Peer-to-Peer Based Multimedia Distribution Service

Zhe Xiang, Qian Zhang, Wenwu Zhu, Zhensheng ZhangIEEE Transactions on Multimedia, Vol. 6, No. 2, April 2004

Presented by Ho Tsz Kin14/04/2004

Agenda

Introduction Architecture Topology-ware Overlay Replication Strategies

Intergroup Replication Intragroup Replication

Performance Evaluation Conclusion

Multimedia distribution services Centralized multimedia distribution

Mirroring, Proxy caching Bottleneck bandwidth problem

Measurement between University of Washington and a set of 13,656 servers

Over 90% is less than 10 Mbps Not scalable

Content distribution network (CDN) Deploys a large number of servers at the edge of the network Objective is to efficiently redirect user requests to appropriate

servers so that request latency is reduced and load among servers are balanced

Multimedia distribution services Capacity of the edge server is not large enough to support

multimedia service Where and when to place those edge servers is a difficult

problem Peer-to-peer network

Some rely on servers to disseminate information Single point of failure

Overlay network in a P2P system is not aware of the underlying topology

Availability depend on peer’s reliability Cannot provide good QoS-provision

Propose a novel framework based on P2P network

Architecture

Determine how many replicas and how they place

Determine grouping among peers

Client join the P2P network, and contribute resources

Topology-aware Overlay

Routing overhead is a key performance metric If randomly constructed, overlay network may

actually be far away in the underlying network Nearby peers in the underlying network are

clustered into groups A group consists of a set of nodes that are close

to each other Close means if the distance is less than some

predefined value Distance can be network latency, or round trip time

Topology-aware Overlay Two different groups are communicating with each other

through the shortest distance Predefined distance threshold

Given a certain transmission delay requirement

Content delivery

When a request to obtain certain content is issued Found within the same group

Content can be directly distributed to the requesting peer Peer may decide to replicate according to the replication

strategies

Not found, flooding search is carried out A shortest communication path is setup between two groups The content in source will first be sent to some host in target

group, that host in target group will send the content to requester

Replication Strategies

Global level replication decision relies on complete information about the network

such as distances between groups or between peers, storage capacity of each group, and each peer

such global information is difficult to obtain in a distributed environment

Divide the problem into two sub-problems Intergroup and Intragroup replication

Intergroup Replication Provide low latency and QoS-aware service within group

level Seed

Group-level replica Number of seeds = number of groups holding this Seed capacity is the total capacity of a group to store different seed

Minimize the average distance between requesting group and the group

providing content Subject to

the constraint of each group’s seed capacity

Intergroup Replication

Variation of K-center problem NP-Complete

Ignore seed capacity of each group, and only consider the totally seed capacity

Idea of heuristic

L

L

2D Euclidean spaceAverage distance

Seed of each content ci should be uniformly distributed over the network, let number be i

Intergroup Replication

Average access distance

Modified problem, with S is total capacity, popularity of content ci is ri

Weighted average minimum distance

Storage capacity constraints

Applying Lagrange Function

Intergroup Replication

Proposed heuristic If distance between the requestor and the peer who has a

replica is larger than , then replicate

Substitute back to find the average distance

Estimated using local information

Intragroup Replication

Improving the availability of the content Replica is copies of the content within the

group Replica replication matrix

Availability of content ci N peers

Reliability of pj

Intragroup Replication

Optimization problem

Variation of the knapsack problem NP-complete

size of content ci

storage capacity of peer pj

Intragroup Replication

Proposed heuristic Climb-hill based algorithm Adding a new replica for content cr will improve its

availability Deleting the stored contents cj also decreases its

availability A(cr): availability of content cr

A’(ci):availability of content ci if we delete this content If A’(ci) > A(cr)

Deleting ci does not conflict with the objective

Performance Evaluation

Network topology Euclidean space model

Nodes are randomly located Edge longitudes are fixed as 3000 ms 200 groups are generated Latency within group are very small

Packet loss model mainly due to the congestion occurred at routers

Number of hops between two peers increases linearly to the distance between two peers

Largest hop is ten Bandwidth of link range from 800 Kbps to 1.4 Mbps, and average

is about 1.2 Mbps

Performance Evaluation

Content distribution 10,000 MPEG-4 format video clips encoded in 1.28

Mbps Length follows a normal distribution in range of 3 min

to 5 min, correspondingly to 37.8 MB to 48 MB in files sizes

Request distribution Zipf distribution Truncated Geometric Distribution (TGD) Truncated Pareto Distribution (TPD)

Performance Evaluation

Peer Storage capacity and reliability Storage contributed by a peer follows a normal distribution

in the range of 300 MB and 2 GB, which approximately supports 8 to 50 video clips

Peer reliability of sustaining service follows normal distribution in the range of 0.1 to 0.9

Comparison Freenet

Always makes a replica for each requested content LRU replacement policy

Random replication system Contents are uniformly distributed into peer’s storage

Performance Evaluation

Performance metrics Average latency

Average access distance between the requestor peer and the content provider peer

Video quality Perceived video quality by the client PSNR

Weighted availability Represents the service availability provided by contents in a

certain area (within distance d) Defined as:

Performance Evaluation

Average latency Varying number of content from 8000 to 12500 Varying skew factor with 10000 content

Performance Evaluation

Video quality Varying peer storage Varying average packet loss ratio of network

links with peer storage capacity as 960 MB

Performance Evaluation

Availability Varying distance d

Conclusion Propose and analyze

A topology-aware overlay Replication strategies

Intergroup replication Intragroup replication

Comments: Assume equal sizes in intergroup replication, but

different sizes in intragroup replication Topology-aware techniques can also be applied to

clustering in SLVoD How to formulate and resolve stripping strategies