Multicast Routing

Papers: DVMRP and MOSPF

Computer Networks
Dr. Jorge A. Cobb
 Bridges and Extended LANs
 LANs have physical limitations (e.g., 2500m)
 Connect two or more LANs with a bridge
• accept and forward strategy
• transparent (does not add packet header)

A B C

Port 1
Bridge
Port 3
Port 2

X Y Z
H I J

2
Learning Bridges
 Do not forward when unnecessary
 Maintain forwarding table
Host Port
A B C A 1
B 1
Port 1 C 1
X 2
Bridge
Port 3 Y 2
Port 2 H 3

X Y Z
H I J

 Learn table entries based on source address

 Table is an optimization; need not be complete
 Always forward broadcast frames

3
Learning Bridges
Host Port
A B C A 1
B 1
Port 1 C 1
Bridge
Port 3 X 2
Port 2 Y 2
H 3
X Y Z
H I J

 I sends a message to Z
 Bridge “learns” I comes from port 3, sends the message over
BOTH port 1 and port 2
 Z sends a message to I (e.g. some reply)
 Bridge “learns” Z comes from port 2, sends the message over port
3 (it learned above where I is)

4
 Spanning Tree Algorithm

 Problem: loops
A
A
B
B B3
B3
C B5
C B5
D B7
D B7 K
B2 K
E F
E F

B1
B1
G H
G H

B6 B4
B4
I
I
J
J

 Bridges run a distributed spanning tree algorithm

• Logically delete certain bridges to have a spanning tree
• We will assume we have a spanning tree
5
 Multicast Routing in Datagram
Internetworks and Extended LANs

S. Deering and D. Cheriton

Computer Networks
Dr. Jorge A. Cobb
ACM Transactions on Computer Systems
1990
Efficient multicast in Extended LANs
 How to do efficient multicast in Extended LANs?
 Extended LANs are formed by connecting LANs with link layer
bridges
 Bridges propagate broadcast (and multicast) packets across
every segment of the extended LAN
• Way too inefficient, especially for multicast applications with
sparsely located receivers
• Find an efficient way of doing multicast and convert
applications to use multicast

How to do multicast better in Extended LANs?

7
Single Spanning-Tree Multicast Routing
 If bridges know which interface led to members of a given group, they
would forward the packets on those interfaces only
 But, in general, how do bridges learn which interface lead to individual
hosts?
• When a packet arrives from a host, bridge records the (source-
host addr, interface, age) into a table
• If an entry is too old, it is removed (self-cleaning table)
 Can we do something similar for multicast packets?
• We want to forward multicast packets only over the needed LANs
• Problem: bridge does not learn about a host (or group member)
unless it receives a packet from the host (receivers don’t send
data)
• Thus, group members periodically transmit a membership-report

8
 Multicast Table
Arrow indicates
where group
members are
located

 Multicast table of each bridge has the following row for each multicast group:
(muticast addr, (outgoing interface, age), (outgoing interface, age), … )
 A group member of a group G sends a membership-report, the destination address is
the “ALL-BRIDGES” multicast address, the source address is G.
 Bridge receiving this message records this interface as an outgoing interface for
group G
 It then forwards this report to other bridges (out all of its interfaces) in the extended
LAN

9
 Single Spanning-Tree Multicast Routing
 Bridge algorithm
• If source address is a mcast group address (i.e. a report),
• record arriving interface as an outgoing-interface with an age of zero
into the table entry for this mcast address
• and forward to all other interfaces.

• Periodically increment age of outgoing interface

• when age=expiry threshold, delete this outgoing-interface info from
the table’s entry
• if no outgoing-interfaces remain, delete the entire entry

• If a packet arrives with multicast dest addr

• forward a copy on every outgoing-interface recorded in the table entry
(if any) excluding the arriving interface

10
Another example
A

B
B3

C B5

G
D B7
K

E F
•Arrows indicate the
location of the group
B1
member
G
•Membership reports are
H
forwarded over all links of
a bridge
B4

I
G
J

G = group G member (receiver)

11
Suppressing Membership Reports
 An efficiency improvement to suppress unnecessary
membership reports
• Hosts send membership-reports as (G,G)

• This suppresses #membership reports to 1 per report

interval
• because all other group members (hosts) on the same LAN
heard the first membership-report (G,G), and will not send a
report

• Bridge, on receiving a pkt with address (G,G) (bridges receive

ALL packets, remember?)

