         | with reserv. |      |
   |                  | without reservation |              |      |
   +------------------+---------------------+--------------+------+
   |                  | AF with & without reservation share| 20 % |
   |                  | 40% of reserved EF aggregate.      |      |
   |       40%        | -> EF packets with reservation and |      |
   |                  |    without reservation will be     |      |
   |                  |    discarded!                      |      |
   +------------------+---------------------+--------------+------+

               (b) Excess flow has AF codepoint

        Figure 4: Resulting share of bandwidth in a egress
                  boundary node with a neglected reservation of
                  (a) an Expedited Forwarding flow or (b) an
                  Assured Forwarding flow.

      Fig. 4 shows the resulting share of bandwidth in cases when (a)
      Expedited Forwarding and (b) Assured Forwarding is used by the
      additional multicast branch causing the NRS Problem.  Assuming
      that the additional traffic would use another 30% of the link
      bandwidth, Fig. 4 (a) illustrates that the resulting aggregate of
      Expedited Forwarding (70% of the outgoing link bandwidth) is
      throttled down to its originally reserved 40%.  In this case, the
      amount of dropped EF bandwidth is equal to the amount of excess
      bandwidth.  Consequently, the original Expedited Forwarding



Bless & Wehrle               Informational                      [Page 8]



RFC 3754              IP Multicast in DS Networks             April 2004


      aggregate (which had 40% of the link bandwidth reserved) is also
      affected by packet losses.  The other services, e.g., Assured
      Forwarding or Best-Effort, are not disadvantaged.

      Fig. 4 (b) shows the same situation for Assured Forwarding.  The
      only difference is that now Assured Forwarding is solely affected
      by discards, as the other services will still get their
      guarantees.  In either case, packet losses are restricted to the
      misbehaving service class by the traffic meter and policing
      mechanisms in boundary nodes.  Moreover, the latter problem (case
      1) occurs only in egress boundary nodes because they are
      responsible for ensuring that the traffic leaving the
      Differentiated Services domain is not more than the following
      ingress boundary node will accept.  Therefore, those violations of
      SLAs will already be detected and processed in egress boundary
      nodes.

   o  Case 2: The Neglected Reservation Subtree problem can also occur
      if the branching point between the previous multicast tree and the
      new subtree is located in an interior node (as shown in Fig. 5).
      In Fig. 5, it is assumed that receivers A and B have already
      joined the multicast group and have reserved resources
      accordingly.  The interior node in the second domain starts
      replication of multicast packets as soon as receiver C joins.
      Because the router is not equipped with metering or policing
      functions, it will not recognize any amount of excess traffic and
      will forward the new multicast flow.  If the latter belongs to a
      higher priority service, such as Expedited Forwarding, bandwidth
      of the aggregate is higher than the aggregate's reservation at the
      new branch and will use bandwidth from lower priority services.





















Bless & Wehrle               Informational                      [Page 9]



RFC 3754              IP Multicast in DS Networks             April 2004


    Sender
     +---+
     | S |                 DS domains
     +---+                  /       \
      .||...............   /         \   ................
     . ||               .<-           ->.                .
    .  ||                .             .                  .
    . +---+   +--+     +--+           +--+    +--+      +--+   +------+
    . |FHN|===|IN|=====|BN|===========|BN|====|IN|======|BN|===|Recv.B|
    . +---+   +--+     +--+\\         +--+    +--+      +--+   +------+
    .   \\       \        . \\         .         #        .
    .  +--+     +--+      .  \\        .          # *)    .
    .  |IN|-----|IN|      .   \\        .          +--+  .
    .  +--+     +--+      .    \\        ..........|BN|..
    .   ||        \      .     +------+            +--+
     .  ||         \    .      |Recv.A|              #
      .+--+        +--+.       +------+              #
       |BN|........|BN|                            +------+
       +--+        +--+                            |Recv.C|
        ||                                         +------+

    FHN: First-Hop Node, BN: Boundary Node, Recv.x: Receiver x
    S: Sender, IN: Interior Node
    ===: Multicast branch with reserved bandwidth
    ###: Multicast branch without reservation
    *) Bandwidth of EF aggregated on the output link is higher than
       actual reservation, EF aggregate will be limited in bandwidth
       without considering the responsible flow

         Figure 5: Neglected Reservation Subtree problem case 2
                   after join of receiver C




















Bless & Wehrle               Informational                     [Page 10]



RFC 3754              IP Multicast in DS Networks             April 2004


      The additional amount of EF without a corresponding reservation is
      forwarded together with the aggregate which has a reservation.
      This results in no packet losses for Expedited Forwarding as long
      as the resulting aggregate is not higher than the output link
      bandwidth.  Because of its higher priority, Expedited Forwarding
      gets as much bandwidth as needed and as is available.  The effects
      on other PHBs are illustrated by the following example in Fig. 6.

   +------------------+---------------------+--------------+------+
   | Expedited Forw.  | Expedited Forw.     | Assured Forw.|  BE  |
   |                  |                     |              |      |
   | with reservation | excess flow         | with reserv. |      |
   |                  | without reservation |              |      |
   +------------------+---------------------+--------------+------+
   |      40%         |        30%          |     30%      |  0%  |
   +------------------+---------------------+--------------+------+
     EF with reservation and the excess flow use together 70%
     of the link bandwidth because EF, with or without reservation,
     has the highest priority.

               (a) Excess flow has EF codepoint

   +------------------+---------------------+--------------+------+
   | Expedited Forw.  | Assured Forw.       | Assured Forw.|  BE  |
   |                  |                     |              |      |
   | with reservation | excess flow         | with reserv. |      |
   |                  | without reservation |              |      |
   +------------------+---------------------+--------------+------+
   |      40%         |                   60%              |  0%  |
   |                  |                (10% loss)          |      |
   +------------------+---------------------+--------------+------+
     AF with reservation and the excess flow use together 60%
     of the link bandwidth because EF has the highest priority
     (-> 40%).  10% of AF packets will be lost.

               (b) Excess flow has AF codepoint

        Figure 6: Resulting share of bandwidth in an interior
                  node with a neglected reservation of (a) an
                  Expedited Forwarding flow or (b) an Assured
                  Forwarding flow

      The result of case 2 is that there is no restriction for Expedited
      Forwarding, but as Fig. 6 (a) shows, other services will be
      extremely disadvantaged by this use of non-reserved resources.
      Their bandwidth is used by the new additional flow.  In this case,
      the additional 30% Expedited Forwarding traffic preempts resources
      from the Assured Forwarding traffic, which in turn preempts



Bless & Wehrle               Informational                     [Page 11]



RFC 3754              IP Multicast in DS Networks             April 2004


      resources from the best-effort traffic, resulting in 10% packet
      losses for the Assured Forwarding aggregate, and a complete loss
      of best-effort traffic.  The example in Fig. 6 (b) shows that this
      can also happen with lower priority services like Assured
      Forwarding.  When a reservation for a service flow with lower
      priority is neglected, other services (with even lower priority)
      can be reduced in their quality (in this case the best-effort
      service).  As shown in the example, the service's aggregate
      causing the NRS problem can itself be affected by packet losses
      (10% of the Assured Forwarding aggregate is discarded).  Besides
      the described problems of case 2, case 1 will occur in the DS
      boundary node of the next DS domain that performs traffic metering
      and policing for the service aggregate.

