Constructing efficient self-organising application layer multicast overlays

Application Layer Multicast (ALM) is an alternative to IP multicast, which has yet to achieve a widespread deployment in the Internet. ALM places multicast primitives directly in the multicast members, i.e. end systems, which use an overlay topology on top of the physical network for multicasting. The overlay consists of unicast connections between the members, hence bypasses the need for multicast support at the infrastructure level. The overlay structure used is a key factor that determines the efficiency of an ALM solution. This thesis investigates efficient techniques to build both low cost (i.e. low resource usage) and low delay ALM trees. We focus on self-organising distributed proposals that use limited information about the underlying physical network, limited coordination between the members, and construct overlays with bounded branching degree subject to the bandwidth constraint of each individual member. This work begins with a detailed simulation evaluation of existing ALM proposals chosen from different classes. This has resulted in enhancements to some existing proposals as well as a set of observations that could be used to assist future development of ALM proposals. As part of the evaluation, we devise a simple centralised greedy heuristic for creating low diameter degree-bounded mesh overlays, for benchmarking the class of distributed proposals. With the insights collected from the evaluation effort, we develop proposals for a distributed heuristic to build low delay delivery trees for both one-tomany and many-to-many multicast. For one-to-many multicast, we propose MeshTree, which is based on the observation that distributed delay optimisation can be trapped by the greedy problem and delaycost trade-off. MeshTree addresses the problems by embedding the delivery tree in a degree-bounded mesh containing many short links. For many-to-many multicast, we consider a multiple shared trees approach to strike a balance between the performance and quality trade-off of the conventional single shared tree and multiple source-specific trees approaches. We show that both our proposals perform well compared to existing proposals. Both of these proposals are based on a mesh-based overlay creation and maintenance framework which we have developed. The framework offers quick failure recovery mechanism to address the inherent dynamic characteristic of the ALM system.