RFC1584 - Multicast Extensions to OSPF

Network Working Group J. Moy
Request for Comments: 1584 Proteon, Inc.
Category: Standards Track March 1994
Multicast Extensions to OSPF
Status of this Memo
This document specifies an Internet standards track protocol for the
Internet community, and requests discussion and suggestions for
improvements. Please refer to the current edition of the "Internet
Official Protocol Standards" (STD 1) for the standardization state
and status of this protocol. Distribution of this memo is
unlimited.
Abstract
This memo documents enhancements to the OSPF protocol enabling the
routing of IP multicast datagrams. In this proposal, an IP multicast
packet is routed based both on the packet"s source and its multicast
destination (commonly referred to as source/destination routing). As
it is routed, the multicast packet follows a shortest path to each
multicast destination. During packet forwarding, any commonality of
paths is eXPloited; when multiple hosts belong to a single multicast
group, a multicast packet will be replicated only when the paths to
the separate hosts diverge.
OSPF, a link-state routing protocol, provides a database describing
the Autonomous System"s topology. A new OSPF link state
advertisement is added describing the location of multicast
destinations. A multicast packet"s path is then calculated by
building a pruned shortest-path tree rooted at the packet"s IP
source. These trees are built on demand, and the results of the
calculation are cached for use by subsequent packets.
The multicast extensions are built on top of OSPF Version 2. The
extensions have been implemented so that a multicast routing
capability can be introdUCed piecemeal into an OSPF Version 2
routing domain. Some of the OSPF Version 2 routers may run the
multicast extensions, while others may continue to be restricted to
the forwarding of regular IP traffic (unicasts).
Please send comments to mospf@gated.cornell.edu.
Table of Contents
1 Introduction ........................................... 4
1.1 Terminology ............................................ 5
1.2 Acknowledgments ........................................ 6
2 Multicast routing in MOSPF ............................. 6
2.1 Routing characteristics ................................ 6
2.2 Sample path of a multicast datagram .................... 8
2.3 MOSPF forwarding mechanism ............................ 10
2.3.1 IGMP interface: the local group database .............. 10
2.3.2 A datagram"s shortest-path tree ....................... 14
2.3.3 Support for Non-broadcast networks .................... 16
2.3.4 Details concerning forwarding cache entries ........... 16
3 Inter-area multicasting ............................... 18
3.1 Extent of group-membership-LSAs ....................... 19
3.2 Building inter-area datagram shortest-path trees ...... 22
4 Inter-AS multicasting ................................. 27
4.1 Building inter-AS datagram shortest-path trees ........ 28
4.2 Stub area behavior .................................... 30
4.3 Inter-AS multicasting in a core Autonomous System ..... 31
5 Modelling internal group membership ................... 31
6 Additional capabilities ............................... 33
6.1 Mixing with non-multicast routers ..................... 34
6.2 TOS-based multicast ................................... 35
6.3 Assigning multiple IP networks to a physical network .. 36
6.4 Networks on Autonomous System boundaries .............. 37
6.5 Recommended system configuration ...................... 38
7 Basic implementation requirements ..................... 40
8 Protocol data structures .............................. 40
8.1 Additions to the OSPF area structure .................. 41
8.2 Additions to the OSPF interface structure ............. 42
8.3 Additions to the OSPF neighbor structure .............. 43
8.4 The local group database .............................. 43
8.5 The forwarding cache .................................. 44
9 Interaction with the IGMP protocol .................... 45
9.1 Sending IGMP Host Membership Queries .................. 46
9.2 Receiving IGMP Host Membership Reports ................ 46
9.3 Aging local group database entries .................... 47
9.4 Receiving IGMP Host Membership Queries ................ 47
10 Group-membership-LSAs ................................. 48
10.1 Constructing group-membership-LSAs .................... 49
10.2 Flooding group-membership-LSAs ........................ 52
11 Detailed description of multicast datagram forwarding . 52
11.1 Associating a MOSPF interface with a received datagram 55
11.2 Locating the source network ........................... 55
11.3 Forwarding locally originated multicasts .............. 57
12 Construction of forwarding cache entries .............. 58
12.1 The Vertex data structure ............................. 59
12.2 The SPF calculation ................................... 60
12.2.1 Candidate list Initialization: Case SourceIntraArea ... 65
12.2.2 Candidate list Initialization: Case SourceInterArea1 .. 66
12.2.3 Candidate list Initialization: Case SourceInterArea2 .. 66
12.2.4 Candidate list Initialization: Case SourceExternal .... 67
12.2.5 Candidate list Initialization: Case SourceStubExternal 70
12.2.6 Processing labelled vertices .......................... 70
12.2.7 Merging datagram shortest-path trees .................. 71
12.2.8 TOS considerations .................................... 72
12.2.9 Comparison to the unicast SPF calculation ............. 74
12.3 Adding local database entries to the forwarding cache 75
13 Maintaining the forwarding cache ...................... 76
14 Other additions to the OSPF specification ............. 77
14.1 The Designated Router ................................. 77
14.2 Sending Hello packets ................................. 78
14.3 The Neighbor state machine ............................ 78
14.4 Receiving Database Description packets ................ 78
14.5 Sending Database Description packets .................. 79
14.6 Originating Router-LSAs ............................... 79
14.7 Originating Network-LSAs .............................. 79
14.8 Originating Summary-link-LSAs ......................... 80
14.9 Originating AS external-link-LSAs ..................... 80
14.10 Next step in the flooding procedure ................... 81
14.11 Virtual links ......................................... 81
15 References ............................................ 83
Footnotes ............................................. 84
A Data Formats .......................................... 88
A.1 The Options field ..................................... 89
A.2 Router-LSA ............................................ 91
A.3 Group-membership-LSA .................................. 93
B Configurable Constants ................................ 95
B.1 Global parameters ..................................... 95
B.2 Router interface parameters ........................... 95
C Sample datagram shortest-path trees ................... 97
C.1 An intra-area tree .................................... 98
C.2 The effect of areas .................................. 100
C.3 The effect of virtual links .......................... 101
Security Considerations .............................. 102
Author"s Address ..................................... 102
1. Introduction
This memo documents enhancements to OSPF Version 2 to support IP
multicast routing. The enhancements have been added in a backward-
compatible fashion; routers running the multicast additions will
interoperate with non-multicast OSPF routers when forwarding regular
(unicast) IP data traffic. The protocol resulting from the addition
of the multicast enhancements to OSPF is herein referred to as the
MOSPF protocol.
