Described are the major features of the completely decentralized adaptive file system Melody, which runs on top of the distributed operating system Dragon Slayer. A definition of all major design requirements is presented, along with a heuristic algorithm for resource scheduling which allows for replicating, relocating, or deleting file copies in a local area network. At every site a local task scheduler tries to schedule the arriving critical tasks such that no deadlines are missed, based on the availability of resources at this site. Depending on the task-deadline failure history and file-access request history at each node, file copies may be relocated, replicated to other sites, or deleted. Such copies may also assume a different status (public or private) depending on how recent their information is. Only public copies are maintained under a strong concurrency-control protocol. The idea is to maintain optimal distribution of file copies in the network regarding reliability and real-time responsiveness of the system. Experiments performed to analyze Melody's capability of minimizing deadline failures of time-critical tasks were compared to structurally simpler file-system models. A report is presented on the distributed simulation results which show Melody's superior performance and its built-in sensitivity to changes in the environment.>

The authors analyze current and future needs in building adaptive real-time systems and formulate a series of design requirements for their hardware/interconnection, operating system, and application levels that could be satisfied together only in an integrated system design. They describe the major features of the corresponding model. On the operating-system level, novel services are rendered by their own file system, Dragon Slayer. The file system, in order to both meet real-time constraints and provide for high availability in a hazardous environment, allows for replicating, relocating, or deleting of file copies. A distributed implementation of the integrated system design model is currently being completed in order to serve as a distributed real-time testbed. The authors also report on a sensitivity study that shows how promptly the model file system reacts to changing environmental situations (changing request patterns).>