A Case for End System Multicast Yang-hua Chu, Sanjay G. Rao, Srinivasan Seshan and Hui Zhang Presentation by Warren Cheung Some Slides from ar/2000/%EC%9D%B4%EB%AF%BC%ED% 98%B8/5
Multicast Unicast: – Point to Point delivery – One Host to One Client Multicast: – Deliver to multiple destinations – One Host to Many Clients
IP vs. End System Multicast IP multicast – Implemented at Internetworking layer Routers and switches End System multicast – Implemented in the hosts and clients
Examples
Examples – IP Multicast
Examples – naive Unicast
Examples – End System Multicast
Considerations Transmission Redundancy of Data – Unicast: many copies per link – IP Multicast: one copy per link – End System Multicast: slightly more inefficient than IP Multicast Delay – Unicast same as IP Multicast – End System Multicast incurs penalty
IP multicast Pros – Possibly large performance benefits Cons – Needs to maintain “group state” – Infrastructure level changes are slow to deploy
End System Multicast Pros – Can be implemented now Hosts (Peer-to-Peer) Proxy Cons – Performance degradation
Narada Protocol Self-Organising – Constructs Overlay – Adapt to Network/Group Dynamics Efficient – Latency vs. Bandwidth – Self-improving
Group Management Everyone keeps the member list – Target Medium-Sized Groups – Everyone periodically exchanges group information with neighbours (refresh) Join – Bootstrapping Leave – Partition repair
Mesh Performance Mesh may be suboptimal due to: – Network conditions – Group dynamics Adding random neighbours Dropping low “cost” links
Open Issues Group size on Average overlay hops Short-term Effects of events on Performance Overlay construction/maintenance costs when group sizes get very large
Related Work on Overlays Mesh-based Tree-based overlays Delaunay Triangulations – Map addresses to coordinate space – Find closest neighbours Hierarchies of Clusters
Related Works End System Multicast, Narada, Video Streaming – Comparison of some Application Layer Multicast solutions – mparative.html
Not-So-Closely Related Works BitTorrent (File Swarming) – Herbivore (Anonymity/Security) – ore.html
Discussion IP Multicast – Basically a failure – deployment issues – Any fundamental/low level changes to Internet infrastructure unlikely to succeed End System Multicast – Overhead/performance impact no longer looks as disadvantageous as it originally appeared
Discussion(2) Applications – Limitations to video-conferencing more due to limitation on the number of people you can communicate with simultaneously – For large number of clients, more likely to be a broadcast Narada – Implemented and used – Broadcasts the annual SIG networking conference