1. Digital Fountains, and Their Application to Informed Content Delivery over Adaptive Overlay Networks Michael Mitzenmacher Harvard University

2. The Talk Survey of the area –My work, and work of others –History, perspective –Less on theoretical details, more on big ideas Start with digital fountains –What they are –How they work –Simple applications Content delivery –Digital fountains, and other tools

3. Data in the TCP/IP World Data is an ordered sequence of bytes –Generally split into packets Typical download transaction: –“I need the file: packets 1-100,000.” –Sender sends packets in order (windows) –“Packet 75 is missing, please re-send.” Clean semantics –File is stored this way –Reliability is easy –Works for point-to-point downloads

4. Problem Case: Multicast One sender, many downloaders –Midnight madness problem – new software –Video-on-demand (not real time) Can download to each individual separately –Doesn’t scale Can “broadcast” –All users must start at the same time? –Heterogeneous packet loss –Heterogeneous download rates

5. Digital Fountain Paradigm Stop thinking of data as an ordered stream of bytes. Data is like water from a fountain –Put out your cup, stop when the cup is full. –You don’t care which drops of water you get. –You don’t care what order the drops get to your cup.

6. What is a Digital Fountain? For this talk, a digital fountain is an ideal/paradigm for data transmission. –Vs. the standard (TCP) paradigm: data is an ordered finite sequence of bytes. Instead, with a digital fountain, a k symbol file yields an infinite data stream; once you have received any k symbols from this stream, you can quickly reconstruct the original file.

7. Digital Fountains for Multicast Packets sent from a single source along a tree. Everyone grabs what they can. –Starting time does not matter – start whenever. –Packet loss does not matter – avoids feedback explosion of lost packets. –Heterogeneous download rates do not matter – drop packets at routers as needed for proper rate. When a user has filled their cup, they leave the multicast session.

8. Digital Fountains for Parallel Downloads Download from multiple sources simultaneously and seamlessly. –All sources fill the cup – since each fountain has an “infinite” collection of packets, no duplicates. –Relative fountain speeds unimportant; just need to get enough. –No coordination among sources necessary. Combine multicast and parallel downloading. –Wireless networks, multiple stations and antennas.

9. Digital Fountains for Point-to-Point Data Transmission TCP has problems over long-distance connections. –Packets must be acknowledged to increase sending window (packets in flight). –Long round-trip time leads to slow acks, bounding transmission window. –Any loss increases the problem. Using digital fountain + TCP-friendly congestion control can greatly speed up connections. Separates the “what you send” from “how much” you send. –Do not need to buffer for retransmission.

10. One-to-Many TCP Setting: Web server with popular files, may have many open connections serving same file. –Problem: has to have a separate buffer, state for each connection to handle retransmissions. –Limits number of connections per server. Instead, use a digital fountain to generate packets useful for all connections for that file. Separates the “what you send” from “how much” you send. –Do not need to buffer for retransmission. Keeps TCP semantics, congestion control.

11. Digital fountains seem great! But do they really exist?

12. How Do We Build a Digital Fountain? We can construct (approximate) digital fountains using erasure codes. –Including Reed-Solomon, Tornado, LT, fountain codes. Generally, we only come close to the ideal of the paradigm. –Streams not truly infinite; encoding or decoding times; coding overhead.

13. Digital Fountains through Erasure Codes Message Encoding Received Message Encoding Algorithm Decoding Algorithm Transmission n cn n n

14. Reed-Solomon Codes In theory, can produce an unlimited number of encoding symbols, only need k to recover. In practice, limited by: –Field size (usually 256 or 65,536) –Quadratic encoding/decoding times These problems ameliorated by striping data. –But raises overhead; now many more than k packets required to recover. Conclusion: may be suitable for some applications, but far from practical or theoretical goals of a digital fountain.

15. Tornado Codes Irregular low-density parity check codes. Based on graphs: k input symbols lead to n encoding symbols, using XORs. –Sparse set of equations derived from input symbols. –Solve received set of equations using back substitution. Properties: –Graph of size n agreed on by encoder, decoder, and stored. –Need k(1+  ) symbols to decode, for some  > 0. –Encoding/decoding time proportional to n ln (1/  ).

