In this thesis, we address the problem of distributing large contents in the Internet. We focus on two interesting and very common services, video on demand (VoD) and file replication.
We first investigate how to provide an efficient VoD service to a large number of clients in a dedicated overlay network (e.g. Akamai). Our contribution here is a new video distribution architecture that is highly scalable and very cost effective. The novelty of our architecture is an analytical cost model that allows us to find the minimum delivery cost of the video. In contrast to previous cost models, our model includes both, the network bandwidth cost and the server cost consumed for storing and scheduling the video at the different servers.
Using the cost model, we study many interesting scenarios like the dimensioning problem of a VoD system from scratch or the evaluation of architectural choices. One architectural choice that we consider is the use of satellite to broadcast a part of the video.
We then continue our work on VoD but in a P2P network (rather than overlay network). While previous work constructs multicast trees to deliver the video to clients, we prove that we can achieve a similar efficiency in a simpler way. To this purpose, we introduce a new approach, called PBA, where clients perform locally an algorithm to find available servants from which they retrieve the video. An available servant is an existing client that has already received or is currently receiving the video and has free upload capacity.
We evaluate PBA over a wide range of scenarios and compare it to an existing multicast tree-based approach. Our results prove that PBA not only simplifies the server, but it also consumes low network resources.
Finally, we study the use of P2P networks for file replication. Existing solutions for this service can be largely classified into tree-based and mesh-based approaches. Our first contribution here is a comparison between the scaling behavior of both kinds of approaches. Throughout this comparison, we prove that mesh approaches can be at least as efficient as tree ones.
Our second contribution is a complete analysis of mesh approaches where we identify the main parameters that influence their performance. Our results and conclusions provide new insights for the design of new cooperative architectures.
