FatNemo: Multisource Multicast Overlay Fat-Trees

High-bandwidth multisource multicast among widely distributed nodes is critical for a wide range of important applications including audio and video conferencing, multi-party games and content distribution. Multicast decouples the size of the receiver set from the amount of state kept at any single node and potentially avoids redundant communication in the network. The limited deployment of IP Multicast [10], [11] has led to considerable interest in alternate approaches implemented at the application layer that rely only on end systems [9], [12], [2], [15], [8]. In an end system multicast approach, participating peers organize themselves into an overlay topology for data delivery. Each edge in the topology corresponds to a unicast path between two peers in the underlying Internet. All multicast-related functionality is implemented at the peers, instead of at the routers of the underlying network, and the goal of the multicast protocol is to construct and maintain an efficient overlay for data transmission. Among the proposed end system multicast protocols, tree-based systems have proven to be highly scalable and efficient in terms of physical link stress, state and control overhead, and end-to-end latency [12], [2], [7]. However, normal tree structures have inherent problems in terms of resilience and bandwidth capacity. Trees are highly dependent on the reliability of non-leaf nodes. Resilience is especially relevant to the application-layer approach, as overlays are composed of autonomous, unpredictable end systems. The high degree of transiency of end systems is one of the main challenges for these architectures [3]. Furthermore, trees are likely to be bandwidth constrained since bandwidth availability monotonically decreases as one ascends from the leaves. The bandwidth limitations of normal tree structures are particularly problematic for multisource, bandwidth-intensive applications. For a set of randomly placed sources in a tree, higher-level paths (those near the root) will become the bottleneck under high load, and tend to dominate delivery latencies. Once these links become heavily loaded or overloaded, packets will be buffered or dropped. We have addressed the resilience issue of tree-based systems by exploiting the alternative paths introduced through co-leaders [4]. In this work we address the bandwidth constraints of conventional trees by importing Leiserson’s fat-trees from parallel computing into overlay