   Directly applying RSVP to Differentiated Services would also result
   in temporary occurrence of the NRS Problem.  A receiver has to join
   the IP multicast group to receive the sender's PATH messages, before
   being able to send a resource reservation request (RESV message).
   Thus, the join message on the link for receiving PATH messages can
   cause the NRS Problem, if this situation is not handled in a special
   way (e.g., by marking all PATH messages with codepoint 0 and dropping
   or re-marking all other data packets of the multicast flow).

2.2.  Heterogeneous Multicast Groups

   Heterogeneous multicast groups contain one or more receivers, which
   would like to get another service or quality of service as the sender
   provides or other receiver subsets currently use.  A very important
   characteristic which should be supported by Differentiated Services
   is that participants requesting a best-effort quality only should
   also be able to participate in a group communication which otherwise
   utilizes a better service class.  The next better support for
   heterogeneity provides concurrent use of more than two different
   service classes within a group.  Things tend to get even more complex
   when not only different service classes are required, but also
   different values for quality parameters within a certain service
   class.

   A further problem is to support heterogeneous groups with different
   service classes in a consistent way.  It is possible that some
   services will not be comparable to each other so that one service
   cannot be replaced by the other, and both services have to be
   provided over the same link within this group.

   Because an arbitrary new receiver that wants to get the different
   service can be grafted to any point of the current multicast delivery
   tree, even interior nodes may have to replicate packets using the
   different service.  At a first glance, this seems to be a



Bless & Wehrle               Informational                     [Page 12]



RFC 3754              IP Multicast in DS Networks             April 2004


   contradiction with respect to simplicity of the interior nodes,
   because they do not even have a profile available and should now
   convert the service of quality of individual receivers.
   Consequently, in order to accomplish this, interior nodes have to
   change the codepoint value during packet replication.

2.3.  Dynamics of Any-Source Multicast

   Basically, within an IP multicast group, any participant (actually,
   it can be any host not even receiving packets of this multicast
   group) can act as a sender.  This is an important feature which
   should also be available in case a specific service other than best-
   effort is used within the group.  Differentiated Services possess,
   conceptually, a unidirectional character.  Therefore, for every
   multicast tree implied by a sender, resources must be reserved
   separately if simultaneous sending should be possible with a better
   service.  This is even true if shared multicast delivery trees are
   used (e.g., with PIM-SM or Core Based Trees).  If not enough
   resources are reserved for a service within a multicast tree allowing
   simultaneous sending of more than one participant, the NRS problem
   will occur again.  The same argument applies to half-duplex
   reservations which would share the reserved resources by several
   senders, because it cannot be ensured by the network that exactly one
   sender sends packets to the group.  Accordingly, the corresponding
   RSVP reservation styles "Wildcard Filter" and "Shared-Explicit
   Filter" [4] cannot be supported within Differentiated Services.  The
   Integrated Services approach is able to ensure the half-duplex nature
   of the traffic, because every router can check each packet for its
   conformance with the installed reservation state.

3.  Solutions for Enabling IP-Multicast in Differentiated Services
    Networks

   The problems described in the previous section are mainly caused by
   the simplicity of the Differentiated Services architecture.
   Solutions that do not introduce additional complexity need to be
   introduced so as to not diminish the scalability of the DiffServ
   approach.  This document suggests a straightforward solution for most
   of the problems.

3.1.  Solution for the NRS Problem

   The proposed solution consists conceptually of the following three
   steps that are described in more detail later.







Bless & Wehrle               Informational                     [Page 13]



RFC 3754              IP Multicast in DS Networks             April 2004


      1. A new receiver joins a multicast group that is using a DiffServ
         service.  Multicast routing protocols accomplish the connection
         of the new branch to the (possibly already existing) multicast
         delivery tree as usual.

      2. The unauthorized use of resources is avoided by re-marking at
         branching nodes all additional packets departing down the new
         branch.  At first, the new receiver will get all packets of the
         multicast group without quality of service.  The management
         entity of the correspondent DiffServ domain may get informed
         about the extension of the multicast tree.

      3. If a pre-issued reservation is available for the new branch or
         somebody (receiver, sender or a third party) issues one, the
         management entity instructs the branching router to set the
         corresponding codepoint for the demanded service.

   Usage of resources which were not previously reserved must be
   prevented.  In the following example, we consider a case where the
   join of a new receiver to a DS multicast group requires grafting of a
   new branch to an already existing multicast delivering tree.  The
   connecting node that joins both trees converts the codepoint (and
   therefore the Per-Hop Behavior) to a codepoint of a PHB which is
   similar to the default PHB in order to provide a best-effort-like
   service for the new branch.  More specifically, this particular PHB
   can provide a service that is even worse than the best-effort service
   of the default PHB.  See RFC 3662 [16] for a corresponding Lower
   Effort Per-Domain Behavior.

   The conversion to this specific PHB could be necessary in order to
   avoid unfairness being introduced within the best-effort service
   aggregate, and, which results from the higher amount of resource
   usage of the incoming traffic belonging to the multicast group.  If
   the rate at which re-marked packets are injected into the outgoing
   aggregate is not reduced, those re-marked packets will probably cause
   discarding of other flow's packets in the outgoing aggregate if
   resources are scarce.

   Therefore, the re-marked packets from this multicast group should be
   discarded more aggressively than other packets in this outgoing
   aggregate.  This could be accomplished by using an appropriately
   configured PHB (and a related DSCP) for those packets.  In order to
   distinguish this kind of PHB from the default PHB, it is referred to
   as the Limited Effort (LE) PHB (which can be realized by an
   appropriately configured AF PHB [9] or Class Selector Compliant PHB
   [1]) throughout this document.  Merely dropping packets more
   aggressively at the re-marking node is not sufficient, because there
   may be enough resources in the outgoing behavior aggregate (BA) to



Bless & Wehrle               Informational                     [Page 14]



RFC 3754              IP Multicast in DS Networks             April 2004


   transmit every re-marked packet without having to discard any other
   packets within the same BA.  However, resources in the next node may
   be short for this particular BA.  Those "excess" packets, therefore,
   must be identifiable at this node.

   Re-marking packets is only required at branching nodes, whereas all
   other nodes of the multicast tree (such with outdegree 1) replicate
   packets as usual.  Because a branching node may also be an interior
   node of a domain, re-marking of packets requires conceptually per-
   flow classification.  Though this seems to be in contradiction to the
   DiffServ philosophy of a core that avoids per-flow states, IP
   multicast flows are different from unicast flows: traditional IP
   multicast forwarding and multicast routing are required to install
   states per multicast group for every outgoing link anyway.
   Therefore, re-marking in interior nodes is scalable to the same
   extent as IP multicast (cf. section 4).

   Re-marking with standard DiffServ mechanisms [10] for every new
   branch requires activation of a default traffic profile.  The latter
   accomplishes re-marking by using a combination of an MF-classifier
   and a marker at an outgoing link that constitutes a new branch.  The
   classifier will direct all replicated packets to a marker that sets
   the new codepoint.  An alternative implementation is described in
   section 7.

   The better service will only be provided if a reservation request was
   processed and approved by the resource management function.  That
   means an admission control test must be performed before resources
   are actually used by the new branch.  In case the admission test is
   successful, the re-marking node will be instructed by the resource
   management to stop re-marking and to use the original codepoint again
   (conceptually by removing the profile).

   In summary, only those receivers will obtain a better service within
   a DiffServ multicast group, which previously reserved the
   corresponding resources in the new branch with assistance of the
   resource management.  Otherwise, they get a quality which might be
   even lower than best-effort.

3.2.  Solution for Supporting Heterogeneous Multicast Groups

   In this document, considerations are limited to provisioning
   different service classes, but not different quality parameters
   within a certain service class.

