Multicast Data Dissemination

Abstract

Distributing large amounts of information on a computer network requires a data-dissemination server that is able to accommodate many interested clients making varied requests for data. For such a server, its outgoing network connection is a bottleneck that severely limits the number of clients and the size of shared data that the server can handle. To ameliorate this bottleneck, multicast allows a server to address a single transmission of data to multiple clients at once, reducing the repeated transmission of the same data items to different clients. Consequently, the server is able to provide faster transmissions to more clients than otherwise possible over the same network connection using unicast transmissions alone. To handle varied requests from many clients, the server must have an efficient way to decide which of the many requested items to send at any one time. This dissertation considers a variety of data schedulers to optimize this task for several performance metrics and a variety of client loads, including downloaders receiving sets of data as quickly as possible, and subscribers striving to keep changing data items current. Using modelling and simulation of such a multicast facility, we show the dramatic performance difference between various schedulers and the effects of batching clients over time. Because clients can have differing network capacities, the server must be able to deliver its data items reliably over client connections of varying loss rates. This dissertation shows how to exploit the unique properties of this system to extend traditional reliability techniques and reduce average client wait time by over 30%. Also, the server must decide how to slice its available outgoing network capacity into data channels, how to assign its data items to those channels, and how to assign clients to the channels given clients' varied requests and