Real-time operating systems: Predictable threads and support for multi-level scheduling

In hard real-time systems, if computations do not execute before their explicit deadlines, catastrophic or life-threatening failure may occur. The traditional approach to real-time operating systems is to use a small set of primitives to obtain predictable response. However, by restricting the available primitives, the flexibility of the user is greatly diminished. There are two areas in which this dissertation increases flexibility for user applications in dynamic, hard real-time environments. The first contribution is the UMass Spring threads package specification, which is unique because of its emphasis on predictability, support for new semantics ( guaranteed and nonguaranteed thread group spawning) and explicit support for semantically-correct premature thread (and thread group) termination under hard real-time constraints. Implementation of this threads package in the UMass Spring kernel is discussed. The UMass Spring kernel is a parallel and distributed operating system featuring an on-line, planning-based scheduler that makes it suitable for dynamic environments. The evaluation of the implementation of the threads package in the UMass Spring kernel is discussed both in terms of primitive operations and higher-level system activities. In addition to new flexibility for user and kernel applications, the adoption of the threads package in the UMass Spring kernel significantly reduces kernel overhead in the context switch and memory access time. The second contribution involves multi-level scheduling for real-time systems. Multi-level scheduling consists of scheduling and planning at more than one level of abstraction, generally in terms of time and resources. The contribution of this dissertation is operating system support in the form of dynamic function and timing composition and performance studies that are the first steps towards quantifying the overall performance of the multi-level scheduling system in dynamic, hard real-time environments. Results in a flexible manufacturing domain show that high performance can be achieved if the multiple schedulers have similar performance goals; performance is significantly hindered otherwise