   The proposed concept from section 3.1 provides a limited solution of
   the heterogeneity problem.  Receivers are allowed to obtain a Limited
   Effort service without a reservation, so that at least two different



Bless & Wehrle               Informational                     [Page 15]



RFC 3754              IP Multicast in DS Networks             April 2004


   service classes within a multicast group are possible.  Therefore, it
   is possible for any receiver to participate in the multicast session
   without getting any quality of service.  This is useful if a receiver
   just wants to see whether the content of the multicast group is
   interesting enough, before requesting a better service which must be
   paid for (like snooping into a group without prior reservation).

   Alternatively, a receiver might not be able to receive this better
   quality of service (e.g., because it is mobile and uses a wireless
   link of limited capacity), but it may be satisfied with the reduced
   quality, instead of getting no content at all.

   Additionally, applying the RSVP concept of listening for PATH
   messages before sending any RESV message is feasible again.  Without
   using the proposed solution, this would have caused the NRS Problem.

   Theoretically, the proposed approach in section 7 also supports more
   than two different services within one multicast group, because the
   additional field in the multicast routing table can store any DSCP
   value.  However, this would work only if PHBs can be ordered, so that
   the "best" PHB among different required PHBs downstream is chosen to
   be forwarded on a specific link.  This is mainly a management issue
   and is out of the scope of this document.

   More advanced concepts may also support conditional re-marking in
   dependence on the group address and DSCP value.  This is useful if
   the group uses different PHBs (e.g., for flows to different transport
   protocol ports) and the re-marking should thus additionally depend on
   the DSCP value of an incoming packet.

3.3.  Solution for Any-Source Multicast

   Every participant would have to initiate an explicit reservation to
   ensure the possibility of sending to the group with a better service
   quality, regardless of whether other senders within the group already
   use the same service class simultaneously.  This would require a
   separate reservation for each sender-rooted multicast tree.

   However, in the specific case of best-effort service (the default
   PHB), it is nevertheless possible for participants to send packets to
   the group anytime without requiring any additional mechanisms.  The
   reason for this is that the first DS-capable boundary node will mark
   those packets with the DSCP of the default PHB because of a missing
   traffic profile for this particular sender.  The first DS capable
   boundary nodes should therefore always classify multicast packets
   based on both the sender's address and the multicast group address.

4.  Scalability Considerations



Bless & Wehrle               Informational                     [Page 16]



RFC 3754              IP Multicast in DS Networks             April 2004


   The proposed solution does not add complexity to the DS architecture
   or to a DS node, nor does it change the scalability properties of
   DiffServ.  With current IP multicast routing protocols, a multicast
   router has to manage and hold state information per traversing
   multicast flow.  The suggested solution scales to the same extent as
   IP multicast itself, because the proposed re-marking may occur per
   branch of a multicast flow.  This re-marking is logically associated
   with an addition to the multicast routing state that is required
   anyway.  In this respect, re-marking of packets for multicast flows
   in interior nodes is not considered as a scalability problem or to be
   in contradiction to the DiffServ approach itself.  It is important to
   distinguish the multicast case from existing justifiable scalability
   concerns relating to re-marking packets of unicast flows within
   interior routers.  Moreover, the decision of when to change a re-
   marking policy is not performed by the router, but by some management
   entity at a time scale which is different from the time scale at the
   packet forwarding level.

5.  Deployment Considerations

   The solution proposed in section 3.1 can be deployed on most router
   platforms available today.  Architectures that perform routing and
   forwarding functions in software could be updated by a new software
   release.

   However, there may be some specialized hardware platforms that are
   currently not able to deploy the proposed solution from section 7.
   This may be the case when a multicast packet is directly duplicated
   on the backplane of the router, so that all outgoing interfaces read
   the packet in parallel.  Consequently, the codepoint cannot be
   changed for a subset of these outgoing interfaces and the NRS problem
   can not be solved directly in the branching point.

   In this case, there exist several alternative solutions:

      1. As mentioned in section 3.1, if traffic conditioning mechanisms
         can be applied on the outgoing packets at the individual output
         interfaces, a combination of classifier and marker may be used
         for each branch.

      2. The change of the codepoint for subtrees without properly
         allocated resources could take place in the following
         downstream router.  There, for every incoming packet of the
         considered multicast group, the codepoint would be changed to
         the value that the previous router should have set.  If a LAN
         (e.g., a high-speed switching LAN) is attached to the
         considered outgoing interface, then on every router connected
         to the LAN, packets of the considered group should be changed



Bless & Wehrle               Informational                     [Page 17]



RFC 3754              IP Multicast in DS Networks             April 2004


         on the incoming interface by standard DiffServ mechanisms.

   Future releases of router architectures may support the change of the
   codepoint directly in the replication process as proposed in section
   7.

6.  Security Considerations

   Basically, the security considerations in [1] apply.  The proposed
   solution does not imply new security aspects.  If a join of arbitrary
   end-systems to a multicast group is not desired (thereby receiving a
   lower than best-effort quality) the application usually has to
   exclude these participants.  This can be accomplished by using
   authentication, authorization, or ciphering techniques at the
   application level -- like in traditional IP multicast scenarios.

   Moreover, it is important to consider the security of corresponding
   management mechanisms, because they are used to activate re-marking
   of multicast flows.  On the one hand, functions for instructing the
   router to mark or re-mark packets of multicast flows are attractive
   targets to perform theft of service attacks.  On the other hand,
   their security depends on the router management mechanisms which are
   used to realize this functionality.  Router management should
   generally be protected against unauthorized use, therefore preventing
   those attacks as well.

7.  Implementation Model Example

   One possibility of implementing the proposed solution from section
   3.1 is described in the following.  It has to be emphasized that
   other realizations are also possible, and this description should not
   be understood as a restriction on potential implementations.  The
   benefit of the following described implementation is that it does not
   require any additional classification of multicast groups within an
   aggregate.  It serves as a proof of concept that no additional
   complexity is necessary to implement the proposed general solution
   described in section 3.

   Because every multicast flow has to be considered by the multicast
   routing process (in this context, this notion signifies the multicast
   forwarding part and not the multicast route calculation and
   maintenance part, cf. Fig. 1), the addition of an extra byte in each
   multicast routing table entry containing the DS field, and thus its
   DS codepoint value per output link (resp. virtual interface, see Fig.
   8), results in nearly no additional cost.  Packets will be replicated
   by the multicast forwarding process, so this is also the right place
   for setting the correct DSCP values of the replicated packets.  Their
   DSCP values are not copied from the incoming original packet, but



Bless & Wehrle               Informational                     [Page 18]



RFC 3754              IP Multicast in DS Networks             April 2004


   from the additional DS field in the multicasting routing table entry
   for the corresponding output link (only the DSCP value must be
   copied, while the two remaining bits are ignored and are present for
   simplification reasons only).  This field initially contains the
   codepoint of the LE PHB if incoming packets for this specific group
   do not carry the codepoint of the default PHB.

   When a packet arrives with the default PHB, the outgoing replicates
   should also get the same codepoint in order to retain the behavior of
   current common multicast groups using the default PHB.  A router
   configuration message changes the DSCP values in the multicast
   routing table and may also carry the new DSCP value which should be
   set in the replicated packets.  It should be noted that although re-
   marking may also be performed by interior nodes, the forwarding
   performance will not be decreased, because the decision of when and
   what to re-mark is made by the management (control plane).