IP multicasting is an extension of LAN multicasting to a TCP/IP
internet. Multicasting support for TCP/IP hosts has been specified
in [RFC1112]. In that document, multicast groups are represented by
IP class D addresses. Individual TCP/IP hosts join (and leave)
multicast groups through the Internet Group Management Protocol
(IGMP, also specified in [RFC1112]). A host need not be a member of
a multicast group in order to send datagrams to the group. Multicast
datagrams are to be delivered to each member of the multicast group
with the same "best-effort" delivery accorded regular (unicast) IP
data traffic.
MOSPF provides the ability to forward multicast datagrams from one
IP network to another (i.e., through internet routers). MOSPF
forwards a multicast datagram on the basis of both the datagram"s
source and destination (this is sometimes called source/destination
routing). The OSPF link state database provides a complete
description of the Autonomous System"s topology. By adding a new
type of link state advertisement, the group-membership-LSA, the
location of all multicast group members is pinpointed in the
database. The path of a multicast datagram can then be calculated by
building a shortest-path tree rooted at the datagram"s source. All
branches not containing multicast members are pruned from the tree.
These pruned shortest-path trees are initially built when the first
datagram is received (i.e., on demand). The results of the shortest
path calculation are then cached for use by subsequent datagrams
having the same source and destination.
OSPF allows an Autonomous System to be split into areas. However,
when this is done complete knowledge of the Autonomous System"s
topology is lost. When forwarding multicasts between areas, only
incomplete shortest-path trees can be built. This may lead to some
inefficiency in routing. An analogous situation exists when the
source of the multicast datagram lies in another Autonomous System.
In both cases (i.e., the source of the datagram belongs to a
different OSPF area, or to a different Autonomous system) the
neighborhood immediately surrounding the source is unknown. In these
cases the source"s neighborhood is approximated by OSPF summary link
advertisements or by OSPF AS external link advertisements
respectively.
Routers running MOSPF can be intermixed with non-multicast OSPF
routers. Both types of routers can interoperate when forwarding
regular (unicast) IP data traffic. Obviously, the forwarding extent
of IP multicasts is limited by the number of MOSPF routers present
in the Autonomous System (and their interconnection, if any). An
ability to "tunnel" multicast datagrams through non-multicast
routers is not provided. In MOSPF, just as in the base OSPF
protocol, datagrams (multicast or unicast) are routed "as is" --
they are not further encapsulated or decapsulated as they transit
the Autonomous System.
1.1. Terminology
This memo uses the terminology listed in section 1.2 of [OSPF].
For this reason, terms such as "Network", "Autonomous System"
and "link state advertisement" are assumed to be understood. In
addition, the abbreviation LSA is used for "link state
advertisement". For example, router links advertisements are
referred to as router-LSAs and the new link state advertisement
describing the location of members of a multicast group is
referred to as a group-membership-LSA.
[RFC1112] discusses the data-link encapsulation of IP multicast
datagrams. In contrast to the normal forwarding of IP unicast
datagrams, on a broadcast network the mapping of an IP multicast
destination to a data-link destination address is not done with
the ARP protocol. Instead, static mappings have been defined
from IP multicast destinations to data-link addresses. These
mappings are dependent on network type; for some networks IP
multicasts are algorithmically mapped to data-link multicast
addresses, for other networks all IP multicast destinations are
mapped onto the data-link broadcast address. This document
loosely describes both of these possible mappings as data-link
multicast.
The following terms are also used throughout this document:
o Non-multicast router. A router running OSPF Version 2, but
not the multicast extensions. These routers do not forward
multicast datagrams, but can interoperate with MOSPF routers
in the forwarding of unicast packets. Routers running the
MOSPF protocol are referred to herein as either multicast-
capable routers or MOSPF routers.
o Non-broadcast networks. A network supporting the attachment
of more than two stations, but not supporting the delivery
of a single physical datagram to multiple destinations
(i.e., not supporting data-link multicast). [OSPF] describes
these networks as non-broadcast, multi-Access networks. An
example of a non-broadcast network is an X.25 PDN.
o Transit network. A network having two or more OSPF routers
attached. These networks can forward data traffic that is
neither locally-originated nor locally-destined. In OSPF,
with the exception of point-to-point networks and virtual
links, the neighborhood of each transit network is described
by a network links advertisement (network-LSA).
o Stub network. A network having only a single OSPF router
attached. A network belonging to an OSPF system is either a
transit or a stub network, but never both.
1.2. Acknowledgments
The multicast extensions to OSPF are based on Link-State
Multicast Routing algorithm presented in [Deering]. In addition,
the [Deering] paper contains a section on Hierarchical Multicast
Routing (providing the ideas for MOSPF"s inter-area multicasting
scheme) and several Distance Vector (also called Bellman-Ford)
multicast algorithms. One of these Distance Vector multicast
algorithms, Truncated Reverse Path Broadcasting, has been
implemented in the Internet (see [RFC1075]).
The MOSPF protocol has been developed by the MOSPF Working Group
of the Internet Engineering Task Force. Portions of this work
have been supported by DARPA under NASA contract NAG 2-650.
2. Multicast routing in MOSPF
This section describes MOSPF"s basic multicast routing algorithm.
The basic algorithm, run inside a single OSPF area, covers the case
where the source of the multicast datagram is inside the area
itself. Within the area, the path of the datagram forms a tree
rooted at the datagram source.
2.1. Routing characteristics
As a multicast datagram is forwarded along its shortest-path
tree, the datagram is delivered to each member of the
destination multicast group. In MOSPF, the forwarding of the
multicast datagram has the following properties:
o The path taken by a multicast datagram depends both on the
datagram"s source and its multicast destination. Called
source/destination routing, this is in contrast to most
unicast datagram forwarding algorithms (like OSPF) that
route based solely on destination.
o The path taken between the datagram"s source and any
particular destination group member is the least cost path
available. Cost is expressed in terms of the OSPF link-state
metric. For example, if the OSPF metric represents delay, a
minimum delay path is chosen. OSPF metrics are configurable.
A metric is assigned to each outbound router interface,
representing the cost of sending a packet on that interface.
