| Current approaches to utilizing computational resources on large, unstructured networks like the Internet, require either centralized servers or extensive system knowledge, limiting their scalability. In a radical departure from current models, we follow a fully-decentralized approach in which mobile agents autonomously organize themselves on a peer-to-peer network so as to maximize resource utilization. Computation and scheduling behavior are encapsulated into agents and thus decoupled from the underlying infrastructure.; The following design objectives are achieved: near-zero knowledge of network topology, zero knowledge of system status, distributed computation, lack of specialized roles. Every node is equally responsible for scheduling and computation, both of which are performed with minimal information about the system. This model avoids the extensive system knowledge requirements of traditional Grid scheduling approaches. Unlike popular master/worker organizations of current large-scale grids, our model is not reliant on specialized super-servers or on application-specific clients.; Strong mobility allows the asynchronous and transparent migration of multi-threaded agents, and is a powerful mechanism for realizing a peer-to-peer computing environment. We have implemented and evaluated an agent system that endows Java with strong mobility.; Through the careful design of agent behavior, the global organization of computation can be customized for different classes of applications. We describe our scheduling framework and demonstrate the flexibility of our approach by applying it to two problems that are vastly different in terms of the synchrony of their communication patterns: an Independent Task Application and Cannon's Matrix Multiplication. |
