P2P vs. ISP

1. P2P vs. ISP Ofir Israel Guy Paskar

2. An Internet Tale • Once upon a time.. • Users unhappy (slow connection) • ISPs unhappy (poor revenues) • Then came Broadband access... • And everybody were happy

3. The Villain arrives • P2P File-Sharing Applications (Kazaa, eMule, BitTorrent, etc..) • Users love it! • Good and free content, overnight downloads • ISPs hate it! • Users using their entire link • Internet link utilization gone wild • More bandwidth costs more money!

4. But is it really a villain? • Users love it • Driving force for broadband adoption • Increased revenues for ISPs • What should the ISPs do??

5. Some Ideas • User friendly ideas • Acquire more BW • Network caching • User unfriendly ideas • Increase subscription cost • Volume-based pricing • Block/shape P2P traffic (priority for non-P2P packets)

6. Today.. • Generally understand the problem – DONE! (? ) • Describe an analytical model to help us understand situation better • Describe one practical solution and it’s empirical results (Hint: it works)

7. Research Goals • Modeling framework to analyze interactions between P2P File-Sharing users and their ISPs • Basic insight about system dynamics • Used to evaluate different strategies to manage P2P traffic

8. Meet the Players • User • Generates queries (P2P application finds the object and retrieves it) • Pays a subscription price, has QoS expectations • What’s popular, what’s not • ISP • Goal: TO MAKE MONEY • Sets subscription price • Controls bandwidth • Influences P2P app behavior

9. System Setting N users in the world n users inside ISP User-ISP ul., dl. bandwidth ISP-ISP ul., dl. bandwidth User generates query Gets a response from within the ISP Or from a user in another ISP

10. The Simple System Model Unconstrained downloads from within the ISP Prob. Object is located inside ISP Prob. P2P App. locates object Aggregate query rate Average query rate System throughput Model for “Internet to ISP” link Object retrieval prob. (QoS):

11. User Utility Function Cost Benefit Subscription cost Shape parameter Object retrieval prob. Users subscribe only if: Equivalently, if: is the minimal service level acceptable by user i

12. ISP Utility Function Fixed cost Cost Benefit Cost per unit of BW Revenues from subscribers’ fee ISP starts service only if:

13. Traffic Locality • Probability that there exists at least one internal replica of object replicated r times in the system • Probability to download from internal replica Number of files inside ISP Number of files outside ISP Locality parameter

14. Minimum BW • Reminder: • So: • Assuming • We get

15. Minimum BW • Non-linear behavior (on n) • More users  more locality  less BW needed • Can be zero if n large enough (self-sustainability) • Dependant on multiple parameters • Self-Sustainability

16. Simple(?!) Model

17. Impact of Object Replication (r) • More replicas  Better locality  Lower Bd needed • Bmin has two roots: x1 – No users, x2 – Enough users for self-sustainability

18. Impact of external QoS ( ) • Higher external QoS More BW needed (because there are more replicas externally)

19. Impact of prob. to locate objects (q) • Some ISPs drop queries. This graph shows them different.

20. Impact of prob. to retrieve objects internally (Gamma) • Det. by the ability to find a local object given that it exists. • Can be influenced by the ISP – this graph shows it should.

21. Model Refinements • Refined Model • Relax these assumptions • Propose object popularity and replication model • Simple Model • Users’ access BW are unconstrained • Object popularity is identical • Users availability identical

22. Model Refinements • We adopt a processor sharing model with rate limit bd to describe the sharing of Bd Now each user is limited by it’s own BW. Queue Model

23. Model Refinements • We introduce a new parameter: that describes user patience • Denote b as the initial download rate, and assuming the decision to abort is made at the beginning then the prob. pg to continue the transfer is: • Larger eta  user claims to get a rate close to what they paid for

24. Bmin as a function of bd=bu with different values of gamma • Higher gamma  smaller bu needed for self-sustainability • Optimal gamma is not gamma=1 !!! • For bu < 250 the BW available inside the ISP is not enough to satisfy demanding users

