Chunkyspread: Multi-tree Unstructured Peer to Peer Multicast

1. Chunkyspread: Multi-treeUnstructured Peer to PeerMulticast Vidhyashankar Venkataraman (Vidhya) Paul Francis (Cornell University) John Calandrino (University of North Carolina)

2. Introduction • Increased interest in P2P live streaming in recent past • Existing multicast approaches • Swarming style (tree-less) • Tree-based

3. Swarming-based protocols • Data-driven tree-less multicast (swarming) getting popular • Neighbors send notifications about data arrivals • Nodes pull data from neighbors • Eg. Coolstreaming, Chainsaw  Simple, unstructured  Latency-overhead tradeoff • Not yet known if these protocols can have good control over heterogeneity (upload volume)

4. Tree-based solutions  Low latency and low overhead • Tree construction/repair considered complex • Eg. Splitstream (DHT, Pushdown, Anycast) [SRao] • Tree repair takes time • Requires buffering, resulting in delays • Contribution: A multi-tree protocol that • Is simple, unstructured • Gives fine-grained control over load • Has low latencies, low overhead, robust to failures [SRao] Sanjay Rao et. Al. The Impact of Heterogeneous Bandwidth Constraints on DHT-Based Multicast Protocols, IPTPS February 2005.

5. Chunkyspread – Basic Idea • Build heterogeneity-aware unstructured neighbor graph • Tree building: • Sliced data stream: one tree per slice (Splitstream) • Simple and fast loop avoidance and detection • Parent/child relationships locally negotiated to optimize criteria of interest • Load, latency, tit-for-tat, node-disjointness, etc.

6. Heterogeneity-aware neighbor graph • Neighbor graph built with simple random walks • Using “Swaplinks”, developed at Cornell [SWAP] • Degree of node in graph proportional to its desired transmit load • Notion of heterogeneity-awareness • So that higher-capacity nodes have more children in multicast trees [SWAP] V. Vishnumurthy and P. Francis. On Heterogeneous Overlay Construction and Random Node Selection in Unstructured P2P Networks. To appear in INFOCOMM, Barcelona 2006.

7. Sliced Data Stream Source selects random neighbors using Swaplinks Slice Source 2 Multicasts the slice Slice 2 Slice Source 1 Source Slice 1 Slice 3 Slice Source 3 Source sends one slice to each node - acts as slice source to a tree

8. Building trees • Initialized by flooding control message • Pick parents subject to capacity (load) constraints • Produces loop-free but bad trees • Subsequently fine-tune trees according to criteria of interest

9. Simple and fast loop avoidance/ detection • Proposed by Whitaker and Wetherall [ICARUS] • All data packets carry a bloom filter • Each node adds its mask to the filter • Small probability of false positives • Avoidance: advertise per-slice bloom filters to neighbors • Detection: by first packet that traverses the loop [ICARUS] A. Whitaker and D. Wetherall. Forwarding without loops in Icarus. In OPENARCH, 2002.

10. Parent/child selection based on load & latency Never enter this load region ML = Maximum Load Sheds children Higher Load (more children) TL+δ Children Improve Latency TL = Target Load TL-δ Lower Load (less children) Adds children 0

11. Parent/Child Switch Parent for slice k (Load>Satisfactory threshold) Potential Parents (Load<Satisfactory Threshold) 2)Gets info from all children A B 5)A says yes if still underloaded 3) Chooses A and asks child to switch • Child sends info • about A and B 4)Child requests A Child

12. Chunkyspread Evaluation • Discrete event-based simulator implemented in C++ • Run over transit-stub topologies having 5K routers • Heterogeneity : Stream split into 16 slices • TL uniformly distributed between 4 & 28 slices • ML=(1.5)TL: Enough capacity in network • Two cases • No latency improvement (δ=0) • With latency improvement: • δ=(2/16).TL (or 12.5% of TL) ML=(1.5).(TL) TL+δ TL = Target Load TL-δ 0

13. Control over load Flash crowd scenario 2.5K nodes join a 7.5K node network at 100 joins/sec Nodes within 20% of TL even with latency reduction (δ=12.5%) With latency Peak of 40 control messages per node per second with latency reduction Median of 10 messages per node per second during period of join Snap shot of system after nodes finished fine-tuning trees Trees optimized ~95s after all nodes join with latency reduction [(Load-TL)/TL]%

14. Latency Improvement Maximum latency ~ Buffer capacity without node failures With Latency Flash crowd scenario No Latency 90th percentile network stretch of 9 ~ small buffer

15. Burst Failure Recovery Time CDF of Disconnect duration with latency reduction FEC codes improve recovery time Failure Burst: 1K nodes fail in a 10K-node network at the same time instant 3 Redundancy 0 Redundant slices Neighbor failure timeout set at 4 seconds 1 Redundant slice Recovery time within a few seconds Buffering: Dominant over effects of latency Shown with various Redundancy levels

16. Conclusion • Chunkyspread is a simple multi-tree multicast protocol • A design alternative to swarming style protocols • Achieves fine-grained control over load with good latencies • Suited to non-interactive live streaming applications • Need to do apples-to-apples comparisons with swarming protocols and Splitstream