| Emerging applications, such as the electronic commerce integrate large amounts of data that are heterogeneous and/or time sensitive. These data are typically disseminated over the Internet and target a potentially large number of users. As the number of users increases, the capacity of single server architectures becomes insufficient to handle the load and server replication is often used to improve scalability. This has led to the development of architectures for distributed clusters of replicated servers. An important problem in connection with server replication is load distribution among servers. Architectural support for load balancing has to take into account the collection of performance data on servers and the control over server assignment to clients. Of particular interest to our investigation is how a client may locate an appropriate server without being aware of specific details of replica organization, and how can this process scale to a large number of clients.; In this thesis, we investigate an architecture where server selection and QoS management are delegated to an independent brokerage service. Brokers are placed on the server side, and their responsibilities include (1) performance monitoring of servers, (2) implementation of server selection policies, and (3) the interaction with users for the selection of servers and the negotiation of QoS related issues.; We first study the relevant QoS issues in the context of web-based applications. Our objective is to provide a framework in which a client's QoS requirements can be considered. We then present our basic architecture, its different components, the server selection algorithms under consideration, and performance results for load balancing that we obtained through simulation and experimentation.; The ability to provide different levels of service to different classes of clients is very appealing for applications such as electronic commerce. Clients often do not have the same QoS requirements, expectations and/or capabilities. Applications may also classify users based on specific objectives or business models. We extend our basic architecture to support service differentiation by including priority class as a key factor in the server selection process. Issues that need to be addressed in order to support service differentiation are discussed and simulation results that show the effectiveness of our server selection algorithm are presented.; Scalability is a main objective in this work. In our proposed architecture, scalability can further be improved by replicating the brokers, in case the capacity of a single broker is not sufficient to handle the load. We discuss alternative organizations to support access to multiple brokers and the needed cooperation between independent brokers in order to achieve server selection effectively. Algorithms to select a server globally among multiple clusters are developed, and their performance evaluated by simulation. |