25. Impact of asymmetric access BWs • Cost for ISP increases as ratio increases (what about ADSL??) • Larger bu Better locality  lower Bd

26. Conclusions • Locality is good for the ISPs • More replicas, larger querying probability, larger upload bandwidth for users’ access, larger probability to retrieve objects internally (gamma)  SELF SUSTAINABILITY == GOOD • Reading slow leads to better understanding 

27. Further Reading • Original paper of course: Garetto et al, “A modeling framework to understand the tussle between ISPs and peer-to-peer file-sharing users” in PerformanceEvaluation, June 2007 • Same as the original paper but talks about ISP-ISP connections: Wang et al, “Modeling the Peering and Routing Tussle between ISPs and P2P Applications” in the proceedings of IWQoS 2006

28. BREAK ?

29. Academic Work • Oracle-based vs. non-Oracle-based (e.g., with ISP cooperation or without) • Legality issues, reluctance issues • Improvements via locality research • Network location or Geographic location? • Which method of network location? • Improvements via redirection research • Can we redirect traffic to inexpensive links? • Many more

30. Part 2 Taming the Torrent - A Practical Approach to Reducing Cross-ISP Traffic in Peer-to-Peer Systems David R. Choffnes and Fabián E. Bustamante

31. The problem • Over 66% of P2P users & growing • But how do we know which peer to choose? • Which peers? Trackers provide a random subset of peers in the torrent • Random peer connections → growing ISP operation costs. • So , how do we know if a suggested peer is inside our Isp or outside? • We want to reduce cross isp transport. Meaning use the “closest” peers. • But , how can we do that?

32. The ISP Perspective • P2P performance - key factor for service upgrade & selection by users • A major engineering challenge for ISPs • ≈70% of the Internet traffic • But , a lot of cross isp ,means a lot of cost for the Isp. • What can the isp do in order to fight the p2p users?

33. Isp methods and its problems • ISPs shape traffic directed to standard ports • P2Ps move to dynamic, non-standard ports • ISPs turn to deep-packet inspection to identify & shape P2P flows • P2Ps encrypt their connections • ISPs place caches and/or spoofs TCP RST msgs • Legality issues. (Some ways to overcome this – in Israel!) So good solution must be agreed by the p2p users!

34. One solution: Oracles. • Suggestion – the isp’s itself will have to implement an oracle, this oracles will guide the user which peers to choose. • Help reduce cross-ISP traffic • This solution looks appealing But: • Assumes P2P users & ISPs trust each other • Misses incentive for user adoption • Therefore not so good after all

35. The authors suggestion • CDN’S – content distribution network. What is it? • CDNs attempt to improve web performance by redirecting requests to replica servers • The goal is to help content providers (i.e. CNN) to distribute content by redirecting requests to replica servers that are: • Topologically proximate • Provide lower-latency • But how do they do that?

36. How does CDN work? • There are some ways that a CDN works by for example: • Way 1 : I want to go to cnn.com  dns lookup , directs me to the domain name of the CDN(cnn.akamai.com) CDN sends me to the right replica. • Way 2: I want to go to cnn.com dns lookup  first page from original cnn.com, directs me to CDN server  sends to right replica.

37. Reusing CDNs’ network views • Client’s request redirected to “nearby” server • Client gets web site’s DNS CNAME entry with domain name in CDN network • Hierarchy of CDN’s DNS servers direct client to nearby servers Multiple redirections to find nearby edge servers Hierarchy of CDN DNS servers Internet Customer DNS servers Web replica servers (3) Another web client (4) Clients and replica servers are “nearby”] (5) (2) (6) Client is given 2 web replica servers (fault tolerance) Client gets CNAME entry with domain name in Akamai LDNS (1) Client requests translation for Yahoo Web client

