1. P2P systems: epidemic scheduling, content placement and user profiling Laurent Massoulié Thomson, Paris Research Lab

2. 2 Outline Epidemic schemes for live streaming – Rate-optimality – Delay-optimality Content placement – Optimisation framework – Adaptive replication User profiling – Spectral clustering – Linear programming

3. 3 Outline Epidemic schemes for live streaming – Rate-optimality – Delay-optimality Content placement – Optimisation framework – Adaptive replication and 3/4 - competitivity User profiling – Spectral clustering – Linear Programming

4. 4 Context P2P systems for live streaming on the Internet – PPLive, CoolStreaming, Sopcast, TVants,TVUPlay, Joost…

5. 5 Network constraints ● Graph connecting nodes ● Capacities assigned to edges Achievable broadcast rate [Edmonds, 73]:  Equals maximal number of edge-disjoint spanning trees that can be packed in graph  Coincides with minimum over receivers of max-flow ( = min-cut) between source and receiver

6. 6 Based on local informations No explicit construction of spanning trees Random Useful chunk selection and Edmonds’ theorem [LM, A. Twigg, C. Gkantsidis & P. Rodriguez] 14 5 124578 When injection rate at source is strictly feasible, Markov process is ergodic.  Chunks successfully broadcast with bounded delay ? ? ? ? ? ? ?? ?

7. 7 Network with access (node) constraints … ● Scarce resource: access capacity ● Complete communication graph: Everyone can send to anyone ●Bound on maximum streaming rate λ: Let c i = uplink b/w of node i Necessary condition for feasibility:

8. 8 Deprived Peer / Random Useful Chunk [LM, A. Twigg, C. Gkantsidis & P. Rodriguez] 124578 Sender’s packets 157814 Potential receiver 1Potential receiver 2 5 Source policy: sends “fresh” packets if any (fresh = not sent yet to anyone)

9. 9 Deprived Peer / Random Useful Chunk [LM, A. Twigg, C. Gkantsidis & P. Rodriguez] 124578 Sender’s packets 157814 Potential receiver 1Potential receiver 2 5 Neighborhood management: Periodically add random neighbor & suppress least deprived neighbor  Fixed neighborhood sizes

10. 10 Main result When λ < λ*, Markov process is ergodic.  Hence all packets are received at all nodes after time bounded in probability

11. 11 Multiple commodities Several sources s, Dedicated receiver sets V(s) Can overlap Sources are not receivers Nodes cannot relay commodities they don’t consume …

12. 12 Multiple commodities Necessary conditions for feasibility: Bundled most deprived / random useful: do not distinguish between commodities when – measuring deprivation – Chosing random useful packet System is ergodic when Conditions hold with strict inequality

13. 13 Symmetric Networks (c 1 = c 2 =... = c N = 1 chunk / sec ) Previous lower bound reads log 2 (N) Achievable [J. Mundinger & R. Weber]: source t t-1 t-2 t-3 t+1 Makes use of log 2 (N) trees; not robust against churn

14. 14 A look at the corresponding trees N=4 N=8 N=16 N=32

15. 15 Random target / latest useful packet ? Sender’s packets Receiver’s packets Latest useful pkt ??? 124578 1238

16. 16 I.e:Diffusion at rates arbitrarily close to optimal feasible under optimal delay ( plus constant) Random target / latest useful packet For arbitrary injection rate λ<1 and constant x>0, Each peer receives fraction 1- 1/x of packets in time log 2 (N)+O(x). [T. Bonald, LM, F. Mathieu, D. Perino & A. Twigg]

17. 17 Open questions Delay optimality in heterogeneous environments Cost optimality Convergence time scale

18. 18 Outline Epidemic schemes for live streaming – Rate-optimality – Delay-optimality Content placement – Optimisation framework – Adaptive replication User profiling – Spectral clustering – Linear programming

19. 19 Outline Epidemic schemes for live streaming – Rate-optimality – Delay-optimality Content placement – Optimisation framework – Adaptive replication User profiling – Spectral clustering – Linear programming

20. 20 Problem statement N users Storage capacity: m objects Service capacity: B requests Local accesses are free Request rate: f for object f Request duration: 1 Aim: minimize number of lost requests

