1. Paper Group: 20 Overlay Networks 2 nd March, 2004 Above papers are original works of respective authors, referenced here for academic purposes only Chetan Hiremath CSE525 – Advanced Networking, Winter 2004 Oregon Graduate Institute Papers Discussed: - A Routing Underlay for Overlay networks - Topologically-Aware Overlay Construction and Server Selection - Routing in Overlay Multicast Networks

2. Paper Group Objectives Design overlay services for making informed application specific routing decisions Propose a binning scheme for network overlay construction and server selection Introduce routing algorithms for overlay multicast networks & study their performance.

3. Routing Overlay Networks Used to deploy network services that cannot leverage underlying Internet directly. –File sharing, CDN, Object Location, QoS overlays, etc. Overlays often use “traceroute” and “ping” for network probing –Experiments have limited scalability up to 50 nodes –Architecture not scalable Introduce Routing Underlay –Sits between overlay networks & underlying Internet

4. Routing Underlay Service Overlay Networks Library of Routing Services Topology Probing Kernel Raw Topology Information Non scalability of ping and traceroute to “encourage” designers to adopt this approach Goals: –Provide a graph of known network connectivity at specific resolution –Expose actual route taken by packet from SRC to DST –Report topological facts about specific paths between pair of points Topology Probing Kernel exposed primitives –GetGraph(), GetPath(src, dst), GetDistance(target, metric) Library of routing services –DisjointPaths(..), NearestNodes(...), BuildMesh(..)

5. Performance and Observations BGP routers need to export routing tables to overlay networks Transit AS force usage of latency probes Discourage pushing any dynamic capability into BGP routers Finding k smallest-latency neighbors

6. Topologically-Aware Overlay Construction Propose a practical and scalable method for gathering topological information –Non scalability of traceroute and ping –Scalability & Practicality more important than accuracy Applications do not require exact topological information, but need sufficient hints on relative position of Internet hosts –Tie Overlay construction with underlying Internet topology Latency is direct indicator of performance seen by end nodes; can be measured in light weight non intrusive manner

7. Distributed Binning Set of nodes independently partition into disjoint “bin” –Nodes within a single bin are relatively closer to one another than to nodes not in their bin Small set of Landmark machines geographically distributed over the Internet to “measure” latency Check average inter-bin and intra-bin latencies to ensure binning does the job

8. Distributed Binning Example

9. TS-10K and TA-1K: Transit-sub topologies with 10,000 and 1000 nodes PLRG1 and PLRG2: Power-Law Random Graphs with 1166 and 1779 nodes NLANR: National Lab for Applied Network Research based Active Measurement Project –Consisting of 100 active monitors that exchange information

10. Distributed Binning Example TS-10K and TA-1K: Transit-sub topologies with 10,000 and 1000 nodes PLRG1 and PLRG2: Power-Law Random Graphs with 1166 and 1779 nodes NLANR: National Lab for Applied Network Research based Active Measurement Project –Consisting of 100 active monitors that exchange information

11. Binning based Server Selection If there exists one or more servers within the same bin as the client, then the client is redirected to a random server from its own bin If no server exists within the same bin as the client, an existing server from another similar bin

12. Latency Stretch Comparison

13. Routing in Overlay Multicast Networks Set of distributed Multicast Service Nodes (MSN), communicating with hosts or with each other over standard unicast mechanisms –Optimization of interface BW is primary focus Balanced Compact Tree (BCT) Algorithm –New node is always attached to the tree at the point that yields smallest diameter in the resulting tree Iterative Closest Pair (ICP) Algorithm –Select eligible pairs, for which connecting edge has minimum cost Iterative Compact Component (ICC) Algorithm –Select eligible pairs to minimize diameter of resulting component Iterative Compact Tree (ICT) Algorithm –Select eligible pairs, with one vertex of each pair in a single tree being constructed, other to minimize tree diameter

14. ICT Example Geo distance as routing cost and diameter bound of 8000 Km BDA o/p creates star topology with NY of degree 7 In second round, degree allocation is loosened by 1, resulting in smaller diameter tree satisfying diameter bound

15. Algorithm Comparison Geo distance as routing cost and diameter bound of 8000 Km ICP & ICC benefit in initial rounds of degree adjustment; allowing nearby nodes to be joined together ICT utilizes increased degree allocation at centrally located nodes and form smaller diameter trees –ICT, combined with degree loosening procedure, is more effective at producing smaller diameter trees Rejection: Session rejected if required MSN interface BW exceeds total unused BW at all MSNs

16. Simulation setup

17. Performance Results

18. Paper Group Objectives Design overlay services for making informed application specific routing decisions –Adoption is a concern Propose a binning scheme for network overlay construction and server selection –Binning performs similar to Hotz model; not much improvement Introduce routing algorithms for overlay multicast networks & study their performance. Better than above mechanisms

19. Questions ?