1. Pseudo-DHT: Distributed Search Algorithm for P2P Video Streaming
December 15, 2008
Jeonghun Noh
Stanford University
Sachin Deshpande
Sharp Laboratories of America

2. P2P Live Video Streaming
Relaying video using uplink bandwidth
Define leech clearly.
Video source
Live video: no needs for locating video chunks

3. P2P Time-Shifted Streaming
Local storage to store fractions of the video
To locate video at arbitrary point, a query server may be used
7~11m
3~6m
0~4m
Seeking video of position 5m
The server can become a bottleneck as peer population increases
No dedicated server may be available
A scalable distributed content search is desired

4. Outline P2TSS system Pseudo-DHT: Distributed Search
Performance evaluation
Numerical analysis
Simulation study
Carol: I think this slide could be eliminated

5. Introduction to P2TSS P2TSS (P2P Time-shifted Streaming System)
Serves both live streaming and time-shifted streaming
Time-shifted stream (TSS) is the same as the original stream except being delayed in time
Peers store a fraction of video
Cached video chunks are later served to other peers
Source peer
Peer 1 stores Video [0m, 4m) from 9:00am
Peer 2 watches from 0m at 9:10am
[Deshpande et al., 2008]

6. Caching Live Stream
Peers cache live stream with another video connection
Peer3
Cached portion t2 x2
Video position
Playback trajectory
Peer 2
Peers locally determine which portion to cache
Distributed Stream Buffer (DSB)
Peer’s local buffer to hold a fraction of video
A finite size of cache (e.g., size of 2 to 4 minutes of video)
Independent of playback buffer
Static contents in cache
No content change once the cache is full
Provides a bottom line performance t3 x3 x1 t1
Peer 1
Live stream
Time

7. Chord: Distributed Lookup
Chord: a distributed lookup overlay [Stoica et al., 2001]
Nodes (peers) are connected as a circle
Keys are mapped to successor node
Fast lookup by a finger table. Lookup latency: O( log(Np) )
An example key/node space
Nodes and keys sorted by IDs
ID length: 6bits
Can be normalized to [0,1)
Each peer participates in the search process
Computational resources” contributed by peers
N: node
K: key

8. Building Lookup Overlay
Node is entered to overlay as peer joins
Node ID: uniformly drawn between [0,1)
Node is placed in a distributed manner
(Key, Value) is entered as peer registers buffer status
Key (K): hashed video chunk ID ( 0 ≤ K < 1)
Value (V): peer network address
(K,V) is mapped to the successor node
Example 1

9. Pseudo-DHT: Registration
Register (K, V) may result in “collision”
(i, Peer 1)
(i, Peer 2)
Chunk ID i
Repeat Register (K’, V) until there is no collision
K’ = K + (n-1)Δ (n=# attempts, Δ=offset base)
Another consequence : low penalty due to node disconnect. But this is low level (Chord functionality.. Or DHT)
Increases lookup performance.
1st attempt
2nd attempt
Chunk ID i
i-1
empty
: Occupied ID
Unlike original DHT, single key-multiple values is discouraged
Leads to better load balancing and low latency retrieval

10. Pseudo-DHT: Retrieval
Seeking Video Chunk I
Retrieve(i) may return a “miss”
K’ = i - (n-1)Δ (n=# attempts, Δ=offset base)
Repeat Retrieve(K’) with different keys until a hit occurs
Best-effort search, different from original DHT
3rd
2nd
1st attempt
Chunk ID i
i-1
: Video chunks available
i-2
: Video Chunk Header
miss
miss
Chunks delivered to the consumer peer

11. Outline P2TSS system Pseudo-DHT: Distributed Search
Performance evaluation
Numerical analysis
Simulation study
Carol: I think this slide could be eliminated

12. Preliminaries for Analysis
Symbols
Np: Number of peers (or nodes)
L: Video length (in secs)
Q: Video chunk size (in secs)
: Ratio between Np and available slots M (=L/Q)

13. Registration Latency Independence model for the number of collision, C
More sophisticated model
where A denotes the number of peer arrival in Q seconds

14. Experimental Setup
Pseudo-DHT implemented in PlanetSim [Garcia et al, 2004]
Simulation setup
Maximum successive lookup: 4
Video length L: 7200s
Buffer size D: 240s
Chunk size Q: 5, 10, 15, 30s

