Resource management for real-time fault-tolerant distributed systems

Computer applications are increasingly being built as real-time distributed object systems. As such systems become larger and more complex, it becomes more difficult to predict the needs of the applications, particularly because those needs are likely to change dynamically while the applications execute. To respond to transient changes in the load on the processors or in the availability of the resources, profiling, scheduling and migration algorithms are required.; This dissertation presents a Resource Management System for a soft real-time distributed object system that is based on a three-level feedback loop. The three levels of the feedback loop comprise: (i) a dynamic scheduling algorithm that schedules the methods of the objects invoked by the tasks on a timescale of a few milliseconds, (ii) a profiling algorithm that measures the usage of the resources and monitors the behavior of objects averaged over a second or so, and (iii) distribution and migration algorithms that allocate and reallocate objects to balance the load on the resources over many seconds.; The Resource Management System consists of a single (but possibly replicated and distributed) Resource Manager, and Profilers and Schedulers located on each of the processors within the distributed system. The Resource Manager allocates objects to processors. The Profilers monitor the behavior of the objects, measure the current load on the processors, and supply the collected data to the Resource Manager. The Schedulers exploit information collected by the Resource Manager to schedule the methods comprising the tasks, to meet soft real-time deadlines.