Hierarchical Geometric Overlay Multicast Network

In this poster proposal, we present Bos: a hierarchical geometric overlay multicast network that is built on a 2-tier hierarchical architecture called Lightweight SuperPeer Topologies (LST) [1], [2]. Bos makes use of the geometric connectivity and Minimum Spanning Trees (MST) properties of the LST overlay network to provide efficient overlay multicast. We evaluate Bos’s performance using the large-scale network models that were used in Scribe [3] and SplitStream [4]. The results show that Bos performs reasonably well in large size networks with reasonable link and node stress. I. THE LST OVERLAY NETWORK The construction of 2-tier hierarchical architecture of LST overlay network consists of three key steps: Step 1: SuperPeer election. When a new overlay node joins the LST overlay network, the following criteria are evaluated to determine whether this overlay node will be elected as a SuperPeer or normal Peer: A) The SuperPeer should have enough resources to serve other SuperPeers and peers. B) The SuperPeer should be reliable or stable and it is not joining and leaving the LST overlay network very frequently. With the above criteria imposed, the SuperPeers layer will consist of SuperPeers acting as backbone high-speed gateway for the peers in the peers layer. The list of SuperPeers are being classified as the list of landmark nodes (lighthouses) for the procedure in Step 2. Step 2: Highways [5]. Highways is an overlay network control plane service [5] that performs scalable network embedding [6] to map overlay nodes in metric space onto geometric points in geometric space and assign geometric coordinates to the overlay nodes to represent their geometric position for the construction of the geometric overlay network. If accurate, such techniques would allow us to predict Internet distances without extensive measurements. We use landmark-based and Singular Value Decomposition (SVD) embedding techniques for low-dimensional network embedding. Firstly, Round-Trip-Time’s (RTT’s) measurements of each overlay node to at least d + 1 landmark nodes (SuperPeers) are performed for embedding into d-dimensional geometric space. Network superspace embedding embeds the whole set of overlay nodes in the system as one large set into Global geometric space while subspace embedding embeds all small partitioned clusters of overlay nodes into Local geometric space. The rationale for performing network subspace embedding arises from the scalability (meta-) metric observations in [6], subspace embeddings into Euclidean space of various partitioned clusters of overlay nodes achieve better accuracy in geometric distance estimation. Using the overlay nodes’ Global geometric position information, all overlay nodes in the overlay network are partitioned into clusters by adopting a simplistic approach of the K-means method (we use K = 3). Network subspace embedding is done to overlay nodes within these clusters. Therefore, all overlay nodes will have both Global and Local geometric position information. The local geometric position information helps to provide a more accurate geometric distance estimation among overlay nodes in the cluster while the global geometric position estimates the geometric distances between overlay nodes in different clusters. We recognize the fact that slim possibility of inaccuracy in overlay nodes’ rank ordering through geometric distance estimation may happen. In order to mitigate this, from the perspective of each node, a sanitary check is done for the list of closest g = 10 nodes derived from its geometric distances. That is, a comparison is performed with its measured RTTs and re-ordering of the list of closest nodes is done if distance ordering errors were found. This sanitary check helps in ensuring the list of closest nodes is identified. Step 3: SuperPeers and Peers Topology Construction. In the SuperPeers layer, we use Yao-Graphs [7] to construct the overlay network connectivity among the SuperPeers by using their geometric position information and estimated geometric distances with other SuperPeers as computed from Step 2. Since the geometric space around the SuperPeer is cut into six sectors of equal angle θ < π/3), every SuperPeer choose the six closest SuperPeers in terms of their geometric distances to connect to. These SuperPeer-SuperPeer YaoGraphs routes serve as the reliable high-bandwidth backbone network connectivity. In the Peers layer, Peers are directly connected to the first closest SuperPeers that are capable of serving an additional Peer and this connectivity is called the Peer-SuperPeer 1-Hop route. Among the Peers being served by their closest SuperPeer, direct connectivity between these Peers can be established if there exists a shortcut route between the Peers. That is, a Peer-Peer Shortcut route is established between two Peers belonging to a SuperPeer, if the direct connectivity between these two Peers is the shortest route compared to their Peer-SuperPeer 1-Hop routes.