Transcript Host Multicast: A Framework for Delivering Multicast to

Host Multicast: A Framework for
Delivering Multicast to End Users
Beichuan Zhang (UCLA)
Sugih Jamin (UMich)
Lixia Zhang (UCLA)
1
June, 2002
INFOCOM
Motivation


Fast increasing need for scalable and efficient
group communication
Slow deployment of IP Multicast
–
–

Emerging End-host based Multicast
–
–
2
Implemented in most OS and routers
But not widely enabled
Member hosts duplicate and forward packets
Easy to deploy, but less efficient
Host Multicast



A hybrid approach
Goal: Ubiquitous Multicast
Design Requirements
–
Deployable on the current Internet



–
Compatible with IP Multicast to the furthest extent



3
Install a user-space program at end hosts
No support is required from OS, routers and servers
Enable multicast applications
Use IP Multicast where available
Keep IP Multicast service model
Provide incentive to future deploy
Architecture
Rendezvous Point (HMRP)
RP
Designated Member (DM)
DM
Host
Host
Host
DM
Host
Unicast Tunnel
DM
Host
Host
Normal Member
IP Multicast Island
4
Host
Group
Management
Protocol
(HGMP)
for
intra-island
A network
Each
To
bootstrap
member
Multicast
of
new
any
runs
Tree
size
members
our
Protocol
that
daemon
supports
(HMTP)
program
IPtoMulticast,
atbuild
user-space
inter-island
e.g. single
tunnels
host,
management
Ethernet, campus network etc.
Components
IP Multicast
Host Multicast
Host Extension
OS kernel support
A user-space daemon
Local Membership
Management
IGMP
HGMP
Intra domain/island DVMRP, MOSPF,
multicast routing
PIM, CBT
Using deployed IP
Multicast
Inter domain/island MASC/BGMP,
multicast routing
MBGP/MSDP/PIM
HMTP
Rest of the talk is focused on HMTP
5
Host Multicast Tree Protocol (HMTP)


Build a bi-directional shared-tree connecting all
islands
The tree should be congruent to physical
network topology to be efficient
–

The tree should be robust
–
6
use member-to-member round-trip time as distance
metric in current design
be able to handle node failure, dynamic join/leave
etc.
Join Group
root
HMRP always knows
the root of the tree.
B
A newcomer does a
depth-first search of
the tree to find a close
member as its parent.
A
C
D
F
E
7
Clustering nearby
members makes the
tree congruent to
physical network
topology to the first
order.
H
G
Where is my group?
Root of your group is A
RP
Tree Maintenance


8
Each member keeps
its children list and
root path up to date
by exchanging
REFRESH and
PATH messages
with neighbors.
Root sends
REFRESH message
to HMRP.
B
A
C
D
F
E
H
G
RP
Member Leave and Partition
Recovery


Parent deletes the
leaving node from
children list.
Direct Children repair
the tree by running
join procedure in
the reverse order.

9
If root is leaving, the
first node contacting
HMRP is assigned as
new root.
B
A
C
D
F
E
H
G
RP
Tree Improvement

Periodically re-run the join procedure
–
–
–
10
To accommodate changes in network conditions
and group membership
Start from a randomly picked node in the root path.
Less frequent than REFRESH and PATH messages.
Loop Detection and Resolution

Loop is possible:
–

11
A
C
D
One’s root path contains itself
Resolution:
–

B
Detection:
–

Multiple conflicting joins
happen at the same time.
Leave the current parent and
re-join the tree from the root.
Loop is rare.
E
F
G
Performance Metrics

Tree Cost: sum of all tree link delays
–

Tree Delay: delay from one member to another along
the tree
–

12
Delay ratio to unicast delay
Link Load
–

Cost ratio to IP Multicast
Number of duplicate packets carried by a physical link.
Compared to Unicast Star and IP Multicast
Convergence
1.6
stabilized
Cost Ratio
1.5
leave
done
stabilized
1.4
done
1.3
1.2
1.1
join
1
0
13
60
120
180
240
300
Time (second)
360
420
480
Tree Cost
4
Unicast Star
Cost Ratio
3.5
Host Multicast
3
2.5
2
1.5
1
0.5
0
14
100
200
300
Group Size
400
500
Tree Delay
9
Host Multicast Mean
8
IP Multicast Mean
Delay Ratio
7
6
5
4
3
2
1
0
15
100
200
300
Group Size
400
500
Link Load
Number of Physical Links
1000
Unicast Star
Host Multicast
100
10
1
1
16
10
Physical Link Load
100
Internet Experiment

A real topology of 96 hosts and 978 routers,
derived from NLANR’s traceroute data.
–
–
–
17
Tree cost ratio
Mean delay ratio
Worst link load8
0.99
1.76
Conclusion

Related Work of HMTP
–
Centralized

–
Mesh-based

–
Yoid, BTP etc.
Future Work
–
–
18
Narada, Gossamar etc.
Tree-based


ALMI etc.
Reducing HMTP tree delay
Supporting Multiple DMs in one island
~ The End ~
19