Resource provisioning in large-scale self-organizing distributed systems

This dissertation presents on-line solutions to the challenge of scheduling large numbers of web services on significant numbers of servers. The services' resource demands change over time in response to surges and lapses in demand. The servers' resource supplies change dynamically to both expected and unexpected availability such as scheduled maintenance and network faults. This work presents a decentralized numeric method for quickly obtaining a solution to assign services to servers. Unlike other existing methods, this approach allows service availability and performance policy to be implemented through tunable parameters. The method is further enhanced by the use of control theory. Control is applied to the system's perceived demand of the services in response to their performance. By inflating the perceived demand of the service as a reaction to poor response time, services are reassigned and performance is shown to improve substantially. The research in this dissertation is unique because the services' demand and servers' resources are not considered fixed and the services are allowed to be reassigned