Inside the New Coolstreaming: Principles, Measurements and Performance Implications Bo Li, Susu Xie, Yang Qu, Gabriel Y. Keung, Chuang Lin, Jiangchuan Liu and Xinyan Zhang

Paper overview  Introduces an improved version the original CoolStreaming protocol:  Partitions streams into substreams  For load balancing, not fault-tolerance  Switches to a push policy for parents  Presents extensive data on CoolStreaming's performance

Single tree-based multicast Three nodes out of seven do all the work!

Single tree-based multicast  Bad choice for P2P multicasting  Interior nodes do all the work  Leaf nodes do nothing  Does not tolerate interior node failures  Even worse for high bandwidth applications  Most peers do not have enough resource to transmit the data to act as leaf nodes

Multi-tree approach  Media source encodes video stream over multiple sub-streams  Each sub-stream is distributed over a separate overlay tree  Advantages  Better resilience  Fairer bandwidth utilization  Disadvantage:  More complex to manage (cf. SplitStream)

CoolStream  Basic components  Multiple sub-streams  Buffer partitioning  Push content delivery  Parent reselection

Key choices  Two basic functionalities  Locating the node from which a userobtains its video data  How the video stream is delivered  BitTorrent uses  Random peer selection (BitTorrent)  Hybrid push and pull (NEW)  Decisions based on which node has which data  Nodes periodically exchange information

Advantages  Easy to deploy  No need to maintain any global state information  Efficient  Data forwarding is dictated by data availability  Robust and resilient  Periodic exchanges of data availability information

Basic modules  Membership manager:  maintains partial view of the overlay  Partnership manager:  establishes and maintains partnerships  exchanges availability of information usin Buffer Maps  Stream manager:  handles actual data delivery

How it works  Incoming peer contacts bootstrap node  Gets a list of nodes and stores it in its mCache  Randomly selects a few nodes to initiate partnerships  Exchanges block availability information with partners  Results in parent-child relationships  Parent sends video data to child

Multiple sub-streams  Partition video stream into fixed size blocks  Assign these blocks to K substreams  Block 1 to stream 1  …  Block K+1 to stream 1  …  Block nK+i to stream i

Buffer partitioning  Buffer map comprises two elements  Sequence number of latest received block in each sub-stream  List of requested sub-streams  New need for substream synchronization

Contribution index  Ratio of Aggregate upload bandwidth over Aggregate download bandwidth  Measure contribution of each user  Four levels  Level 0: contribution index > 1  Level 1: 0.5 < contribution index < 1  Level 2: 1/6 < contribution index < 0.5  Level 3: contribution index < 1/6  65.1 percent of peers!

Push and pull scheme  Specific to the new CoolStreaming  When a peer subscribes to a substream,  It connects to a potential parent via a single request (pull )  When potential parent receives the request,  It starts pushing blocks from the specified substream to its new child

Selecting a potential parent  Assume a potential parent has blocks n to m  Which blocks should the peer request?  Not blocks in vicinity of block m  Continuity issues if parent does not receive successive blocks in time  Not blocks in vicinity of block n  Blocks will be the first ones to be dumped by the parent

Parent reselection  When a parent peer lags too much behind in the blocks and a better alternative exists

Log and data collection  Not discussed

Conclusions  Critical performance issue in P2P streaming is "excessive start-up time and high join failure rates during flash crowds  "There is a highly unbalanced distribution in terms of uploading contributions from nodes."