38. The authors suggestion • So how do we use CDN? • We are going to recycle data that is already being collected by Content Distribution Networks, and use it. • But how? • By simply comparing DNS redirections. • Assumptions : • Links between “nearby” hosts cross few ISPs • If two hosts are close to the same CDN replica servers, they are close to each other

39. Reducing cross-ISP traffic • So we can use the CDN’s data, what are the advantages for this recycilng? • Does not requires any trust between isp and p2p users • The infrastructure is already exist • And most importantly reduces cross isp traffic without harming the p2p users (even improving)

40. An approach to reducing cross isp • Introducing “Ono” • Extension (plugin) to AzureusBitTorrent client • Will use CDN and ratio maps(?) to determine who is “closer” • We will describe its implementation • Uses multiple CDN customers • Only DNS resolution, no content download needed • Adaptive lookup rates on CDN names • Represent redirection behavior using ratio-maps

41. Ratio Maps • A ratio map represents the frequency of redirecting to a specific replica • Number of replicas is usually small (max 31) • Keep a time window about 24 hours • How does it looks?

42. Ratio maps represantation • The ratio map of a peer (a) is a set of (replica server, ratio) • for peer a • Specifically, if peer a is redirected toward replica server r1 75% of the time window, and toward replica server r2 25% of the time window, then the corresponding ratio-map is • The sum of all in a given ratio map equals one • For each peer there exist a ratio map • But what can we do with it?

43. choosing peers by ratio map • 2 peers has close ratio map , than we say that they are close. ( possibly in the same network), and the ooposit • So , we need a calculation that will determine for 2 peers if they are “close” or not. • Than we can check for all available peers and choose the “closest” one • For that we define cosine similarity for 2 peers

44. Cosine-similarity • the cosine similarity of two maps will range from 0 to 1, since the term frequencies cannot be negative • If cos_sim(a,b) = 0 , the vectors are orthogonal • if cos_sim(a,b) = 1 than they are the same • This is very close to dot product • And we determine a threshold currently 0.15 , if cos_sim(a,b)>0.15 than we recommend these peers as close

45. Implementation • Ono, an extension to Azureusclient • Performs periodic DNS lookups on popular CDN and create a ratio map • Periodically updates the ratio-maps • Exchanging ratio maps for cos-sim(a,b): • On Handshake • From trackers • But how do we deal with peers not using Ono? • Ono also attempts to perform DNS lookups on behalf of other peers that it encounters, to determine their ratio maps • How? • Taken from Ono code : getting the other peer DNS server • And querying it

46. So what's now? • Get ratio information from other peers that got from tracker , and understand who is close • When determine similar redirection behavior, attempts to bias traffic towards that peer by ensuring that the connection is always on • Sends Ono information to supporting trackers(in case of supporting trackers) • But what is the cost? How much is our overhead? • 18KB upstream, 36KB downstream per day • Computation of cosine-similarity is easy

47. Important notes • CDN names being used: • Initialization of ratio map: • DNS on each CDN name at most once every 30 sec. for 2 min. this gives basis ratio map • After this phase • If the redirection info for CDN name similar to prev. query  the interval between queries increases by 1 min. • Otherwise the interval is halved(to a min. of 30 sec.)

48. Some statistics regarding Ono • Details for 2007 : • > 195,000 users worldwide • … collecting ~15GB of data per day

49. Empirical results • Over 120,000 peers use Ono • Ono collects extra network data • Samples transfer rates for each connection every 5 sec. • Get RTT for endpoint using pings • Get Trace-route between end points • Note : • Not easy to determine cross-ISP hops • IP hops is easy and gives some measure

50. Empirical resualts • So in practice Trace Route gives a router level views of path between hosts. BUT an ISP can contain many routers, we wish for a metric that is closely correspond to ISP hops. • How do we get this metric? • Autonomous systems , how? • Although there is no 1 to 1 correlation between AS and ISPs, the number of AS hops gives us an upper bound estimate on the number of cross-ISP hops • So in practice we generate AS level path info from our trace-routes using mapping that can be provided • Example :