• changes it to (G, all-bridges) and forwards to other interfaces

12
 Why change the destination?
 Why does the router change from (G,G) to (G, “ALL-
BRIDGES”)?  Assume member G1 sends a
membership report (G,G) over
lan A
 Assume the bridge forwards it
A B as (G,G) over lan B
 This will suppress the
membership report of G2
G1  The bridge will never know
there is a group member in lan
bridge B

G2

13
 Done with extended LANS (bridges)

 Now we do multicast over multiple LANs (across

routers)

14
Distance Vector Multicast Routing
 How to support multicast routing in a distance-vector
environment? (this is general enough for ANY unicast routing
protocol)
• Compute a spanning tree across all the links and use a
multicast routing algorithm similar to the one above
• Specifically, a source-based shortest path spanning trees
• Tree is rooted at the source site
• It corresponds to shortest path from each receiver to the
source
• Main assumption is path symmetry
(links have the same costs in both directions)
• Observation:
Every shortest-path multicast tree rooted at the sender is a
subtree of a single shortest-path spanning (i.e. broadcast) tree
rooted at this sender
15
 Broadcast Trees and Multicast Trees in Point-to-Point
Networks

Network Broadcast Tree Multicast Tree

S S
S

S = source (host) node, also root of tree

R = receiver
For any router U, parent(U) = next-hop(U,S) = V

16
 Reverse Path Flooding (RPF) (broadcast in point-
to-point networks)

 Routers forward a broadcast packet originated at S on all of their

outgoing interfaces (except the incoming one) if it arrives via the
shortest path from the router back to S

• Note: each router forwards each packet only once.

• Router needs to know the shortest path back to S

• Given by the distance vector routing algorithm

• Forward to all neighbors

17
Example

Broadcast tree S
Broadcast messages
v

Which of all three copies will be forwarded by t?

18
Internetworks
 Original RPF is designed considering point-to-point links
 In the Internet we typically have LANs (e.g. Ethernet)
• There are many routers per LAN
 A multicast packet sent by a source host S will have
• LAN source: LAN address of S
• LAN destination: LAN multicast address for group G
• IP source: IP address of S
• IP destination: IP multicast address for group G
 Routers do not modify the IP src or dst addresses
 Routers send the packet over the LAN with
• LAN source: LAN address of router
• LAN destination: LAN multicast address for group G
 All routers “listen” (receive a copy) of all LAN packets
addressed to any multicast group.

19
 A problem with original RPF
 When shared links (Ethernet) used between routers, two problems

 Identifying the parent on the tree

• In RPF, routers only accept the broadcast packet from the next router on the
shortest path back to S (its parent on the tree)

• How does this router learn which packet is coming from its parent? (many
routers per LAN)
• Looks at the link layer (Ethernet) source address in the incoming
packets, it must be from the next-hop router.

 Multiple copies of the packet is sent on the shared link

• Multiple routers are attached to the same LAN (Ethernet)

• Waste of bandwidth on the link & waste of router resources

20
Reverse Path Broadcasting (RPB)
 Eliminates duplicate broadcasts on shared links in RPF

 Identify a single “parent” router for each link w.r.t. S

• For the link S is attached to, S is considered the parent

• Otherwise, router with min distance to S is the parent
• In case of tie, router w/ lowest IP address is the parent

21
 RPF:

Both x and y inject

packets into a

z only forwards the copy it

receives from x, because x
is its next hop to S

RPB:

x, y and z learn that only x

should inject packets into
a

22
 S

X Y

Z L

Path of broadcast
messages

23
Observations
 How to identify min distance router to S?
• Routers exchange distance vector records with each other
• Therefore, each router independently finds the parent
(break ties by ID)

 Overhead:
• Parent selection requires that routers add a children bitmap
to each routing table entry (i.e. for each possible source)
• the bit-map has one bit for each incident link
• Bit for link λ for destination S is set, if λ is a child link of this
router for broadcasts originating at S

24
 Prunning the Tree
 Both RPF and RPB are broadcast algorithms !

 To provide shortest path multicast delivery from source S to group

members, broadcast tree of S must be pruned back to reach only
links with receivers

 May be accomplished by requiring members of G to send

membership reports back up the broadcast tree of S periodically
• How do you identify the sources?
• Too costly – need to be done for each group and each source
• What if only a few groups/sources are active at a time?
• We want to do this only for active groups/sources

25
Truncated Reverse-Path Broadcast (TRPB)
 An alternative in which only non-member leaf LANs
are deleted from each broadcast tree

• Truncate child link if

• no router uses this link to receive multicast messages
from the source (i.e., it is a leaf link)
• No host is a group member on this link (LAN)

26
 Leaf Truncation example

X Y

Z L

Path of broadcast
messages

 At L, L truncates the lower LAN in the figure if

• if Z does not accept multicast messages from this
LAN
• No host is a group member on this link (LAN)

27
 S

Receiver

Receiver

X Y

Z L
X cannot cut this
child LAN since Z
uses X as its next L truncates this LAN since Z
hop (not leaf link) does not use L to reach S (it
is a leaf link) and there are
no receivers.
28
Algorithm

 If a multicast packet (S, G) arrives from the next-hop-link for S,

• forward a copy of the packet on all child links for S,
• except leaf links (no other router receives from this link) that
have no members of G

 To implement this, we need two things:

• The router needs to learn if the child link is a leaf link
• I.e. If no other router uses this link to receive messages from
the group.
• The router also needs to know if no host group members
are on this link.