16. Tornado Codes An Example

17. Encoding Process

18. Decoding Process: Substitution Recovery indicates right node has one edge

19. Regular Graphs Random Permutation of the Edges Degree 3 Degree 6

20. Decoding Process Analysis Recovered Missing/not yet recovered = = Induced Graph

21. 3-6 Regular Graph Analysis LeftRightLeft

22. 3-6 Regular Graph Equation Want: y < x for all 0 < x <  Works for  < 0.43

23. Irregular Graphs 3-6 regular graphs can correct up to 0.43 fraction of erasures. Best possible, with n/2 constraints for n symbols, would be 0.5. 3-6 gives best performance of all regular graphs. Need irregular graphs, with varying degrees, to reach optimality.

24. Tornado Codes: Weaknesses Encoding size n must be fixed ahead of time. Memory, encoding and decoding times proportional to n, not k. Overhead factor of (1+  ). –Hard to design around. In practice  = 0.05. Conclusion: Tornado codes a dramatic step forward, allowing good approximations to digital fountains for many applications. Key problem: fixed encoding size.

25. Decoding Process: Direct Recovery

26. Digital Fountains through Erasure Codes : Problem Message Encoding Received Message Encoding Algorithm Decoding Algorithm Transmission n cn n n

27. Digital Fountains through Erasure Codes : Solution Message Encoding Received Message Encoding Algorithm Decoding Algorithm Transmission n n n

28. LT Codes Key idea: graph is implicit, rather than explicit. –Each encoding symbol is the XOR of a random subset of neighbors, independent of other symbols. –Each encoding symbol carries a small header, telling what message symbols it is the XOR of. No initial graph; graph derived from received symbols. Properties: –“Infinite” supply of packets possible. –Need k +o(k) symbols to decode. –Decoding time proportional to k ln k. –On average, ln k time to produce an encoding symbol.

29. LT Codes Conclusion: making the graph implicit gives us an almost ideal digital fountain. One remaining issue: why does average degree need to be around ln k? –Standard coupon collector’s problem: for each message symbol to be hit by some equation, need k ln k variables in the equations. Can remove this problem by pre-coding.

30. Rateless/Raptor Codes Pre-coding independently described by Shokrollahi, Maymoukov. Rough idea: –Expand original k message symbols to k (1+  ) symbols using (for example) a Tornado code. –Now use an LT code on the expanded message. –Don’t need to recover all of the expanded message symbols, just enough to recover original message.

31. Raptor/Rateless Codes Properties: –“Infinite” supply of packets possible. –Need k(1+  ) symbols to decode, for some  > 0. –Decoding time proportional to k ln (1/  ). –On average, ln (1/  ) (constant) time to produce an encoding symbol. –Very efficient. Raptor codes give, in practice, a digital fountain.

32. Impact on Coding These codes are examples of low-density parity-check (LDPC codes). Subsequent work: designed LDPC codes for error-correction using these techniques. Recent developments: LDPC codes approaching Shannon capacity for most basic channels.

33. Putting Digital Fountains To Use Digital fountains are out there. –Digital Fountain, Inc. sells them. Limitations to their use: –Patent issues. –Perceived complexity. Lack of reference implementation. –What is the killer app?

34. Patent Issues Several patents / patents pending on irregular LDPC codes, LT codes, Raptor codes by Digital Fountain, Inc. Supposition: this stifles external innovation. –Potential threat of being sued. –Potential lack of commercial outlet for research. Suggestion: unpatented alternatives that lead to good approximations of a digital fountain would be useful. –There is work going on in this area, but more is needed to keep up with recent developments in rateless codes.

35. Perceived Complexity Digital fountains are now not that hard… …but networking people do not want to deal with developing codes. A research need: –A publicly available, easy to use, reasonably good black box digital fountain implementation that can be plugged in to research prototypes. Issue: patents. –Legal risk suggests such a black box would need to be based on unpatented codes.

36. What’s the Killer App? Multicast was supposed to be the killer app. –But IP multicast was/is a disaster. –Distribution now handled by contend distributions companies, e.g. Akamai. Possibilities: –Overlay multicast. –Big wireless: e.g. automobiles, satellites. –Others???

37. Conclusions, Part I Stop thinking of data as an ordered stream of bytes. Think of data as a digital fountain. Digital fountains are implementable in practice with erasure codes.

