1. On Reducing Mesh Delay for Peer- to-Peer Live Streaming Dongni Ren, Y.-T. Hillman Li, S.-H. Gary Chan Department of Computer Science and Engineering The Hong Kong University of Science and Technology INFOCOM 2008 Junction

2. Design goals overlay –Low delay –Robust to user churn –Accommodation of asymmetric bandwidths –Distributed, simple and adaptive

3. Tree & Mesh Tree –Achieve low delay –Cannot accommodate well network dynamics and asymmetric bandwidth Mesh –Robust to user churn –Asymmetric bandwidth is overcomed by aggregating the bandwidth of multiple parents to guarantee a certain streaming rate.

4. Delay in Mesh Mesh delay –Due to the longest path from the node to the source out of all its parents. –The number indicated in the square boxes Packet scheduling delay –Due to packet transmission time and scheduling policy of a peer with its parents of heterogeneous bandwidth.

5. Optimization of mesh delay Problem formulation and a centralized heuristic –To form a mesh which minimizes the maximum delay of the peers in the network while meeting a certain streaming rate requirement. –Centralized heuristic as a benchmark A distributed protocol for low-delay mesh –Power concept –Adaptation mechanism Performance study on the algorithms –Simulation –Compare with traditional and state-of-the-art approaches Outreach Closest-parents

6. Problem Formulation Minimum Delay Mesh Problem (MDM problem) : –To find a mesh which minimizes the maximum of the peer delay MDM problem is NP-Hard –Traveling Salesman Problem (TSP) can be reduced to MDM problem

7. A Centralized Heuristic Shallow streaming mesh –Peers are close to the source with low hop count –Put the nodes with high uplink bandwidth close to the source to increase the fanout of the mesh towards the uplink. Power –Achieve a balance between the delay and uplink bandwidth –Power is defined as the throughput divided by delay Power between a peer i and its parent j The rate that node j is serving node i The delay of i via parent j

8. Algorithm Rank all the nodes according to their uplink capacities divided by their delay to the source Push them into the mesh in descending order node i is pushed into the mesh –calculate the power P i (j) for all the nodes already in the mesh –connect node i to node j with the largest P i (j) value If node i is not fully served by node j, connect node i to one more parent with the second largest P i (j) value …

9. Power-Based Distributed Algorithm Rendezvous Point –Caches a list of recently arrived peers –Returns a few of them to the newcomer (potential parents) Same as centralized heuristic If the peers returned by the Rendezvous Point cannot fully serve the newcomer, the newcomer request the neighbor of those peers.

10. Adaptation With high probability, there are some low-bandwidth peers occupying the areas in between the source and the powerful ones. –Request Step –Grant Step –Accept Step child parent REQUEST : child’s uplink BW time-to-live (TTL) Its residual BW > streaming rate TTL > 0, decrease and forward to its parent Its uplink BW > sender’s, GRANT GRANT Among these the ancestor with shortest distance from the source is picked.

11. Simulation Results Simulation setup and metrics –Brite generate 10 two levels top-down hierarchical topology –8 autonomous systems each of which has 624 routers –Bandwidth distribution

12. Evaluation Metrics Delay –The time taken for data to travel from the streaming server to the peers –Average : source to end among all peers –Maximum : source to end delay of the mesh Hop Count –The number of intermediate peers involved on the overlay path form the source to a peer. Source Workload –The amount of bandwidth consumed at the source

13. Simulation Results Adaptation : –Average delay reduces –Variance narrows down

14. Simulation Results Delay –Power scheme outperforms the other two as the number of peers grows

15. Simulation Results Hop Count –The power scheme gives a more compact mesh than Outreach –High BW peers in Power scheme are aggressively promoted upwards and thus more branches occur near the source.

16. Simulation Results Source Workload –Outreach actively places peers under source -> rely on source –Power and closest parent scheme, the source contribution roughly the same

17. Simulation Results TTL –Number of Adaptation Change : the number existing connections that are broken in the adaptation phase before the mesh reaches a static state –The cost of adaptation proportional to the number of adaptation happened

18. Simulation Results Delay Reduction : –Average : ratio of average delay reduced by adaptation to average delay without adaptation –Maximum : ratio of maximum delay reduced by adaptation to maximum delay without adaptation –Having large TTL value only gives slightly better benefit. –Risk of flooding the overlay

19. Contribution & Conclusion The first body of work addressing the optimization of mesh delay for P2P streaming Not mention much about how the algorithm tolerate the churn ( peer join or leave )