1. 1 Insertion of ISP-owned Peer & Locality Awareness in BitTorrent Ioanna Papafili, George D. Stamoulis, Sergios Soursos AUEB EuroNF workshop, Athens October 16-17, 2008

2. 2 Outline BitTorrent Optimization Potential Incentives for All Players Simulations Results Conclusions Open Issues

3. 3 BitTorrent Optimization Potential

4. 4 Problem & Objectives BitTorrent responsible for great amount of traffic – Implies great cost for ISPs in terms of interconnection agreements Choking algorithm: tit-for-tat principle for social fairness Exploit tit-for-tat to achieve performance improvements in terms of – Reduction of inter-domain traffic – Reduction of end-users completion times (if possible…)

5. 5 Insertion of ISP-owned Peers (I) Idea: Insertion of ISP-owned peers (IoPs) – Controlled by the ISP – Equipped with high download / upload capacity – Equipped with large storage space – Participate actively in the swarm – Run the overlay protocol – More unchokes available than regular peers – Only licensed content due to legal issues Worst case: IoP initially has no content Best case: Content is downloaded only once per ISP Clarification: “IoP insertion is not insertion of a cache combined with interception of peer’s messages”

6. 6 Insertion of ISP-owned Peers (II) Motivation: “Reduce inter-domain traffic while maintain good peer performance or improve it if possible” Issue #1: IoP also selected from peers outside the ISP – Due to high upload rates & tit-for-tat mechanism – As a result extra egress inter-domain traffic generated Issue #2: Increase intra-domain traffic – In same cases means increase of costs for the ISP

7. 7 Insertion of ISP-owned Peers (III) IoP preferred by local peers in 1.Pure BT network: due to ‘tit-for-tat’ mechanism 2.Locality-aware BT network: due to locality mechanism* * R. Bindal, P. Cao, W. Chan, J. Medval, G. Suwala, T. Bates, A. Zhang, “Improving Traffic Locality in BitTorrent via Biased Neighbor Selection”, 26 th IEEE International Conference on Distributed Computing Systems, p. 66, 2006 ISP 1 ISP 2 : inter-domain link : intra-domain link : overlay link

8. 8 Incentives for All Players

9. 9 Incentives ISP – Reduction of ingress inter-domain traffic – Reduction of interconnection costs – Improvement of customers’ QoE – What about intra-domain costs? Content Provider – Improvement of customers’ QoE – Possible interconnection agreement with ISP? End-Users – Transparent to the end-users

10. 10 Simulations* *K. Eger, T. Hoßfeld, A. Binzenhöfer, G. Kunzmann, "Efficient Simulation of Large-Scale P2P Networks: Packet-level vs. Flow-level Simulations", 2 nd Workshop on the Use of P2P, GRID and Agents for the Development of Content Networks (UPGRADE-CN'07) in conjunction with IEEE HPDC, Monterey Bay, USA, June 2007

11. 11 Scenarios (I) 1.Pure BT 2.Locality-aware BT 3.BT with Insertion of ISP-owned peer 4.LA BT with Insertion of ISP-owned peer Symmetric or Asymmetric – Symmetric: 25 peer per AS, e.g. 2 Tier-4 ISPs – Asymmetric: 35 and 15 peers in each AS, e.g. Tier-3 and Tier-4 ISPs respectively All-together or Split – All-together: Joining time of all peers ~U(0,10) – Split: Joining time of 5 peers in each AS ~U(150,300), whereas joining time of the rest of the peers in each AS and the ISP-owned peer ~U(0,10)

12. 12 Scenarios (II) Insertion of IoP in BitTorrent without locality awareness Insertion of IoP in BitTorrent combined with locality awareness

13. 13 Simulation parameters DescriptionValue Number of peers50 Number of seeds1 Number of ASes2 Number of peers per AS(25,25), (35,15) Upload capacity of regular peers512K Download capacity of regular peers4096K File size20M Number of peers requested from tracker (Size of tracker’s list) 25 Number of local peers replied by tracker20 Number of connections20 Choking interval10 Number of unchoked connections permitted per peer 4, 10 (in case of IoP) Number of ISP-owned peers1 Upload/download capacity of ISP-owned peers40960K @ AS 0 @ AS 1

14. 14 Results

15. 15 Insertion of IoP without LA (Symmetric, All-together) – Ingress inter-domain traffic Important reduction – up to 30% - of ingress inter-domain traffic to AS 1 Similar results apply for symmetric and asymmetric, and all-together and split cases