     Multicast   Other    List
     Destination Fields   of
     Address              virtual                   Inter-   DS
                          interfaces                face ID  Field
    +--------------------------------+             +-------------------+
    |    X      | .... |     *-------------------->|   C   | (DSCP,CU) |
    |--------------------------------|             +-------------------+
    |    Y      | .... |     *-----------+         |   D   | (DSCP,CU) |
    |--------------------------------|   |         +-------------------+
    |   ...     | .... |    ...      |   |
    .           .      .             .   |         +-------------------+
    .   ...     . .... .    ...      .   +-------->|   B   | (DSCP,CU) |
    +--------------------------------+             +-------------------+
    |   ...     | .... |    ...      |             |   D   | (DSCP,CU) |
    +--------------------------------+             +-------------------+
                                                   |  ...  |   ...     |
                                                   .       .           .
                                                   .       .           .

         Figure 8: Multicast routing table with additional
                   fields for DSCP values

8.  Proof of the Neglected Reservation Subtree Problem

   In the following, it is shown that the NRS problem actually exists
   and occurs in reality.  Hence, the problem and its solution was
   investigated using a standard Linux Kernel (v2.4.18) and the Linux-
   based implementation KIDS [11].

   Furthermore, the proposed solution for the NRS problem has been
   implemented by enhancing the multicast routing table, as well as the



Bless & Wehrle               Informational                     [Page 19]



RFC 3754              IP Multicast in DS Networks             April 2004


   multicast routing behavior in the Linux kernel.  In the following
   section, the modifications are briefly described.

   Additional measurements with the simulation model simulatedKIDS [12]
   will be presented in section 9.  They show the effects of the NRS
   problem in more detail and also the behavior of the BAs using or not
   using the Limited Effort PHB for re-marking.

8.1.  Implementation of the Proposed Solution

   As described in section 3.1, the proposed solution for avoiding the
   NRS Problem is an extension of each routing table entry in every
   Multicast router by one byte.  In the Linux OS, the multicast routing
   table is implemented by the "Multicast Forwarding Cache (MFC)".  The
   MFC is a hash table consisting of an "mfc-cache"-entry for each
   combination of the following three parameters: sender's IP address,
   multicast group address, and incoming interface.

   The routing information in a "mfc-cache"-entry is kept in an array of
   TTLs for each virtual interface.  When the TTL is zero, a packet
   matching to this "mfc-cache"-entry will not be forwarded on this
   virtual interface.  Otherwise, if the TTL is less than the packet's
   TTL, the latter will be forwarded on the interface with a decreased
   TTL.

   In order to set an appropriate codepoint if bandwidth is allocated on
   an outgoing link, we added a second array of bytes -- similar to the
   TTL array -- for specifying the codepoint that should be used on a
   particular virtual interface.  The first six bits of the byte contain
   the DSCP that should be used, and the seventh bit indicates whether
   the original codepoint in the packet has to be changed to the
   specified one (=0) or has to be left unchanged (=1).  The default
   entry of the codepoint byte is zero; so initially, all packets will
   be re-marked to the default DSCP.

   Furthermore, we modified the multicast forwarding code for
   considering this information while replicating multicast packets.  To
   change an "mfc-cache"-entry we implemented a daemon for exchanging
   the control information with a management entity (e.g., a bandwidth
   broker).  Currently, the daemon uses a proprietary protocol, but
   migration to the COPS protocol (RFC 2748) is planned.










Bless & Wehrle               Informational                     [Page 20]



RFC 3754              IP Multicast in DS Networks             April 2004


8.2.  Test Environment and Execution

   Sender
    +---+             FHN: First Hop Node
    | S |             BN: Boundary Node
    +---+
      +#
      +#
      +#
     +---+            +--+           +------+
     |FHN|++++++++++++|BN|+++++++++++| host |
     |   |############|  |***********|  B   |
     +---+            +--+##         +------+
       +#                   #
        +#                   #
         +#                   #
         +------+           +------+
         |host A|           |host C|
         +------+           +------+

   +++  EF flow (group1) with reservation
   ###  EF flow (group2) with reservation
   ***  EF flow (group2) without reservation

         Figure 8.1: Evaluation of NRS-Problem described in
                     Figure 3

   In order to prove case 1 of the NRS problem, as described in section
   2.1, the testbed shown in Figure 8.1 was built.  It is a reduced
   version of the network shown in Figure 5 and consists of two DS-
   capable nodes, an ingress boundary node and an egress boundary node.
   The absence of interior nodes does not have any effect on to the
   proof of the described problem.

   The testbed is comprised of two Personal Computers (Pentium III at
   450 Mhz, 128 MB Ram, 3 network cards Intel eepro100) used as DiffServ
   nodes, as well as one sender and three receiver systems (also PCs).
   On the routers, KIDS has been installed and an mrouted (Multicast
   Routing Daemon) was used to perform multicast routing.  The network
   was completely built of separate 10BaseT Ethernet segments in full-
   duplex mode.  In [11], we evaluated the performance of the software
   routers and found out that even a PC at 200Mhz had no problem
   handling up to 10Mbps DS traffic on each link.  Therefore, the
   presented measurements are not a result of performance bottlenecks
   caused by these software routers.






Bless & Wehrle               Informational                     [Page 21]



RFC 3754              IP Multicast in DS Networks             April 2004


   The sender generated two shaped UDP traffic flows of 500kbps (packets
   of 1.000 byte constant size) each and sent them to multicast group 1
   (233.1.1.1) and 2 (233.2.2.2).  In both measurements, receiver A had
   a reservation along the path to the sender for each flow, receiver B
   had reserved for flow 1, and C for flow 2.  Therefore, two static
   profiles are installed in the ingress boundary node with 500kbps EF
   bandwidth and a token bucket size of 10.000 bytes for each flow.

   In the egress boundary node, one profile has been installed for the
   output link to host B and one related for the output link to host C.
   Each of them permits up to 500kbps Expedited Forwarding, but only the
   aggregate of Expedited Forwarding traffic carried on the outgoing
   link is considered.

   In measurement 1, the hosts A and B joined to group 1, while A, B,
   and C joined to group 2.  Those joins are using a reservation for the
   group towards the sender.  Only the join of host B to group 2 has no
   admitted reservation.  As described in section 2.1, this will cause
   the NRS problem (case 1).  Metering and policing mechanisms in the
   egress boundary node throttle down the EF aggregate to the reserved
   500kbps, and do not depend on whether or not individual flows have
   been reserved.

                +--------+--------+--------+
                | Host A | Host B | Host C |
      +---------+--------+--------+--------+
      | Group 1 | 500kbps| 250kbps| 500kbps|
      +---------+--------+--------+--------+
      | Group 2 | 500kbps| 250kbps|        |
      +---------+--------+--------+--------+

          Figure 8.2: Results of measurement 1 (without the
                      proposed solution): Average bandwidth of
                      each flow.
                      --> Flows of group 1 and 2 on the link to
                      host B share the reserved aggregate of
                      group 1.

   Figure 8.2 shows the obtained results.  Host A and C received their
   flows without any interference.  But host B received data from group
   1 with only half of the reserved bandwidth, so one half of the
   packets have been discarded.  Figure 8.2 also shows that receiver B
   got the total amount of bandwidth for group 1 and 2, that is exactly
   the reserved 500kbps.  Flow 2 got Expedited Forwarding without
   actually having reserved any bandwidth and additionally violated the
   guarantee of group 1 on that link.





Bless & Wehrle               Informational                     [Page 22]



RFC 3754              IP Multicast in DS Networks             April 2004


   For measurement 2, the previously presented solution (cf. section
   3.1) has been installed in the boundary node.  Now, while duplicating
   the packets, whether the codepoint has to be changed to Best-Effort
   (or Limited Effort) or whether it can be just duplicated is checked.
   In this measurement, it changed the codepoint for group 2 on the link
   to Host B to Best-Effort.

