MELODY: A completely decentralized adaptive file system for handling real-time tasks in unpredictable environments

Abstract

We describe the major features of the completely decentralized adaptive file system MELODY which was designed for realizing anintegrated system design for a distributed real-time system working in a hazardous and unpredictable environment. MELODY's adaptivity mechanisms are based on novel services rendered by the distributed operating system DRAGON SLAYER. The file system, in order to both meet real-time constraints and provide for high availability, allows for replicating, relocating, or deleting file copies. Such copies may also bepublic orprivate. At every site aLocal Task Scheduler tries to schedule the arriving critical tasks, based on the availability of resources at this site such that deadline failures are minimized. Depending on the deadline failure history, status changes as well as file replication, deletion, or relocation are analyzed and managed by the cooperatingLocal File Assigners. In order to analyze MELODY's real-time performance we report on simulation experiments in which its capability of minimizing deadline failures of time-critical tasks was compared to other file system models: an idealbest-case model, abaseline model with no file replication, a file system allowingonly for replication ofprivate copies, and a model which allows forreplication and relocation of public copies only. While the best-case is unrealistic for a distributed implementation, the other models embody only part of MELODY's mechanisms yet have the benefit of a considerably smaller communication overhead. We report on the distributed simulation results which unambiguously show MELODY's superior performance, in addition to the built-in sensitivity to changes in the environment. A DRAGON SLAYER/MELODY prototype has been completed in our labs in order to serve as a distributed real-time testbed in our future work with MELODY.