The cost of a path is the sum of its constituent (outbound)
router interfaces[1].
o MOSPF takes advantage of any commonality of least cost paths
to destination group members. However, when members of the
multicast group are spread out over multiple networks, the
multicast datagram must at times be replicated. This
replication is performed as few times as possible (at the
tree branches), taking maximum advantage of common path
segments.
o For a given multicast datagram, all routers calculate an
identical shortest-path tree. There is a single path between
the datagram"s source and any particular destination group
member. This means that, unlike OSPF"s treatment of regular
(unicast) IP data traffic, there is no provision for equal-
cost multipath.
o On each packet hop, MOSPF normally forwards IP multicast
datagrams as data-link multicasts. There are two exceptions.
First, on non-broadcast networks, since there are no data-
link multicast/broadcast services the datagram must be
forwarded to specific MOSPF neighbors (see Section 2.3.3).
Second, a MOSPF router can be configured to forward IP
multicasts on specific networks as data-link unicasts, in
order to avoid datagram replication in certain anomalous
situations (see Section 6.4).
While MOSPF optimizes the path to any given group member, it
does not necessarily optimize the use of the internetwork as a
whole. To do so, instead of calculating source-based shortest-
path trees, something similar to a minimal spanning tree
(containing only the group members) would need to be calculated.
This type of minimal spanning tree is called a Steiner tree in
the literature. For a comparison of shortest-path tree routing
to routing using Steiner trees, see [Deering2] and [Bharath-
Kumar].
2.2. Sample path of a multicast datagram
As an example of multicast datagram routing in MOSPF, consider
the sample Autonomous System pictured in Figure 1. This figure
has been taken from the OSPF specification (see [OSPF]). The
larger rectangles represent routers, the smaller rectangles
hosts. Oblongs and circles represent multi-access networks[2].
Lines joining routers are point-to-point serial connections. A
cost has been assigned to each outbound router interface.
All routers in Figure 1 are assumed to be running MOSPF. A
number of hosts have been added to the figure. The hosts
labelled Ma have joined a particular multicast group (call it
Group A) via the IGMP protocol. These hosts are located on
networks N2, N6 and N11. Similarly, using IGMP the hosts
labelled Mb have joined a separate multicast group B; these
hosts are located on networks N1, N2 and N3. Note that hosts can
join multiple multicast groups; it is only for clarity of
presentation that each host has joined at most one multicast
group in this example. Also, hosts H2 through H5 have been
added to the figure to serve as sources for multicast datagrams.
Again, the datagrams" sources have been made separate from the
group members only for clarity of presentation.
To illustrate the forwarding of a multicast datagram, suppose
that Host H2 (attached to Network N4) sends a multicast datagram
to multicast group B. This datagram originates as a data-link
layer multicast on Network N4. Router RT3, being a multicast
router, has "opened up" its interface data-link multicast
filters. It therefore receives the multicast datagram, and
attempts to forward it to the members of multicast group B
(located on networks N1, N2 and N3). This is accomplished by
sending a single copy of the datagram onto Network N3, again as
a data-link multicast[3]. Upon receiving the multicast datagram
from RT3, routers RT1 and RT2 will then multicast the datagram
on their connected stub networks (N1 and N2 respectively). Note
that, since the datagram is sent onto Network N3 as a data-link
multicast, Router RT4 will also receive a copy. However, it will
not forward the datagram, since it does not lie on a shortest
path between the source (Host H2) and any members of multicast
group B.
Note that the path of the multicast datagram depends on the
datagram"s source network. If the above multicast datagram was
instead originated by Host H3, the path taken would be
identical, since hosts H2 and H3 lie on the same network
(Network N4). However, if the datagram was originated by Host
H4, its path would be different. In this case, when Router RT3
+
3+---+ +--+ +--+ N12 N14
N1--RT11 Mb H4 N13 /
_ +---+ +--+ /+--+ 8 8/8
+ ___/ /
+--+ +--+ / 1+---+8 8+---+6
Mb Mb * N3 *---RT4------RT5--------+
+--+ /+--+ ____/ +---+ +---+
+ / 7
3+---+ /
N2--RT2/1 1 6
__ +---+ +---+8 6+---+
+ RT3--------------RT6
+--+ +--+ +---+ +--+ +---+
Ma H3_ 2 _H2 Ia7
+--+ +--+ / +--+
+---------+
N4


N11
+---------+
N12
3 +--+ 6 2/
+---+ Ma +---+/
RT9 +--+ RT7---N15
+---+ +---+ 9
1 + 1
___ Ib5 ___ +--+
/ 1+----+2 3+----+1 / --Ma
* N9 *------RT11-------RT10---* N6 * +--+
____/ +----+ +----+ ____/

1 + 1
+--+ 10+----+ N8 +---+
H1-----RT12 RT8
+--+SLIP +----+ +---+ +--+
2 4 _H5
/ +--+
+---------+ +--------+
N10 N7
Figure 1: A sample MOSPF configuration
receives the datagram, RT3 will drop the datagram instead of
forwarding it (since RT3 is no longer on the shortest path to
any member of Group B).
Note that the path of the multicast datagram also depends on the
destination multicast group. If Host H2 sends a multicast to
Group A, the path taken is as follows. The datagram again starts
as a multicast on Network N4. Router RT3 receives it, and
creates two copies. One is sent onto Network N3 which is then
forwarded onto Network N2 by RT2. The other copy is sent to
Router RT10 (via RT6), where the datagram is again split,
eventually to be delivered onto networks N6 and N11. Note that,
although multiple copies of the datagram are produced, the
datagram itself is not modified (except for its IP TTL) as it is
forwarded. No encapsulation of the datagram is performed; the
destination of the datagram is always listed as the multicast
group A.
2.3. MOSPF forwarding mechanism
Each MOSPF router in the path of a multicast datagram bases its
forwarding decision on the contents of a data cache. This cache
is called the forwarding cache. There is a separate forwarding
cache entry for each source/destination combination[4]. Each
cache entry indicates, for multicast datagrams having matching
source and destination, which neighboring node (i.e., router or
network) the datagram must be received from (called the upstream
node) and which interfaces the datagram should then be forwarded
out of (called the downstream interfaces).
A forwarding cache entry is actually built from two component
pieces. The first of these components is called the local group
database. This database, built by the IGMP protocol, indicates
the group membership of the router"s directly attached networks.
The local group database enables the local delivery of multicast
datagrams. The second component is the datagram"s shortest path
tree. This tree, built on demand, is rooted at a multicast
datagram"s source. The datagram"s shortest path tree enables the
delivery of multicast datagrams to distant (i.e., not directly
attached) group members.
