Managing distributed objects in peer-to-peer systems

Peer-to-peer systems that dynamically organize, interact and share resources are increasingly being deployed in large-scale environments. The location, intermittent connectivity, and organization of the peers have significant impact on meeting the quality of service requirements of distributed applications. This article presents the design, implementation, and empirical evaluation of a middleware architecture for managing distributed objects in peer-to-peer systems. The architecture consists of a self-organizing infrastructure that uses only local knowledge to dynamically form overlays of multiple peers and respond to changing processing and networking conditions; and a management layer that monitors the behavior of the applications transparently, schedules object invocations over multiple machines, and obtains accurate resource projections. The system works in a two-level feedback loop structure that uses measurements of elapsed time and resource loads to refine the initial estimates and revise the peer connections. Our empirical evaluation shows that the system manipulates the peer connections dynamically in response to changes in resource utilization to meet application end-to-end soft real-time requirements.