Achieving Performance in Networks-On-Chip for Real-Time Systems

Abstract

In many new applications, such as in automatic driving, high performance requirements have reached safety critical real-time systems. Consequently, Networks-on-Chip (NoCs) must efficiently host new sets of highly dynamic workloads e.g., high resolution sensor fusion and data processing, autonomous decision’s making combined with machine learning. The static platform management, as used in current safety critical systems, is no more sufficient to provide the needed level of service. A dynamic platform management could meet the challenge, but it usually suffers from a lack of predictability and the simplicity necessary for certification of safety and real-time properties. In this work, we propose a novel, global and dynamic arbitration for NoCs with real-time QoS requirements. The mechanism decouples the admission control from arbitration in routers thereby simplifying a dynamic adaptation and real-time analysis. Consequently, the proposed solution allows the deployment of a sophisticated contract-based QoS provisioning without introducing complicated and hard to maintain schemes, known from the frequently applied static arbiters. The presented work introduces an overlay network to synchronize transmissions using arbitration units called Resource Managers (RMs), which allows global and work-conserving scheduling. The description of resource allocation strategies is supplemented by protocol design and verification methodology bringing adaptive control to NoC communication in setups with different QoS requirements and traffic classes. For doing that, a formal worst-case timing analysis for the mechanism has been proposed which demonstrates that this solution not only exposes higher performance in simulation but, even more importantly, consistently reaches smaller formally guaranteed worst-case latencies than other strategies for realistic levels of system's utilization. The approach is not limited to a specific network architecture or topology as the mechanism does not require modifications of routers and therefore can be used together with the majority of existing manycore systems. Indeed, the evaluation followed using the generic performance optimized router designs, as well as two systems-on-chip focused on real-time deployments. The results confirmed that the proposed approach proves to exhibit significantly higher average performance in simulation and execution.