2.3.1. IGMP interface: the local group database
The local group database keeps track of the group membership
of the router"s directly attached networks. Each entry in
the local group database is a [group, attached network]
pair, which indicates that the attached network has one or
more IP hosts belonging to the IP multicast destination
group. This information is then used by the router when
deciding which directly attached networks to forward a
received IP multicast datagram onto, in order to complete
delivery of the datagram to (local) group members.
The local group database is built through the operation of
the Internet Group Management Protocol (IGMP; see [RFC
1112]). When a MOSPF router becomes Designated Router on an
attached network (call the network N1), it starts sending
periodic IGMP Host Membership Queries on the network. Hosts
then respond with IGMP Host Membership Reports, one for each
multicast group to which they belong. Upon receiving a Host
Membership Report for a multicast group A, the router
updates its local group database by adding/refreshing the
entry [Group A, N1]. If at a later time Reports for Group A
cease to be heard on the network, the entry is then deleted
from the local group database.
It is important to note that on any particular network, the
sending of IGMP Host Membership Queries and the listening to
IGMP Host Membership Reports is performed solely by the
Designated Router. A MOSPF router ignores Host Membership
Reports received on those networks where the router has not
been elected Designated Router[5]. This means that at most
one router performs these IGMP functions on any particular
network, and ensures that the network appears in the local
group database of at most one router. This prevents
multicast datagrams from being replicated as they are
delivered to local group members. As a result, each router
in the Autonomous System has a different local group
database. This is in contrast to the MOSPF link state
database, and the datagram shortest-path trees (see Section
2.3.2), all of which are identical in each router belonging
to the Autonomous System.
The existence of local group members must be communicated to
the rest of the routers in the Autonomous System. This
ensures that a remotely-originated multicast datagram will
be forwarded to the router for distribution to its local
group members. This communication is accomplished through
the creation of a group-membership-LSA. Like other link
state advertisements, the group-membership-LSA is flooded
throughout the Autonomous System. The router originates a
separate group-membership-LSA for each multicast group
having one or more entries in the router"s local group
database. The router"s group-membership-LSA (say for Group
A) lists those local transit vertices (i.e., the router
itself and/or any directly connected transit networks) that
should not be pruned from Group A"s datagram shortest-path
trees. The router lists itself in its group-membership-LSA
for Group A if either 1) one or more of the router"s
attached stub networks contain Group A members or 2) the
router itself is a member of Group A. The router lists a
directly connected transit network in the group-membership-
LSA for Group A if both 1) the router is Designated Router
on the network and 2) the network contains one or more Group
A members.
Consider again the example pictured in Figure 1. If Router
RT3 has been elected Designated Router for Network N3, then
Table 1: lists the local group database for the routers
RT1-RT4.
In this case, each of the routers RT1, RT2 and RT3 will
originate a group-membership-LSA for Group B. In addition,
RT2 will also be originating a group-membership-LSA for
Group A. RT1 and RT2"s group-membership-LSAs will list
solely the routers themselves (N1 and N2 are stub networks).
RT3"s group-membership-LSA will list the transit Network N3.
Figure 2 displays the Autonomous System"s link state
database. A router/transit network is labelled with a
multicast group if (and only if) it has been mentioned in a
group-membership-LSA for the group When building the
shortest-path tree for a particular multicast datagram, this
labelling enables those branches without group members to be
pruned from the tree. The process of building a multicast
datagram"s shortest path tree is discussed in Section 2.3.2.
Note that none of the hosts in Figure 1 belonging to
multicast groups A and B show up explicitly in the link
state database (see Figure 2). In fact, looking at the link
state database you cannot even determine which stub networks
Router local group database
_____________________________________
RT1 [Group B, N1]
RT2 [Group A, N2], [Group B, N2]
RT3 [Group B, N3]
RT4 None
Table 1: Sample local group databases
**FROM**
RTRTRTRTRTRTRTRTRTRTRTRT
1 2 3 4 5 6 7 8 9 101112N3N6N8N9
----- ---------------------------------------------
RT1 0
RT2 0
RT3 6 0
RT4 80
RT5 8 6 6
RT6 8 7 5
RT7 6 0
* RT8 0
* RT9 0
T RT10 70 0
O RT11 0 0
* RT12 0
* N13
N2 3
N31 1 1 1
N4 2
N6 1 1 1
N7 4
N8 3 2
N91 1 1
N10 2
N113
N12 8 2
N13 8
N14 8
N15 9
H1 10
Figure 2: The MOSPF database.
Networks and routers are represented by vertices.
An edge of cost X connects Vertex A to Vertex B iff
the intersection of Column A and Row B is marked
with an X. In addition, RT1, RT2 and N3 are labelled
with multicast group A and RT1, N6 and RT9 are
labelled with multicast group B.
contain multicast group members. The link state database
simply indicates those routers/transit networks having
attached group members. This is all that is necessary for
successful forwarding of multicast datagrams.
2.3.2. A datagram"s shortest-path tree
While the local group database facilitates the local
delivery of multicast datagrams, the datagram"s shortest-
path tree describes the intermediate hops taken by a
multicast datagram as it travels from its source to the
individual multicast group members. As mentioned above, the
datagram"s shortest-path tree is a pruned shortest-path tree
rooted at the datagram"s source. Two datagrams having
differing [source net, multicast destination] pairs may
have, and in fact probably will have, different pruned
shortest-path trees.
A datagram"s shortest path tree is built "on demand"[6],
i.e., when the first multicast datagram is received having a
particular [source net, multicast destination] combination.
To build the datagram"s shortest-path tree, the following
calculations are performed. First, the datagram"s source IP
network is located in the link state database. Then using
the router-LSAs and network-LSAs in the link state database,
a shortest-path tree is built having the source network as
root. To complete the process, the branches that do not
contain routers/transit networks that have been labelled
with the particular multicast destination (via a group-
membership-LSA) are pruned from the tree.
As an example of the building of a datagram"s shortest path
tree, again consider the Autonomous System in Figure 1. The
Autonomous System"s link state database is pictured in
Figure 2. When a router initially receives a multicast
datagram sent by Host H2 to the multicast group A, the
following steps are taken: Host H2 is first determined to be
on Network N4. Then the shortest path tree rooted at net N4
is calculated[7], pruning those branches that do not contain
routers/transit networks that have been labelled with the
multicast group A. This results in the pruned shortest-path
tree pictured in Figure 3. Note that at this point all the
leaves of the tree are routers/transit networks labelled
with multicast group A (routers RT2 and RT9 and transit
Network N6).
