1. Distributed Session Announcement Agents for Real-time Streaming Applications Keio University, Graduate School of Media and Governance Kazuhiro Mishima This work is partly supported by National Institute of Information and Communications Technology (NICT), Japan.

2. Background Development of Multimedia contents –Demand of increasing multimedia contents –High-quality streaming architecture Session Information –Contains… Media Locater (Streaming Source Information) Meta-level Information (Title, Description etc…) –Receiver MUST obtain the information about the specified multimedia content (Require some announcement method)

3. Requirements of Session Announcement Scalability –System MUST manage a large number of session information Dynamicity –Streaming sessions change frequently –System MUST announce information fast Policy Definition –Session Information MUST be limited to specified receivers by sender (Secret or Private contents)

4. Purpose of our work Building a new Session Announcement or Distribution Architecture as an alternative to present announcement system for the future media streaming services

5. Related Works SAP (Announcement with ASM) Pros: Broadly used on multicast environment Pros: Low latency (UDP multicast) Cons: No reliability Cons: Scalability issue (SAP DoS/SAP Storm) Cons: No access control Search Engine Pros: Scalability using server distribution Pros: Contents information aggregation Cons: Latency issue (occurred by robot crawling) Cons: Central management (no scope or user control)

7. System Operation Overview

8. Architecture Tree-based System –Multiple Agent –Agent Tree Transfer the info. through the tree structure DHT-based System –Multiple Agent –Each agents create the DHT global storage –DHT Global Cloud All information are stored into this cloud

9. Tree-based system: Registration WebUI LocalDB Agent Registration Store All information are stored into LocalDB by agent Agent stores information and transfer to the neighbor agents Store Hop-by-Hop Transfer

10. Tree-based system: Retrieve WebUI LocalDB Agent Request Retrieve Search Use only LocalDB to retrieve the information

11. DHT-based system: Registration WebUI LocalDB DHT Engine Registration Store Replicate selected by DHT algorithm for local session information (if specify “Local”) for global session information (if specify “Global”)

12. DHT-based system: Retrieve WebUI LocalDB DHT Engine Request Search Retrieve from selected node Search results are combined LocalDB’s and DHT’s data

13. Policy-aware Information Distribution To fulfill the “Policy Definition” requirement, we propose a new Information Distribution technique –Tree-based system At the Hop-by-Hop transfer time, each agent confirm the agent has “Group Label”. If the agent has no specified “Group Label”, transferred information will discard. –DHT-based system Next slide… Using this method, secret or private information are distributed only specified agent

14. Policy-aware Information Distribution DHT-based system –Information Filtering method using the DHT’s characteristic (Exact Match) –DHT key: Secret! ff02:9a8c:aacd::3 –DHT value: session_title ff02:9a8c:aacd::3 2001:200::13ca 2008-10-03... By putting the secret value on top of the DHT key, users who know the secret value, only can retrieve the information

15. Policy-aware Information Distribution

16. Implementation gSOAP XMLRPC-C MySQL –for storing session and management data DHT –OpenDHT Algorithm: Pastry

17. Evaluation Perform on PlanetLab Tree-based system –51 nodes on US site 6 root 45 neighbor –All sites are statically connected DHT-based system –Using OpenDHT –About 200 nodes Tree-based evaluation connection

18. Performance of Tree-based system Registration time –Message size: 800B –Min.: 2.39s –Avg.: 5.18s –Std.: 3.9 Retrieval time –Message size: 800B –Min: 0.045s –Avg.: 0.047s –Std.: 0.001 Registration Retrieve

19. Performance of DHT-based system Registration time –Message size: 400B –Min.: 0.55s –Avg.: 0.59s –Std.: 0.079 Retrieval time –Message size: 800B –Min: 0.23s –Avg.: 0.33s –Std.: 0.088 Registration Retrieve

20. Tree-based v.s. DHT-based On the “Latency” view –Tree-based: Longer registration time since hop-by-hop transfer, but very short time to retrieve –DHT-based: Less registration and retrieval time On the “Scalability” view –DHT-based system’s transmitted message size is half of the Tree-based system On the “Management” view –Tree-based system must maintain the tree statically by site administrator (Very High Cost!!)

21. Conclusion New Session Announcement Architecture –Cooperation with multiple agents –Actual implementation of our agent system with two methods of operation Tree-based system DHT-based system –Performance evaluation using real network environment (on the PlanetLab)

22. Future works Promote our agent system Agent and DHT bootstrap node address acquisition method Adaptive to end-user’s communication environment –Computing Resource of end-user’s machine –Authentication of end-user