Sharing Non-Processor Resources in Multiprocessor Real-Time Systems

Computing devices are increasingly being leveraged in cyber-physical systems, in which computing devices sense, control, and interact with the physical world. Associated with many such real-world interactions are strict timing constraints, which if unsatisfied, can lead to catastrophic consequences. Modern examples of such timing constraints are prevalent in automotive systems, such as airbag controllers, anti-lock brakes, and new autonomous features. In all of these examples, a failure to correctly respond to an event in a timely fashion could lead to a crash, damage, injury and even loss of life. Systems with imperative timing constraints are called real-time systems, and are broadly the subject of this dissertation. Much previous work on real-time systems and scheduling theory assumes that computing tasks are independent, i.e., the only resource they share is the platform upon which they are executed. In practice, however, tasks share many resources, ranging from more overt resources such as shared memory objects, to less overt ones, including data buses and other hardware and I/O devices. Accesses to some such resources must be synchronized to ensure safety, i.e., logical correctness, while other resources may exhibit better run-time performance if accesses are explicitly synchronized. The goal of this dissertation was to develop new synchronization algorithms and associated analysis techniques that can be used to synchronize access to many classes of resources, while improving the overall resource utilization, specifically as measured by real-time schedulability. Towards that goal, the Real-Time Nested Locking Protocol (RNLP), the first multiprocessor real-time locking protocol that supports lock nesting or fine-grained locking is proposed and analyzed. Furthermore, the RNLP is extended to support reader/writer locking, as well as k-exclusion locking. All presented RNLP variants are proven optimal. Furthermore, experimental results demonstrate the schedulability-related benefits of the RNLP. Additionally, three new synchronization algorithms are presented, which are specifically motivated by the need to manage shared hardware resources to improve real-time predictability. Furthermore, two new classes of shared resources are defined, and the first synchronization algorithms for them are proposed. To analyze these new algorithms, a novel analysis technique called idleness analysis is presented, which can be used to incorporate the effects of blocking into schedulability analysis.Doctor of Philosoph