1. “Scalable and Topologically-aware Application-layer Multicast” 2004.1.29 Yusung Kim yskim@cosmos.kaist.ac.kr Korea Advanced Institute of Science and Technology 17 th APAN meetings / Jt Techs workshop Korea Advanced Institute of Science and Technology

2. Outline 1. Introduction 2. Related works 3. Problem definition 4. Model 5. Performance evaluation 6. Analysis 7. Conclusion 8. Future work Reference Korea Advanced Institute of Science and Technology

3. 1. Introduction : Logistical Networking Logistical Networking is an end-to-end approach for globally scalable network storage [Beck 02]  It is applied to large-scale distributed network storage system such as web caching, FTP mirroring, Content Distribution Network (CDN), and Data Grid etc.  A scalable and efficient one-to-many data transfer mechanism is necessary when moving data to large-scale distributed nodes on Logistical Networking Logistical Backbone Korea Advanced Institute of Science and Technology

4. 1. Introduction : Application-layer Multicast IP multicast is an efficient mechanism for multipoint data transfer, but deployment has not been widely adopted yet [Banerjee 02] Application-layer Multicast does not change the network infrastructure, instead it implements multicast forwarding functionality at end-host. Korea Advanced Institute of Science and Technology H1 H2 R1R2 H3 H4 1 2 1 1 25 H1 H2 H3 H4 3 2 27 a. Physical network topologyb. Application-layer data path H : Host R : RouterNumber : latency Total latency : 30Total latency : 32

5. 2. Related works - Application-layer Multicast approaches H1H2 H3H4 H1 H2 H3 H4 H5 1) Centralized approach [Pendarakis 01] 2) Tree first approach [Zhang 02] Using global topology information Not scalable Using partial topology information Scalable H Host : multicast participant Hx ? ? Korea Advanced Institute of Science and Technology

6. 3. Problem definition Lack of global topology information cause data-paths to include unnecessary high-latency hops, it increases the usage of network resource and delays data transfer time H1 H2 R3 R4 H3 R2R1 H4 5 5 90 5 5 190 5 H: Host, R: Router, Number: latency a. Physical network topology Topologically-aware path 100 200 10 Non topologically-aware path H1 H2 H3 H4 300 100 b. Comparison between two application-layer data paths Korea Advanced Institute of Science and Technology

7. 4. Model H Host : multicast participant H1 L(1,2,3) H4 H2 H3 H5 H6 H7 H9 H8 H10 H11 L(1,3,2) L(2,1,3) L(2,3,1) L(3,2,1) L(3,1,2) L ( landmark 1, landmark 2, landmark 3 ) : order of near landmarks Adding landmark scheme to tree-first approach for the scalable application-layer multicast, construct topologically-aware data paths Source landmark 1 landmark 3 landmark 2 Korea Advanced Institute of Science and Technology Seoul Univ. KAIST Tokyo Univ. KOREA JAPAN

8. 5. Performance evaluation 5.1 Topology generation methodology Methodology Logistical Backbone experiment [Lbone] Inet : topology generator [Inet] Number of nodes 170 10,000 Network cost Latency : 10 KB transfer time instead of RTT [Jannotti 00] Latency : weight ( allocated bye Inet ) Number of landmarks 6 20 Number of receivers 5 ~ 155 10 ~ 8,000 Korea Advanced Institute of Science and Technology

9. 5. Performance evaluation 5.1 Topology generation methodology 5.2 Result I : link latency 5.3 Result II : path stretch 5.4 Reuslt III : number of control message * Stretch (relative delay penalty) [Chu 00] : the ratio of the delay from the source to the member along the application-layer data path, to the delay of the direct unicast path Control messages (control overhead) [Banerjee 02] : messages exchanged between all nodes to construct data path Korea Advanced Institute of Science and Technology

10. 5.2 Result I : link latency - on Logistical Backbone experiment Comparison on average link latency ( with 6 landmarks ) less average link latency than that of tree first approach Unicast Tree first approach Centralized approach Landmark based approach Korea Advanced Institute of Science and Technology

11. 5.2 Result I : link latency - on Inet simulation Comparison on average link latency (with 20 landmarks) IP multicast less average link latency than that of tree first approach Unicast Tree first approach Centralized approach Landmark based approach Korea Advanced Institute of Science and Technology