In order to forward the multicast datagram, each router must
find its own position in the datagram"s shortest path tree.
o RT3 (N4, origin)
/
1/ 8
/
N3 (Mb) o o RT6
/
0/ 7
/
RT2 (Ma,Mb) o o RT10
/
3/ 1
/
N8 o o N6 (Ma)
/
0/
/
RT11 o
/
1/
/
N9 o
/
0/
/
RT9 (Ma) o
Figure 3: Sample datagram"s shortest-path tree,
source N4, destination Group A
The router"s (call it Router RTX) position in the datagram"s
pruned shortest-path tree consists of 1) RTX"s parent in the
tree (this will be the forwarding cache entry"s upstream
node) and 2) the list of RTX"s interfaces that lead to
downstream routers/transit networks that have been labelled
with the datagram"s destination (these will be added to the
forwarding cache entry as downstream interfaces). Note that
after calculating the datagram"s shortest path tree, a
router may find that it is itself not on the tree. This
would be indicated by a forwarding cache entry having no
upstream node or an empty list of downstream interfaces.
As an example of a router describing its position on the
datagram"s shortest-path tree, consider Router RT10 in
Figure 3. Router RT10"s upstream node for the datagram is
Router RT6, and there are two downstream interfaces: one
connecting to Network N6 and the other connecting to Network
N8.
2.3.3. Support for Non-broadcast networks
When forwarding multicast datagrams over non-broadcast
networks, the datagram cannot be sent as a link-level
multicast (since neither link-level multicast nor broadcast
are supported on these networks), but must instead be
forwarded separately to specific neighbors. To facilitate
this, forwarding cache entries can also contain downstream
neighbors as well as downstream interfaces.
The IGMP protocol is not defined over non-broadcast
networks. For this reason, there cannot be group members
directly attached to non-broadcast networks, nor do non-
broadcast networks ever appear in local group database
entries.
As an example, suppose that Network N3 in Figure 1 is an
X.25 PDN. Consider Router RT3"s forwarding cache entry for
datagrams having source Network N4 and multicast destination
Group B. In place of having the interface to Network N3
appear as the downstream interface in the matching
forwarding cache entry, the neighboring routers RT1 and RT2
would instead appear as separate downstream neighbors. In
addition, in this case there could not be a Group B member
directly attached to Network N3.
2.3.4. Details concerning forwarding cache entries
Each of the downstream interface/neighbors in the cache
entry is labelled with a TTL value. This value indicates the
number of hops a datagram forwarded out of the interface (or
forwarded to the neighbor) would have to travel before
encountering a router/transit network requesting the
multicast destination. The reason that a hop count is
associated with each downstream interface/neighbor is so
that IP multicast"s expanding ring search procedure can be
more efficiently implemented. By expanding ring search is
meant the following. Hosts can restrict the frowarding
extent of the IP multicast datagrams that they send by
appropriate setting of the TTL value in the datagram"s IP
header. Then, for example, to search for the nearest server
the host can send multicasts first with TTL set to 1, then
2, etc. By attaching a hop count to each downstream
interface/neighbor in the forwarding cache, datagrams will
not be forwarded unless they will ultimately reach a
multicast destination before their TTL expires[8]. This
avoids wasting network bandwidth during an expanding ring
search.
As an example consider Router RT10"s forwarding cache in
Figure 3. Router RT10"s cache entry has two downstream
interfaces. The first, connecting to Network N6, is labelled
as having a group member one hop away (Network N6). The
second, which connects to Network N8, is labelled as having
a group member two hops away (Router RT9).
Both the datagram shortest path tree and the local group
database may contribute downstream interfaces to the
forwarding cache entries. As an example, if a router has a
local group database entry of [Group G, NX], then a
forwarding cache entry for Group G, regardless of
destination, will list the router interface to Network NX as
a downstream interface. Such a downstream interface will
always be labelled with a TTL of 1.
As an example of forwarding cache entries, again consider
the Autonomous System pictured in Figure 1. Suppose Host H2
sends a multicast datagram to multicast group A. In that
case, some routers will not even attempt to build a
forwarding cache entry (e.g, router RT5) because they will
never receive the multicast datagram in the first place.
Other routers will receive the multicast datagram (since
they are forwarded as link-level multicasts), but after
building the pruned shortest path tree will notice that they
themselves are not a part of the tree (routers RT1, RT4,
RT7, RT8 and RT12). These latter routers will install an
empty cache entry, indicating that they do not participate
in the forwarding of the multicast datagram. A sample of the
forwarding cache entries built by the other routers in the
Autonomous System is pictured in Table 2.
A MOSPF router must clear its entire forwarding cache when
the Autonomous System"s topology changes, because all the
datagram shortest-path trees must be rebuilt. Likewise, when
the location of a multicast group"s membership changes
(reflected by a change in group-membership-LSAs), all cache
entries associated with the particular multicast destination
group must be cleared. Other than these two cases,
forwarding cache entries need not ever be deleted or
otherwise modified; in particular, the forwarding cache
entries do not have to be aged. However, forwarding cache
entries can be freely deleted after some period of
inactivity (i.e., garbage collected), if router memory
Router Upstream Downstream interfaces
node (interface:hops)
___________________________________________
RT10 Router RT6 (N6:1), (N8:2)
RT11 Net N8 (N9:1)
RT3 Net N4 (N3:1), (RT6:3)
RT6 Router RT3 (RT10:2)
RT2 Net N3 (N2:1)
Table 2: Sample forwarding cache entries,
for source N4 and destination Group A.
resources are in short supply.
3. Inter-area multicasting
Up to this point this memo has discussed multicast forwarding when
the entire Autonomous System is a single OSPF area. The logic for
when the multicast datagram"s source and its destination group
members belong to the same OSPF area is the same. This section
explains the behavior of the MOSPF protocol when the datagram"s
source and (at least some of) its destination group members belong
to different OSPF areas. This situation is called inter-area
multicast.
Inter-area multicast brings up the following issues, which are
resolved in succeeding sections:
o Are the group-membership-LSAs specific to a single area? And if
they are, how is group membership information conveyed from one
area to the next?
o How are the datagram shortest-path trees built in the inter-area
case, since complete information concerning the topology of the
datagram source"s neighborhood is not available to routers in
other areas?
o In an area border router, multiple datagram shortest-path trees
are built, one for each attached area. How are these separate
datagram shortest-path trees combined into a single forwarding
cache entry?