                +--------+--------+--------+
                | Host A | Host B | Host C |
      +---------+--------+--------+--------+
      | Group 1 | 500kbps| 500kbps| 500kbps|
      +---------+--------+--------+--------+
      | Group 2 | 500kbps| 500kbps|        |
      +---------+--------+--------+--------+

          Figure 8.3: Results of measurement 1 (with the
                      proposed solution): Average bandwidth of
                      each flow.
                      --> Flow of group 1 on the link to host B
                      gets the reserved bandwidth of group 1.
                      The flow of group 2 has been re-marked to
                      Best-Effort.

   Results of this measurement are presented in Figure 8.3.  Each host
   received its flows with the reserved bandwidth and without any packet
   loss.  Packets from group 2 are re-marked in the boundary node so
   that they have been treated as best-effort traffic.  In this case,
   they got the same bandwidth as the Expedited Forwarding flow, and
   because there was not enough other traffic on the link present, there
   was no need to discard packets.

   The above measurements confirm that the Neglected Reservation Subtree
   problem is to be taken seriously and that the presented solution will
   solve it.

9.  Simulative Study of the NRS Problem and Limited Effort PHB

   This section shows some results from a simulative study which shows
   the correctness of the proposed solution and the effect of re-marking
   the responsible flow to Limited Effort.  A proof of the NRS problem
   has also been given in section 8 and in [13].  This section shows the
   benefit for the default Best Effort traffic when Limited Effort is
   used for re-marking instead of Best Effort.  The results strongly
   motivate the use of Limited Effort.







Bless & Wehrle               Informational                     [Page 23]



RFC 3754              IP Multicast in DS Networks             April 2004


9.1.  Simulation Scenario

   In the following scenario, the boundary nodes had a link speed of 10
   Mpbs and Interior Routers had a link speed of 12 Mbps.  In boundary
   nodes, a 5 Mbps aggregate for EF has been reserved.

   When Limited Effort was used, LE got 10% capacity (0.5Mpbs) from the
   original BE aggregate and BE 90% (4.5Mbps) of the original BE
   aggregate capacity.  The bandwidth between LE and BE is shared by
   using WFQ scheduling.

   The following topology was used, where Sx is a sender, BRx a boundary
   node, IRx an interior node, and Dx a destination/receiver.

     +--+ +--+                       +---+     +--+
     |S1| |S0|                     /=|BR5|=====|D0|
     +--+ +--+                    // +---+     +--+
       \\  ||                    //
        \\ ||                   //
   +--+  \+---+     +---+     +---+      +---+     +--+
   |S2|===|BR1|=====|IR1|=====|IR2|======|BR3|=====|D1|
   +--+   +---+    /+---+     +---+      +---+     +--+
                  //                       \\        +--+
                 //                         \\     /=|D2|
   +--+   +---+ //                           \\   // +--+
   |S3|===|BR2|=/                            +---+/
   +--+   +---+                            /=|BR4|=\
           ||                        +--+ // +---+ \\ +--+
          +--+                       |D4|=/         \=|D3|
          |S4|                       +--+             +--+
          +--+

              Figure 9.1: Simulation Topology

   The following table shows the flows in the simulation runs, e.g., EF0
   is sent from Sender S0 to Destination D0 with a rate of 4 Mbps using
   an EF reservation.

   In the presented cases (I to IV), different amounts of BE traffic
   were used to show the effects of Limited Effort in different cases.
   The intention of these four cases is described after the table.

   In all simulation models, EF sources generated constant rate traffic
   with constant packet sizes using UDP.

   The BE sources also generated constant rate traffic, where BE0 used
   UDP, and BE1 used TCP as a transport protocol.




Bless & Wehrle               Informational                     [Page 24]



RFC 3754              IP Multicast in DS Networks             April 2004


   +----+--------+-------+----------+-----------+-----------+---------+
   |Flow| Source | Dest. |  Case I  |  Case II  |  Case III | Case IV |
   +----+--------+-------+----------+-----------+-----------+---------+
   | EF0|   S0   |  D0   |  4 Mbps  |   4 Mbps  |   4 Mbps  |  4 Mbps |
   +----+--------+-------+----------+-----------+-----------+---------+
   | EF1|   S1   |  D1   |  2 Mbps  |   2 Mbps  |   2 Mbps  |  2 Mbps |
   +----+--------+-------+----------+-----------+-----------+---------+
   | EF2|   S2   |  D2   |  5 Mbps  |   5 Mbps  |   5 Mbps  |  5 Mbps |
   +----+--------+-------+----------+-----------+-----------+---------+
   | BE0|   S3   |  D3   |  1 Mbps  | 2.25 Mbps | 0.75 Mbps |3.75 Mbps|
   +----+--------+-------+----------+-----------+-----------+---------+
   | BE1|   S4   |  D4   |  4 Mbps  | 2.25 Mbps | 0.75 Mbps |3.75 Mbps|
   +----+--------+-------+----------+-----------+-----------+---------+

   Table 9.1: Direction, amount and Codepoint of flows in the four
              simulation cases (case I to IV)

   The four cases (I to IV) used in the simulation runs had the
   following characteristics:

   Case I:   In this scenario, the BE sources sent together exactly 5
             Mbps, so there is no congestion in the BE queue.

   Case II:  BE is sending less than 5 Mbps, so there is space available
             in the BE queue for re-marked traffic.  BE0 and BE1 are
             sending together 4.5 Mbps, which is exactly the share of BE
             when LE is used.  So, when multicast packets are re-marked
             to LE because of the NRS problem, the LE should get 0.5
             Mbps and BE 4.5 Mbps, which is still enough for BE0 and
             BE1.  LE should not show a greedy behavior and should not
             use resources from BE.

   Case III: In this case, BE is very low.  BE0 and BE1 use together
             only 1.5 Mbps.  So when LE is used, it should be able to
             use the unused bandwidth resources from BE.

   Case IV:  BE0 and BE1 send together 7.5 Mbps so there is congestion
             in the BE queue.  In this case, LE should get 0.5 Mbps (not
             more and not less).

   In each scenario, loss rate and throughput of the considered flows
   and aggregates have been metered.









Bless & Wehrle               Informational                     [Page 25]



RFC 3754              IP Multicast in DS Networks             April 2004


9.2.  Simulation Results for Different Router Types

9.2.1.  Interior Node

   When the branching point of a newly added multicast subtree is
   located in an interior node, the NRS problem can occur as described
   in section 2.1 (Case 2).

   In the simulation runs presented in the following four subsections,
   D3 joins to the multicast group of sender S0 without making any
   reservation or resource allocation.  Consequently, a new branch is
   added to the existing multicast tree.  The branching point issued by
   the join of D3 is located in IR2.  On the link to BR3, no bandwidth
   was allocated for the new flow (EF0).

   The metered throughput of flows on the link between IR2 and BR3 in
   the four different cases is shown in the following four subsections.
   The situation before the new receiver joins is shown in the second
   column.  The situation after the join without the proposed solution
   is shown in column three.  The fourth column presents the results
   when the proposed solution of section 3.1 is used and the responsible
   flow is re-marked to LE.

