1. Temporal-DHT and its Application in P2P-VoD Systems Abhishek Bhattacharya, Zhenyu Yang & Shiyun Zhang

2. Roadmap Introduction Indexing Buffer Management Content Distribution Results Summary

3. Introduction Peer-to-Peer (P2P) applications gained prominence due to decentralized/self-organizing behavior, scalability, tolerance dynamics (churn/flash crowd). P2P-based multimedia streaming services poses challenges that are different from file-sharing applications. Live Streaming applications already popular with Internet- scale deployments (PPLive, CoolStreaming, Uusee, etc.) Video-on-Demand (VoD) systems present unique challenges with asynchronous/dynamic user interactivity.

4. Introduction c1c1 c2c2 c3c3 c4c4 c5c5 c6c6 c7c7 c8c8 p1p4p1p4 p5p5 p3p3 p2p2 Content Discovery:  Tracking Server  Decentralized Indexing Structures Content Distribution:  Overlay Tree/Multi-Tree/Mesh

5. Introduction Current P2P-VoD systems are classified into 2 categories:  Cache-and-Relay : indexing based on playing position.  Content Dissemination based on playing position proximity  Smooth and easy content availability from parent  Not resilient to asynchronous jumps and peer dynamics (leave/failure)  High streaming efficiency due to seamless in-order playback continuity  Dynamic/In-Order Caching

6. Introduction  Static-Cache : indexing based on a static distributed storage contributed by all the peers and are independent of playing position.  Content dissemination is independent of playing position.  Frequent parent change: after each segment playback.  Increased resiliency to peer dynamics and efficient support for random access patterns.  Avoids exploiting playing position proximity thereby reduces streaming efficiency.  Out-of-Order/Static Caching.

7. Introduction Distributed Hash Tables (DHT) are stable substrates for P2P based applications with decentralized operations. DHT s are efficient for indexing static data like file-sharing applications where the content remains same throughout peer lifetime. DHTs are efficiently used by Static-Cache based systems since the cached segments remain constant. DHTs are unable to efficiently handle data with temporal dynamics (Cache-and-Relay based systems) thereby invoking high update overheads.

8. Indexing Temporal-DHT(t-DHT) for dynamic indexing with respect to playing position. t-DHT augments the generic DHT interface for indexing dynamic content by modifying query resolution and indexing record structure. t-DHT reduce update overhead by using lazy updation with a coarser granularity of periodicity. t-DHT estimates the playing position by exploiting predictive temporal dynamics of the content.

9. Indexing: t-DHT semantics Generic DHT indexing record: where p i is the peer currently hosting the content c j t-DHT indexing record: where TTL is the time-to-live for the particular indexing record. t-DHT also utilizes a publish interval (T) which signifies the periodicity of the lazy updates. T is a design choice and can be defined as a multiple of playback data rate.

10. Indexing: Query Resolution … C i+1 CiCi C i+2 C i+z ……… CiCi CiCi CiCi T Range Query Reformulation in t-DHT

11.  Given the size of playback buffer k, the position of representative segment r, the publish interval z, a peer that searches for dynamic segment c i needs to perform a range query for segments in [c i-k+r-z, c i+r-1 ] ∩ [c 1, c M ] Detailed proof in the paper as Theorem IV.1  Given the size of playback buffer k, the position of representative segment r, the publish interval z, and the target segment c i, an index record of can be filtered out of if id(c)+k-r+z-TTL i. Proof in the paper as Theorem IV.3 Indexing: Query Resolution

12. Indexing  A semantic overlay is layered on top of DHT network for allowing easy and efficient in-order access.  In-order consecutive segments are highly correlated with high access probabilities.  Each t-DHT peer maintains content predecessor/successor pointers which link to the next and previous in-ordered segments respectively in the stream.  The range query is accelerated with the help of these links and also lowers messaging cost instead of invoking multiple exact- match DHT queries.

13. Indexing  Dynamic Indexing is implemented by t-DHT augmented semantics with each record as.  Indexing based on playing position and thereby updating the t- DHT after every publish interval. Query resolution is performed by range query reformulation.  Static Indexing is also implemented by t-DHT only but mostly retains generic DHT operations with each record as.  Indexing is independent of playing position and the records are constant with no updations required. Query resolution is performed similar to exact-match generic DHT routing.

14. Buffer Management  Continuous playback achieved by dynamic caching and random access by static caching are conflicting features.  t-DHT organizes an integrated approach by utilizing static and dynamic caching in a single framework.  Buffer is divided into 2 parts: static and dynamic. Static buffer contents remain same but dynamic buffer contents keep changing.  Temporal-DHT based Mesh (TDM) is constructed on top of static-dynamic buffers by forming the content linkage pointers and parent-child pointers.

15. Buffer Management  The static buffer is filled up by randomly downloading b segments from other peers or server.  The static buffer contents are published to the t-DHT by a one- time post operation with the format for static indexing records.  The dynamic buffer is filled up by caching k segments around the current playing position.  The dynamic buffer contents are published to the t-DHT by a chosen representative segment and performs continuous update operations with a certain interval.

16. Content Distribution  VoD users frequently perform random seeks during the initial period after joining and then either leaves the system or stabilizes to continuous in-order playback mode [15].  TDM is motivated from the above observation by dividing the users duration into 2 modes: random seek and continuous playback.  TDM follows adaptive content distribution by associating randomized and synchronous dissemination.  Continuous playback mode is handled by the construction of an overlay tree after the peer enters this mode and the distribution follows along the synchronous parent-child pointers.

17. Content Distribution  After joining the system, each VoD peer is placed into random access mode where static indexing/querying is used for content location/distribution.  Decision on transition from random access mode to continuous playback mode!  TDM relies on user workload profiling to perform seamless transition from random to continuous mode.  After entering the continuous mode, VoD peer joins an overlay tree and streams content always from tree parent. This process avoids unnecessary parent search-switch.

18. 18 Results

19. Summary  We propose Temporal-DHT which is a novel augmentation to generic DHT for indexing contents with temporal dynamics.  Temporal-DHT applies lazy updates to reduce update overhead and query reformulation/TTL filtering techniques for query resolution.  An integrated static/dynamic caching and buffer management mechanism is presented to efficiently harness the advantages of both the approaches.  Adaptive Content Distribution is utilized by exploiting user request pattern to efficiently support random and continuous content access within a single framework.

20. Please send all your questions to: abhat002@cis.fiu.edu