[Paper Review] Tree dynamics for peer-to-peer streaming

Ji Zhu, Bruce Hajek|arXiv (Cornell University)|Aug 8, 2013

Peer-to-Peer Network Technologies21 references3 citations

TL;DR

This paper proposes a distributed, asynchronous algorithm for managing multiple multicast trees in peer-to-peer (P2P) streaming systems to ensure low-delay, high-coverage substream delivery. By dynamically updating tree topologies based on node degree and upload capacity, the algorithm achieves logarithmic tree depth, cycle elimination, and fast convergence—demonstrated theoretically and via simulations even under relaxed assumptions.

ABSTRACT

This paper presents an asynchronous distributed algorithm to manage multiple trees for peer-to-peer streaming in a flow level model. It is assumed that videos are cut into substreams, with or without source coding, to be distributed to all nodes. The algorithm guarantees that each node receives sufficiently many substreams within delay logarithmic in the number of peers. The algorithm works by constantly updating the topology so that each substream is distributed through trees to as many nodes as possible without interference. Competition among trees for limited upload capacity is managed so that both coverage and balance are achieved. The algorithm is robust in that it efficiently eliminates cycles and maintains tree structures in a distributed way. The algorithm favors nodes with higher degree, so it not only works for live streaming and video on demand, but also in the case a few nodes with large degree act as servers and other nodes act as clients. A proof of convergence of the algorithm is given assuming instantaneous update of depth information, and for the case of a single tree it is shown that the convergence time is stochastically tightly bounded by a small constant times the log of the number of nodes. These theoretical results are complemented by simulations showing that the algorithm works well even when most assumptions for the theoretical tractability do not hold.

Motivation & Objective

• Address the challenge of efficient, scalable, and robust distribution of video substreams in P2P streaming systems with dynamic peer participation.
• Overcome limitations of unstructured and mesh-based P2P streaming, which suffer from high overhead or poor playback continuity.
• Design a distributed algorithm that maintains tree diversity, balance, and low depth without centralized control.
• Ensure robustness to peer churn, link failures, and heterogeneous upload capacities, including server-client topologies.
• Achieve fast convergence and minimal delay even when the total demand approaches total upload capacity (ρ = 1).

Proposed method

• Use a complete underlying network for control messaging, where nodes periodically sample random peers to exchange topology and depth information.
• Apply a distributed, asynchronous update rule—CombinedUpdate—that adjusts parent-child relationships based on depth and degree to balance load and minimize tree depth.
• Favor nodes with higher upload capacity (higher degree) by preferentially assigning them as parents, promoting efficient tree formation.
• Maintain tree structure by detecting and eliminating cycles through depth and parent information exchange, ensuring acyclic, tree-like topologies.
• Support heterogeneous upload capacity by allowing nodes to have different maximum fan-out degrees, including cases where a few high-capacity nodes act as servers.
• Use source coding (e.g., MDC) to provide redundancy across substreams, and manage multiple independent trees per substream to enhance resilience.

Experimental results

Research questions

• RQ1 Can a fully distributed algorithm manage multiple multicast trees in P2P streaming to ensure all nodes receive sufficient substreams with low delay?
• RQ2 How can tree depth be kept logarithmic in the number of peers under dynamic peer arrivals and departures?
• RQ3 Can the algorithm maintain tree diversity and balance while efficiently eliminating cycles in a distributed manner?
• RQ4 How does the algorithm perform when the total demand is close to or equal to the total upload capacity (ρ = 1)?
• RQ5 To what extent does the algorithm favor high-capacity nodes, and how well does it scale in server-client or polarized degree scenarios?

Key findings

• The algorithm ensures that each node receives all required substreams with a delay bounded by O(log N), where N is the number of peers.
• Convergence time is stochastically tightly bounded by O(log N) for a single tree, with simulations showing fast convergence even when theoretical assumptions are relaxed.
• Adding just 10% extra capacity (α = 0.1) eliminates the long tail in coverage, achieving full node coverage by time 25 in simulations.
• In the server-client case, increasing the number of high-capacity server nodes (e.g., r = 2) significantly reduces maximum tree depth and accelerates convergence.
• The algorithm maintains robust performance under heterogeneous upload capacities, favoring high-degree nodes and achieving near-exponential convergence in coverage and depth reduction.
• Simulations confirm that cycles are eliminated efficiently, and the algorithm remains stable and scalable even when assumptions like instantaneous depth updates do not hold.