9.2.1.1.  Case I:

   +--------+-----------------+-----------------+------------------+
   |        |  before join    | after join      |after join,       |
   |        |                 | (no re-marking) |(re-marking to LE)|
   +--------+-----------------+-----------------+------------------+
   |        | EF0:   ---      | EF0: 4.007 Mbps | LE0: 0.504 Mbps  |
   |achieved| EF1: 2.001 Mbps | EF1: 2.003 Mbps | EF1: 2.000 Mbps  |
   |through-| EF2: 5.002 Mbps | EF2: 5.009 Mbps | EF2: 5.000 Mbps  |
   |put     | BE0: 1.000 Mbps | BE0: 0.601 Mbps | BE0: 1.000 Mbps  |
   |        | BE1: 4.000 Mbps | BE1: 0.399 Mbps | BE1: 3.499 Mbps  |
   +--------+-----------------+-----------------+------------------+
   |BA      | EF:  7.003 Mbps | EF: 11.019 Mbps | EF:  7.000 Mbps  |
   |through-| BE:  5.000 Mbps | BE:  1.000 Mbps | BE:  4.499 Mbps  |
   |put     | LE:    ---      | LE:    ---      | LE:  0.504 Mbps  |
   +--------+-----------------+-----------------+------------------+
   |        | EF0:   ---      | EF0:     0 %    | LE0:  87.4 %     |
   |packet  | EF1:     0 %    | EF1:     0 %    | EF1:     0 %     |
   |loss    | EF2:     0 %    | EF2:     0 %    | EF2:     0 %     |
   |rate    | BE0:     0 %    | BE0:  34.8 %    | BE0:     0 %     |
   |        | BE1:     0 %    | BE1:  59.1 %    | BE1:     0 %     |
   +--------+-----------------+-----------------+------------------+
    (*) EF0 is re-marked to LE and signed as LE0





Bless & Wehrle               Informational                     [Page 26]



RFC 3754              IP Multicast in DS Networks             April 2004


9.2.1.2.  Case II:

   +--------+-----------------+-----------------+------------------+
   |        |  before join    | after join      |after join,       |
   |        |                 | (no re-marking) |(re-marking to LE)|
   +--------+-----------------+-----------------+------------------+
   |        | EF0:   ---      | EF0: 4.003 Mbps | LE0: 0.500 Mbps  |
   |achieved| EF1: 2.000 Mbps | EF1: 2.001 Mbps | EF1: 2.001 Mbps  |
   |through-| EF2: 5.002 Mbps | EF2: 5.005 Mbps | EF2: 5.002 Mbps  |
   |put     | BE0: 2.248 Mbps | BE0: 0.941 Mbps | BE0: 2.253 Mbps  |
   |        | BE1: 2.252 Mbps | BE1: 0.069 Mbps | BE1: 2.247 Mbps  |
   +--------+-----------------+-----------------+------------------+
   |BA      | EF:  7.002 Mbps | EF: 11.009 Mbps | EF:  7.003 Mbps. |
   |through-| BE:  4.500 Mbps | BE:  1.010 Mbps | BE:  4.500 Mbps  |
   |put     | LE:    ---      | LE:    ---      | LE:  0.500 Mbps  |
   +--------+-----------------+-----------------+------------------+
   |        | EF0:   ---      | EF0:     0 %    | LE0:  87.4 %     |
   |packet  | EF1:     0 %    | EF1:     0 %    | EF1:     0 %     |
   |loss    | EF2:     0 %    | EF2:     0 %    | EF2:     0 %     |
   |rate    | BE0:     0 %    | BE0:  58.0 %    | BE0:     0 %     |
   |        | BE1:     0 %    | BE1:  57.1 %    | BE1:     0 %     |
   +--------+-----------------+-----------------+------------------+
    (*) EF0 is re-marked to LE and signed as LE0

9.2.1.3.  Case III:

   +--------+-----------------+-----------------+------------------+
   |        |  before join    | after join      |after join,       |
   |        |                 | (no re-marking) |(re-marking to LE)|
   +--------+-----------------+-----------------+------------------+
   |        | EF0:   ---      | EF0: 3.998 Mbps | LE0: 3.502 Mbps  |
   |achieved| EF1: 2.000 Mbps | EF1: 2.001 Mbps | EF1: 2.001 Mbps  |
   |through-| EF2: 5.000 Mbps | EF2: 5.002 Mbps | EF2: 5.003 Mbps  |
   |put     | BE0: 0.749 Mbps | BE0: 0.572 Mbps | BE0: 0.748 Mbps  |
   |        | BE1: 0.749 Mbps | BE1: 0.429 Mbps | BE1: 0.748 Mbps  |
   +--------+-----------------+-----------------+------------------+
   |BA      | EF:  7.000 Mbps | EF: 11.001 Mbps | EF:  7.004 Mbps  |
   |through-| BE:  1.498 Mbps | BE:  1.001 Mbps | BE:  1.496 Mbps  |
   |put     | LE:    ---      | LE:    ---      | LE:  3.502 Mbps  |
   +--------+-----------------+-----------------+------------------+
   |        | EF0:   ---      | EF0:     0 %    | LE0:  12.5 %     |
   |packet  | EF1:     0 %    | EF1:     0 %    | EF1:     0 %     |
   |loss    | EF2:     0 %    | EF2:     0 %    | EF2:     0 %     |
   |rate    | BE0:     0 %    | BE0:  19.7 %    | BE0:     0 %     |
   |        | BE1:     0 %    | BE1:  32.6 %    | BE1:     0 %     |
   +--------+-----------------+-----------------+------------------+
    (*) EF0 is re-marked to LE and signed as LE0




Bless & Wehrle               Informational                     [Page 27]



RFC 3754              IP Multicast in DS Networks             April 2004


9.2.1.4.  Case IV:

   +--------+-----------------+-----------------+------------------+
   |        |  before join    | after join      |after join,       |
   |        |                 | (no re-marking) |(re-marking to LE)|
   +--------+-----------------+-----------------+------------------+
   |        | EF0:   ---      | EF0: 4.001 Mbps | LE0: 0.500 Mbps  |
   |achieved| EF1: 2.018 Mbps | EF1: 2.000 Mbps | EF1: 2.003 Mbps  |
   |through-| EF2: 5.005 Mbps | EF2: 5.001 Mbps | EF2: 5.007 Mbps  |
   |put     | BE0: 2.825 Mbps | BE0: 1.000 Mbps | BE0: 3.425 Mbps  |
   |        | BE1: 2.232 Mbps | BE1:   ---      | BE1: 1.074 Mbps  |
   +--------+-----------------+-----------------+------------------+
   |BA      | EF:  7.023 Mbps | EF: 11.002 Mbps | EF:  7.010 Mbps  |
   |through-| BE:  5.057 Mbps | BE:  1.000 Mbps | BE:  4.499 Mbps  |
   |put     | LE:    ---      | LE:    ---      | LE:  0.500 Mbps  |
   +--------+-----------------+-----------------+------------------+
   |        | EF0:   ---      | EF0:     0 %    | LE0:  75.0 %     |
   |packet  | EF1:     0 %    | EF1:     0 %    | EF1:     0 %     |
   |loss    | EF2:     0 %    | EF2:     0 %    | EF2:     0 %     |
   |rate    | BE0:  23.9 %    | BE0:  73.3 %    | BE0:     0 %     |
   |        | BE1:  41.5 %    | BE1:   ---      | BE1:     0 %     |
   +--------+-----------------+-----------------+------------------+
   (*) EF0 is re-marked to LE and signed as LE0

   NOTE: BE1 has undefined throughput and loss in situation "after join
   (no re-marking)", because TCP is going into retransmission back-off
   timer phase and closes the connection after 512 seconds.

9.2.2.  Boundary Node

   When the branching point of a newly added multicast subtree is
   located in a boundary node, the NRS problem can occur as described in
   section 2.1 (Case 1).

