A Distributed Software Agent Approach to High Available Clustered Computing

70384 High performance computing systems, built from clusters of inexpensive, commodity personal computers, are being used increasingly at DOE laboratories to support nuclear physics computing. However, the low reliability of these systems is a serious concern. While tools exist that could improve the fault tolerance of a given cluster, there exists no easy scalable way to manage and compose these tools. This project will use artificial intelligence techniques to automate the management and integration of heterogeneous fault tolerance tools. The system will allow a system administrator to design a fault tolerance policy for a cluster using a graphical user interface. It will then implement the policy by dispatching a set of software agents that execute those policies. Phase I will develop a limited prototype using distributed computing facilities at the Lawrence Berkeley National Laboratory as a testbed. The prototype will be elevated in terms of its ability to notice and recover from node failures, its ability to recover from application failures, and its ease of use. Commercial Applications and Other Benefits as described by the awardee: Clustered computing environments are being used increasingly in many commercial areas, e.g., bioinformatics, internet services, and financial service industries. As these enterprises turn to clusters to reduce cost, the automated management system would allow administrators to rapidly integrate commercially available fault tolerance tools with custom applications.