Rapid Deterministic Fault Detection in Distributed Systems

A novel approach to deterministic fault detection is proposed, based on the idea of augmenting component interface contracts with behavioral and timing semantic tags. Since the component interfaces are registered with the middleware infrastructure,violations of the interface contract semantics can be detected and managed by the middleware. The proposed solution integrates the common middleware communications models, including distributed-object, publish-subscribe, and group-ordered communication.It ensures fault-delivery guarantees, such as bounded detection latencies. This fault-detection approach can be integrated with commercially-available system management products to deliver a robust solution that meets a wide range of needs by distributedapplications in both the commercial and defense sectors.The proposed effort will characterize faults and isolate contract semantics relevant to fault detection, define an integrated middleware communication model that is cognizant of interface contract semantics, develop techniques for integrating withcommercial system-management products, and demonstrate a prototype implementation of the approach for selected fault conditions. The new approach will ensure fault notification guarantees (e.g., bounded detection latencies) previously not possible. Since it incorporates all three communication models, the solution will be able to handle fault situations previously not accessible,while providing an optimized and efficient implementation. In addition, the integrated middleware can keep better track of fault statistics for later analysis and debugging purposes.The integrated middleware will reduce the burden of fault detection on the application programmer by pushing the responsibility of fault detection and notification to the distributed infrastructure. This will free the application designer to instead focuson developing appropriate fault handling and recovery schemes. Furthermore, the faults detection will be easier to evolve and maintain as requirements change.If this work is not done, the current state of affairs with disparate middleware products will continue. Fault detection would not be guaranteed with bounded latencies. Fault semantics will need to be developed on a per application basis by theapplication programmer. Overlapping middleware semantics with their subtle differences will require careful handling, which lead to greater potential for errors. Opportunities for optimization will be missed, and internal middleware faults status will bemade available on an ad-hoc basis. Thus, fault management will remain a thorny issue placing considerable burden on the application programmer and resource management algorithms.