1. On the Optimal Scheduling for Media Streaming in Data-driven Overlay Networks Meng ZHANG with Yongqiang XIONG, Qian ZHANG, Shiqiang YANG Globecom 2006

2. Outline Background Related Work Problem Statement and Formulation Global Optimal Solution Distributed Algorithm Performance Evaluation Conclusion & Future Work

3. Background The Internet has witnessed a rapid growth in deployment of data-driven (swarming based) overlay/peer-to-peer network based IPTV systems during recent years. These products are based on data-driven protocol Facts of concurrent online users GridMedia: over 230,000, rate 310kbps (achieved by one server) (developed by our lab) PPLive: 500,000, rate 300-500kbps QQLive: 1,460,000, rate 300-500kbps (not one server)

4. Background - Data-Driven Protocol Review Aiming to enable large-scale live broadcasting in the Internet environment Very simple and very similar to that of Bit- Torrent Two steps in data-driven protocol The overlay construction The block scheduling

5. Background - Data-Driven Protocol Review The first step – overlay construction All the nodes self-organize into a random graph I have block 1,2,4 I have block 1,2,3 I have block 1,2I have block 2,3 Request block 4 Request block 3 Request block 1Request block 2 Send block 4 Send block 3 Send block 1Send block 2 The second step – block scheduling The streaming is divided into blocks Each node has a sliding window containing all the blocks it is interested in currently

6. Related Work To improve data-driven protocol, most recent efforts focus on optimizing overlay construction (i.e. the first step ): Vishnumurthy & Francis (INFOCOM2006): random graph building under heterogeneous overlay Liang & Nahrstedt (INFOCOM2006): propose RandPeer, a peer-to-peer QoS-sensitive membership management protocol

7. Related Work An problem not well addressed is how to optimize the second step, that is, how to do optimal block scheduling and maximize the throughput of data-driven protocol under a constructed overlay Most existent methods are straight forward and ad hoc Chainsaw: pure random way DONet: greedy local rarest-first PALS: round-robin method

8. Problem Statement and Formulation How to do optimal scheduling to maximize the throughput of the whole overlay? The real situation is more complicated because different blocks may have different importance and the bottlenecks are not only at the last mile. Our basic approach: Define priority to different blocks due to their importance Maximize the sum of priorities of all requested blocks Throughput is 4Optimal scheduling, throughput gain is 25% Some requests congestion at node 1 Local Rarest First (LRF) strategy

9. Problem Statement and Formulation - Priority Definition We use two factors to represent the significance of a block: rarity factor emergency factor We define the priority of block j ∈ A i for node i ∈ R as follow: P j i = βP R ( Σ k ∈ Nbr(i) h kj )+(1-β)P E (C i +W T -d j i ), Where 0≤β≤1, functions P R (*) (rarity factor) and P E (*) (emergency factor) are both monotonously non-increasing ones

10. Problem Statement and Formulation - Formulation Decision variable Global block scheduling problem: s.t. NotationDefinition NN+1 is the number of overlay nodes, where node 0 is the source node Ii,Ii,the inbound bandwidth of node i Oi,Oi,the outbound bandwidth of node i E ik,the end-to-end available bandwidth between node i and node k h kj ∈ {0,1} Blocks availability: “a kj =1” denotes node k holds block j; otherwise, “a kj =0” NBR i set of neighbors of node i τperiod of requesting new blocks WTWT the exchanging windows size CiCi the current play out time of node i djidji play out time of block j at node i DiDi set of all absent blocks in the current exchanging window of node i

11. Global Optimal Solution Convert the global block scheduling formulation into an equivalent Min-Cost Flow Problem

12. Global Optimal Algorithm Proposition: The optimal goal of global block scheduling problem has the same absolute value as the minimum flow amount of its corresponding min-cost network flow problem. The flow amount on arc (v ki n, v ij b ) ∈ {0, 1} is just the value of x kj i, which is the solution to the optimal block scheduling. Algorithm complexity: O(nm(loglogU)log(nC)), where n and m are the number of vertices and arcs while U and C is the largest magnitude of arc capacity and cost

13. Distributed Algorithm We first use a simple way to estimate the bandwidth that is available from each neighbor with historical information. q ki (m) : the total number of blocks arrived at node i from neighbor k in the m th period. W ki (m+1) : the estimated bandwidth from node k to node i

14. Distributed Algorithm With the estimated available bandwidth, a local block scheduling is performed on each node It can be also transformed into an equivalent min-cost network flow problem for local optimal request

15. Distributed Algorithm Heuristic distributed algorithm: Node i estimates the bandwidth W ki (m+1) that its neighbor k can allocate it in the (m+1) th period with the traffic received from that neighbor in the previous M periods, as shown in equation (3); Based on W ki (m+1), node i performs the local block scheduling (2) using min-cost network flow model. The results x kj i ∈ {0,1} represent whether node i should request block j from neighbor k; Send requests to every neighbor.

16. Performance Evaluation - Compared Scheduling Methods Random Strategy: each node will assign each desired block randomly to a neighbor which holds that block. Chainsaw uses this simple strategy. Local Rarest First (LRF) Strategy: A block that has the minimum owners among the neighbors will be requested first. DONet adopts this strategy. Round Robin (RR) Strategy: All the desired blocks will be assigned to one neighbor in a prescribed order in a round-robin way. If there is multiple available senders, it is assigned to a sender that has the maximum surplus available bandwidth.

17. Simulation Configuration For a fair comparison, all the experiments use the same simple algorithm for overlay construction Delivery ratio: to represent the number of blocks that arrive at each node before playback deadline over the total number of blocks encoded. DSL nodes: Download bandwidth: 40% 512K, 30% 1M, 30% 2M Upload bandwidth: half of download bandwidth 500 nodes Each node has 15 neighbors Request period: 2 second

18. Simulation Results All are DSL nodes with exchanging window of 10 sec and bottlenecks only at the last mile. Group size is 500

19. Simulation Results All are DSL users with exchanging window of 10 sec and end-to-end available bandwidth 10~150Kbps. Group size is 500

20. Conclusion & Future Work The contributions of this paper are twofold. First, to the best of our knowledge, we are the first to theoretically address the streaming scheduling problem in data-driven (swarming based) streaming protocol. Second, we give the optimal scheduling algorithm under different bandwidth constraints, as well as a distributed asynchronous algorithm which can be practically applied in real system and outperforms existent methods by about 10%~80% Future work How to do optimization over a horizon of several periods, taking into account the inter-dependence between the periods. How to do optimal scheduling with scalable video coding (such as layered video coding) or multiple description coding

21. Thanks Q&A