A two-layer agent-based system for large-scale distributed computation

The paper is devoted to a two-layered, agent-based system that can support and manage large-scale distributed computation. Each managing agent, being a part of the lower layer, carries individual computational tasks and examines the environment to find the best place of execution for it. Effective scheduling policies based on the molecular diffusion phenomena may be applied. They only utilize local information about the computing node characteristics, which saves communication time. The lower level of the system constitutes the software platform that implements the information, communication and migration policies for managing agents. The architecture and a formal description of the platform and the managing agent, as well as scheduling algorithms (especially agent diffusion rules involving inter-task communication), are presented. A sample Java implementation called OCTOPUS/SmartSolid and test results are also reported.