A Third-Party Value-Added Network Service Approach to Reliable Multicast
Kunwadee Sripanidkulchai Andy Myers Hui Zhang Carnegie Mellon University Sigmetrics '99
[email protected]
1
Outline • Previous work in Reliable multicast • Our approach: Waypoint recovery service – Third-party value-added network services – Example implementation using STORM
• Performance evaluation • Conclusions • Future work
Sigmetrics '99
[email protected]
2
IP Multicast
X
X
X IP Multicast is best-effort. Sigmetrics '99
[email protected]
3
Source-Based Reliable Multicast
X
X
X
Implosion at the source Sigmetrics '99
[email protected]
4
End Host-Based Reliable Multicast
X
X
X • Limited by end hosts in the group • SRM, TMTP, RMP, LORAX, STORM Sigmetrics '99
[email protected]
5
End Host-Based With Router Support
X
X
X • Scalability and deployment issues • [Levine97], PGM, LMS, OTERS, [Lehman98] Sigmetrics '99
[email protected]
6
Server-Based Reliable Multicast
X
X
X • Static binding of servers to multicast group members • RMTP, LBRM, [Kasera98] Sigmetrics '99
[email protected]
7
Our Approach
•
• •
• •
Sigmetrics '99
No prior knowledge of locations of senders and receivers in multicast group Receivers invoke service Some waypoints will join the group and participate in error recovery Waypoints form a separate recovery structure Waypoints are additional recovery resources
[email protected]
8
Third-Party Value-Added Network Services •
•
•
Architecture – Third-party value-added service provider places servers around the network – Incremental service deployment – Change not required at the router level Service Definition – Dynamic service invocation – Dynamic binding of server and service – Distributed control Waypoint Recovery Service – Minimal change to existing reliable multicast protocols – Support multiple multicast groups and multiple reliable multicast protocols
Sigmetrics '99
[email protected]
9
STORM: Structure-Oriented Resilient Multicast [Xu, Myers, Zhang, Yavatkar 1997]
• Resilient multicast – continuous media applications – receivers do not need all packets – packets must arrive by playback time
• Application level recovery structure
Sigmetrics '99
[email protected]
10
Waypoint Recovery Service for STORM
Without waypoints
• • • •
With waypoints
Additional recovery resources Participate in error recovery Form a separate recovery structure To receivers, waypoints appear to be the same as normal receivers
Sigmetrics '99
[email protected]
11
Waypoint Protocol • • • •
Waypoint invocation algorithm Join algorithm Leave algorithm Considerations – Limit join implosion – Minimize useless joins – Minimize oscillations
Sigmetrics '99
[email protected]
12
Waypoint Invocation Heuristic When a receiver’s loss rate is high, it does an expanding ring search to locate waypoints
Sigmetrics '99
[email protected]
13
Waypoint Join and Leave Algorithms Heuristics • Join algorithm – If the waypoint has joined the group before, it will join again only if it has been useful. – If the waypoint has never joined the group before, it will always join.
• Leave algorithm – If the waypoint does not improve loss rates, it will leave the group.
Sigmetrics '99
[email protected]
14
Outline • Previous work: Reliable multicast • Our approach: Waypoint recovery service – Third-party value-added network services – Example implementation using STORM
• Performance evaluation • Conclusions • Future work
Sigmetrics '99
[email protected]
15
Simulation Environment • Packet-level event-based simulator • All packets traversing the same link are subjected to the same network delay and loss characteristics • Dynamic cross traffic was not modelled for simulation efficiency • Network links are characterized by – loss rate li – delay [di , αdi ]
Sigmetrics '99
[email protected]
16
Simulation Methodology • • • • •
Backbone topology emulating the Internet based on 3 Internet Service Providers’ backbones (MCI, Sprint, and BBN) Waypoints are connected to backbone routers One constant bit rate ‘audio’ source Receiver distribution Bottleneck Scenarios – – – –
Backbone Backbone hotspots Network access point Local domain
Sigmetrics '99
Backbone
Network access point
Local domain
[email protected]
17
Performance Indices • Loss rates observed by receivers – without any recovery – with STORM – with STORM and waypoints
• Change in recovery protocol processing load • Additional network overhead incurred by waypoints
Sigmetrics '99
[email protected]
18
Simulation Experiment
Network access point bottleneck 23 receivers “Sparse receivers” Waypoints are attached to all backbone routers
Sigmetrics '99
[email protected]
19
Loss rates after recovery (%)
Waypoints Helped: Network Access Point Bottleneck 15
STORM
10
STORM with waypoints 5
0 0 Sigmetrics '99
5
10
Loss rates without recovery (%)
[email protected]
15 20
Backbone Bottleneck
Loss rates after recovery (%)
30 25 20 STORM 15
STORM with waypoints
10 5 0 0
Sigmetrics '99
10
20
Loss rates without recovery (%)
[email protected]
30 21
Maximum Help Across Simulations
Loss rates after recovery (%)
30 25 20 STORM 15
STORM with waypoints
10 5 0 0
Sigmetrics '99
10
20
Loss rates without recovery (%)
[email protected]
30 22
Loss Rates Across All Simulations
Loss rates after recovery (%)
30 25 20
STORM
15
STORM with waypoints
10 5 0 0
5
10
15
20
25
30
Loss rates without recovery (%) Sigmetrics '99
[email protected]
23
Network Overhead from Waypoints: Number of Additional Links
Number of additional links introduced by waypoints Total number of links
Sigmetrics '99
[email protected]
=
(2%,35%)
24
Results Summary: When do waypoints help? • Waypoints improve receivers’ loss rates when – Isolated, sparse or skewed receivers – Parents and children share losses – Congestion between parents and children
• Waypoints reduce receivers’ protocol processing load • Waypoints do not significantly increase the source’s processing load. • Waypoints incur moderate network overhead
Sigmetrics '99
[email protected]
25
Conclusions • Proposed a general architecture for introducing services into the network without requiring changes to routers or existing infrastructure • Waypoints enhance error recovery in reliable multicast – More parent choices – More advantageous positions
Sigmetrics '99
[email protected]
26
Future Work • Evaluation – – – –
Effects of dynamic receivers Effects of multiple groups More sophisticated loss/delay models Implementation of a prototype service for the MBONE
• Algorithm – Improve waypoint invocation algorithm – Improve waypoint join and leave algorithms
• Architecture – Extension to other reliable multicast protocols – Extension to other services Sigmetrics '99
[email protected]
27