38. A Short Breather We’ve covered digital fountains. Next up: –Digital fountains for overlay networks. –And other tricks! Pause for questions, 30 second stretch.

39. Overlays for Content Delivery A substitute for IP multicast. Build distribution topology out of unicast connections (tunnels). Requires active participation of end-systems. Native IP multicast unnecessary. Saves considerable bandwidth over N * unicast solution. Basic paradigm easy to build and deploy. Bonus: Overlay topology can adapt to network conditions by self-reconfiguration. SOURCE

40. Limitations of Existing Schemes Tree-like topologies. –Rooted in history (IP Multicast). –Limitations: bandwidth decreases monotonically from the source. losses increase monotonically along a path. Does this matter in practice? –Anecdotal and experimental evidence says yes: Downloads from multiple mirror sites in parallel. Availability of better routes. Peer-to-peer: Morpheus, Kazaa and Grokster.

41. An Illustrative Example 1. A basic tree topology. 1 2. Harnessing the power of parallel downloads. 2 3. Incorporating collaborative transfers. 3

42. Our Philosophy Go beyond trees. –Use additional links and bandwidth by: downloading from multiple peers in parallel taking advantage of “perpendicular” bandwidth –Has potential to significantly speed up downloads… But only effective if: –collaboration is carefully orchestrated –methods are amenable to frequent adaptation of the overlay topology

43. Suitable Applications Prerequisite conditions: –Available bandwidth between peers. –Differences in content received by peers. –Rich overlay topology. Applications –Downloads of large, popular files. –Video-on-demand or nearly real-time streams. –Shared virtual environments.

44. Use Digital Fountains! Intrinsic resilience to packet loss, reordering. Better support for transient connections via stateless migration, suspension. Peers with full content can always generate useful symbols. Peers with partial content are more likely to have content to share. ButBut using a digital fountain comes at a price: –Content is no longer an ordered stream. –Therefore, collaboration is more difficult.

45. Informed Content Delivery: Definitions and Problem Statement Peers A and B have working sets of symbols S A, S B drawn from a large universe U and want to collaborate effectively. Key components: 1) 1) Summarize: Furnish a concise and useful sample of a working set to a peer. 2) 2) Approximately Reconcile: Compute as many elements in S A - S B as possible and transmit them. Do so with minimal control messaging overhead.

46. Summarization Goal: each peer has a 1 packet calling card. –Can be used to estimate |S A – S B |. One possibility: random sampling. –B sends A a random sample of k elements of S B. –Each element is in S A with probability –Negative: must search S A. –Negative: hard to work with multiple summaries. Alternative: min-wise independent sampling.

47. Min-Wise Summaries Let U be the set of 64 bit numbers, and  be a random permutation on U. Then Calling card for A: keep vector of k values min  j (A), j=1…k. To estimate, count the j for which min  j (A) = min  j (B), divide by k.

48. Min-Wise Summaries : Example

49. Recoding: An Intermediate Solution Problem: What to transmit when peers have similar content? Solution: Allow peers to probabilistically “hedge their bets,” minimizing chance of transmission of useless content. Example: S A Suppose the resemblance between S A and S B is 0.9. If A sends a symbol at random the probability of it being useful to B is 0.1. A better strategy is to XOR 10 random symbols together. B can extract one useful symbol with probability: 10 x ( 1 / 10 ) x ( 9 / 10 ) 9 > 1/e  0.37

50. Approximate Reconciliation Suppose summarization suggests collaboration is worthwhile. Goal: compute as many elements in S A - S B as possible, with low communication. Idea: we do not need all of S A - S B, just as much as possible. –Use Bloom filters.

51. Lookup Problem Given a set S A = {x 1,x 2,x 3,…x n } on a universe U, want to answer queries of the form: Bloom filter provides an answer in –“Constant” time (time to hash). –Small amount of space. –But with some probability of being wrong.

52. Bloom Filters Start with an m bit array, filled with 0s. Hash each item x j in S k times. If H i (x j ) = a, set B[a] = 1. 0000000000000000 B 0100101001110110 B To check if y is in S, check B at H i (y). All k values must be 1. 0100101001110110 B 0100101001110110 B Possible to have a false positive; all k values are 1, but y is not in S.

