Timing Predictable and High-Performance Hardware Cache Coherence Mechanisms for Real-Time Multi-Core Platforms

Abstract

Multi-core platforms are becoming primary compute platforms for real-time systems such as avionics and autonomous vehicles. This adoption is primarily driven by the increasing application demands deployed in real-time systems, and the cost and performance benefits of multi-core platforms. For real-time applications, satisfying safety properties in the form of timing predictability, is the paramount consideration. Providing such guarantees on safety properties requires applying some timing analysis on the application executing on the compute platform. The timing analysis computes an upper bound on the application’s execution time on the compute platform, which is referred to as the worst-case execution time (WCET).

However, multi-core platforms pose challenges that complicate the timing analysis. Among these challenges are timing challenges caused due to simultaneous accesses from multiple cores to shared hardware resources such as shared caches, interconnects, and off-chip memories. Supporting timing predictable shared data communication between real-time applications further compounds this challenge as a core’s access to shared data is dependent on the simultaneous memory activity from other cores on the shared data. Although hardware cache coherence mechanisms are the primary high-performance data communication mechanisms in current multi-core platforms, there has been very little use of these mechanisms to support timing predictable shared data communication in real-time multi-core platforms. Rather, current state-of-the-art approaches to timing predictable shared data communication sidestep hardware cache coherence. These approaches enforce memory and execution constraints on the shared data to simplify the timing analysis at the expense of application performance.

This thesis makes the case for timing predictable hardware cache coherence mechanisms as viable shared data communication mechanisms for real-time multi-core platforms. A key takeaway from the contributions in this thesis is that timing predictable hardware cache coherence mechanisms offer significant application performance over prior state-of-the-art data communication approaches while guaranteeing timing predictability.

This thesis has three main contributions. First, this thesis shows how a hardware cache coherence mechanism can be designed to be timing predictable by defining design invariants that guarantee timing predictability. We apply these design invariants and design timing predictable variants of existing conventional cache coherence mechanisms. Evaluation of these timing predictable cache coherence mechanisms show that they provide significant application performance over state-of-the-art approaches while delivering timing predictability. Second, we observe that the large worst-case memory access latency under timing predictable hardware cache coherence mechanisms questions their applicability as a data communication mechanism in real-time multi-core platforms. To this end, we present a systematic framework to design better timing predictable cache coherence mechanisms that balance high application performance and low worst-case memory access latency. Our systematic framework concisely captures the design features of timing predictable cache coherence mechanisms that impacts their WCET, and identifies a spectrum of approaches to reduce the worst-case memory access latency. We describe one approach and show that this approach reduces the worst-case memory access latency of timing predictable cache coherence mechanisms to be the same as alternative approaches while trading away minimal performance in the original cache coherence mechanisms. Third, we design a timing predictable hardware cache coherence mechanism for multi-core platforms used in mixed-critical real-time systems (MCS). Applications in MCS have varying performance and timing predictability requirements. We design a timing predictable cache coherence mechanism that considers these differing requirements and ensures that applications with no timing predictability requirements do not impact applications with strict predictability requirements.