Supporting Suspension-based Locking Mechanisms for Real-Time Networks-on-chip

Abstract

In the majority of safety critical systems, suspension-based locking protocols e.g. MPCP, OMLP, FMLP are used to efficiently and safely coordinate accesses to shared resources. However, existing architectures do not support such arbitration for Networks-on-Chip (NoCs) although they must resolve conflicts between concurrent transmissions. Enabling suspensions requires not only predictable transmission latencies but also to provide feedback about the global state of the interconnect which is difficult in NoCs where arbitration is done locally and independently in routers. This leads to pessimistic formal guarantees, decreased utilization and unfulfilled design requirements as network blocking unnecessarily propagates to other tasks scheduled on cores. In this work, we evaluate existing NoC architectures and propose extensions allowing to benefit from real-time tasks multithreading to increase performance while achieving predictability. Consequently, we describe how to improve the processor's utilization and more importantly, how to consistently reach lower worst case latencies for other tasks running in the system. We demonstrate the effectiveness of our approach using formal analysis and scenario-based simulation results.