12. 5.3 Result II : path stretch - on Inet simulation Tree first approach Centralized approach Landmark based approach Comparison on average path streth (with 20 landmarks) significantly reduced average path stretch Korea Advanced Institute of Science and Technology

13. 5.4 Result III : number of control messages - on Inet simulation Comparison on number of control messages (with 20 landmarks) Case of receivers above 30, the number of control messages of landmark based approach is similar to that of tree first approach Tree first approach Centralized approach Landmark based approach Korea Advanced Institute of Science and Technology

14. 6. Analysis Centralized approach [Pendarakis 01] Tree-first approach [Zhang 02] Landmark based approach Data path construction Mesh based measurement Tree based measurement with landmark Scalability O (n²) at one node O (logn) at each node O (logn) at each node Topology awareness Global awarenessPartial awareness Limited global awareness Additional infrastructure None Setting constant number of landmarks Korea Advanced Institute of Science and Technology

15. 7. Conclusion + Contribution 1) We applied landmark scheme to Tree first approach 2) In results of performance evaluation, Landmark based approach can reduce the average link latency and path stretch of Tree first approach 3) Landmark based approach needs the constant number of landmarks and low number of control messages as much as Tree first approach does => Landmark based approach can offer the scalability of tree first approach and construct the topologically-aware data paths using landmarks. Using topologically-aware paths, we can reduce bandwidth consumption and data transfer time + Limitation => An additional landmark infrastructure is necessary Korea Advanced Institute of Science and Technology

16. 8. Future work Existing researches for application-layer multicast considered network latency to construct data paths Bandwidth-awareness is also important to construct data paths Korea Advanced Institute of Science and Technology H1 H3 1 ms 100 Kbps H2 100 ms 1 Gbps H Host

17. Reference [Akamai] [Banerjee 02] [Beck 02] [Faloutsos 99] [Inet] [Jannotti 00] [Lbone] [Pendarakis 01] [Ratnasamy 02] [Zhang 02] Akamai, http://www.akamai.com (Accessed: 8 December 2003). S. Banerjee, B. Bhattacharjee, and C. Kommareddy, "Scalable application layer multicast," in Proc. ACM SIGCOMM, Pittsburgh, PA, USA, August 2002. M. Beck, T. Moore, and J. Plank, "An end-to-end approach to globally scalable network storage,“ in Proc. ACM SIGCOMM, Pittsburgh, PA, USA, August 2002. M. Faloutsos, P. Faloutsos, and C. Faloutsos, “On power-law relationships of the Internet topology,” in Pro. ACM SIGCOMM, Cambridge, MA, USA, September 1999. Inet, http://topology.eecs.umich.edu (Accessed: 8 December 2003). J. Jannotti, D. K. Gifford, K. L. Johnson, M. F. Kaasheok, and J. W. O’Toole, “Overcast: reliable multicasting with an overlay network,” in Proc. 4th USENIX OSDI, San Diego, CA, USA, October 2000. Lbone, http://loci.cs.utk.edu (Accessed: 9 December 2003). D. Pendarakis, S. Shi, D. Verma, and M. Waldvogel, “ALMI: an application level multicast infrastructure,” in Proc. 3rd USENIX Symp. Internet Tech. and Sys., San Francisco, CA, USA, March 2001. S. Ratnasamy, M. Handley, Richard Karp, and S Shenker, “Topologically-aware overlay construction and server selection,” in Proc. IEEE INFOCOM, New York, NY, USA, June 2002. B. Zhang, S. Jamin, and L. Zhang, “Host multicast: a framework for delivering multicast to end users,” in Proc. IEEE INFOCOM, New York, NY, USA, June 2002.

18. Appendix : number of control messages - on Logistical Backbone experiment Tree first approach Centralized approach Landmark based approach

19. Appendix : link stress * stress: defines the stress of a physical link as the number of identical packets it carries landmark based tree-first centralized IP multicast

20. Appendix: path stretch a. average path stretch on Logistical Backbone b. average path stretch on Inet Tree first approach Centralized approach Landmark based approach Tree first approach Centralized approach Landmark based approach

21. Appendix: metric [Chu 00] Stress: defines the stress of a physical link as the number of identical packets it carries Stretch: is the ratio of the delay between a source and a member along the overlay distribution topology, to the delay of the direct unicast path Resource usage: defines this metric as the sum of the delay * stress product over all the links that participate in data transmissions