Efficient Universal Recovery in Broadcast Networks

Transcription

1 Efficient Universal Recovery in Broadcast Networks Thomas Courtade and Rick Wesel UCLA September 30, 2010 Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

2 System Model and Problem Statement Universal recovery in a broadcast network Consider a network of N nodes. Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

3 System Model and Problem Statement Universal recovery in a broadcast network Consider a network of N nodes. Node i begins with a set of packets P i {p 1,..., p k }. p 1 p 2 p 3 Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

4 System Model and Problem Statement Universal recovery in a broadcast network Consider a network of N nodes. Node i begins with a set of packets P i {p 1,..., p k }. Each node broadcasts packets to its neighbors. One transmission consists of one packet. p 1 p 3 p 1 p 2 p 3 Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

5 System Model and Problem Statement Universal recovery in a broadcast network Consider a network of N nodes. Node i begins with a set of packets P i {p 1,..., p k }. Each node broadcasts packets to its neighbors. One transmission consists of one packet. Links are noiseless and interference-free. Each node knows the network topology and which packets each node already knows. p 1 p 1, p 2, p 3 p 3 Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

6 System Model and Problem Statement Universal recovery in a broadcast network Consider a network of N nodes. Node i begins with a set of packets P i {p 1,..., p k }. Each node broadcasts packets to its neighbors. One transmission consists of one packet. Links are noiseless and interference-free. Each node knows the network topology and which packets each node already knows. p 2, p 1 + p 3 p 2, p 1 + p 3 p 1 p 1, p 2, p 3 p 3 Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

7 System Model and Problem Statement Universal recovery in a broadcast network Consider a network of N nodes. Node i begins with a set of packets P i {p 1,..., p k }. Each node broadcasts packets to its neighbors. One transmission consists of one packet. Links are noiseless and interference-free. Each node knows the network topology and which packets each node already knows. Universal Recovery is achieved if some transmission scheme permits each node to recover all k packets. p 1, p 2, p 3 p 1, p 2, p 3 p 1, p 2, p 3 Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

8 System Model and Problem Statement Universal recovery in a broadcast network Consider a network of N nodes. Node i begins with a set of packets P i {p 1,..., p k }. Each node broadcasts packets to its neighbors. One transmission consists of one packet. Links are noiseless and interference-free. Each node knows the network topology and which packets each node already knows. Universal Recovery is achieved if some transmission scheme permits each node to recover all k packets. Question What is the minimum number of transmissions required for universal recovery? Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

9 Transmission Strategies The scheduling of transmissions Definition (Transmission Strategy) Let b j i be the number of transmissions by node i during round j. The set of values {b j i } i,j is called a transmission strategy. Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

10 Transmission Strategies The scheduling of transmissions Definition (Transmission Strategy) Let b j i be the number of transmissions by node i during round j. The set of values {b j i } i,j is called a transmission strategy. For the previous example, b 1 1 = 1, b1 3 = 1, and b2 2 = 2 is a transmission strategy that permits universal recovery. Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

11 Transmission Strategies The scheduling of transmissions Definition (Transmission Strategy) Let b j i be the number of transmissions by node i during round j. The set of values {b j i } i,j is called a transmission strategy. For the previous example, b 1 1 = 1, b1 3 = 1, and b2 2 = 2 is a transmission strategy that permits universal recovery. Given a transmission strategy that permits universal recovery, the encoding operations can be efficiently computed using the algorithm developed by Jaggi et al. Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

12 Transmission Strategies in R r Permit Universal Recovery Theorem For a fixed number of communication rounds r, a transmission strategy in which node i makes at most b j i transmissions during round j permits universal recovery if and only if {b j i } R r. Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

13 Transmission Strategies in R r Permit Universal Recovery Theorem For a fixed number of communication rounds r, a transmission strategy in which node i makes at most b j i transmissions during round j permits universal recovery if and only if {b j i } R r. Proof. 1 Create an equivalent single-source network coding problem. Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

14 Transmission Strategies in R r Permit Universal Recovery Theorem For a fixed number of communication rounds r, a transmission strategy in which node i makes at most b j i transmissions during round j permits universal recovery if and only if {b j i } R r. Proof. 1 Create an equivalent single-source network coding problem. 2 Take any source-sink cut in the network coding graph and reduce to a minimal cut. Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