15. Registration Latency When  is small, models match simulation results
Registration latency of both forward/backward key change is identical
Load factor = Num of peers.
As alpha increases, higher-order model (sophisticated) matches the result better.
Video chunk size (Q): 5s

16. Retrieval Latency Empirical statistics Modeling retrieval latency N:
Xk: 1 if video chunk Bk is occupied. 0 otherwise
Conditional hit probability Pr (Xi-j=1| Xi=0), ( j: offset base )
With a larger offset base Δ, the correlation between successive retrievals becomes weaker
(Np=500, L=7200s, Q=5s, =0.346)
Each trial is assumed to be independent (with large offset Δ)
Modeling retrieval latency N:

17. Offset Base and Retrieval Latency
Retrieval latency decreases as an offset base increases
Q=5. Load Factor = 300 / (7200/5+1).
Step size (offset) : 5, 10, 15 seconds.
A penalty is more frequent supplier switching
Video chunk size Q: 5s

18. Video Chunk Size and Retrieval Latency
As chunk size Q increases, retrieval latency decreases
Potential discussion: Larger Q is always better? 1. Larger Q takes peers to take more time before registration (slow contribution). 2. More dense key accumulation -> Losing one node results in many key information (can be avoided by key replication, though) 3. Only a few of peers will hold all keys. Bad load balancing..

19. Overhead of Key Storage
Balls and bins problem
How many nodes (bins) hold k keys (balls)?
Bins are created by nodes
Random keys fall into [0,1) uniformly
Statistically, larger bins will receive more balls
Sample space: N=90, K=12. 1
Key/Node space

20. Overhead of Key Storage
One node’s probability of storing k keys
Observations
Low overhead: keys are spread out on the overlay
50% of nodes store no keys when N=K
Number of keys stored in a node
N=300
K=300

21. Conclusion and Future Work
Proposed Pseudo-DHT
Allows peers to register / retrieve video chunk in a scalable way
Slightly different from original DHT due to video continuity
Spreads out (key, value) items over the overlay
P2TSS and Pseudo-DHT
Application to a P2P system
Thorough evaluation with analysis and simulations
Future research topics
Changing Q dynamically according to peer population size
Considering heterogeneous peer uplink capacity for registration

22. Thank you!

23. Backup Slides

24. Preliminaries for Analysis
Video chunk ID space
M-1 X0 X2
XM-1 X1

25. Sample Key/Node Space Interval between nodes is not constant
Nodes (peers)
Keys (DSB info)

26. Chord: Distributed Lookup
Chord: a distributed lookup overlay [Stoica et al., 2001]
Nodes (peers) and keys are mapped to a common space
Fast lookup by a finger table. Lookup time: O( log(Np) )
An example key/node space
Each peer participates in the search process
Computational resources” contributed by peers
N: node
K: key

27. Caching Video Peers locally determine which portion to cache
Distributed Stream Buffer (DSB)
Peer’s local buffer to hold a fraction of video
A finite size of cache (e.g., size of 2 to 4 minutes of video)
Independent of playback buffer
Peers locally determine which portion to cache
Distributed Stream Buffer (DSB)
Peer’s local buffer to hold a fraction of video
A finite size of cache (e.g., size of 2 to 4 minutes of video)
Independent of playback buffer
Static contents in cache
No content change once the cache is full
Provides a bottom line performance
Static contents in cache
No content change once the cache is full
Provides a bottom line performance

28. A lookup on Chord Overlay
Fast search using a finger table.
Each node has more detailed knowledge about nodes closer to them.

29. Registration Approach 2: Simplified dependence model
Ai : Number of insertions for key i
( Number of arrivals during the slot i)
Ai: Arrival in slot I. Not the number of key! Be careful.

30. Simultaneous Retrieval
To reduce a retrieval latency

31. Retrieval Analysis With larger offsets
(+) The number of lookups decreases slightly.
(-) Peers have to switch earlier to another peer.
The simulation results match the model
Parameters
Np = 500
L = 7200
D = 480s
Q = 5s
Number of slots: (M = 1440)

32. Analysis: Overhead for Key Storage
Poisson process property:
Given N(t) = N, N arrival times are independently uniformly distributed.
In Poisson process, interarrival time between events is exponentially distributed.
The converse is also true.
Finally,