21. 21 Optimal placement structure Let M f = number of replicas of object f Schedulable region: request rates x f verifying Effective arrival rates: times K if objects can be split into K size (1/K) sub-objects

22. 22 Hot/Warm/Cold partition Sort objects according to popularity : 1  2  … Replicate everywhere (M f =N) top popular objects 1…,f(1) Partial replication of objects f(1)+1,…f(2) : No replication of objects for f>f(2) f(1) and f(2) : such that “warm objects” generate requests at rate BN, and all memory is used

23. 23 Adaptive replication Replication policy: – Create new replica for object f after each dropped request – Remove object chosen at random Ignoring object-specific capacity constraints, caricature dynamics:  Equilibrium:

24. 24 Adaptive replication (ctd) Compare to full replication of only top popular objects, i.e. Then reductions to offered rates verify  “Value of foresight” is less than 25%...

25. 25 Outline Epidemic schemes for live streaming – Rate-optimality – Delay-optimality Content placement – Optimisation framework – Adaptive replication User profiling – Spectral clustering – Linear programming

26. 26 Outline Epidemic schemes for live streaming – Rate-optimality – Delay-optimality Content placement – Optimisation framework – Adaptive replication User profiling – Spectral clustering – Linear programming

27. 27 User profiling Aim: predict tastes of users Applications: – Further optimization of placement – Recommender Systems

28. 28 Netflix dataset 17, 770 movies, rated by 480, 000 users

29. 29 The planted partition model Users partitioned into clusters k=1,…,K Each pair of users (i,j) : conflict level C(i,j) in [0,1] (e.g., fraction of movies rated differently) Statistical assumptions: – C(i,j) independent over i<j – E(C(i,j)) = b kl D/N if users i,j belong clusters k, l

30. 30 A spectral algorithm Step 1: find suitable “de-noised” descriptors of users  Form normalized eigenvectors x(1),…,x(K) associated to K largest (in absolute value) eigenvalues of conflict matrix  To each user i, assign vector z i =(x i (1),…,x i (K))

31. 31 A spectral algorithm Step 2: do crude clustering on descriptors  Pick a random set of A users u(1),…,u(A)  Identify pair with closest descriptors (for L 2 norm) and remove one of them, until only K users are left, say v(1),…,v(K)  Cluster the nodes according to proximity of their descriptors to the cluster exemplars v(1),…,v(K)

32. 32 Theorem Assume that – Fixed number K of clusters, each of size  (N) – Matrix (b kl ) has full rank K – D  C log(N) for some constant C Then with probability 1-o(1), Algorithm partitions correctly fraction 1-o(1) of nodes for suitable A ( 1<< A << D 1/2 ) Main tool: control of spectral structure of E-R graph adjacency matrix when average degree D  C log(N) [Feige-Ofek]

33. 33 Open question Brute force Maximum Likelihood: retrieves clusters when D>>1  Efficient procedure under this assumption?

34. 34 Another algorithmic version of Netflix Objective: for user n, find inference of all unknown ratings that maximizes number of users fully agreeing with user n  NP-hard (badly so) Probabilistic model – Users belong to clusters k=1,…,K, with sizes a(k) N – Within a cluster, identical ratings (i.i.d., +1 or -1 w.p. ½ for each movie, F movies in total) – Each rating of each user: revealed w.p. p

35. 35 Proposed algorithm (inspiration: compressive sensing; see [Decoding by linear programming, Candes&Tao]) Consider user 1 For suitable cost function g, determine full rating vectors X(n), compatible with known ratings (i.e. P n X(n)=Y(n) ), that minimize A proxy to (intractable) minimization of

36. 36 Conditions for optimality Assume optimum of (II) : “clustered” reconstruction X**(n) such that X**(n)=X**(1) for all indices n  A Then optimum of (I) such that X*(n)=X*(1), n  A provided:

37. 37 Application to probabilistic model Necessary condition for hidden cluster to be optimal: Sufficient condition for LP algorithm to retrieve hidden cluster, under choice g= |.|  :  Differ by factor at most K-1

38. 38 Outlook Clustering – Robustness of proposed schemes to statistical modeling assumptions – Efficient (distributed?) implementations