Adaptive Dual-Mode Arbitration for High-Performance Real-Time Embedded Systems

Multi-core platforms can deliver substantial computational power together with minimum costs, compact size, weight, and power usage. However, multi-core architectures are shaking the very foundation of modern real-time systems, i.e. deriving the Worst-Case Execution Time (WCET) of the tasks. Modern embedded systems such as those deployed in the automotive and avionic fields face two difficult-to-resolve conflicting requirements due to the interference problem on the shared hardware components amongst cores: delivering high average-case performance and providing tight WCET. This challenge exists in different shared hardware resources including on-chip shared cache, hardware prefetchers, buses, and memory controller. The problem is mainly because various cores in the system interfere with each other while competing to access the aforementioned hardware components. While dedicated real-time controllers provide timing guarantees, they do so at the cost of significantly degrading system performance. This dissertation overcomes this trade-off by introducing Duetto, a general hardware resource management paradigm that pairs a real-time arbiter with a high-performance arbiter and a latency estimator module. Based on the observation that the resource is rarely overloaded, Duetto executes the high-performance arbiter most of the time, switching to the real-time arbiter only in the rare cases when the latency estimator deems that timing guarantees risk being violated. In this thesis, the Duetto paradigm is realized for different shared hardware resources. In the first part, I demonstrate Duetto on the case study of a multi-bank on-chip memory and discuss the foundation of the methodology. The methodology is concerned about designing the real-time arbiter in such a way that it is compatible with Duetto, deriving latency analysis, and designing the latency estimator module. In the second part, this thesis addresses the trade-off between maintaining cache coherence in multi-core real-time systems and improving average-case performance by proposing a novel coherency arbiter infrastructure and employing it in the context of Duetto. This is achieved by precisely engineering the multi-core hardware architecture and its underlying interconnect infrastructure such that data sharing is feasible for real-time systems in a manner amenable for timing analysis. The proposed solution provides near-to Commercial-Off-The-Shelf (COTS) performance and does not impose any coherency protocol modifications. The third part of this dissertation proposes DuoMC by applying Duetto to off-chip Memory Controller (MC) which is crucial since Dynamic Random-Access Memory (DRAM) main memory is one of the most complex shared resources in multi-core architectures and it is one of the critical bottlenecks both from latency as well as performance perspectives. As part of the MC evaluation, we release MCsim, an open-source, cycle-accurate simulator for memory controllers