An autonomous machine operates on its own by continuously sensing the environment, making decisions, and executing actions. Due to this attribute, it is well-suited for managing critical tasks that demand rapid responses without human intervention such as avionics, surgical robots, and autonomous driving. Given that failures in such tasks can have catastrophic consequences, these autonomous machines are meticulously designed with stringent timing safety guarantees, which have been extensively studied and integrated into real-time systems. As hardware and software complexities increase, the fundamental concepts in real-time systems, namely worst-case behaviors and determinism, are becoming extremely challenging to analyze. The recent trend in autonomous driving exacerbates this challenge even further. As tasks previously performed by humans are now delegated to real-time systems, the workloads are increasingly demanding in terms of performance, which in turn necessitates the use of modern complex hardware. The widening gap between worst case scenarios and typical operating conditions results in pessimistic resource planning and high energy consumption. Furthermore, design-time analyses and static design struggle to address the dynamic nature of the physical world, replete with ever-changing situations.
In this dissertation, I delve into the concept of operation modes as an approach to tackle the challenges in designing real-time systems that interact with the dynamic physical world. First, I introduce the EASYR framework, which leverages adaptive system reconfiguration to enhance energy efficiency. EASYR reacts to velocity changes and associated end-to-end deadline changes, holistically adapting scheduling parameters such as task period and speed factor while satisfying timing constraints. The possible deadline range is divided into several discrete operation modes, each with its maximum end-to-end latency guarantees. EASYR not only optimizes the energy efficiency within each mode but also provides a safe transition between these modes. Consequently, in real-world driving scenario, simulation results show that EASYR achieved 62% reduction in energy consumption compared to classical static deadline optimization.
Next, I introduce the FRTS framework, which exploits adaptive system reconfiguration to improve resilience of timing safety against WCET changes. Carefully designed real-time systems are confronting diverse faults and aging effects resulting from their increased lifespan. Safety critical systems must be safe when designed, and continue to be safe as conditions change. In FRTS, faults are manifested by a statistical distribution changes in execution times. Testing the assumption of independent & identical distribution (i.i.d) is well-suited for proactively sensing these changes prior to actual violations. Once a fault is detected, the system enters the emergency mode, during which system utilization is secured as a buffer against WCET changes. With the buffer, the system is capable of assessing the fault impact and performing reconfiguration to an optimized operating point. As a result, the system exhibits greater resilience compared to a conventional fault-tolerance technique with an improvement of 18% with a small energy overhead in simulation results.
These two frameworks demonstrate a great potential to improve autonomous operations by facilitating adaptation to dynamic situations. This is a stepping stone toward a true autonomous reconfiguration of safety-critical machines.