29
Implementation
 Leaf pruning requires that routers add a leaf bitmap to
each routing table entry (i.e. for each possible source)
• the bit-map has one bit for each incident link
• bit for link λ for destination S is set, if λ is a leaf link of this
router for multicasts originating from S

 How to know if the link is a leaf link?

30
Implementation (continued…)
 Distance vectors tell me the distance from routers on this
link to S, but not if I am their next hop.

 However, if DV with split-horizon and poisoned-reverse is

used then
• If a router gives me a distance of infinity then it uses
me as the next hop
• I.e., the link is not a leaf link

31
Implementation (contd)…
 Hosts send a membership report message over their
LAN using the mcast group address as the
destination (all hosts group members listen to this,
and only one report is sent per interval)

 Routers maintains a table with one entry per link

(LAN)
• The entry contains a bit-map field, link-groups, with one
bit per group
• It identifies for each group if there are members on this
link.

32
Overview of “bitmaps” required
 For each source S and each link λ,
• Is λ a child link of the router on the broadcast tree of
S?
• Is λ a leaf link of the router on the broadcast tree of S?
• Both of these are obtained via the DV routing
algorithm

 For each link λ and group G,

• Does λ have any members of G?
• Obtained via membership reports (IGMP)

33
Reverse Path Multicasting (RPM)
 Prune shortest path multicast tree as follows

• First packet for (S,G) is forwarded everyone on the truncated

shortest path broadcast tree acc. to TRPB

• Leaf routers w/ no attached members send nonmembership report

(NMR) to their parent on the tree

• If a router receives NMR from all of its children routers and itself
has no directly attached members, it also sends NMR to its parent
router on the tree.

34
 S

NMR

R NMR NMR NMR

X Y

NMR Z L

35
 S

X Y
Cannot be
truncated because
there is a receiver
here.
Z L

36
NMR expiration

 NMR reports include an age field, when it expires data

flows all the way to leaves again and gets re-pruned back

 Routers remember NMR reports that they sent

• When a new host joins G, they send a cancellation
message to undo the effect of NMR

37
Pros and Cons
 Reverse path multicasting, when used with distance vector
routing, is known as distance-vector multicast routing protocol
(DVMRP)
 Advantages: good when there are many receivers, since
multicast messages are initially flooded to the entire
network.
 Disadvantages
• Bad if there are few receivers
• Again, the first multicast messages are sent throughout the
network unnecessarily.
• Routers need to remember the “prune” state, i.e. they need to
maintain state even when there are no receivers below them
on the tree
• The path from source to receiver may not be optimal if the
cost of links is not bi-directional.
38
Link-State Multicast Routing
 This is covered in the book, section 4.4.1

 Extend OSPF – MOSPF

• Send group membership information in OSPF link state
advertisement (LSA) messages

• I.e., each router learns the entire set of receivers.

• Each router can compute the minimum cost path from every
source to the current set of receivers of the multicast group.

39
 R(G)
A B
R(G)

S(G) C

Routers A and B mention in their LSA that they have receivers in their
adjacent LANs.
Hence, C can recreate the above picture in detail.
40
 R(G)
A B
R(G)

S(G) C
∞ ∞

C computes the shortest path tree from S(G) to the receivers

Each link is tagged with its distance to the closest R(G)

41
How do you know who are the sources?
 You could precompute, for every group G, a tree for every
source S
 This is way too expensive
 Instead, use caching

42
MOSPF Cache
 The cache has entries of the following form:
• (S, G, iif, min-hops)
• Min-hops is a VECTOR with an entry per output link
• This vector contains the number of hops needed to reach a
group member via the link
• If a link does not reach a group member (not on tree) use
infinity
• If a packet is received from S to G,
• The packet is sent over all links such that the time-to-live of the
packet is at most the link’s entry in min-hops
 If no cache-entry of (S,G) compute the tree on the fly
 Cache entries are not timed-out, they are just flushed out if new
ones are needed.

43