It should be noted in the following that the basic protocol
mechanisms in the inter-area case are the same as for the intra-area
case. Forwarding of multicasts is still defined by the contents of
the forwarding cache. The forwarding cache is still built from the
same two components: the local group database and the datagram
shortest-path trees. And while the calculation of the datagram
shortest-path trees is different in the inter-area case (see Section
3.2), the local group database is built exactly the same as in the
intra-area case (i.e., MOSPF"s interface with IGMP remains unchanged
in the presence of areas). Finally, the forwarding algorithm
described in Section 11 is the same for both the intra-area and
inter-area cases.
The following discussion uses the area configuration pictured in
Figure 4 as an example. This figure, taken from the OSPF
specification, shows an Autonomous System split into three areas
(Area 1, Area 2 and Area 3). A single backbone area has been
configured (everything outside of the shading). Since the backbone
area must be contiguous, a single virtual link has been configured
between the area border routers RT10 and RT11. Additionally, an area
address range has been configured in Router RT11 so that Networks
N9-N11 and Host H1 will be reported as a single route outside of
Area 3 (via summary-link-LSAs).
3.1. Extent of group-membership-LSAs
Group-membership-LSAs are specific to a single OSPF area. This
means that, just as with OSPF router-LSAs, network-LSAs and
summary-link-LSAs, a group-membership-LSA is flooded throughout
a single area only[9]. A router attached to multiple areas
(i.e., an area border router) may end up originating several
group-membership-LSAs concerning a single multicast destination,
one for each attached area. However, as we will see below, the
contents of these group-membership-LSAs will vary depending on
their associated areas.
Just as in OSPF, each MOSPF area has its own link state
database. The MOSPF database is simply the OSPF link state
database enhanced by the group-membership-LSAs. Consider again
the area configuration pictured in Figure 4. The result of
adding the group-membership-LSAs to the area databases yields
the databases pictured in Figures 6 and 7. Figure 6 shows Area
1"s MOSPF database. Figure 7 shows the backbone"s MOSPF
database. Superscripts indicate which transit vertices have been
advertised as requesting particular multicast destinations. A
superscript of "w" indicates that the router is advertising
itself as a wild-card multicast receiver (see below). The dashed
lines are OSPF summary-link-LSAs or AS external-link-LSAs. Note
in Figure 7 that Router RT11 has condensed its routes to
Networks N9-N11 and Host H1 into a single summary-link-LSA.
..................................
. + .
. 3+---+ +--+ +--+ . N12 N14
. N1--RT11 Mb H4 . N13 /
. _ +---+ +--+ /+--+ . 8 8/8
. + ___/ . /
. +--+ +--+ / 1+---+8. 8+---+6
. Mb Mb * N3 *---RT4------RT5--------+
. +--+ /+--+ ____/ +---+ . +---+
. + / . 7
. 3+---+ / .
. N2--RT2/1 1 . 6
. __ +---+ +---+8 . 6+---+
. + RT3--------------RT6
. +--+ +--+ +---+ +--+. +---+
. Ma H3_ 2 _H2. Ia7
. +--+ +--+ / +--+.
. +---------+ .
.Area 1 N4 .
..................................
................................
. N11 .
. +---------+ .
. . N12
. 3 +--+ . 6 2/
. +---+ Ma . +---+/
. RT9 +--+ . RT7---N15
. +---+ ....... +---+ 9
. 1 .. + .............1........
. ___ .. Ib5 ___ +--+.
. / 1+----+2.. 3+----+1 / --Ma.
. * N9 *------RT11-------RT10---* N6 * +--+.
. ____/ +----+ .. +----+ ____/ .
. !************! .
. 1 Virtual + Link 1 .
. +--+ 10+----+ ..N8 +---+ .
. H1-----RT12 .. RT8 .
. +--+SLIP +----+ .. +---+ +--+.
. 2 .. 4 _H5.
. .. / +--+.
. +---------+ .. +--------+ .
. N10 Area 3..Area 2 N7 .
.............................................................
Figure 4: A sample MOSPF area configuration
Suppose an OSPF router has a local group database entry for
[Group Y, Network X]. The router then originates a group-
membership-LSA for Group Y into the area containing Network X.
For example, in the area configuration pictured in Figure 4,
Router RT1 originates a group-membership-LSA for Group B. This
group-membership-LSA is flooded throughout Area 1, and no
further. Likewise, assuming that Router RT3 has been elected
Designated Router for Network N3, RT3 originates a group-
membership-LSA into Area 1 listing the transit Network N3 as
having group members. Note that in the link state database for
Area 1 (Figure 6) both Router RT1 and Network N3 have
accordingly been labelled with Group B.
In OSPF, the area border routers forward routing information and
data traffic between areas. In MOSPF. a subset of the area
border routers, called the inter-area multicast forwarders,
forward group membership information and multicast datagrams
between areas. Whether a given OSPF area border router is also a
MOSPF inter-area multicast forwarder is configuration dependent
(see Section B.1). In Figure 4 we assume that all area border
routers are also inter-area multicast forwarders.
In order to convey group membership information between areas,
inter-area multicast forwarders "summarize" their attached
areas" group membership to the backbone. This is very similar
functionality to the summary-link-LSAs that are generated in the
base OSPF protocol. An inter-area multicast forwarder
calculates which groups have members in its attached non-
backbone areas. Then, for each of these groups, the inter-area
multicast forwarder injects a group-membership-LSA into the
backbone area. For example, in Figure 4 there are two groups
having members in Area 1: Group A and Group B. For that reason,
both of Area 1"s inter-area multicast forwarders (Routers RT3
and RT4) inject group-membership-LSAs for these two groups into
the backbone. As a result both of these routers are labelled
membership +------------------+ datagrams
+ > > > >> Backbone < < < < +
^ +------------------+ ^
^ / ^
^ / ^
+----^-----+/ +----------+ +----^-----+
Area 1 Area 2 Area 3
+----------+ +----------+ +----------+
Figure 5: Inter-area routing architecture
with Groups A and B in the backbone link state database (see
Figure 7).
However, unlike the summarization of unicast destinations in the
base OSPF protocol, the summarization of group membership in
MOSPF is asymmetric. While a non-backbone area"s group
membership is summarized to the backbone, this information is
not then readvertised into other non-backbone areas. Nor is the
backbone"s group membership summarized for the non-backbone
areas. Going back to the example in Figure 4, while the presence
of Area 3"s group (Group A) is advertised to the backbone, this
information is not then redistributed to Area 1. In other Words,
routers internal to Area 1 have no idea of Area 3"s group
membership.