53. Errors Assumption: We have good hash functions, look random. Given m bits for filter and n elements, choose number k of hash functions to minimize false positives: –Let As k increases, more chances to find a 0, but more 1’s in the array. Find optimal at k = (ln 2)m/n by calculus.

54. Example m/n = 8 Opt k = 8 ln 2 = 5.45...

55. Bloom Filters for Reconciliation B transmits a Bloom filter of its set to A; A then sends packets from the set difference. –All elements will be in difference: no false negatives. –Not all element in difference found: false pos. Improvements –Compressed Bloom filters –Approximate Reconciliation Trees

56. Experimental Scenarios Three methods for collaboration Uninformed – Uninformed: A transmits symbols at random to B. Speculative – Speculative: B transmits a minwise summary to A; A then sends recoded symbols to B. Reconciled – Reconciled: B transmits a Bloom filter of its set to A; A then sends packets from the set difference. Overhead: –Decoding overhead: with erasure codes, fixed 2.5%. –Reception overhead: useless duplicate packets. –Recoding overhead: useless recoding packets. symbols received - symbols needed symbols needed

57. Pairwise Reconciliation Containment of B in A: |S A  S B | |S B | 128MB file 96K input symbols 115K distinct symbols in system initially

58. Four peers in parallel 128MB file 96K input symbols 105K distinct symbols in system initially Containment of B in A: |S A  S B | |S B |

59. Four peers, periodic updates 128MB file 96K input symbols 105K distinct symbols in system initially Filters updated at every 10%. Containment of B in A: |S A  S B | |S B |

60. Subsequent Work Maymounkov: each source sends a stream of consecutive encoded packets. –Possibly simplifies collaboration, with loss of flexibility. Bullet (SOSP ’03): –An implementation with our ideas, plus purposeful distribution of different content. Network coding –Nodes inside the network can compute on the input, rather than just the endpoints. –Potentially more powerful paradigm –Practice?

61. Conclusions whatEven with ultimate routing topology optimization, the choice of what to send is paramount to content delivery. Digital fountain model ideal for fluid and ephemeral network environments. Collaborations with coded content worthwhile. Richly connected topologies are key to harnessing perpendicular bandwidth. Wanted: more algorithms for intelligent collaboration.

62. Why regular graphs are bad d Left node has on average neighbors of degree one. Right degree 2d implies Pr[right degree =1] =

63. Irregular Graphs Random Permutation of the Edges Degree 2 Degree 3 Degree 4 Degree 1 Degree 4 Degree 5 Degree 6 Degree 10

64. Degree Sequence Functions Left Side – fraction of edges of degree i on the left in the original graph. Right Side – fraction of edges of degree i on the right in the original graph.

65. Irregular Graph Analysis LeftRightLeft

66. Irregular Graph Condition Want: y < x for all 0 < x < 

67. Good Left Degree Sequence: Truncated Heavy Tail D = 9, N = Fraction of nodes of degree i is Average node degree is

68. Good Right Degree Sequence: Poisson Average node degree is

69. Good Degree Sequence Functions Want: y < x for all 0 < x <  Works for

70. Tornado Code Performance Time overhead (Average left degree) Reception Efficiency

71. Why irregular graphs are good D+1 Left node of max degree has on average one neighbor of degree one. Average right degree 2ln(D) implies Pr[right degree =1]  1/(D+1)

72. Digital Fountains: A Survey and Look Forward Michael Mitzenmacher

73. Goals of the Talk Explain the digital fountain paradigm for network communication. Examine related advances in coding. Summarize work on applications. Speculate on what comes next.

74. How Do We Build a Digital Fountain? We can construct (approximate) digital fountains using erasure codes. –Including Reed-Solomon, Tornado, LT, fountain codes. Generally, we only come close to the ideal of the paradigm. –Streams not truly infinite; encoding or decoding times; coding overhead.

75. Reed-Solomon Codes In theory, can produce an unlimited number of encoding symbols, only need k to recover. In practice, limited by: –Field size (usually 256 or 65,536) –Quadratic encoding/decoding times These problems ameliorated by striping data. –But raises overhead; now many more than k packets required to recover. Conclusion: may be suitable for some applications, but far from practical or theoretical goals of a digital fountain.

