Informed Content Delivery Across Adaptive Overlay Networks John Byers Dept. of Computer Science, Boston University Joint work with Jeffrey Considine, Michael Mitzenmacher and Stanislav Rost

Overlays for Content Delivery 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. adapt Bonus: Overlay topology can adapt to network conditions by self-reconfiguration. SOURCE

Killer apps:  Millions of users want to download a new movie watch the SIGCOMM technical sessions.  CDNs want to populate thousands of servers with new movies for those users. Research directions to date:  Considerable effort on optimizing overlay layout (Narada, Overcast, RON, etc.).  Scalable solutions for indexing/locating content using overlays (CAN, Chord, etc.). Our focus:  Maximize throughput of large transfers across overlays. Use of Overlays

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? yes  Anecdotal and experimental evidence says yes: Downloads from multiple mirror sites in parallel [BLM ’99, RKB ’00] Availability of better routes [SCHSA ’99, ABKM ’01]. Peer-to-peer: Morpheus, Kazaa and Grokster.

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

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

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.

Erasure Codes I need packets 1-1,000.We typically think of data as an ordered stream. I need packets 1-1,000. Using erasure codes, data is like water:  Can generate a pool of redundant data from full original content.  You don’t care what droplets you get.  You don’t care if some spills. I need any 1,000 packets.  You just want enough to get through the pipe. I need any 1,000 packets. The digital fountain model [BLMR ’98] is ideal for use in a fluid overlay environment.

Erasure Codes Offer Freedom 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 erasure codes comes at a price:  Content is no longer an ordered stream.  Therefore, collaboration is more difficult.

Informed Content Delivery: Definitions and Problem Statement working sets S A S B UPeers 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) Summarize 1) Summarize: Furnish a concise and useful sample of a working set to a peer. 2) Approximately Reconcile S A - S B 2) Approximately Reconcile: Compute as many elements in S A - S B as possible and transmit them. Do so with minimal control messaging overhead.

Min-Wise Summaries Problem: Neighboring peers may have similar content. Solution: Give peers a “calling card” (fits in 1 packet) to summarize the content they have, check similarity.

Recoding 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 S B A B 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 B can extract one useful symbol with probability: 10 x ( 1 / 10 ) x ( 9 / 10 ) 9 > 1/e  0.37

Approximate Reconciliation Trees Problem: Collaborating peers have overlapping content. Solution: Efficient data structures for reconciliation.

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 digest 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

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

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 |

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

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. Richly connected topologies are key to harnessing perpendicular bandwidth. Wanted: more algorithms for intelligent collaboration.