Peer Centrality in Socially-Informed P2P Topologies Nicolas Kourtellis, Adriana Iamnitchi Department of Computer Science & Engineering University of South Florida Tampa, USA 11 th IEEE International Conference on Peer-to-Peer Computing Kyoto, Japan, 2011

Social and Socially-aware Applications  Applications collect social information: Location, collocation, history of interactions, etc.  Use it for recommendations, inferring trust, etc.  How is this information stored and mined? 2 Internet Applications Mobile Applications

Social Graphs and P2P Networks  User social graph over particular activity edges  Users’ peers organized into a P2P network  Users store their data (edges) on particular peers 3

Motivational Example  G’s 2-hop neighborhood?  Social graph traversals translate to many P2P lookups 4 => B, C, E, A, D, F, I

Motivation  Peers acquire particular network properties due to users storing their data on them. E.g., peer 2 more central than peer 1  Application performance affected by projection of social graph on P2P network.  How does the topology of the social graph affect the P2P routing? 5

 Projection Graphs: help us study the network properties of the peers. Projection Graph Model 6 Projection Graph (PG) P2P Overlay Social Graph (SG)

7 Outline  Motivation  Projection Graph Model  Social Network Centrality Metrics Degree Node Betweenness Edge Betweenness  Centrality Calculation Limitations  Research Questions  Experimental Setup  Experimental Results  Impacts on Applications & Systems

Degree Centrality  Direct connections of a node with others  Useful to identify nodes that: Can contact directly many others with a message broadcast and perform as network hubs in a graph 8

Node Betweenness Centrality  Shortest paths between two nodes that pass through a third node, over all shortest paths between the two nodes.  Useful to identify nodes that: Control communication over indirect routes Can host data caches for reduced latency to locate data 9

Edge Betweenness Centrality  Shortest paths between two nodes that pass through an edge, over all shortest paths between the two nodes.  Useful to identify edges that: Connect distant parts of network Can monitor and block malware traffic 10

Centrality Calculation: Not easy!  Limiting factors in calculation of peer measures: Users keep their data private (encrypted, etc.) Users allow access only to their peer Intractable number of shortest paths in large graphs Unavailability of data due to peer churn 11

Research Questions  Assuming that users allow access to their centrality scores in the social graph (SG): How well can we approximate the centrality scores of their peers in the projection graph (PG)? How do the cumulative centrality scores of users associate with the centrality scores of their peers? How does the number of users storing data per peer affect the centrality scores of their peers? 12

13 Outline  Motivation  Projection Graph Model  Social Network Centrality Metrics Degree Node Betweenness Edge Betweenness  Calculation Problems  Research Questions  Experimental Setup  Experimental Results  Impacts on Applications & Systems

Experimental Setup  Community Detection: a recursive version of the Louvain algorithm  Each community mapped on a peer  Merged communities to reach average size 10, 20, …, 1000 users/peer  Community sizes exhibit social structure of power-law nature  Calculate & compare centralities for SGs & PGs 14 Social NetworkNumber of UsersNumber of Edges gnutella0410,87639,994 gnutella3162,561147,878 enron33,696180,811 epinions75,877405,739 slashdot82,168504,230

Comparison of Centrality Scores  Turning point: Centralities of peers reach max P2P network exhibits optimal structuring Maximum opportunity for peers to influence information flows through them. 15 Users/Peer vs. Degree Users/Peer vs. Node Betweenness Users/Peer Vs. Edge Betweenness

Correlation of Centrality Scores  Before turning point: PG resembles closely SG Correlation of SG and PG metrics is highest Degree and Node Betweenness estimated by local info (cumulative scores) 16  After turning point: PG topology loses social properties A highly connected clique Peers acquire equal importance in graph traversal Users/Peer vs. Degree Users/Peer vs. Node Betweenness Users/Peer vs. Edge Betweenness

Finding High Betweenness Peers  Such peers affect system performance and security.  Difficult to identify (network scale, peer churn, etc.)  Can we identify such peers, knowing the top betweenness users?  Top 5% betweenness centrality users => top betweenness centrality peers with 80–90% accuracy 17 Users/Peer (Top-N% users) Users/Peer (Top-N% communities)

Impact on Applications & Systems  Target high degree peers to: Decrease search time Increase breadth of search and diversity of results  Target high betweenness peers to: Monitor information flow and collect traces Place data caches and indexes of data location Quarantine malware outbursts Disseminate software patches  Tackle P2P churn Predict centrality of peers to allocate resources  Reduce overlay overhead Enhance routing tables with P2P edges for faster & more secure peer discovery 18

19 Thank you! This work was supported by NSF Grants: CNS and CNS

Projection Graphs  Community Size Distribution  Degree Distribution 20

Approximation of Peer Betweenness  Top 5% betweenness centrality users => top betweenness centrality peers with 80–90% accuracy 21 1.Pick top-N% betweenness users. 2.Identify set U of their peers. 3.Pick k=|U| top betweenness peers, set P. 4.Compare sets U & P, find peer overlap. 1.Pick set C communities in top-N% cumulative score of betweenness. 2.Pick q=|C| top-N% betweenness peers, set P. 3.Compare sets C & P, find peer overlap.

P2P Social Networks and Services  P2P Systems that could benefit from this work: Commercial Efforts:  Diaspora  FreedomBox  EnThinnai Academic Efforts:  Prometheus  LifeSocial.KOM  Vis-à-Vis  Safebook  PeerSoN  Tribler  F2F  Turtle  Sprout 22