16. 16 Ingress inter-domain traffic (Symmetric, All-together) Up to 30% reduction for both Ases when only LA is employed Up to 60% reduction for AS1 when IoP is also inserted in the LA BT As expected ingress inter-domain traffic of AS 0 is increased, however it is still less than the pure BT case

17. 17 Ingress inter-domain traffic (Asymmetric, All-together) Up to 15% reduction when only LA is employed Up to 45% reduction for AS1 when IoP is also inserted in the LA BT However, the ingress inter-domain traffic for AS 0 (no IoP) is increased up to 30% compared to pure BitTorrent!

18. 18 Completion Times – Symmetric case Up to 10-15% improvement of all peers’ completion times by IoP Up to 35-40% improvement especially for peers that join the swarm later than the IoP (peaks)

19. 19 Completion Times – Asymmetric case Similar results as in previous case

20. 20 Conclusions

21. 21 Conclusions ISP deploying only locality awareness risk to lose customers because of completion time deterioration The insertion of IoP improves both ingress inter-domain traffic and peers’ completion times for the AS that deploys the extra peer Consequently: – Interconnection agreement may be modified in favor of the AS that deploys the extra peer – The AS does not risk losing customers, on the contrary it may attract new ones

22. 22 Open Issues

23. 23 Possible Issues for Future Research Asymmetric case & IoP  increased ingress inter-domain traffic for AS 0 – Is there an incentive for AS 0 to introduce an IoP as well? – Study how the results will be affected So far: One swarm & one IoP – Variations: One swarm and multiple IoPs / Multiple swarms and multiple IoPs – Tradeoff: extra performance improvement vs. extra resources – How many IoPs? – How do their total resources scale? So far: Fixed peer population & only 1 file per simulation – New assumptions: peers leave/enter the swarm, more files exchanged Study how interconnections agreements are affected! – E.g. under the 95 th percentile charging scheme

24. 24 Legal issues Copyrighted content in ISP’s premises Interconnection agreements with content providers? Other idea: Enhance regular peers’ upload capacity – Choose highly active peers and increase their upload (and/or download) capacity instead of inserting ISP-owned peers Incentives for All Players: – Highly active peers: Get better service at no extra cost – Other peers: Experience better QoE – ISP: Achieves reduction of inter-domain traffic without facing legal issues since the content is not stored in its premises!

25. 25 Thank you! Questions?

26. 26 Back up

27. 27 Related Work

28. 28 Topology Awareness Issue: Tracker replies a random list of peers to each peer’s request! – Inefficient use of the underlay – Affects also the performance of the overlay Idea: Alternative peer selection at the tracker – Apply a proximity criterion Potential proximity criteria: Autonomous System, number of hops, Round-Trip-Time, congestion, price Topology information must be available to the tracker – Information provided by the underlay – Or by the peers themselves – Incentives to be truthful?

29. 29 Locality Awareness Proximity criterion: Autonomous System Tracker’s reply list comprises of (Bindal et al.*): – K out of N peers selected within the same AS with the requesting peer – N-K selected from other ASes Important reduction of ingress inter-domain traffic is achieved No improvements on peer’s completion times are observed – For some peers the completion time is increased! * R. Bindal, P. Cao, W. Chan, J. Medval, G. Suwala, T. Bates, A. Zhang, “Improving Traffic Locality in BitTorrent via Biased Neighbor Selection”, 26 th IEEE International Conference on Distributed Computing Systems, p. 66, 2006

30. 30 Simulator

31. 31 bittorrent.patch* for the ns-2 simulator BitTorrent-like protocol, functions simplified Four classes implemented: Application, Tracker, Connection, Message BitTorrent implementation is modular; e.g. peer and piece selection algorithms can be replaced by alternatives Network model: – FullTCP: bidirectional data transfers – Uplink is assumed to be the bottleneck in the whole network – Downlink is neglected * K. Eger, T. Hoßfeld, A. Binzenhöfer, G. Kunzmann, "Efficient Simulation of Large-Scale P2P Networks: Packet-level vs. Flow-level Simulations", 2 nd Workshop on the Use of P2P, GRID and Agents for the Development of Content Networks (UPGRADE-CN'07) in conjunction with IEEE HPDC, Monterey Bay, USA, June 2007