At this point, if no extra functionality was added to MOSPF,
multicast traffic originating in Area 1 destined for Multicast
Group A would never be forwarded to those Group A members in
Area 3. To accomplish this, the notion of wild-card multicast
receivers is introduced. A wild-card multicast receiver is a
router to which all multicast traffic, regardless of multicast
destination, should be forwarded. A router"s wild-card multicast
reception status is per-area. In non-backbone areas, all inter-
area multicast forwarders[10] are wild-card multicast receivers.
This ensures that all multicast traffic originating in a non-
backbone area will be forwarded to its inter-area multicast
forwarders, and hence to the backbone area. Since the backbone
has complete knowledge of all areas" group membership, the
datagram can then be forwarded to all group members. Note that
in the backbone itself there is no need for wild-card multicast
receivers[11]. As an example, note that Routers RT3 and RT4 are
wild-card multicast receivers in Area 1 (see Figure 6), while
there are none in the backbone (see Figure 7).
This yields the inter-area routing architecture pictured in
Figure 5. All group membership is advertised by the non-
backbone areas into the backbone. Likewise, all IP multicast
traffic arising in the non-backbone areas is forwarded to the
backbone. Since at this point group membership information meets
the multicast datagram traffic, delivery of the multicast
datagrams becomes possible.
3.2. Building inter-area datagram shortest-path trees
When building datagram shortest-path trees in the presence of
areas, it is often the case that the source of the datagram and
(at least some of) the destination group members are in separate
areas. Since detailed topological information concerning one
**FROM**
RTRTRTRTRTRT
1 2 3 4 5 7 N3
----- -------------------
RT1 0
RT2 0
RT3 0
* RT4 0
* RT5 148
T RT7 2014
O N13
* N2 3
* N31 1 1 1
N4 2
Ia,Ib 1522
N6 1615
N7 2019
N8 1818
N9-N11,H1 1916
N12 8 2
N13 8
N14 8
N15 9
Figure 6: Area 1"s MOSPF database.
Networks and routers are represented by vertices.
An edge of cost X connects Vertex A to Vertex B iff
the intersection of Column A and Row B is marked
with an X. In addition, RT1, RT2 and N3 are labelled
with multicast group A, RT1 is labelled with multicast
group B, and both RT3 and RT4 are labelled as
wild-card multicast receivers.
**FROM**
RTRTRTRTRTRTRT
3 4 5 6 7 1011
------------------------
RT3 6
RT4 8
RT5 8 6 6
RT68 7 5
RT7 6
* RT10 7 2
* RT11 3
T N14 4
O N24 4
* N31 1
* N42 3
Ia 5
Ib 7
N6 1 1 3
N7 5 5 7
N8 4 3 2
N9-N11,H1 1
N12 8 2
N13 8
N14 8
N15 9
Figure 7: The backbone"s MOSPF database.
Networks and routers are represented by vertices.
An edge of cost X connects Vertex A to Vertex B iff
the intersection of Column A and Row B is marked
with an X. In addition, RT3 and RT4 are labelled
with both multicast groups A and B, and RT7, RT10,
and RT11 are labelled with multicast group A.
OSPF area is not distributed to other OSPF areas (the flooding
of router-LSAs, network-LSAs and group-membership-LSAs is
restricted to a single OSPF area only), the building of complete
datagram shortest-path trees is often impossible in the inter-
area case. To compensate, approximations are made through the
use of wild-card multicast receivers and OSPF summary-link-LSAs.
When it first receives a datagram for a particular [source net,
destination group] pair, a router calculates a separate datagram
shortest-path tree for each of the router"s attached areas. Each
datagram shortest-path tree is built solely from LSAs belonging
to the particular area"s link state database. Suppose that a
router is calculating a datagram shortest-path tree for Area A.
It is useful then to separate out two cases.
The first case, Case 1: The source of the datagram belongs to
Area A has already been described in Section 2.3.2. However, in
the presence of OSPF areas, during tree pruning care must be
taken so that the branches leading to other areas remain, since
it is unknown whether there are group members in these (remote)
areas. For this reason, only those branches having no group
members nor wild-card multicast receivers are pruned when
producing the datagram shortest-path tree.
As an example, suppose in Figure 4 that Host H2 sends a
multicast datagram to destination Group A. Then the datagram"s
shortest-path tree for Area 1, built identically by all routers
in Area 1 that receive the datagram, is shown in Figure 8. Note
that both inter-area multicast forwarders (RT3 and RT4) are on
the datagram"s shortest-path tree, ensuring the delivery of the
datagram to the backbone and from there to Areas 2 and 3.
o Case 2: The source of the datagram belongs to an area other
than Area A. In this case, when building the datagram
shortest-path tree for Area A, the immediate neighborhood of
the datagram"s source is unknown. However, there are
summary-link-LSAs in the Area A link state database
indicating the cost of the paths between each of Area A"s
inter-area multicast forwarders and the datagram source.
These summary links are used to approximate the neighborhood
of the datagram"s source; the tree begins with links
directly connecting the source to each of the inter-area
multicast forwarders. These links point in the reverse
o RT3 (W, origin=N4)

1

N3 (Mb) o
/
0/
/
RT2 (Ma,Mb) o o RT4 (W)
Figure 8: Datagram"s shortest-path tree,
Area 1, source N4, destination Group A
direction (towards instead of away from the datagram source)
from the links considered in Case 1 above. All additional
links added to the tree also point in the reverse direction.
The final datagram shortest-path tree is then produced by,
as before, pruning all branches having no group-members nor
wild-card multicast receivers.
As an example, suppose again that Host H2 in Figure 4 sends
a multicast datagram to destination Group A. The datagram"s
shortest-path tree for the backbone is shown in Figure 9.
The neighborhood around the source (Network N4) has been
approximated by the summary links advertised by routers RT3
and RT4. Note that all links in Figure 9"s datagram
shortest-path tree have arrows pointing in the reverse
direction, towards Network N4 instead of away from it.
The reverse costs used for the entire tree in Case 2 are forced
because summary-link-LSAs only specify the cost towards the
datagram source. In the presence of asymmetric link costs, this
may lead to less efficient routes when forwarding multicasts
o N4
/
2/ 3
/
RT3 (Ma,Mb) o o RT4 (Ma,Mb)
/
6/ 8
/
RT6 o o RT5

5 6

RT10 (Ma) o o RT7 (Ma)

2

RT11 (Ma) o
Figure 9: Datagram shortest-path tree: Backbone,
source N4, destination Group A. Note that
reverse costs (i.e., toward origin) are
used throughout.
between areas.
