UC San Diego

Emerging configurable optical networks and Grid computing create intriguing opportunities for new application capabilities and resource efficiencies. Applications can exploit dedicated, high-speed optical circuits to tightly interconnect remote resources on-demand, and achieve high quality of service. However, they must contend with the complexity of highly distributed and heterogeneous resource environments. In addition, network configurability presents unique challenges, adding the complexity of planning configurations to that of traditional end resource management. To enable efficient and simple development of high performance applications, this dissertation proposes the Distributed Virtual Computer (DVC), a novel integrated architecture for managing configurable networks and wide-area resource sharing. The DVC allows an application to describe and acquire a combined set of communication and end resources, and then automatically manages them for guaranteed, high performance. Such an integrated approach enables coordinated resource management improving both application capabilities and resource efficiencies. In this framework, a key challenge is selecting appropriate sets of resources for individual applications. We formulate the selection problem, explore several approaches, and evaluate each via simulation. Best performance is achieved by techniques that combine the selection of communication and end resources. Such approaches produce high-quality solutions both for application performance and for network efficiency, and scale well for large resource environments. This enables an online service where applications can request and acquire high-quality resources quickly on-demand. In a multi-domain network, a critical tension exists between service providers who are business competitors. As a result, controlled information sharing is required that balances their competitive positions and enables efficient resource selection. We characterize the network information that could be shared between providers and assess how individual information affects applications and service providers. Our results suggest providers should share their internal information as it can improve their resource efficiencies and application performance. We implement a DVC system software prototype and present experimental results with real scientific applications and optical networks. We demonstrate our prototype enables the simple configuration of collaborative data visualization environments that can be flexibly run on different physical resource configurations. Additionally, the applications are able to exploit dedicated optical circuits on-demand and efficiently utilize the network capacity