Maximizing Resilient Throughput in Peer-to-Peer Network: A Generalized Flow Approach

2. Outline Introduction Solution Overview Resilience Factor Generalized Throughput Optimal Peer Selection Optimization Framework Maximizing Generalized Throughput Findings Simulation Study Generalized Throughput vs. Volume Conclusion

3. Introduction Dynamics in Peer-to-Peer Network: A Fact to Live with Unscheduled Peer Departure User Leaving Machine/Network/Software Failure Focus of This Work Modeling the fault resilient peer selection problem under an optimization framework Combining fault resilience with key performance metrics throughput Heuristic: Lack the theoretical foundation to analyze the global performanceHeuristic: Lack the theoretical foundation to analyze the global performance

4. Existing Works Resilience of Random Graph D. Leonard, Z. Yao, V. Rai, and D. Loguinov, On Lifetime-Based Node Failure and Stochastic Resilience of Decentralized Peer-to-Peer Networks, IEEE/ACM Transactions on Networking, 2007. Expected delay before a peer is isolated Prioritization Heuristics M. Bishop, S. Rao, and K. Sripanidkulchai, Considering Priority in Overlay Multicast Protocols under Heterogeneous Environments, IEEE InfoCom 2006. Preempt-Degree, Preempt-Age, Hybrid G. Tan and SA Jarvis, Improving the Fault Resilience of Overlay Multicast for Media Streaming, IEEE Transactions on Parallel and Distributed Systems, 2007. Bandwidth-ordered Tree, Time-ordered Tree, Reliability-Oriented Switching Tree

5. Solution Overview Generalized Flow Model A generalized version of multicommodity flow Broad Range of Applications Financial Network, Commodity Transportation Gain Factor Theft, Loss Interest Rates Application in P2P Network? Modeling peer resilience as gain factor

6. Resilience Factor Measuring Peer Reliability Our optimization model makes no assumption on how resilience factor should be defined As an example r = Pr(T > ?) = 1 � F(?) Chance that the peer will survive until ?

7. Generalized Throughput Resilience Index Rt(v): resilience peer v experiences in tree t Summary of resilience indices for all peers in tree t R(t) = sumv?t Rt(v) Generalized Throughput Product of flow rate of and resilience index of tree t fg(t) = f(t)R(t)

8. Optimal Peer Selection Goal Maximizing Generalized Throughput Under Network Capacity Constraint Flow conservation Problem Settings Network Model General Topology vs. Star Topology Overlay Organization Unlimited Number of Trees vs. Single Tree Why not Mesh? Resilience Index Concatenation vs. Non-concatenation

9. Findings Maximizing Generalized Throughput Maximizing Throughput (in comparison) Y. Cui, B. Li, and K. Nahrstedt, On Achieving Optimized Capacity Utilization in Application Overlay Networks with Multiple Competing Sessions, ACM SPAA 2004.

10. Two Topologies General Network

11. Solutions under Star Topology Multi-Tree Solution At most n+1 trees needed Single Tree Solution Find the maximum flow rate a single tree can afford Greedily construct the tree by prioritizing the most resilient peers

12. Simulation Study Two Experimental Topologies BRITE Topology 1000 nodes, 2000 edges, Waxman Model Bandwidth randomly distributed between 100 and 1000Kbps STAR Topology 1000 nodes Peer outbound bandwidth randomly distributed between 100 and 1000Kbps Peer Resilience 100 peers randomly attached to each of the above networks Lifetime Distribution Mean Lifetime Varying from 1500 to 3500 seconds Exponential and Pareto distributions Resilience Factor r = Pr(T > ?) = 1 � F(?)

13. Generalized Throughput vs. Volume Volume Total amount of data collected by each peer until itself or one of its ancestor dies Multiple Trees, General Network, Non-Concatenation

14. Generalized Throughput vs. Volume Volume Total amount of data collected by each peer until itself or one of its ancestor dies Multiple Trees, Star Network, Concatenation

15. Performance of Single-Tree Algorithms Normalized by the generalized throughput achieved by the optimal multi-tree solution STAR Topology, Concatenation

16. Conclusions Modeling P2P Network Resilience Optimization framework Generalized Flow Theory Maximizing Generalized Throughput Problem Space General Topology vs. Star Topology Unlimited Number of Trees vs. Single Tree Concatenation vs. Non-concatenation Findings Details in Paper Much harder than maximizing throughput problem Future work Better Approximation Algorithms and Heuristics Distributed Solution Combination with other strategies to improve P2P network resilience Preemption Peer Repairing

17. Thank You VANETS Group http://vanets.vuse.vanderbilt.edu Questions?