15 Transmission Strategies in R r Permit Universal Recovery Theorem For a fixed number of communication rounds r, a transmission strategy in which node i makes at most b j i transmissions during round j permits universal recovery if and only if {b j i } R r. Proof. 1 Create an equivalent single-source network coding problem. 2 Take any source-sink cut in the network coding graph and reduce to a minimal cut. 3 Each minimal cut corresponds to a constraint defining R r (and vice versa). Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

16 Transmission Strategies in R r Permit Universal Recovery Theorem For a fixed number of communication rounds r, a transmission strategy in which node i makes at most b j i transmissions during round j permits universal recovery if and only if {b j i } R r. Proof. 1 Create an equivalent single-source network coding problem. 2 Take any source-sink cut in the network coding graph and reduce to a minimal cut. 3 Each minimal cut corresponds to a constraint defining R r (and vice versa). 4 Thus, a transmission strategy permits universal recovery iff {b j i } R r. Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

17 Neighborhood of a set S S Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

18 Neighborhood of a set S Γ(S ) S Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

19 The Region R r A Set of Transmission Strategies Definition (The region R r Z N r + ) {b j i } R r if and only if: S 0 S r [N] satisfying S j Γ(S j 1 ) for each j [r], the following inequalities hold : r b (r+1 j) i. j=1 i Sj c Γ(S j 1) i S r P c i Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

20 An Example 3-Node Line Network Revisited Constraints defining R r r b (r+1 j) i j=1 i Sj c Γ(S j 1) i S r P c i p 1 p 2 p 3 Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

21 An Example 3-Node Line Network Revisited Constraints defining R r r b (r+1 j) i j=1 i Sj c Γ(S j 1) i S r P c i p 1 p 2 p 3 S 0 Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

22 An Example 3-Node Line Network Revisited Constraints defining R r r b (r+1 j) i j=1 i Sj c Γ(S j 1) i S r P c i p 1 p 2 p 3 S 0 S 1 Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

23 An Example 3-Node Line Network Revisited Constraints defining R r r b (r+1 j) i j=1 i Sj c Γ(S j 1) i S r P c i p 1 p 2 p 3 S 0 S 1 S 2 S 0 = {1}, S 1 = S 2 = {1, 2} b Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

24 An Example 3-Node Line Network Revisited Constraints defining R r r b (r+1 j) i j=1 i Sj c Γ(S j 1) i S r P c i p 1 p 2 p 3 Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

25 An Example 3-Node Line Network Revisited Constraints defining R r r b (r+1 j) i j=1 i Sj c Γ(S j 1) i S r P c i p 1 p 2 p 3 S 0 Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

26 An Example 3-Node Line Network Revisited Constraints defining R r r b (r+1 j) i j=1 i Sj c Γ(S j 1) i S r P c i p 1 p 2 p 3 S 0 S 1 Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

27 An Example 3-Node Line Network Revisited Constraints defining R r r b (r+1 j) i j=1 i Sj c Γ(S j 1) i S r P c i p 1 p 2 p 3 S 0 S 1 S 2 S 0 = S 1 = {1}, S 2 = {1, 2} b Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

28 An Example 3-Node Line Network Revisited Constraints defining R r r b (r+1 j) i j=1 i Sj c Γ(S j 1) i S r P c i p 1 p 2 p 3 Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

29 An Example 3-Node Line Network Revisited Constraints defining R r r b (r+1 j) i j=1 i Sj c Γ(S j 1) i S r P c i p 1 p 2 p 3 S 0 Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

30 An Example 3-Node Line Network Revisited Constraints defining R r r b (r+1 j) i j=1 i Sj c Γ(S j 1) i S r P c i p 1 p 2 p 3 S 0 S 1 Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

31 S 0 S 1 S 2 An Example 3-Node Line Network Revisited Constraints defining R r r b (r+1 j) i j=1 i Sj c Γ(S j 1) i S r P c i p 1 p 2 p 3 S 0 = S 1 = S 2 = {1} b b2 2 2 Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

32 An Example 3-Node Line Network Revisited Constraints defining R r r b (r+1 j) i j=1 i Sj c Γ(S j 1) i S r P c i S 0 = {1}, S 1 = S 2 = {1, 2} b S 0 = S 1 = {1}, S 2 = {1, 2} b S 0 = S 1 = S 2 = {1} b b2 2 2 Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