76. Tornado Codes Irregular low-density parity check codes. Based on graphs: k input symbols lead to n encoding symbols, using XORs. –Sparse set of equations derived from input symbols. –Solve received set of equations using back substitution. Properties: –Graph of size n agreed on by encoder, decoder, and stored. –Need k(1+  ) symbols to decode, for some  > 0. –Encoding/decoding time proportional to n ln (1/  ).

77. Tornado Codes: Weaknesses Encoding size n must be fixed ahead of time. Memory, encoding and decoding times proportional to n, not k. Overhead factor of (1+  ). –Hard to design around. In practice  = 0.05. Conclusion: Tornado codes a dramatic step forward, allowing good approximations to digital fountains for many applications. Key problem: fixed encoding size.

78. LT Codes Key idea: graph is implicit, rather than explicit. –Each encoding symbol is the XOR of a random subset of neighbors, independent of other symbols. –Each encoding symbols carries a small header, telling what message symbols it is the XOR of. No initial graph; graph derived from received symbols. Properties: –“Infinite” supply of packets possible. –Need k +o(k) symbols to decode. –Decoding time proportional to k ln k. –On average, ln k time to produce an encoding symbol.

79. LT Codes Conclusion: making the graph implicit gives us an almost ideal digital fountain. One remaining issue: why does average degree need to be around ln k? –Standard coupon collector’s problem: for each message symbol to be hit by some equation, need k ln k variables in the equations. Can remove this problem by pre-coding.

80. Rateless/Raptor Codes Pre-coding independently described by Shokrollahi, Maymoukov. Rough idea: –Expand original k message symbols to k (1+  ) symbols using (for example) a Tornado code. –Now use an LT code on the expanded message. –Don’t need to recover all of the expanded message symbols, just enough to recover original message.

81. Raptor/Rateless Codes Properties: –“Infinite” supply of packets possible. –Need k(1+  ) symbols to decode, for some  > 0. –Decoding time proportional to k ln (1/  ). –On average, ln (1/  ) (constant) time to produce an encoding symbol. Conclusion: these codes can be made very efficient and deliver on the promise of the digital fountain paradigm.

82. Applications Long-distance transmission (avoiding TCP) Reliable multicast Parallel downloads One-to-many TCP Content distribution on overlay networks Streaming video

83. Point-to-Point Data Transmission TCP has problems over long-distance connections. –Packets must be acknowledged to increase sending window (packets in flight). –Long round-trip time leads to slow acks, bounding transmission window. –Any loss increases the problem. Using digital fountain + TCP-friendly congestion control can greatly speed up connections. Separates the “what you send” from “how much” you send. –Do not need to buffer for retransmission.

84. Reliable Multicast Many potential problems when multicasting to large audience. –Feedback explosion of lost packets. –Start time heterogeneity. –Loss/bandwidth heterogeneity. A digital fountain solves these problems. –Each user gets what they can, and stops when they have enough.

85. Downloading in Parallel Can collect data from multiple digital fountains for the same source seamlessly. –Since each fountain has an “infinite” collection of packets, no duplicates. –Relative fountain speeds unimportant; just need to get enough. –Combined multicast/multigather possible.

86. One-to-Many TCP Setting: Web server with popular files, may have many open connections serving same file. –Problem: has to have a separate buffer, state for each connection to handle retransmissions. –Limits number of connections per server. Instead, use a digital fountain to generate packets useful for all connections for that file. Separates the “what you send” from “how much” you send. –Do not need to buffer for retransmission. Keeps TCP semantics, congestion control.

87. Distribution on Overlay Networks Encoded data make sense for overlay networks. –Changing, heterogeneous network conditions. –Allows multicast. –Allows downloading from multiple sources, as well as peers. Problem: peers may be getting same encoded packets as you, via the multicast. –Not standard digital fountain paradigm. Requires reconciliation techniques to find peers with useful packets.

88. Video Streaming For “near-real-time” video: –Latency issue. Solution: break into smaller blocks, and encode over these blocks. –Equal-size blocks. –Blocks increases in size geometrically, for only logarithmically many blocks. Engineering to get right latency, ensure blocks arrive on time for display.