Predictable and Performant Computer Architectures for Time-Critical Systems

Abstract

In real-time systems, computing tasks must complete within a certain time limit. Otherwise, the system might fail with potentially catastrophic consequences. Therefore, it is essential to analyze the timing behavior of these tasks in order to guarantee that they will always complete in time. However, the timing analysis of modern systems is complicated by the lack of temporal predictability of both the software and hardware, to the point where a timing analysis is not feasible at all. The more predictable architectures traditionally used in real-time systems are no longer capable of fulfilling the performance requirements of today’s demanding workloads. This work addresses the problem by investigating new computer architectures, in line with current trends, for use in time-critical cyber-physical systems, with the aim of closing the current performance gap between predictable and high-performance platforms. In particular, it presents an instruction filter, which can be integrated into existing processor architectures with the aim of executing temporally predictable software on a wide variety of architectures. Also, it proposes the use of a timing- predictable vector processor in a real-time system, a processor type capable of efficiently executing data-parallel worloads and provides increased predictability without compromising performance. Comprehensive evaluations and comparisons with existing approaches show that both of these contributions are competitive alternatives that improve the choice of suitable platforms and the performance of hard real-time systems for relevant workloads.