33 An Example 3-Node Line Network Revisited Constraints defining R r r b (r+1 j) i j=1 i Sj c Γ(S j 1) i S r P c i S 0 = {1}, S 1 = S 2 = {1, 2} b S 0 = S 1 = {1}, S 2 = {1, 2} b S 0 = S 1 = S 2 = {1} b b2 2 2 By symmetry: b All other sequences of sets give redundant constraints. Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

34 Sequences of Networks We can consider sequences of networks indexed by k. Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

35 Sequences of Networks We can consider sequences of networks indexed by k. The packet distributions are denoted {P i (k)} N i=1. Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

36 Sequences of Networks We can consider sequences of networks indexed by k. The packet distributions are denoted {P i (k)} N i=1. Naturally, we have N i=1 P i(k) = k. Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

37 Sequences of Networks We can consider sequences of networks indexed by k. The packet distributions are denoted {P i (k)} N i=1. Naturally, we have N i=1 P i(k) = k. Definition (Well-behaved sequence of packet distributions) A sequence of packet distributions is well-behaved if 1 P S lim P i (k) k k exists for all subsets S {1,..., N}. i S Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

38 Sequences of Networks We can consider sequences of networks indexed by k. The packet distributions are denoted {P i (k)} N i=1. Naturally, we have N i=1 P i(k) = k. Definition (Well-behaved sequence of packet distributions) A sequence of packet distributions is well-behaved if 1 P S lim P i (k) k k exists for all subsets S {1,..., N}. i S P S is the limit of the empirical probability that any node in S receives a particular packet. Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

39 Sequences of Networks We can consider sequences of networks indexed by k. The packet distributions are denoted {P i (k)} N i=1. Naturally, we have N i=1 P i(k) = k. Definition (Well-behaved sequence of packet distributions) A sequence of packet distributions is well-behaved if 1 P S lim P i (k) k k exists for all subsets S {1,..., N}. i S P S is the limit of the empirical probability that any node in S receives a particular packet. Let P c S = 1 P S. Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

40 Sequences of Networks We can consider sequences of networks indexed by k. The packet distributions are denoted {P i (k)} N i=1. Naturally, we have N i=1 P i(k) = k. Definition (Well-behaved sequence of packet distributions) A sequence of packet distributions is well-behaved if 1 P S lim P i (k) k k exists for all subsets S {1,..., N}. i S P S is the limit of the empirical probability that any node in S receives a particular packet. Let P c S = 1 P S. Convergence can take place in probability. Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

41 Border Nodes Characterized by simple cuts Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

42 Border Nodes Characterized by simple cuts S Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

43 Border Nodes Characterized by simple cuts (S) S Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

44 Per-round Transmission Constraints Simple Cuts Suffice Theorem For a fixed network topology, τ k/r fixed, and a well-behaved sequence of packet distributions {P i (k)} N i=1, if node i is allowed to transmit b i times per round and b i > τp c S, S [N] i (S) then universal recovery is possible for all sufficiently large k. Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

45 Per-round Transmission Constraints Simple Cuts Suffice Theorem For a fixed network topology, τ k/r fixed, and a well-behaved sequence of packet distributions {P i (k)} N i=1, if node i is allowed to transmit b i times per round and b i > τp c S, S [N] i (S) then universal recovery is possible for all sufficiently large k. Interpretation: If border nodes have the ability to relay the packets S is missing, then universal recovery is possible. Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

46 Per-round Transmission Constraints A corresponding converse Theorem Universal recovery is not possible if node i is allowed to make at most b i transmissions per communication round and there is some set S [N] for which b i < 1 Pi c r. i (S) i S Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

47 Per-round Transmission Constraints A corresponding converse Theorem Universal recovery is not possible if node i is allowed to make at most b i transmissions per communication round and there is some set S [N] for which b i < 1 Pi c r. i (S) Interpretation: The nodes in (S) form a bottleneck which is too tight. i S Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

48 Per-round Transmission Constraints A corresponding converse Theorem Universal recovery is not possible if node i is allowed to make at most b i transmissions per communication round and there is some set S [N] for which b i < 1 Pi c r. i (S) Interpretation: The nodes in (S) form a bottleneck which is too tight. Question What is the minimum number of transmissions required for universal recovery? i S Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

49 Nonsingular, d-regular, d-connected Networks Definitions and Remarks Definition (d-regular Networks) A network is d-regular if each node is adjacent to d other nodes. Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

50 Nonsingular, d-regular, d-connected Networks Definitions and Remarks Definition (d-regular Networks) A network is d-regular if each node is adjacent to d other nodes. Definition (d-connected Networks) A network is d-connected if at least d nodes must be removed in order to disconnect the network. Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

