Bullet: High Bandwidth Data Dissemination Using an Overlay Mesh

1. Bullet: High Bandwidth Data Dissemination Using an Overlay Mesh

2. Introduction • Given a sender and a large set of receivers spread across the Internet, how can we maximize the bandwidth? • Problem domain: • Software or video distribution • Real-time multimedia streaming

3. Existing Solutions • IP multicast does not consider bandwidth when constructing its distribution tree • A promising alternative: Overlay • Attempts to mimic the multicast routing trees • Instead of high-speed routers, use programmable end hosts as interior nodes in the overlay tree

4. Existing Solutions • A tree structure is problematic • Decreasing bandwidth as moving down a tree • Any loss high up the tree will reduce the bandwidth lower down the tree • Bandwidth of a node limited by its single parent

5. A New Approach • Transmit disjoint data set to various points in the network • A node download from multiple sources rather than a single parent • Higher reliability

6. Conceptual Model Root 1 Mbps 1 Mbps 1 Mbps A B

7. Conventional Model Root 1 Mbps 1 Mbps 1 Mbps A B

8. Conventional Model Root 1 Mbps 1 Mbps 1 Mbps A B

9. Bullet Root 1 Mbps 1 Mbps 1 Mbps A B

10. Bullet Root 1 Mbps 1 Mbps 1 Mbps A B

11. Bullet Root 1 Mbps 1 Mbps 1 Mbps A B

12. Bullet Root 1 Mbps 1 Mbps 1 Mbps A B

13. Bullet Root 1 Mbps 1 Mbps 1 Mbps A B

14. Bullet Root 1 Mbps 1 Mbps 1 Mbps A B

15. Bullet Root 1 Mbps 1 Mbps 1 Mbps A B

16. Bullet Root 1 Mbps 1 Mbps 1 Mbps A B

17. Bullet Root 1 Mbps 1 Mbps 1 Mbps A B

18. Bullet Root 1 Mbps 1 Mbps 1 Mbps A B

19. Bullet Properties • TCP friendly • Must respond to congestion signals • Low control overhead • Probing resource • Locating multiple downloading sources • Decentralized and scalable • No global knowledge • Robust to failures, even high up in a tree

20. Bullet Overview • Use meshes as opposed to trees • Bandwidth independent of the underlying overlay tree • Used a 1,000-node overlay and 20,000 network topologies • Up to 2x bandwidth improvement over a bandwidth-optimized tree • Overhead of 30 Kbps

21. System Components • Split the data into packet-sized objects • Disseminate disjoint objects to clients at a rate determined by bandwidth to each client • Nodes need to locate and retrieve disjoint data from their peers • Periodically exchange summary tickets • Minimize overlapping objects from each peer

22. Illustration 1 2 3 4 5 6 7 S 1 2 3 5 1 3 4 6 2 4 5 6 A B C D E 1 2 5 1 3 4

23. Data Encoding • Multimedia: MDC encoding • Large files: erasure encoding • Tornado code • Only need to locate 5% extra packets to reconstruct the original message • Faster encoding and decoding time

24. RanSub • Distributes random subsets of participating nodes • During the collect phase, each node sends a random subset of its descendant nodes up the tree • During the distribute phase, each node sends a random subset of collected nodes down the tree

25. Informed Content Delivery Techniques • Use summary tickets • A summary ticket is an array • array[i] = hashi(working set) • Check ticket elements against a Bloom filter • It is possible to have false positives • It is possible that B will not send a packet to A even though A is missing it

26. TCP Friendly Rate Control (TFRC) • TCP halves the sending rate as soon as one packet loss is detected • Too severe • TFRC is based on loss events, or multiple dropped packets within one round-trip time

27. TCP Friendly Rate Control (TFRC) • Bullet eliminated retransmission from TFRC • Easier to recover from other sources than from the initial sender • TFRC does not aggressively seek newly available bandwidth like TCP

28. Bullet • Layers a mesh on top of an overlay tree to increase overall bandwidth

29. Finding Overlay Peers • RanSub periodically delivers subsets of uniformly random selected nodes • Via summary tickets (120 bytes per node) • The working set from each node is associated with a Bloom filter • Peer with nodes with the lowest similarity ratio

30. Recovering Data from Peers • A receiver assigns a portion of the sequence space to each of its senders, to avoid duplication among senders • A receiver periodically updates each sender with its current Bloom filter and the range of sequences covered in its Bloom filter • Less than 10% of all received packets are duplicates

31. Making Data Disjoint • Given a randomly chosen subset of peer nodes, it is about the same probability that each node has a particular data packet • A parent decides the portion of its data being sent to each child • A function of limiting and sending factors

32. Making Data Disjoint • The portion of data a child should own is proportional to • The number of its descendants • Bandwidth • If not enough bandwidth • Each child receives a completely disjoint data stream

33. Making Data Disjoint • If ample bandwidth • Each child will received the entire parent stream

34. Improving the Bullet Mesh • What can go wrong • Not enough peers • Constant changing network • Use trial senders and receivers • Bullet periodically evaluates the performance of its peers • Places the worst performing sender/receiver

35. Evaluation • Used Internet environments and ModelNet IP emulation • Deployed on the PlanetLab • Built on MACEDON • Specifies the overlay algorithms • Core logic under 1,000 lines of code

36. Evaluation • ModelNet experiments • 50 2 GHz Pentium 4’s running Linux 2.4.20 • 100 Mbps and 1 Gbps Ethernet switches • 1,000 emulated instances • 20,000 INET-generated topologies

37. Offline Bottleneck Bandwidth Tree • Given global knowledge, what is the overlay tree that will deliver the highest bandwidth to a set of overlay nodes? • Finding a tree with a maximum bottleneck • NP hard in general

38. Offline Bottleneck Bandwidth Tree • Assumptions • The path between two overlay nodes is fixed • The overlay tree uses TCP-friendly unicast connections to transfer data point-to-point • In the absence of other flows, we can estimate the throughput of a TCP-friendly flow using a steady-state formula • In the case of sharing, each flow can achieve at most of 1/nth of the total throughput

39. Bullet vs. Streaming • The maximum bottleneck tree achieves 5x bandwidth compared to random trees • Bullet outperforms the bottleneck tree by a factor up to 100%

40. Creating Disjoint Data • Without disjoint transmission of data • Bullet degrades by 25%

41. Epidemic Approaches • Bullet is 60% better than anti-entropy (gossiping) approaches • Epidemic algorithms have an excessive number of duplicate packets

42. Bullet on a Lossy Network • Bullet achieves 2x bandwidth compared to maximum bottleneck trees

43. Performance Under Failure • With failure detection/recovery disabled • 30% performance degradation with a missing child under the root • With failure detection/recovery enabled • Negligible disruption of performance

44. PlanetLab • 47 nodes for deployment • Similar results

45. Related Work • Kazza • Perpendicular downloads • Does not use erasure code • Bandwidth consuming • BitTorrent • Centralized tracker • FastReplica • Not bandwidth aware

46. Related Work • Scalable Reliable Multicast • Difficult to configure • Epidemic approaches • Do not avoid duplicates • Narada • Use overlay meshes • Bandwidth still limited by parent

47. Related Work • Overcast • More heavyweight when nodes leave a tree • Splitstream • Not bandwidth aware • CoopNet • Centralized