Those routers attached to multiple areas must calculate multiple
trees and then merge them into a single forwarding cache entry.
As shown in Section 2.3.2, when connected to a single area the
router"s position on the datagram shortest-path tree determines
(in large part) its forwarding cache entry. When attached to
multiple areas, and hence calculating multiple datagram
shortest-path trees, each tree contributes to the forwarding
cache entry"s list of downstream interfaces/neighbors. However,
only one of the areas" datagram shortest-path trees will
determine the forwarding cache entry"s upstream node. When one
of the attached areas contains the datagram source, that area
will determine the upstream node. Otherwise, the tiebreaking
rules of Section 12.2.7 are invoked.
Consider again the example of Host H2 in Figure 4 sending a
multicast datagram to destination Group A. Router RT3 will
calculate two datagram shortest-path trees, one for Area 1 and
one for the backbone. Since the source of the datagram (Host
H2) belongs to Area 1, the Area 1 datagram shortest-path tree
determines RT3"s upstream node: Network N4. Router RT3
calculates two downstream interfaces for the datagram: the
interface to Network N3 (which comes from Area 1"s datagram
shortest-path tree) and the serial line to Router RT6 (which
comes from the backbone"s datagram shortest-path tree). As for
Router RT10, it calculates two trees, determining its upstream
node from the backbone tree and its two downstream interfaces
from the Area 2 tree. Finally, Router RT11 calculates three
trees, determining its upstream node from the Area 2 tree and
its downstream interface from the Area 3 tree.
4. Inter-AS multicasting
This section explains how MOSPF deals with the forwarding of
multicast datagrams between Autonomous Systems. Certain AS boundary
routers in a MOSPF system will be configured as inter-AS multicast
forwarders. It is assumed that these routers will also be running an
inter-AS multicast routing protocol. This specification does not
dictate the operation of such an inter-AS multicast routing
protocol. However, the following interactions between MOSPF and the
inter-AS routing protocol are assumed:
(1) MOSPF guarantees that the inter-AS multicast forwarders will
receive all multicast datagrams; but it is up to each router so
designated to determine whether the datagram should be forwarded
to other Autonomous Systems. This determination will probably be
made via the inter-AS routing protocol.
(2) MOSPF assumes that the inter-AS routing protocol is forwarding
multicast datagrams in an RPF (reverse path forwarding; see
[Deering] for an explanation of this terminology) fashion. In
other words, it is assumed that a multicast datagram whose
source (call it X) lies outside the MOSPF domain will enter the
MOSPF domain at those points that are advertising (into OSPF)
the best routes back to X. MOSPF calculates the path of the
datagram through the MOSPF domain based on this assumption.
MOSPF designates an inter-AS multicast forwarder as a wild-card
multicast receiver in all of its attached areas. As in the inter-
area case, this ensures that the routers remain on all pruned
shortest-path trees and thereby receive all multicast datagrams,
regardless of destination.
As an example, suppose that in Figure 1 both RT5 and RT7 were
configured as inter-AS multicast forwarders. Then the link state
database would look like the one pictured in Figure 2, with the
addition of a) wild-card status for RT5 and RT7 (they would appear
with superscripts of "w") and b) the external links originated by
RT5 and RT7 being labelled as multicast-capable[12].
As another example, consider the area configuration in Figure 4.
Again suppose RT5 and RT7 are configured as inter-AS multicast
forwarders. Then in Area 1"s link state database (Figure 6), the
external links originated by RT5 and RT7 would again be labelled as
multicast-capable. However, note that in Area 1"s database RT5 and
RT7 are not labelled as wild-card multicast receivers. This is
unnecessary; since Area 1"s inter-area multicast forwarders (RT3 and
RT4) are wild-cards, all multicast datagrams will be forwarded to
the backbone. And in the backbone"s link state database (Figure 7),
RT5 and RT7 will be labelled as wild-cards.
4.1. Building inter-AS datagram shortest-path trees.
When multicast datagrams are to be forwarded between Autonomous
Systems, the datagram shortest-path tree is built as follows.
Remember that the router builds a separate tree for each area to
which it is attached; these trees are then merged into a single
forwarding cache entry. Suppose that the router is building the
tree for Area A. We break up the tree building into three cases.
This first two cases have already been described earlier in this
memo: Case 1 (the source of the datagram belongs to Area A)
having been described in Section 2.3.2 and Case 2 (the source of
the datagram belongs to another OSPF area) having been described
in Section 3.2. The only modification to these cases is that
inter-AS multicast forwarders, as well as group members and
inter-area multicast forwarders, must remain on the pruned
trees. The new case is as follows:
o Case 3: The source of the datagram belongs to another
Autonomous System. The immediate neighborhood of the source
is then unknown. In this case the multicast-capable AS
external links are used to approximate the neighborhood of
the source; the tree begins with links directly attaching
the source to one or more inter-AS multicast forwarders. The
approximating AS external links point in the reverse
direction (i.e., towards the source), just as with the
approximating summary links in Case 2. Also, as in Case 2,
all links included in the tree must point in the reverse
direction. The final datagram shortest-path tree is then
produced (as always) by pruning those branches having no
group members nor wild-card multicast receivers.
As an example, suppose that a host on Network N12 (see
Figure 4) originates a multicast datagram for Destination
Group B. Assume that all external costs pictured are OSPF
external type 1 metrics. Then any routers in Area 1
receiving the datagram would build the datagram shortest-
path tree pictured in Figure 10. Note that all links in the
tree point in the reverse direction, towards the source. The
tree indicates that the routers expect the datagram to enter
the Autonomous System at Router RT7, and then to enter the
area at Router RT4.
Note that in those cases where the "best" inter-AS multicast
forwarder is not directly attached to the area, the
neighborhood of the source is actually approximated by the
concatenation of a summary link and a multicast-capable AS
external link. This is in fact the case in Figure 10.
In Case 3 (datagram source in another AS) the requirement that
all tree links point in the reverse direction (towards the
source) accommodates the fact that summary links and AS external
links already point in the reverse direction. This also leads to
the requirement that the inter-AS multicast routing protocol
operate in a reverse path forwarding fashion (see condition 2 of
Section 4). Note that Reverse path forwarding can lead to sub-
optimal routing when costs are configured asymmetrically. And it