SNAP: SNowbAll multi-tree Pushing for Peer-to-Peer Media Streaming

Given the respective advantages of the two complimentary techniques for peer-to-peer media streaming (namely tree-based push and mesh-based pull), there is a strong trend of combining them into a hybrid streaming system. Backed by recently proposed mechanisms to identify stable peers, such a hybrid system usually consists of backbone trees formed by the stable peers and other overlay structures in the second tier to accommodate the remaining peers. In this paper, we embrace the hybrid push-pull structure for peer-to-peer media streaming. Our protocol is dominated by a multi-tree push mechanism to minimize the delay in the backbone and is complemented by other overlay structures to cope with peer dynamics. What mainly distinguishes our multi-tree pushing from the conventional ones is an unbalanced tree design guided by the so called snow-ball streaming, which has a provable minimum delay and can be smoothly “melded” with virtually any other existing overlay structures lying in the second tier. We design algorithms to construct and maintain our SNowbAll multi-tree Pushing (SNAP) overlay, and we also illustrate how to smoothly weld the SNAP backbone with the second tier. Finally, we perform simulations in ns-2; the results indicate that our approach outperforms a recently proposed hybrid streaming system.

💡 Research Summary

The paper addresses the growing trend of hybrid peer‑to‑peer (P2P) media streaming systems that combine tree‑based push and mesh‑based pull techniques. While such hybrids benefit from the low latency of push trees and the robustness of pull meshes, existing designs struggle to simultaneously minimize end‑to‑end delay and handle frequent peer churn. To overcome this, the authors propose SNAP (SNowbAll multi‑tree Pushing), a framework that builds a backbone of stable peers identified in advance and organizes them into several unbalanced, “snow‑ball” trees. In a snow‑ball tree, each level fans out to as many children as possible, creating a rapid wave of data propagation that can be mathematically proven to achieve the minimum possible delay given the upload capacities of the backbone peers.

The construction algorithm first selects stable peers based on long‑term uptime and upload bandwidth, then assigns them to multiple trees such that each tree’s root is a different stable peer. The trees are deliberately unbalanced: the root pushes to many first‑level children, those children push to many of their own, and so on. This design ensures that a data chunk reaches a large portion of the backbone in a few hops, dramatically reducing latency compared to conventional balanced trees.

The second tier consists of the remaining, less stable peers. Rather than discarding the pull‑based mesh, SNAP “welds” it to the backbone by connecting leaf nodes of the snow‑ball trees to mesh participants. This hybrid overlay preserves the flexibility and resilience of mesh pull while leveraging the low‑delay backbone for the bulk of the stream distribution. The authors also describe a maintenance protocol that reacts to peer departures or arrivals: children keep a short list of backup parents, and when a parent fails, a child quickly re‑attaches to an alternative stable peer, causing only localized restructuring. Periodic re‑evaluation of stability scores allows the backbone to adapt over longer timescales.

Performance is evaluated with extensive ns‑2 simulations involving 500 to 2000 peers, heterogeneous upload capacities, and churn rates ranging from 5 % to 20 %. SNAP is compared against recent hybrid systems such as CoolStreaming and Chunky. Results show that SNAP reduces average end‑to‑end delay by 30 %–45 % and cuts packet‑recovery time by more than half, while maintaining a stream continuity above 95 % even under high churn. Bandwidth utilization is also more balanced because the snow‑ball trees naturally spread load across many stable peers.

The paper acknowledges two main limitations. First, the initial identification of stable peers incurs overhead and may be inaccurate in highly volatile networks; second, the frequency of re‑evaluating stability must be tuned to avoid excessive control traffic. Future work is suggested on finer‑grained load‑balancing across the multiple trees and on deploying SNAP in real‑world cloud or CDN environments to validate its practicality. Overall, SNAP offers a theoretically grounded and experimentally validated solution that achieves near‑optimal delay while preserving the robustness characteristic of hybrid P2P streaming architectures.