51 Nonsingular, d-regular, d-connected Networks Definitions and Remarks Definition (d-regular Networks) A network is d-regular if each node is adjacent to d other nodes. Definition (d-connected Networks) A network is d-connected if at least d nodes must be removed in order to disconnect the network. Theorem (Bollobás): Almost every d-regular network is d-connected. Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

52 Nonsingular, d-regular, d-connected Networks Definitions and Remarks Definition (d-regular Networks) A network is d-regular if each node is adjacent to d other nodes. Definition (d-connected Networks) A network is d-connected if at least d nodes must be removed in order to disconnect the network. Theorem (Bollobás): Almost every d-regular network is d-connected. Definition (Nonsingular Networks) A network is nonsingular if its adjacency matrix is nonsingular. Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

53 Nonsingular, d-regular, d-connected Networks Definitions and Remarks Definition (d-regular Networks) A network is d-regular if each node is adjacent to d other nodes. Definition (d-connected Networks) A network is d-connected if at least d nodes must be removed in order to disconnect the network. Theorem (Bollobás): Almost every d-regular network is d-connected. Definition (Nonsingular Networks) A network is nonsingular if its adjacency matrix is nonsingular. Conjecture (Costello and Vu): For d 3, almost every d-regular network is nonsingular. Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

54 Bounding Excess Transmissions by the Number of Nodes A theorem for structured networks Theorem For all nonsingular d-regular, d-connected networks with ρ > 0 fixed, and P i = ρ < P S, i [N], S : {i} S, (N S ) P c i > d P c S, S : S > N d we can explicitly construct a vector {b j i } that is within N transmissions of optimum. Moreover, for this {b j i }, we have that: 1 d Pi c (k) b j i 1 Pi c (k) + N. d i [N] i [N] j [r] i [N] Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

55 Packets Available to d Nodes Corollary For all nonsingular d-regular, d-connected networks with each packet available to at least d nodes, and P i = ρ < P S, i [N], S : {i} S, we can explicitly construct a vector {b j i } that is within N transmissions of optimum. Moreover, for this {b j i }, we have that: 1 d Pi c (k) b j i 1 Pi c (k) + N. d i [N] i [N] j [r] i [N] If each packet is originally available to d nodes, then almost all transmissions are useful. Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

56 Sparsely Distributed Packets Corollary For all nonsingular d-regular, d-connected networks with 0 < ρ < 1 d+1 fixed, and P i = ρ < P S, i [N], S : {i} S, we can explicitly construct a vector {b j i } that is within N transmissions of optimum. Moreover, for this {b j i }, we have that: 1 d Pi c (k) b j i 1 Pi c (k) + N. d i [N] i [N] j [r] i [N] If packets are sparsely distributed and each node has ρ k packets, then almost every transmission is useful. Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

57 Independent and Identically Distributed Packets Corollary For all nonsingular d-regular, d-connected networks with P S = 1 (1 q) S, S [N] 1 (1 q) N for some 0 < q < 1, we can explicitly construct a vector {b j i } that is within N transmissions of optimum. Moreover, for this {b j i }, we have that: 1 d Pi c (k) b j i 1 Pi c (k) + N. d i [N] i [N] j [r] i [N] If packets are available independently at each node with probability q, then almost every transmission is useful in recovering the set of packets available to the network. Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

58 Examples Two hypothetical scenarios N = 10, Ring Network Ten servers connected in a ring collectively have 1000 files. Each server has a unique 200-file subset. Between 4000 and 4010 file transfers are required for each server to recover all files. N = 1000, Fully Connected Network 1000 peers in a P2P network collectively have a movie consisting of 1,000,000 packets. Each peer has a random half of the movie (i.e. 500,000 randomly selected packets). Between 500,500 and 501,500 packet transmissions are required for universal recovery. Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

59 Conclusions and Extensions Main Results Defined a region R r which characterizes all transmission strategies that permit universal recovery. Proved that it suffices to consider simple cuts in networks with a per-round transmission constraint. Bounded excess transmissions by the number of nodes for many structured networks. Extensions Can include scenarios where helper nodes are present, etc. Convergence in probability can replace Well-behaved and all results hold with arbitrarily high probability as the total number of packets grows. Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19

60 Thank You! Courtade and Wesel (UCLA) Efficient Universal Recovery Allerton / 19