   In the simulation runs presented in the following four subsections,
   D3 joins to the multicast group of sender S1 without making any
   reservation or resource allocation.  Consequently, a new branch is
   added to the existing multicast tree.  The branching point issued by
   the join of D3 is located in BR3.  On the link to BR4, no bandwidth
   was allocated for the new flow (EF1).

   The metered throughput of the flows on the link between BR3 and BR4
   in the four different cases is shown in the following four
   subsections.  The situation before the new receiver joins is shown in
   the second column.  The situation after the join but without the
   proposed solution is shown in column three.  The fourth column
   presents results when the proposed solution of section 3.1 is used
   and the responsible flow is re-marked to LE.



Bless & Wehrle               Informational                     [Page 28]



RFC 3754              IP Multicast in DS Networks             April 2004


9.2.2.1.  Case I:

   +--------+-----------------+-----------------+------------------+
   |        |  before join    | after join      |after join,       |
   |        |                 | (no re-marking) |(re-marking to LE)|
   +--------+-----------------+-----------------+------------------+
   |        | EF0:   ---      | EF0:   ---      | EF0:   ---       |
   |achieved| EF1:   ---      | EF1: 1.489 Mbps | LE1: 0.504 Mbps  |
   |through-| EF2: 5.002 Mbps | EF2: 3.512 Mbps | EF2: 5.002 Mbps  |
   |put     | BE0: 1.000 Mbps | BE0: 1.000 Mbps | BE0: 1.004 Mbps  |
   |        | BE1: 4.000 Mbps | BE1: 4.002 Mbps | BE1: 3.493 Mbps  |
   +--------+-----------------+-----------------+------------------+
   |BA      | EF:  5.002 Mbps | EF:  5.001 Mbps | EF:  5.002 Mbps  |
   |through-| BE:  5.000 Mbps | BE:  5.002 Mbps | BE:  4.497 Mbps  |
   |put     | LE:    ---      | LE:    ---      | LE:  0.504 Mbps  |
   +--------+-----------------+-----------------+------------------+
   |        | EF0:   ---      | EF0:   ---      | EF0:   ---       |
   |packet  | EF1:   ---      | EF1:  25.6 %    | LE1:  73.4 %     |
   |loss    | EF2:     0 %    | EF2:  29.7 %    | EF2:     0 %     |
   |rate    | BE0:     0 %    | BE0:     0 %    | BE0:     0 %     |
   |        | BE1:     0 %    | BE1:     0 %    | BE1:     0 %     |
   +--------+-----------------+-----------------+------------------+
    (*) EF1 is re-marked to LE and signed as LE1

9.2.2.2.  Case II:

   +--------+-----------------+-----------------+------------------+
   |        |  before join    | after join      |after join,       |
   |        |                 | (no re-marking) |(re-marking to LE)|
   +--------+-----------------+-----------------+------------------+
   |        | EF0:   ---      | EF0:   ---      | EF0:   ---       |
   |achieved| EF1:   ---      | EF1: 1.520 Mbps | LE1: 0.504 Mbps  |
   |through-| EF2: 5.003 Mbps | EF2: 3.482 Mbps | EF2: 5.002 Mbps  |
   |put     | BE0: 2.249 Mbps | BE0: 2.249 Mbps | BE0: 2.245 Mbps  |
   |        | BE1: 2.252 Mbps | BE1: 2.252 Mbps | BE1: 2.252 Mbps  |
   +--------+-----------------+-----------------+------------------+
   |BA      | EF:  5.003 Mbps | EF:  5.002 Mbps | EF:  5.002 Mbps  |
   |through-| BE:  4.501 Mbps | BE:  4.501 Mbps | BE:  4.497 Mbps  |
   |put     | LE:    ---      | LE:    ---      | LE:  0.504 Mbps  |
   +--------+-----------------+-----------------+------------------+
   |        | EF0:   ---      | EF0:   ---      | EF0:   ---       |
   |packet  | EF1:   ---      | EF1:  24.0 %    | LE1:  74.8 %     |
   |loss    | EF2:     0 %    | EF2:  30.4 %    | EF2:     0 %     |
   |rate    | BE0:     0 %    | BE0:     0 %    | BE0:     0 %     |
   |        | BE1:     0 %    | BE1:     0 %    | BE1:     0 %     |
   +--------+-----------------+-----------------+------------------+
    (*) EF1 is re-marked to LE and signed as LE1




Bless & Wehrle               Informational                     [Page 29]



RFC 3754              IP Multicast in DS Networks             April 2004


9.2.2.3.  Case III:

   +--------+-----------------+-----------------+------------------+
   |        |  before join    | after join      |after join,       |
   |        |                 | (no re-marking) |(re-marking to LE)|
   +--------+-----------------+-----------------+------------------+
   |        | EF0:   ---      | EF0:   ---      | EF0:   ---       |
   |achieved| EF1:   ---      | EF1: 1.084 Mbps | LE1: 2.000 Mbps  |
   |through-| EF2: 5.001 Mbps | EF2: 3.919 Mbps | EF2: 5.000 Mbps  |
   |put     | BE0: 0.749 Mbps | BE0: 0.752 Mbps | BE0: 0.750 Mbps  |
   |        | BE1: 0.749 Mbps | BE1: 0.748 Mbps | BE1: 0.750 Mbps  |
   +--------+-----------------+-----------------+------------------+
   |BA      | EF:  5.001 Mbps | EF:  5.003 Mbps | EF:  5.000 Mbps  |
   |through-| BE:  1.498 Mbps | BE:  1.500 Mbps | BE:  1.500 Mbps  |
   |put     | LE:    ---      | LE:    ---      | LE:  2.000 Mbps  |
   +--------+-----------------+-----------------+------------------+
   |        | EF0:   ---      | EF0:   ---      | EF0:   ---       |
   |packet  | EF1:   ---      | EF1:  45.7 %    | LE1:     0 %     |
   |loss    | EF2:     0 %    | EF2:  21.7 %    | EF2:     0 %     |
   |rate    | BE0:     0 %    | BE0:     0 %    | BE0:     0 %     |
   |        | BE1:     0 %    | BE1:     0 %    | BE1:     0 %     |
   +--------+-----------------+-----------------+------------------+
    (*) EF1 is re-marked to LE and signed as LE1

9.2.2.4.  Case IV:

   +--------+-----------------+-----------------+------------------+
   |        |  before join    | after join      |after join,       |
   |        |                 | (no re-marking) |(re-marking to LE)|
   +--------+-----------------+-----------------+------------------+
   |        | EF0:   ---      | EF0:   ---      | EF0:   ---       |
   |achieved| EF1:   ---      | EF1: 1.201 Mbps | LE1: 0.500 Mbps  |
   |through-| EF2: 5.048 Mbps | EF2: 3.803 Mbps | EF2: 5.004 Mbps  |
   |put     | BE0: 2.638 Mbps | BE0: 2.535 Mbps | BE0: 3.473 Mbps  |
   |        | BE1: 2.379 Mbps | BE1: 2.536 Mbps | BE1: 1.031 Mbps  |
   +--------+-----------------+-----------------+------------------+
   |BA      | EF:  5.048 Mbps | EF:  5.004 Mbps | EF:  5.004 Mbps  |
   |through-| BE:  5.017 Mbps | BE:  5.071 Mbps | BE:  4.504 Mbps  |
   |put     | LE:    ---      | LE:    ---      | LE:  0.500 Mbps  |
   +--------+-----------------+-----------------+------------------+
   |        | EF0:   ---      | EF0:   ---      | EF0:   ---       |
   |packet  | EF1:   ---      | EF1:  40.0 %    | LE1:  68.6 %     |
   |loss    | EF2:     0 %    | EF2:  23.0 %    | EF2:     0 %     |
   |rate    | BE0:  30.3 %    | BE0:  32.1 %    | BE0:     0 %     |
   |        | BE1:  33.3 %    | BE1:  32.7 %    | BE1:     0 %     |
   +--------+-----------------+-----------------+------------------+
    (*) EF1 is re-marked to LE and signed as LE1




Bless & Wehrle               Informational                     [Page 30]



RFC 3754              IP Multicast in DS Networks             April 2004


10.  Acknowledgements

   The authors wish to thank Mark Handley and Bill Fenner for their
   valuable comments on this document.  Special thanks go to Milena
   Neumann for her extensive efforts in performing the simulations.  We
   would also like to thank the KIDS simulation team [12].

11.  References

11.1.  Normative References

   [1]  Nichols, K., Blake, S., Baker, F. and D. Black, "Definition of
        the Differentiated Services Field (DS Field) in the IPv4 and
        IPv6 Headers", RFC 2474, December 1998.

   [2]  Blake, S., Black, D., Carlson, M., Davies, E., Wang, Z. and W.
        Weiss, "An Architecture for Differentiated Services", RFC 2475,
        December 1998.

11.2.  Informative References

   [3]  Nichols, K. and B. Carpenter, "Definition of Differentiated
        Services Per Domain Behaviors and Rules for their
        Specification", RFC 3086, April 2001.

   [4]  Braden, R., Ed., Zhang, L., Berson, S., Herzog, S. and S. Jamin,
        "Resource ReSerVation Protocol (RSVP) -- Version 1", RFC 2205,
        September 1997.

   [5]  Bernet, Y., "Format of the RSVP DCLASS Object", RFC 2996,
        November 2000.

   [6]  Waitzman, D., Partridge, C. and S. Deering, "Distance Vector
        Multicast Routing Protocol", RFC 1075, November 1988.

   [7]  Estrin, D., Farinacci, D., Helmy, A., Thaler, D., Deering, S.,
        Handley, M., Jacobson, V., Liu, L., Sharma, P. and L. Wei,
        "Protocol Independent Multicast-Sparse Mode (PIM-SM): Protocol
        Specification", RFC 2362, June 1998.

   [8]  Adams, A., Nicholas, J. and W. Siadak, "Protocol Independent
        Multicast - Dense Mode (PIM-DM) Protocol Specification
        (Revised)", Work in Progress.

   [9]  Heinanen, J., Baker, F., Weiss, W. and J. Wroclawski, "Assured
        Forwarding PHB Group" RFC 2597, June 1999.





Bless & Wehrle               Informational                     [Page 31]



RFC 3754              IP Multicast in DS Networks             April 2004


   [10] Bernet, Y., Blake, S., Grossman, D. and A. Smith, "An Informal
        Management Model for DiffServ Routers", RFC 3290, May 2002.

   [11] R. Bless, K. Wehrle. Evaluation of Differentiated Services using
        an Implementation under Linux, Proceedings of the Intern.
        Workshop on Quality of Service (IWQOS'99), London, 1999.

   [12] K. Wehrle, J. Reber, V. Kahmann. A simulation suite for Internet
        nodes with the ability to integrate arbitrary Quality of Service
        behavior, Proceedings of Communication Networks And Distributed
        Systems Modeling And Simulation Conference (CNDS 2001), Phoenix
        (AZ), January 2001.

   [13] R. Bless, K. Wehrle. Group Communication in Differentiated
        Services Networks, Internet QoS for the Global Computing 2001
        (IQ 2001), IEEE International Symposium on Cluster Computing and
        the Grid, May 2001, Brisbane, Australia, IEEE Press.

   [14] Davie, B., Charny, A., Bennett, J.C.R., Benson, K., Le Boudec,
        J.Y., Courtney, W., Davari, S., Firoiu, V. and D. Stiliadis, "An
        Expedited Forwarding PHB (Per-Hop Behavior)", RFC 3246, March
        2002.

   [15] Charny, A., Bennett, J.C.R., Benson, K., Le Boudec, J.Y., Chiu,
        A., Courtney, W., Davari, S., Firoiu, V., Kalmanek, C. and K.K.
        Ramakrishnan, "Supplemental Information for the New Definition
        of the EF PHB (Expedited Forwarding Per-Hop Behavior)", RFC
        3247, March 2002.

   [16] Bless, R., Nichols, K. and K. Wehrle, "A Lower Effort Per-Domain
        Behavior (PDB) for Differentiated Services", RFC 3662, December
        2003.



















Bless & Wehrle               Informational                     [Page 32]



RFC 3754              IP Multicast in DS Networks             April 2004


12.  Authors' Addresses

   Comments and questions related to this document can be addressed to
   one of the authors listed below.

   Roland Bless
   Institute of Telematics
   Universitaet Karlsruhe (TH)
   Zirkel 2
   76128 Karlsruhe, Germany

   Phone: +49 721 608 6413
   EMail: bless@tm.uka.de
   URI: http://www.tm.uka.de/~bless


   Klaus Wehrle
   University of Tuebingen
   WSI - Computer Networks and Internet /
   Protocol Engineering & Distributed Systems
   Morgenstelle 10c
   72076 Tuebingen, Germany

   EMail: Klaus.Wehrle@uni-tuebingen.de
   URI: http://net.informatik.uni-tuebingen.de/~wehrle/


























Bless & Wehrle               Informational                     [Page 33]



RFC 3754              IP Multicast in DS Networks             April 2004


13.  Full Copyright Statement

   Copyright (C) The Internet Society (2004).  This document is subject
   to the rights, licenses and restrictions contained in BCP 78 and
   except as set forth therein, the authors retain all their rights.

   This document and the information contained herein are provided on an
   "AS IS" basis and THE CONTRIBUTOR, THE ORGANIZATION HE/SHE REPRESENTS
   OR IS SPONSORED BY (IF ANY), THE INTERNET SOCIETY AND THE INTERNET
   ENGINEERING TASK FORCE DISCLAIM ALL WARRANTIES, EXPRESS OR IMPLIED,
   INCLUDING BUT NOT LIMITED TO ANY WARRANTY THAT THE USE OF THE
   INFORMATION HEREIN WILL NOT INFRINGE ANY RIGHTS OR ANY IMPLIED
   WARRANTIES OF MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE.

Intellectual Property

   The IETF takes no position regarding the validity or scope of any
   Intellectual Property Rights or other rights that might be claimed to
   pertain to the implementation or use of the technology described in
   this document or the extent to which any license under such rights
   might or might not be available; nor does it represent that it has
   made any independent effort to identify any such rights.  Information
   on the procedures with respect to rights in RFC documents can be
   found in BCP 78 and BCP 79.

   Copies of IPR disclosures made to the IETF Secretariat and any
   assurances of licenses to be made available, or the result of an
   attempt made to obtain a general license or permission for the use of
   such proprietary rights by implementers or users of this
   specification can be obtained from the IETF on-line IPR repository at
   http://www.ietf.org/ipr.

   The IETF invites any interested party to bring to its attention any
   copyrights, patents or patent applications, or other proprietary
   rights that may cover technology that may be required to implement
   this standard.  Please address the information to the IETF at ietf-
   ipr@ietf.org.

Acknowledgement

   Funding for the RFC Editor function is currently provided by the
   Internet Society.









Bless & Wehrle               Informational                     [Page 34]



©2018 Martin Webb